3147 行
728 KiB
JSON
3147 行
728 KiB
JSON
|
|
{
|
|||
|
|
"dimensions": [
|
|||
|
|
{
|
|||
|
|
"dimension": "工作流 / 轮次 / 分支 / 交叉评审 状态机",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A2-工作流与轮次.md",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_generation_cross_review_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_cross_review_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_feedback_interpreter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_feedback_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\branch_lineage_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_job_retry_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\round_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_round_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_cross_review_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\models.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0014_workflow_generation_rounds.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\App.vue"
|
|||
|
|
],
|
|||
|
|
"summary": "轮次状态机最致命的问题是写入口分散且缺少统一状态守卫:`WorkflowRoundRepository.update_round()` 是裸 UPDATE,而 `_ensure_round_enqueued()` 与 `reconcile_round()` 都绕过 `_ensure_writable_round_row()` 无条件改写 status,导致 `confirmed`/`superseded` 两个终态可被普通幂等重放或迟到的 worker 回调回退成 `queued`/`completed`,进而击穿\"已定稿轮次只读\"和跨评审的 `confirmed_candidate_mismatch` / 下游已完成保护。分支层面 `create_next_round` 把父轮次标记 superseded 后既无回写保护也无子轮次唯一性守卫(交叉评审侧有 `next_review_round_id` 守卫,主轮次侧没有),残留候选完成时会把节点 `active_round_id` 指回旧轮次。反馈解释器两条路径都不健壮:LLM 返回扁平结构时候选偏好被静默丢弃且不回退,确定性回退路径把用户文本里的 1~3 位数字当候选编号(实测\"不要 3 个人\"会静默拒掉 TX-01-03,\"分辨率 512\"直接 400)。此外 `list_node_rounds → _persisted_detail` 是 N+1 且每次仓储调用新建一条 SQLite 连接,交叉评审的 `_round_replay` 只按全局唯一的 idempotency_key 匹配而不校验归属,skip 流程由三段独立事务组成会留下\"已跳过未定稿\"的死锁态。",
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P0-致命",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "update_round 无状态机守卫:已确认(confirmed)轮次可被幂等重放静默回退,击穿只读锁与跨评审校验",
|
|||
|
|
"location": "backend/services/workflow_round_service.py:115",
|
|||
|
|
"problem": "create_round 命中 idempotency_key 重放分支时无条件调用 _ensure_round_enqueued(existing, ctx),该方法末尾(:928-939)又无条件把轮次 status 写成 \"queued\",随后 reconcile_round 依据候选聚合改成 completed。整条链路不检查 existing.status 是否已是 confirmed/superseded 终态,而 update_round(workflow_round_repository.py:416-431)只是裸 UPDATE,_ensure_writable_round_row 的 confirmed 只读保护(:1364-1369)完全没有参与。",
|
|||
|
|
"evidence": "# workflow_round_service.py:115-117\n if self.generation_service is not None:\n self._ensure_round_enqueued(existing, ctx)\n return {**self.get_round(existing.round_id).to_dict(), \"created\": False, \"idempotent_replay\": True}\n\n# workflow_round_service.py:928-940\n latest = self.round_repository.get_round(round_record.round_id) or round_record\n self.round_repository.update_round(\n round_record.round_id,\n {\n \"status\": \"queued\",\n \"metadata\": {**latest.metadata, \"planned_only\": False, \"task_ids\": task_ids},\n },\n )\n self.reconcile_round(round_record.round_id, ctx)",
|
|||
|
|
"impact": "客户端用同一 idempotency_key 重试 POST nodes/{id}/rounds(网络超时后的正常重试即可触发,无需并发),已 confirmed 的轮次退回 completed。后果:(1) _ensure_writable_round_row 不再拦截,apply_feedback_atomic 可继续改写已定稿轮次;(2) workflow_cross_review_repository.py:1066-1076 的 confirmed_candidate_mismatch 校验与 :1108-1111 的 _ensure_review_after_confirmation_is_safe 均只在 status=='confirmed' 时生效,被跳过后用户可对与已定稿候选不同的候选发起跨评审;(3) record_skip_restoration_handoff_atomic(:795) 要求 confirmed,回退后 XF 交接永久失败。节点 selected_image_id 仍指旧图,两边彻底不一致。",
|
|||
|
|
"fix": "1) 为 WorkflowRoundRepository 增加 update_round_if_status(round_id, expected_statuses, payload)(可参照已有的 update_workflow_if_status / update_node_if_status),或在 update_round 内对 status 赋值做终态白名单校验,confirmed/superseded/cancelled 一律拒绝降级;2) create_round 幂等重放分支前置判断 `if existing.status in {\"confirmed\",\"superseded\",\"cancelled\"}: return ...` 直接返回不再 enqueue;3) _ensure_round_enqueued 末尾改为 `\"status\": \"queued\" if task_ids else latest.status` 并跳过终态轮次。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "reconcile_round 无条件回写状态与节点指针:迟到的候选回调会把 superseded 父轮次\"复活\"并把节点 active_round_id 指回旧轮次",
|
|||
|
|
"location": "backend/services/workflow_round_service.py:818",
|
|||
|
|
"problem": "create_next_round 只要求父轮次有 1~3 个 succeeded 的 liked 候选,get_consistent_round_snapshot 没有任何 status 校验,因此父轮次仍有 queued/running 候选时即可创建子轮次并把父轮次标为 superseded。这些残留候选完成后,handle_task_terminal → reconcile_round(parent) 会用聚合状态覆盖 superseded,并把节点 metadata.active_round_id 重新指向父轮次、用父轮次的聚合结果覆盖节点 status。同时 _ensure_writable_round_row 的不可写集合里没有 superseded。",
|
|||
|
|
"evidence": "# workflow_round_service.py:818-847\n updated_round = self.round_repository.update_round(\n round_id,\n {\n \"status\": round_status, # 直接覆盖 superseded / confirmed\n ...\n },\n ) or round_record\n if self.generation_service is not None:\n ...\n if node.status not in {\"selected\", \"discarded\"}:\n update_payload = {\n \"status\": node_status,\n \"metadata\": {\n **metadata,\n \"active_round_id\": round_id, # 指回被 reconcile 的那个轮次\n \"last_round_status\": round_status,\n\n# workflow_round_repository.py:1370-1375(不可写集合缺少 superseded)\n if status in {\"cancelled\", \"all_failed\"}:\n raise WorkflowRoundWriteError(\"round_not_writable\", ...)",
|
|||
|
|
"impact": "1) 父轮次从 superseded 变回 completed/partial_success,同一节点出现两个可确认轮次,用户可能在旧轮次上定稿绕过刚创建的优化子轮次;2) 节点 metadata.active_round_id 被改回父轮次,前端 App.vue:2336 的 selectedWorkflowActiveRoundId 直接采信该值(不校验是否在已加载轮次列表中,与 App.vue:7048 的加载逻辑不一致),面板跳回旧轮次;3) 子轮次还在跑时节点已被标 completed;4) superseded 轮次仍可写反馈,且可再次 create_next_round。",
|
|||
|
|
"fix": "1) _reconcile_round_locked 先读 round_record.status,若为 confirmed/superseded/cancelled 则只更新 metadata.execution_summary,不写 status,也不改节点 active_round_id;2) 更新节点前校验 round_id 仍是该节点的活动轮次(== node.metadata['active_round_id'] 或该节点最新的非终态轮次),避免旧轮次抢占节点指针;3) 把 superseded 加入 _ensure_writable_round_row 的不可写状态集合。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "LLM 返回扁平结构时,候选偏好被静默丢弃(不报错、不回退到确定性解析)",
|
|||
|
|
"location": "backend/services/workflow_round_feedback_service.py:178",
|
|||
|
|
"problem": "_model_request 下发的 response_schema 把 liked/rejected/strict_rejected/final 嵌在 feedback 下,_normalize_model_payload 也只从 parsed[\"feedback\"] 取这四个键。模型按常见习惯返回扁平结构(顶层直接给 liked)时 feedback 为空 dict,候选偏好全部丢失,而 positive_patch 因为单独从顶层取被保留,于是 validate_llm_result 通过、source=\"llm_validated\"、fallback_used=False,既不报错也不回退到确定性解析器。",
|
|||
|
|
"evidence": "# workflow_round_feedback_service.py:178-189\n feedback = parsed.get(\"feedback\")\n feedback = dict(feedback) if isinstance(feedback, Mapping) else {}\n merged = {\n **feedback, # 顶层的 liked/rejected/final 在这里被丢掉\n \"positive_patch\": parsed.get(\"positive_patch\", feedback.get(\"positive_patch\", parsed.get(\"next_round_prompt_patch\"))),\n \"negative_patch\": parsed.get(\"negative_patch\", feedback.get(\"negative_patch\")),\n \"material_intent\": parsed.get(\"material_intent\", feedback.get(\"material_intent\")),\n }\n return merged\n\n# 本机实测:\n# flat {'liked': ['TX-01-02'], 'positive_patch': 'more light'} -> source=llm_validated liked=[] fallback_used=False\n# nested {'feedback': {'liked': ['TX-01-02']}, ...} -> source=llm_validated liked=['TX-01-02']",
|
|||
|
|
"impact": "用户说\"保留第 2 张并加点光\",系统只记下 patch、丢掉\"保留第 2 张\"。随后 create_next_round 因 liked 为空直接 400(workflow_round_service.py:504-514 \"select one to three liked candidates\"),用户看不到任何\"模型没理解\"的提示。若同一请求还包含 rejected,用户的拒绝意图同样被无声吞掉。",
|
|||
|
|
"fix": "_normalize_model_payload 对四个动作字段做双层回退 `feedback.get(k, parsed.get(k))`;并在所有动作字段解析结果均为空、但 parsed 顶层存在同名键时抛 WorkflowFeedbackInterpretationError(\"invalid_llm_result\", ...),让 interpret() 的 except Exception 走确定性回退,而不是\"半成功\"。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "确定性反馈解析器把普通数字当候选编号:静默误拒候选 / 合法反馈被 400",
|
|||
|
|
"location": "backend/services/workflow_feedback_interpreter.py:46",
|
|||
|
|
"problem": "_DISPLAY_CODE_OR_NUMBER_RE 里包含裸 \\d{1,3},任何 1~3 位数字都被当成候选序号;interpret 主循环(:172-208)还会把\"没有动作词但有数字\"的后续片段挂到上一个动作(active_prefix_action)。这是 LLM 不可用时的默认路径(round_routes.py:143-150:未配置 prompt_planning 路由时 feedback_callable=None)。",
|
|||
|
|
"evidence": "# workflow_feedback_interpreter.py:46-54\n_DISPLAY_CODE_OR_NUMBER_RE = re.compile(\n r\"(?<![A-Za-z0-9])\"\n r\"(?:\"\n r\"(?=[A-Za-z0-9-]*\\d)[A-Za-z][A-Za-z0-9]*(?:-[A-Za-z0-9]+)+\"\n r\"|\\d{1,3}\" # ← 任意 1~3 位数字\n r\")\"\n r\"(?![A-Za-z0-9])\",\n re.IGNORECASE,\n)\n\n# 本机实测(白名单 TX-01-01..TX-01-04):\n# \"喜欢 TX-01-01,不要 3 个人\" -> liked=['TX-01-01'] rejected=['TX-01-03'] # 静默误拒\n# \"拒绝 02,分辨率 512\" -> WorkflowFeedbackInterpretationError(unknown_candidate_reference, reference='512')",
|
|||
|
|
"impact": "误拒路径:update_feedback 会把 TX-01-03 写成 rejected 并落 pool_state='rejected' + history,用户完全不知情,且该候选此后无法 confirm(confirm_candidate_atomic 拒绝 rejected 候选)。400 路径:只要反馈里出现 3 位以内数字(尺寸、人数、年龄、角度…),整个反馈请求被 ApiError(VALIDATION_FAILED) 拒绝,错误信息对用户毫无意义。",
|
|||
|
|
"fix": "1) 裸数字只在紧邻候选语境时才视为候选引用——要求前置标记(候选/第/#/No./张/号),把无标记的 \\d{1,3} 从 _DISPLAY_CODE_OR_NUMBER_RE 移除,交给 _CHINESE_ORDINAL_RE 这类带标记规则;2) 解析出的数字若不在白名单终端序号集合内应当忽略而非抛 unknown_candidate_reference(仅带显式候选标记时才报错);3) active_prefix_action 跨片段续接时同样只接受带标记 token。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "create_next_round 缺少\"一个父轮次只能有一个子轮次\"的守卫,可分叉出多条优化链",
|
|||
|
|
"location": "backend/services/workflow_round_service.py:620",
|
|||
|
|
"problem": "交叉评审侧有明确守卫(workflow_cross_review_repository.py:171-182 检查 next_review_round_id 已存在则报 cross_review_followup_exists),而主轮次的 create_next_round 只在创建完成后写 superseded_by_round_id,创建前不检查它是否已存在。唯一的并发保护是 parent_feedback_revision 乐观校验(workflow_round_repository.py:162-174),但它只在父轮次 feedback 发生变化时才拒绝;两次内容相同的请求 revision 一致,都会成功。",
|
|||
|
|
"evidence": "# workflow_round_service.py:620-633\n created = self.create_round(parent.node_id, next_payload, ctx)\n superseded_parent = self.round_repository.update_round(\n parent_round_id,\n {\"status\": \"superseded\",\n \"metadata\": {**parent.metadata, \"superseded_by_round_id\": created.get(\"round_id\"), ...}},\n )\n\n# 对照 workflow_cross_review_repository.py:171-182\n existing_child = _optional_text(parent_metadata.get(\"next_review_round_id\"))\n if existing_child:\n raise WorkflowRoundWriteError(\"cross_review_followup_exists\", ...)",
|
|||
|
|
"impact": "同一父轮次产生两个平级子轮次,superseded_by_round_id 被后者覆盖,前者成为无人引用的孤儿优化链;节点 active_round_id 在两条链之间来回跳;feedback_context.parent_round_id 追溯信息失真。",
|
|||
|
|
"fix": "把 superseded_by_round_id 的存在性检查移进 create_round_with_candidates 的同一事务(该事务已对 parent 行加了 FOR UPDATE / BEGIN IMMEDIATE),存在即抛 next_round_exists;或在 workflow_generation_rounds(parent_round_id, round_kind) 上加部分唯一索引。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "list_node_rounds 的 N+1,且每次仓储调用都新建一条 SQLite 连接",
|
|||
|
|
"location": "backend/services/workflow_round_service.py:982",
|
|||
|
|
"problem": "list_node_rounds 对该节点的每一个轮次调用 _persisted_detail,后者对每一个候选再各查一次 image 和 task;而所有仓储方法都是 conn = self.database.connect() … conn.close(),即每次查询都新建一条 sqlite3.connect()(还要跑两条 PRAGMA,db_helper.py:159-170)。",
|
|||
|
|
"evidence": "# workflow_round_service.py:982-998\n def _persisted_detail(self, round_record: WorkflowGenerationRound) -> WorkflowGenerationRoundDetail:\n candidates = {item.candidate_id: item for item in self.round_repository.list_candidates(round_record.round_id)}\n views: list[WorkflowRoundCandidateView] = []\n for pool in self.round_repository.list_pool(round_record.round_id):\n candidate = candidates.get(pool.candidate_id) or self.round_repository.get_candidate(pool.candidate_id)\n if candidate is None:\n continue\n image = self.image_repository.get(candidate.image_id) if candidate.image_id else None\n task = self.job_repository.get(candidate.current_task_id) if candidate.current_task_id else None",
|
|||
|
|
"impact": "一个有 6 轮 × 4 候选的节点,单次 GET nodes/{id}/rounds 约产生 3 + 6×(2 + 4×2) ≈ 63 条独立连接/查询;前端 App.vue 的 scheduleWorkflowRoundPolling 在轮次运行期间还会周期性重拉该接口。SQLite 单写锁下会与 worker 的候选写入争抢 busy_timeout,放大 database is locked 风险。",
|
|||
|
|
"fix": "1) 给 WorkflowRoundRepository 增加批量方法 list_candidates_for_rounds(round_ids) / list_pool_for_rounds(round_ids)(WHERE origin_round_id IN (...)),一次取回全部轮次的候选与池;2) ImageRepository / JobRepository 增加 get_many(ids),_persisted_detail 改为接收预取字典;3) 中期给 Database 加连接复用(线程本地连接或连接池)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "交叉评审的幂等重放只按 idempotency_key 全表匹配,可命中另一个节点/另一种轮次并对其改状态",
|
|||
|
|
"location": "database/repositories/workflow_cross_review_repository.py:969",
|
|||
|
|
"problem": "idempotency_key 在 workflow_generation_rounds 上是全表唯一(0014 迁移的 idx_workflow_rounds_idempotency),主轮次、交叉评审轮次、XF 交接轮次共用同一键空间。但 _round_replay 只按 key 查,不校验命中的轮次是否属于本次请求的 source round / node / round_kind;对照主轮次侧 workflow_round_service.py:94-114 是有 node_id 与 parent_round_id 校验的。",
|
|||
|
|
"evidence": "# workflow_cross_review_repository.py:969-978\n def _round_replay(self, conn: Any, idempotency_key: str | None) -> dict[str, Any] | None:\n if not idempotency_key:\n return None\n row = conn.execute(\n f\"SELECT round_id, display_code FROM workflow_generation_rounds WHERE idempotency_key = {self.placeholder}\",\n (idempotency_key,),\n ).fetchone()\n if row is None:\n return None\n return {\"round_id\": row[\"round_id\"], \"display_code\": row[\"display_code\"]}\n\n# 调用方立刻对命中的轮次做写操作:workflow_generation_cross_review_service.py:389\n self._ensure_enqueued(str(created[\"round_id\"]), ctx) # 内部 update_round(status/metadata) + reconcile_round",
|
|||
|
|
"impact": "客户端在不同操作上复用同一 idempotency_key(UI 用同一个操作 id、聊天侧用消息 id 做 key 都会踩到)时,交叉评审请求会\"成功\"返回一个完全无关的轮次,并覆盖它的 status / metadata.task_ids、触发针对该轮次的 reconcile(进而改写其节点的 active_round_id)。跨用户场景下 round_service.get_round(..., ctx) 的 require_node 会在最后一步拦住读取,但写入已经发生。",
|
|||
|
|
"fix": "_round_replay 增加参数并校验:命中行的 round_kind 必须匹配、node_id/workflow_id 必须等于 round_payload 中的值、parent_round_id 必须等于本次的 source/parent;不匹配时抛 WorkflowRoundWriteError(\"idempotency_key_conflict\", ...) 而不是当作重放。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "跨评审 skip 由三段独立事务组成,中途失败会留下\"已跳过但未定稿\"的死锁态",
|
|||
|
|
"location": "backend/services/workflow_generation_cross_review_service.py:416",
|
|||
|
|
"problem": "跳过流程依次执行三个各自独立提交的事务:① record_skip_atomic(写 generation_cross_review.decision='skip' + history)→ ② round_service.confirm_candidate(另一连接里的 confirm_candidate_atomic)→ ③ record_skip_restoration_handoff_atomic(要求 status == 'confirmed')。② 的失败原因很现实:round_branch_not_active(另一标签页切了分支)、node_already_completed、candidate_image_restricted。",
|
|||
|
|
"evidence": "# workflow_generation_cross_review_service.py:416-445\n try:\n decision = self.cross_review_repository.record_skip_atomic(...) # ① 已提交\n except WorkflowRoundWriteError as exc:\n raise _write_error(exc) from exc\n if decision.get(\"already_confirmed\"):\n result = self.round_service.get_round(source_round_id, ctx).to_dict()\n else:\n result = self.round_service.confirm_candidate( # ② 可能抛错\n source_round_id,\n {\"candidate_id\": candidate_id, \"idempotency_key\": f\"{idempotency_key}:confirm\" if idempotency_key else None},\n ctx,\n )\n try:\n restoration_handoff = (\n self.cross_review_repository.record_skip_restoration_handoff_atomic(...) # ③ 要求 confirmed\n )",
|
|||
|
|
"impact": "轮次 metadata 已写死 decision='skip' 但轮次未 confirmed、XF 输入库无条目。重试时 record_skip_atomic 的 `existing.get(\"decision\") == \"skip\"` 分支会当成幂等重放返回 already_confirmed=False,然后又去 confirm_candidate,永远卡在同一失败点;而改走 review 的 create_review_round_atomic 不检查 decision=='skip',会在已标记跳过的轮次上再建评审轮次,两个状态并存。",
|
|||
|
|
"fix": "1) 把\"记录 skip + 确认候选\"合并为一个事务(record_skip_atomic 内部完成 confirm 所需全部写入,或让 confirm_candidate_atomic 接受 review_decision 参数在同一 conn 内落库);2) 若暂不合并,则在 record_skip_atomic 里区分中间态(如 decision_state='pending_confirm')并允许显式撤销该跳过决定;3) create_review_round_atomic 增加 decision == 'skip' 的互斥校验。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "apply_feedback_atomic 会把旧的 _conversation 无条件带到新快照,导致 create_next_round 复用过期的 prompt patch",
|
|||
|
|
"location": "database/repositories/workflow_round_repository.py:710",
|
|||
|
|
"problem": "每次反馈写入用池里最新偏好重建 feedback_snapshot,然后把上一版快照里所有 _ 前缀的键原样复制回来。用户第一次用自然语言反馈(写入 _conversation),第二次改用结构化 updates(conversation is None,不覆盖)时,旧的 _conversation 被保留,而 create_next_round 直接读取它作为 positive_patch / material_intent / source_text。",
|
|||
|
|
"evidence": "# workflow_round_repository.py:703-716\n feedback_snapshot = {str(row[\"candidate_id\"]): {...} for row in pool_rows}\n for key, value in previous_snapshot.items():\n if str(key).startswith(\"_\"):\n feedback_snapshot[str(key)] = value # ← 旧 _conversation 原样带过来\n ...\n if conversation is not None:\n feedback_snapshot[\"_conversation\"] = conversation\n\n# workflow_round_service.py:534-547\n conversation = parent.feedback_snapshot.get(\"_conversation\")\n conversation = dict(conversation) if isinstance(conversation, dict) else {}\n positive_patch = _bounded_optional_text(conversation.get(\"positive_patch\"), field=\"positive_patch\", maximum=4_000)\n material_intent = _bounded_optional_text(conversation.get(\"material_intent\"), field=\"material_intent\", maximum=2_000)",
|
|||
|
|
"impact": "用户第一轮说\"下一轮加暖光\",之后改口用界面按钮重新勾选偏好(不再提暖光),生成的子轮次仍会带上\"加暖光\"的 prompt_patch 与 material_intent,且 feedback_context.source_text 显示的是早已作废的原话;用户无法通过 UI 清除该 patch。",
|
|||
|
|
"fix": "在 _conversation 里记录写入序号/时间戳及其依据的 pool revision,create_next_round 只在 _conversation 对应最近一次反馈写入时才采用其 patch;或更简单:结构化 updates 写入时显式 feedback_snapshot.pop(\"_conversation\", None),并在 API 上暴露\"清除对话补丁\"的能力。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "反馈幂等日志按字典序而非时间序裁剪,会淘汰掉较新的 key",
|
|||
|
|
"location": "database/repositories/workflow_round_repository.py:726",
|
|||
|
|
"problem": "request_log 从 metadata_json 反序列化得到,而写库用的是 json.dumps(..., sort_keys=True)(workflow_round_repository.py:1318-1321),所以重新读出的插入顺序是字典序而非写入时间顺序;[-20:] 因此保留的是\"字典序最大的 20 个 key\"。",
|
|||
|
|
"evidence": "# workflow_round_repository.py:719-726\n if idempotency_key:\n request_log[idempotency_key] = {\n \"request_fingerprint\": request_fingerprint,\n \"history_id\": history_id,\n \"updated_by\": updated_by,\n \"updated_at\": datetime.now(timezone.utc).isoformat(),\n }\n request_log = dict(list(request_log.items())[-20:])\n\n# workflow_round_repository.py:1318-1321\ndef _json_or_none(value: Any | None) -> str | None:\n if value is None:\n return None\n return json.dumps(value, ensure_ascii=False, sort_keys=True)",
|
|||
|
|
"impact": "同一轮次上做超过 20 次带 key 的反馈写入后,被淘汰的是字典序小的 key(例如以 a 开头的 UUID)而非最旧的。用户对刚提交过的请求做重试时幂等记录可能已被淘汰,导致重复写入 round_feedback_updated history 并重复触发审计事件。",
|
|||
|
|
"fix": "记录里已有 updated_at 字段,裁剪时按它排序:request_log = dict(sorted(request_log.items(), key=lambda kv: str(kv[1].get(\"updated_at\") or \"\"))[-20:])"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "approve 响应里对同一个 next_node_id 重复查询节点",
|
|||
|
|
"location": "backend/services/workflow_generation_cross_review_service.py:262",
|
|||
|
|
"problem": "条件表达式里把 get_node 调了两次(一次判空、一次取值),既是多余的 DB 往返(每次都新建一条连接),也存在 TOCTOU——两次之间节点被删则第二次返回 None,.to_dict() 抛 AttributeError。",
|
|||
|
|
"evidence": "# workflow_generation_cross_review_service.py:259-267\n return {\n **self.round_service.get_round(round_id, ctx).to_dict(),\n \"approval\": approval,\n \"next_node\": (\n self.workflow_repository.get_node(str(approval[\"next_node_id\"])).to_dict()\n if approval.get(\"next_node_id\") and self.workflow_repository.get_node(str(approval[\"next_node_id\"]))\n else None\n ),\n }",
|
|||
|
|
"impact": "多一次数据库连接;极端竞态下返回 500 而非正常响应。",
|
|||
|
|
"fix": "先取一次再判空:next_node = self.workflow_repository.get_node(str(approval[\"next_node_id\"])) if approval.get(\"next_node_id\") else None,然后 \"next_node\": next_node.to_dict() if next_node else None"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "生成流水线 / 任务队列 / 重试 / 并发",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A1-生成与队列.md",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\worker_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_job_retry_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\cost_guard_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\cost_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\job_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\queue_manager.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\fsm.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\event_bus.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\job_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\single_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\image_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py"
|
|||
|
|
],
|
|||
|
|
"summary": "队列的状态机与 CAS 原语(update_if_status / delete_if_status / cancel_job 重读循环)设计得相当认真,这一层没有明显的丢更新。真正的问题全部出在队列语义之外:QueueService.update_task() 是无条件 UPDATE 却被 worker 的成功与失败路径全量使用,导致并发取消会被静默覆盖;worker 认领任务复用了展示用的 ORDER BY created_at DESC 查询,队列实际是 LIFO,积压超过 claim_batch_size 后老任务永久饿死。生命周期上,QueueWorker.stop() 的 join 超时(5s)远小于 provider 超时(最长 3600s),而 Main.terminate() 无视 running=True 直接丢弃 runtime,插件热重载会造成新旧两套 worker 池并发抢占同一队列并重复计费。重试链路整体断裂:round pool 传 RetryPolicy(max_attempts=1) 等于关闭重试,Web 重试又因 task_type=\"workflow_round_candidate\" 被 _validate_executable 硬拒,候选一次网络抖动即终态失败且无恢复通道;_BoundedWorkerHandler 的限流在生产配置下(per_route_limit == worker_count)完全是空操作。此外同步生成路径在 asyncio 默认线程池中长时间阻塞且无并发上限,recover_stale_tasks 只在 worker 启动时执行一次并且只认 source=workflow_round,同步生成崩溃残留会把节点永久锁死。共 12 条发现(P1×5、P2×5、P3×3 中合并计数为 12 条)。",
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "QueueService.update_task() 无 CAS,worker 完成写入会覆盖并发取消",
|
|||
|
|
"location": "backend/services/queue_service.py:406",
|
|||
|
|
"problem": "update_task 只做「读 → 校验迁移 → 无条件 UPDATE」,底层 JobRepository.update 的 WHERE 子句只有 task_id,没有任何状态谓词。而 QueueWorker.run_once 的成功路径(worker_service.py:116)与 _handle_failure 的失败/重排队路径(worker_service.py:272、308)全部调用它。仓库里明明已有 update_task_if_status / update_if_status 这套 CAS 原语,worker 却没有使用。",
|
|||
|
|
"evidence": "# queue_service.py:406\ndef update_task(self, task_id, payload, ctx) -> GenerationTask:\n before = self.get_job(task_id) # ← 读\n status = payload.get(\"status\")\n if status is not None:\n self._validate_task_status_transition(before.status, str(status))\n after = self.repository.update(task_id, payload) # ← 无条件写\n\n# job_repository.py:215\nconn.execute(f\"\"\"UPDATE generation_tasks SET {\", \".join(assignments)}\n WHERE task_id = {placeholder}\"\"\", tuple(params)) # ← 只有 task_id",
|
|||
|
|
"impact": "用户点取消、worker 恰好同时出图完成:取消在 update_task 的 get_job 之后、UPDATE 之前提交,则 cancelled 被直接改写为 succeeded(或被 _handle_failure 改写为 failed / queued)。取消语义失效,已取消的轮次重新变成运行中,cancel_workflow_jobs 也可能被同样方式回滚,用户已明确取消却仍被扣费并产生图片记录。",
|
|||
|
|
"fix": "worker 三处写入全部换成 CAS:run_once 成功路径改为 update_task_if_status(task_id, {\"running\"}, payload, ctx, require_workflow_not_cancelled=True),返回 None 时按「任务已被他人接管」处理并清理已生成产物;_handle_failure 的 failed 写入用 {\"running\"} 作期望态、重排队写入用 {\"failed\"} 作期望态。长期把 QueueService.update_task 标记为管理用途,禁止 worker 调用。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "队列实际是 LIFO:ORDER BY created_at DESC + 固定 batch 导致老任务永久饿死",
|
|||
|
|
"location": "database/repositories/job_repository.py:150",
|
|||
|
|
"problem": "worker 认领任务复用了给前端做「最新在前」的通用列表接口 list_jobs(status=\"queued\", page=1, page_size=claim_batch_size),该接口使用 ORDER BY created_at DESC。于是队列变成后进先出,且每次只看最新的 claim_batch_size(round pool 为 25)条。",
|
|||
|
|
"evidence": "# job_repository.py:145-154\nrows = conn.execute(f\"\"\"\n SELECT {TASK_COLUMNS}\n FROM generation_tasks\n {where_clause}\n ORDER BY created_at DESC # ← 队列消费用了倒序\n LIMIT {placeholder} OFFSET {placeholder}\n\"\"\", tuple(params + [normalized_page_size, offset])).fetchall()\n\n# worker_service.py:207-212\ndef _claim_next(self, ctx):\n queued = self.queue_service.list_jobs(status=\"queued\", page=1, page_size=self.claim_batch_size).items\n for candidate in queued:\n if _retry_deferred(candidate):\n continue",
|
|||
|
|
"impact": "只要待处理队列长度稳定 > 25,最早入队的候选永远排不到(新任务不断插到窗口顶端)。用户表现为某个轮次一直卡在 queued,而后创建的轮次先出图。_retry_deferred 被延迟的任务还会白占这 25 个名额,进一步缩小可认领窗口。SQLite 的 created_at 是秒级 CURRENT_TIMESTAMP,同一秒创建的 4~8 个候选顺序完全不确定。",
|
|||
|
|
"fix": "给 JobRepository 增加专用认领查询,ORDER BY created_at ASC, task_id ASC(FIFO + 稳定次序),让 _claim_next 调用它而非复用展示用的 list;同时把 created_at 精度提升到毫秒或增加自增 seq 列以消除同秒歧义。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "插件重载后旧 worker 线程仍存活,与新 worker 池并发抢占同一队列",
|
|||
|
|
"location": "backend/services/worker_service.py:180",
|
|||
|
|
"problem": "QueueWorker.stop() 只是 set() 停止事件后 join(timeout=5s)。停止事件仅在 _run_loop 循环顶部被检查,而 run_once 内部会阻塞在 provider HTTP 调用上(engines/image_provider.py:308-313 允许 10~3600 秒)。join 超时后 stop() 返回 running: True,而 Main.terminate() 无视该返回值把 runtime 置 None。",
|
|||
|
|
"evidence": "# worker_service.py:180-185\ndef stop(self, *, timeout_seconds: float = 5.0) -> dict[str, Any]:\n self._stop_event.set()\n thread = self._thread\n if thread is not None:\n thread.join(timeout=max(float(timeout_seconds), 0.1))\n return {\"worker_id\": self.worker_id, \"running\": self.is_running()} # ← 可能是 True\n\n# main.py:436-439\nif self.round_worker_runtime is not None:\n self.round_worker_result = self.round_worker_runtime.stop()\n self.round_worker_runtime = None # ← 不管 running 是否 True 都丢弃",
|
|||
|
|
"impact": "AstrBot 热重载时旧的 5 个 daemon 线程仍在跑(进程未退出),新的 create_workflow_round_worker_runtime() 又起 5 个。CAS 能防止同一条任务被同时认领,但旧线程完成时会用无条件 update_task 写回结果,而新进程的 recover_stale_tasks() 又会把超过 300 秒仍 running 的任务(可能正被旧线程执行)标记 failed 后重新入队 → 同一候选生成两次、重复计费重复落图。worker 数量随每次重载线性膨胀。",
|
|||
|
|
"fix": "1) stop() 在 join 超时后返回明确的 stopped: False,QueueWorkerPool.stop() 聚合该标志;2) Main.terminate() 在 stopped=False 时保留 runtime 引用并拒绝再次 start(),或将 stop 超时提升到略大于 provider 超时;3) 给每个 worker 分配进程级唯一 worker_uuid,update_task 时把 metadata.worker.worker_id 作为 CAS 条件之一,旧进程写入直接被拒;4) recover_stale_tasks 增加当前进程 worker 心跳判断。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "轮次候选任务无任何重试路径:worker 侧 max_attempts=1,Web 侧 task_type 被硬拒",
|
|||
|
|
"location": "backend/services/workflow_round_worker_runtime.py:45",
|
|||
|
|
"problem": "两条重试通道同时被关死。worker 侧 round pool 显式传 RetryPolicy(max_attempts=1),而 should_retry 判定是 if attempt >= max(int(self.max_attempts), 1): return False;首次执行 attempt 已经是 1(claim_queued_task 先自增 worker_attempt),所以永远返回 False。Web 侧候选任务 task_type 是 workflow_round_candidate,而 GenerationJobRetryService._validate_executable 只接受 generation。",
|
|||
|
|
"evidence": "# workflow_round_worker_runtime.py:39-49\nself.pool = QueueWorkerPool(\n self.queue_service, self._execute,\n retry_policy=RetryPolicy(max_attempts=1), # ← 等于关闭重试\n task_filter=_is_workflow_round_task, ...)\n\n# generation_service.py:361\n\"task_type\": \"workflow_round_candidate\", # ← 非 \"generation\"\n\n# generation_job_retry_service.py:181\nif task.task_type != \"generation\" or not task.node_id or not isinstance(task.request, dict):\n raise ApiError(ErrorCode.TASK_NOT_RETRYABLE, \"task has no executable generation request\", ...)",
|
|||
|
|
"impact": "详细工作流一个候选遇到 429 / 连接重置 / provider 5xx 立即变成终态 failed;/jobs/{id}/retry 返回 TASK_NOT_RETRYABLE;workflow_round_service 里也没有候选级重试接口。用户唯一出路是重建整轮(重新走 prepare_persisted_generation,成功的候选一起重做),成本与语义都错。worker_service 的 _retry_deferred / next_retry_at 退避逻辑和 workflow_round_service.py:750 的 retry_scheduled 分支全部成为死代码。",
|
|||
|
|
"fix": "1) round pool 改为 RetryPolicy(max_attempts=3, base_delay_seconds=2.0, backoff_multiplier=2.0, retryable_error_types=(\"ImageProviderError\",)),并统一 should_retry 中 attempt 与 max_attempts 的语义;2) _validate_executable 放开 task_type in {\"generation\", \"workflow_round_candidate\"},候选类型走 enqueue_prepared_generation 重新入队而非 generate_node 同步执行;或在 WorkflowRoundService 上新增 retry_candidate(round_id, candidate_id)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "同步生成占满 asyncio 默认线程池,全插件无并发上限与背压",
|
|||
|
|
"location": "backend/services/generation_service.py:119",
|
|||
|
|
"problem": "nodes/generate(Web 与聊天共用)走完全同步的 _generate_node_locked:任务刚 create_task(status=\"queued\") 就立刻 CAS 成 running 并在同一调用栈里发起 provider 请求,整个过程持有进程内 per-node RLock。所有 Web/桥接请求又都通过 asyncio.to_thread(dispatch_web_request, ...) 派发到 asyncio 默认线程池(min(32, cpu+4))。没有任何信号量或专用 executor 限制并发生成数量。",
|
|||
|
|
"evidence": "# generation_service.py:119-126\ndef generate_node(self, node_id, payload, ctx) -> dict[str, Any]:\n lock_entry = _retain_node_generation_lock(node_id)\n lock_entry.lock.acquire()\n try:\n return self._generate_node_locked(node_id, payload, ctx) # ← 内含 provider 阻塞调用\n\n# main.py:333\nresponse = await asyncio.to_thread( # ← 默认 executor,全插件共用\n dispatch_web_request, ctx, dispatch_method, dispatch_endpoint, ...)\n\n# chat_workflow_service.py:216\nwith ThreadPoolExecutor(max_workers=4, thread_name_prefix=\"imagestudio-chat-fast\") as executor:\n futures = {executor.submit(self.generate_node, str(node[\"node_id\"]), chat): node for node in nodes}",
|
|||
|
|
"impact": "单次 provider 调用默认 180 秒、最高允许 3600 秒。十几个并发 nodes/generate 就能把默认 executor 占满,此后所有 ImageStudio Web API(列表、状态轮询、取消)一起挂起——包括用来取消这些生成的接口。快速角色命令还会为每个聊天用户额外派生 4 个线程池线程,且 with 退出时 shutdown(wait=True) 不可中断,N 个用户 = 4N 路并发 provider 请求,成本无上限。",
|
|||
|
|
"fix": "1) 为生成建立独立的、容量可配的 ThreadPoolExecutor(4~8),dispatch_web_request 继续用默认池,二者隔离;2) 在 GenerationService 入口加进程级 threading.BoundedSemaphore(max_concurrent_generations),取不到时快速返回 429 / QUEUE_BUSY 而非排队阻塞;3) chat_workflow_service 的 4 图并发改为复用全局生成池而非每次新建。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "_BoundedWorkerHandler 的并发限流在生产配置下完全是空操作,且限流点位于认领之后",
|
|||
|
|
"location": "backend/services/worker_service.py:400",
|
|||
|
|
"problem": "全局信号量容量等于 worker 线程数,per-route 信号量在唯一的生产构造里也等于 worker 线程数(workflow_round_worker_runtime.py:48 传 per_route_limit=worker_count)。线程数 == 许可数,acquire() 永远不会阻塞。更根本的是限流器包在 handler 上,而 run_once 是先 _claim_next() 把任务 CAS 成 running 再调用 handler。",
|
|||
|
|
"evidence": "# worker_service.py:400-405\ncount = min(max(int(worker_count), 1), 16)\nbounded_handler = _BoundedWorkerHandler(\n handler,\n global_limit=count, # ← 等于线程数\n per_route_limit=min(max(int(per_route_limit or count), 1), count),\n)\n\n# worker_service.py:451-459\ndef __call__(self, task):\n route_key = _task_route_key(task)\n with self._route_guard:\n route_semaphore = self._route_semaphores.setdefault(route_key, threading.BoundedSemaphore(self.per_route_limit))\n with self.global_semaphore, route_semaphore: # ← 任务此时已是 running\n return self.handler(task)",
|
|||
|
|
"impact": "对同一 provider/model 的突发保护形同虚设(本意应是防止把单个模型打爆触发 429)。若日后按注释意图调小 per_route_limit,则立刻引入「已认领任务排队等待信号量」的队头阻塞——被阻塞线程拿着 running 状态的任务空等,updated_at 停滞,最终被 recover_stale_tasks 误判为上个进程遗留并重复执行。",
|
|||
|
|
"fix": "把限流前移到认领之前:_claim_next 里按 route_key 先 acquire(blocking=False),拿不到就跳过该候选看下一条(不认领);per_route_limit 默认值改为小于 worker_count 的值(如 max(1, worker_count // 2))并接入配置。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "recover_stale_tasks 只在 worker 启动时执行且只认 source=workflow_round,同步生成崩溃后节点永久锁死",
|
|||
|
|
"location": "backend/services/queue_service.py:470",
|
|||
|
|
"problem": "全仓库对 recover_stale_tasks 的唯一调用点是 WorkflowRoundWorkerRuntime.start(),且固定传 source=\"workflow_round\";其它来源任务被直接 skipped: source_not_managed。而 nodes.generate 创建的任务 metadata.source = \"nodes.generate\"。进程生命周期内也没有周期性 stale 巡检。",
|
|||
|
|
"evidence": "# queue_service.py:470-480\nfor candidate in candidates:\n task_source = str(candidate.metadata.get(\"source\") or \"\").strip()\n if source is not None and task_source != source:\n skipped.append({\"task_id\": candidate.task_id, \"status\": candidate.status,\n \"reason\": \"source_not_managed\"})\n continue\n\n# generation_service.py:574-585\nactive_task = self._active_generation_task(node)\nif node.status in {\"queued\", \"running\"} and active_task is not None:\n raise ApiError(ErrorCode.VALIDATION_FAILED, \"generation is already in progress for this node\", ...)",
|
|||
|
|
"impact": "进程在同步生成中途被杀(AstrBot 重启、容器 OOM)后,该任务永远停在 running,节点也停在 running。用户再点生成得到「generation is already in progress」,且没有自动恢复——只能手动调用 /jobs/{id}/cancel 才能解锁节点。运行中挂死的任务只能等下一次插件重载。",
|
|||
|
|
"fix": "1) source 过滤改成可选白名单,nodes.generate 也纳入恢复(恢复动作为标记 failed + 释放节点,因为它没有 worker 可重新执行);2) 增加后台巡检线程或复用 worker 循环 idle 分支周期性调用 recover_stale_tasks;3) 恢复时同步修复 workflow_nodes.status / metadata.last_task_id。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "recover_stale_tasks 无重试次数上限,会跨进程重启无限重排队",
|
|||
|
|
"location": "backend/services/queue_service.py:481",
|
|||
|
|
"problem": "恢复逻辑只在 metadata 里写一个会被覆盖的 recovery 字典,没有累加计数,也没有「超过 N 次进 dead_letter」的出口。dead_letter 状态在这条路径上从未被使用。",
|
|||
|
|
"evidence": "# queue_service.py:481-521\nrecovery = {\n \"source\": \"plugin_startup_stale_recovery\",\n \"observed_at\": observed_at.astimezone(timezone.utc).isoformat(),\n \"updated_before\": cutoff_value,\n \"previous_status\": candidate.status,\n}\nfailed = self.update_task_if_status(candidate.task_id, {candidate.status},\n {\"status\": \"failed\", ..., \"metadata\": {**candidate.metadata, \"recovery\": recovery}}, ctx)\n...\nqueued = self.update_task_if_status(failed.task_id, {\"failed\"},\n {\"status\": \"queued\", ..., \"metadata\": {**failed.metadata,\n \"recovery\": {**recovery, \"status\": \"requeued\"}}}, ctx, require_workflow_not_cancelled=True)",
|
|||
|
|
"impact": "某条任务如果稳定地让 worker 卡死(例如 provider 对该 prompt 永远超时、或处理时崩线程),每次插件重启都会把它重新入队,形成跨重启的无限重试。既浪费额度,也可能反复拖垮同一个 worker。",
|
|||
|
|
"fix": "在 metadata 维护 recovery.attempts(int(previous.get(\"attempts\") or 0) + 1),超过阈值(如 3)时不再 requeue,直接迁移到 dead_letter 并在 error_message 写明原因;recover_stale_tasks 返回值中额外报告 dead_lettered 列表供诊断 UI 展示。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "轮次候选的 timeout_seconds 只写进 metadata,执行时从未被读取",
|
|||
|
|
"location": "backend/services/generation_service.py:355",
|
|||
|
|
"problem": "enqueue_prepared_generation 计算出路由级超时并存进 metadata[\"timeout_seconds\"],但 execute_claimed_generation 构造 runtime_request 时没有把它回填进请求,而 single_flow_engine 只读 request.get(\"timeout_seconds\")。全仓库 grep 显示该 metadata 键没有任何读取点。",
|
|||
|
|
"evidence": "# generation_service.py:355 & 373\ntimeout_seconds = _round_timeout_seconds(request, prepared.get(\"route_resolution\"))\nreturn self.queue_service.create_task({\n ...,\n \"metadata\": {..., \"timeout_seconds\": timeout_seconds, ...}, # ← 只写不读\n}, ctx)\n\n# engines/single_flow.py:163\ntimeout_seconds=_optional_timeout_seconds(request.get(\"timeout_seconds\")), # ← 读 request 而非 metadata",
|
|||
|
|
"impact": "轮次准备阶段基于路由 settings.timeout_seconds 算出的超时被静默丢弃,实际生效的是 provider 全局配置(provider.timeout_seconds or 180)。为长耗时模型单独调大/为快模型调小的意图无效;metadata 上的数字会误导排障。此外整个 worker 执行路径没有任何 wall-clock 上限,只有 HTTP socket 超时,provider 慢速滴数据可远超预期地占住 worker 线程。",
|
|||
|
|
"fix": "execute_claimed_generation 构造 runtime_request 时加 runtime_request.setdefault(\"timeout_seconds\", task.metadata.get(\"timeout_seconds\"));并在 QueueWorker.run_once 外层加基于该值的整体看门狗(独立线程执行 handler 并 join(timeout),超时写 failed 并记录 timeout 错因)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "worker 执行成功但回写失败时,已落库的图片记录被孤立且不清理",
|
|||
|
|
"location": "backend/services/worker_service.py:104",
|
|||
|
|
"problem": "execute_claimed_generation 内部一旦成功就已通过 image_service.create_image 把图片元数据写进数据库并返回;随后 worker 才调用 update_task 回写任务。若这一步抛异常(任务被并发取消导致状态迁移校验失败、数据库瞬时错误),run_once 的 except Exception 只会走 _handle_failure 把任务标记 failed,不会回收刚创建的图片记录与磁盘文件。",
|
|||
|
|
"evidence": "# generation_service.py:492 —— 成功路径已落库\ncreated = self.image_service.create_image(image_payload, ctx)\n\n# worker_service.py:104-146\ntry:\n update = _normalize_update(self.handler(task)) # ← handler 已建好 image 记录\n ...\n completed = self.queue_service.update_task(...) # ← 这里抛错\nexcept Exception as exc:\n return self._handle_failure(task, exc, ctx) # ← 只改任务状态,不清理图片",
|
|||
|
|
"impact": "候选在 UI 上显示失败(handle_task_terminal 写入 image_id=None),但图库里多出一张无人引用的图片和磁盘文件,且计费统计(CostService.usage_summary 按 generated_images 行统计)仍会把它计入成本。反复发生会污染图库。",
|
|||
|
|
"fix": "把「落库图片」与「回写任务」放进同一补偿边界:WorkflowRoundWorkerRuntime._execute 捕获回写失败后调用 generation_service._cleanup_generation_outputs;或让 _execute 返回 image_ids,由 _on_terminal 在 current_task.status == \"failed\" 且 claimed_task 曾产出图片时执行清理。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "cancel_job 的重读循环无次数上限与退避",
|
|||
|
|
"location": "backend/services/queue_service.py:135",
|
|||
|
|
"problem": "while True 在 CAS 失败时无限重试,既无最大轮数也无 sleep 退避。",
|
|||
|
|
"evidence": "# queue_service.py:135-172\nwhile True:\n before = self.get_job(task_id)\n ...\n after = self.repository.update_if_status(task_id, {before.status}, {...})\n if after is not None:\n break\n # The task changed after it was read. Re-evaluate the current\n # state instead of applying cancellation from a stale snapshot.",
|
|||
|
|
"impact": "若某条任务被高频写入(例如 worker 无条件 update_task 与取消互相打架),该循环会变成忙等,持续开关 SQLite 连接并占住调用线程。正常情况 1~2 轮即可收敛,属低概率但无兜底。",
|
|||
|
|
"fix": "加最大重试轮数(如 5)与极短退避;超出后抛 ApiError(TASK_NOT_CANCELABLE, \"task status is changing too fast\"),让调用方重试而不是把线程锁死在服务层。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "_ensure_round_enqueued 的轮次元数据回写在 _reconcile_lock 之外,存在丢更新",
|
|||
|
|
"location": "backend/services/workflow_round_service.py:928",
|
|||
|
|
"problem": "reconcile_round 明确用 self._reconcile_lock 串行化「读候选 → 聚合 → 写轮次」,但 _ensure_round_enqueued 末尾的「读 latest.metadata → 写回 status=queued」这一对读写没有走同一把锁。",
|
|||
|
|
"evidence": "# workflow_round_service.py:928-940\nlatest = self.round_repository.get_round(round_record.round_id) or round_record # ← 读\nself.round_repository.update_round(round_record.round_id, {\n \"status\": \"queued\",\n \"metadata\": {**latest.metadata, \"planned_only\": False, \"task_ids\": task_ids}, # ← 写,中间无锁\n})\nself.reconcile_round(round_record.round_id, ctx)",
|
|||
|
|
"impact": "入队线程(HTTP 请求线程)与 worker 线程的 reconcile_round 交错时,worker 刚写入的 execution_summary 会被这里的 {**latest.metadata} 快照覆盖;轮次 status 也会被短暂回退到 queued。后面的 reconcile_round 通常能纠正 status,但 metadata 里的聚合信息可能丢失一轮。",
|
|||
|
|
"fix": "把 _ensure_round_enqueued 末尾这段读-改-写包进 with self._reconcile_lock:,或改用仓储层的条件更新(只更新 task_ids / planned_only 两个键而不是整块 metadata 覆盖)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "_record_generation_failure 与 _clear_retry_preparation 里的裸 except Exception: pass",
|
|||
|
|
"location": "backend/services/generation_service.py:1865",
|
|||
|
|
"problem": "失败记录与重试准备清理这两条补偿路径里的异常被完全静默吞掉,没有 logger.exception 也没有任何可观测输出(generation_service.py 甚至没有 import logging)。对比 worker_service.py 里同类补偿路径都规范写了 logger.exception,这里是明确的可观测性缺口。另见 generation_service.py:1887、generation_job_retry_service.py:388、generation_job_retry_service.py:399。",
|
|||
|
|
"evidence": "# generation_service.py:1852-1888\ntry:\n failed_task = self.queue_service.update_task_if_status(task.task_id, {\"queued\", \"running\"}, {...}, ctx)\nexcept Exception:\n pass # ← 任务没能标记 failed,无任何痕迹\n...\n except Exception:\n pass # ← 节点没能标记 failed,无任何痕迹\n\n# generation_job_retry_service.py:386-389\n metadata.pop(\"retry_preparation\", None)\n workflow_service.update_node(node.node_id, {\"metadata\": metadata}, ctx)\nexcept Exception:\n pass # ← retry_preparation 残留在节点/工作流 metadata 上",
|
|||
|
|
"impact": "任务/节点未能落到 failed 时,前端会看到一个永远停在 running 的节点,而日志里没有任何线索可定位;retry_preparation 残留会让后续 _prepare_node 的所有权判定读到过期 task_id,从而误判「node is already owned by another active generation task」。",
|
|||
|
|
"fix": "在 generation_service.py / generation_job_retry_service.py 引入 logger = logging.getLogger(__name__),把这 4 处 pass 换成 logger.exception(\"...\", task_id, node_id),并在返回给调用方的 failure 字典中加 bookkeeping_failed: True 标志,让 UI 能提示「状态可能不一致,请手动刷新/取消」。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "core/ 核心模块 + 安全",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A5-核心与安全.md",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\security.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\secret_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\storage.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\remote_fetch.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\media_proxy.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\config_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\config_writer.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\runtime_paths.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\identity.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\chat_access_policy.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\audit_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\webhook.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\model_router.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\model_route_chain.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\prompt_pipeline.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\interop.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\interactions.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\image_formats.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\astrbot_provider_bridge.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\generation_errors.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\integration_token_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_command_guard_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\settings_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\model_provider_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\media_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\storage_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\config.yaml"
|
|||
|
|
],
|
|||
|
|
"summary": "密码学原语本身写得扎实:HMAC 全部用 compare_digest、令牌带过期校验、remote_fetch 对协议/凭据/私网/重定向降级做了完整校验、storage.split_uri 双重拦截 .. 与绝对路径、SQL 全参数化、无 pickle/yaml.load、core/ 下无任何日志打印。真正的问题在信任边界设计:所有登录 WebUI 的用户被无差别授予 can_manage_config(main.py:104),而 storage.*_path 允许配置为任意绝对路径且 SettingsService.validate() 对 storage 段零校验,两者叠加把\"配置管理\"升级为任意文件读取(经未认证的 media/fetch 出站)与任意文件写入(经 storage/migrate 非 dry-run),我已用真实代码路径完整复现读取链路。第二类系统性问题是配置持久化:自研 dump_yaml + _load_yaml_subset 往返不等价,用仓库自带 config.yaml 实测 6 个空列表字段被静默变成空字典、15 行注释全丢;save() 还会把打包默认值固化进可写配置,导致插件升级后新默认值永久失效。此外 _record_snapshot 把未脱敏的完整配置(含 api_key_plaintext、Postgres 密码、令牌盐与哈希)明文写进数据库,write_text_atomic 的 os.replace 会把 config.yaml 权限位重置为 umask 默认。共 12 条发现:3 条 P1、5 条 P2、4 条 P3。",
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "存储根可配为任意绝对路径 + 未认证的 media/fetch = 任意文件读取",
|
|||
|
|
"location": "core/runtime_paths.py:43",
|
|||
|
|
"problem": "storage.*_path 是普通配置项,SettingsService.validate() 从头到尾没有任何针对 storage 段的校验(只校验 generation/vision/prompt_governance/advanced_materials)。resolve_data_path 对绝对路径直接原样返回,VfsStorage.resolve_uri 的越界检查只保证路径在 root 之内——而 root 本身已被改成 C:/ 或 /。签发出的令牌由 GET media/fetch 消费,该路由权限为 public_token(auth_required: False)。",
|
|||
|
|
"evidence": "# core/runtime_paths.py:43\nresolved = Path(path).expanduser()\nif resolved.is_absolute():\n return resolved # 绝对路径不做任何限制\n\n# core/storage.py:105\nresolved_root = resolve_data_path(root_path).resolve()\nresolved_path = (resolved_root / relative).resolve()\nif resolved_path != resolved_root and resolved_root not in resolved_path.parents:\n raise ApiError(ErrorCode.FORBIDDEN, \"storage uri escapes its root\", {\"uri\": uri})\n\n# backend/route_registry.py:173\nRouteSpec(\"GET\", \"media/fetch\", \"backend.routes.media_routes.fetch_media\", \"public_token\"),\n# backend/route_registry.py:28\n\"public_token\": {\"auth_required\": False, \"required_permission\": None, ...}\n\n# 实测复现(未改任何源码):\n# StorageConfig(temp_path='C:/Windows')\n# VfsStorage(FakeCS()).resolve_uri('storage://temp/System32/drivers/etc/hosts')\n# -> C:\\Windows\\System32\\drivers\\etc\\hosts exists=True\n# MediaProxy(s).sign(...) -> {'url': '/api/plug/.../media/fetch?token=eyJleHBpcmVzX2F0...', 'size_bytes': 827}",
|
|||
|
|
"impact": "攻击链:任一已登录 WebUI 用户 → POST settings/save 把 storage.temp_path 改为 / → GET media/index?root=temp 枚举全盘 → GET media/sign 签名任意文件 → 把 URL 交给任何未认证第三方去 GET media/fetch 取走内容。宿主机任意文件读取并可无认证外传(AstrBot 的 data/config/*.json、.env、SSH 私钥、其他插件凭据)。",
|
|||
|
|
"fix": "1) 在 SettingsService.validate() 增加 storage 段校验:拒绝绝对路径,或强制 resolve_data_path(p).resolve() 必须落在 get_runtime_data_root().resolve() 之内;2) 在 VfsStorage._root_paths() 加二次防线,对越出运行时数据根的 root 直接抛 FORBIDDEN;3) resolve_data_path 不再无条件放行绝对路径,改为白名单;4) media/fetch 令牌 payload 增加 actor_id 并在 fetch 时校验,避免签名 URL 成为可转发的无主凭据。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "storage/migrate 非 dry-run 可在可配置根之间任意复制文件/目录(任意写)",
|
|||
|
|
"location": "core/storage.py:52",
|
|||
|
|
"problem": "migrate_storage 仅要求 require_config_manager(ctx)(实际等价于\"已登录\"),dry_run 由请求体控制且可传 false。_copy_path 会把 source 整棵目录树 shutil.copy2 到 target。结合 root 可为任意绝对路径,即任意文件/目录写入。will_overwrite 只被计算出来放进返回值,没有任何一处用它阻断覆盖。",
|
|||
|
|
"evidence": "# core/storage.py:52\nif not dry_run:\n _copy_path(source_path, target_path)\n\n# core/storage.py:183\nif source_path.is_file():\n target_path.parent.mkdir(parents=True, exist_ok=True)\n shutil.copy2(source_path, target_path) # 目标已存在时静默覆盖\n return\n\n# backend/routes/storage_routes.py:46\ndry_run=as_bool(payload.get(\"dry_run\"), default=True),",
|
|||
|
|
"impact": "覆盖 AstrBot 配置、写入计划任务/启动脚本目录、投毒插件目录 → 从\"配置管理员\"提升到宿主机代码执行。即使 root 全部保持默认,也能无提示覆盖用户素材/存档。该操作全程无审计记录(无 audit_service.record 调用)。",
|
|||
|
|
"fix": "(a) 修完 root 限制后 root 天然受限;(b) migration_plan 在 will_overwrite and not payload.get(\"confirm_overwrite\") 时抛 ApiError;(c) 把非 dry-run 迁移提升为独立权限而非复用 can_manage_config;(d) 为迁移动作补审计日志。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "配置快照把未脱敏的明文密钥写进数据库 config_snapshots",
|
|||
|
|
"location": "backend/services/settings_service.py:614",
|
|||
|
|
"problem": "save() / rollback() 在写盘前调用 _record_snapshot(before_file, ...)(:444、:556),该方法把 before_file(load_config_file_payload 的原始返回值)原样 json.dumps 存入 config_snapshots.config_json。读取路径 list_snapshots / get_snapshot 都调了 _mask_sensitive,唯独写入侧没有。config.yaml 中确实存在 api_key_plaintext(model_provider_runtime.py:366 读取该字段)、database.postgresql.password/dsn(config.yaml:162,167)、security.integration_token_hash 与 _salt。",
|
|||
|
|
"evidence": "# backend/services/settings_service.py:610-615\n(\n snapshot_id,\n reason,\n ctx.actor_id,\n json.dumps(dict(config), ensure_ascii=False, sort_keys=True), # 未脱敏\n),",
|
|||
|
|
"impact": "每保存一次配置就把当时的全部明文凭据复制一份进 DB,无上限、无保留期、无清理。DB 备份 / SQLite 文件泄露 / Postgres 只读账号即可拿到历史全部 API Key 与数据库口令;令牌盐+哈希外泄后可离线爆破集成令牌。",
|
|||
|
|
"fix": "写入前脱敏,同时保留回滚能力——把敏感字段替换为哨兵值(如 \"***PRESERVED***\"),在 rollback() 里遇到哨兵时从当前 before_file 取回原值(_merge_masked 已有同类逻辑,可直接复用)。另外给 config_snapshots 加保留期清理任务。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "自研 YAML 往返不等价:空列表退化为字典、注释全丢(已用仓库 config.yaml 实测)",
|
|||
|
|
"location": "core/config_writer.py:50",
|
|||
|
|
"problem": "dump_yaml 对空 list 只输出 \"key:\",_load_yaml_subset 把\"键后无值\"一律解析成空 dict,类型从 list 翻成 dict。注释在 _strip_yaml_comment 里被丢弃且写回时不还原。触发点:settings_service.py:445/:557、integration_token_service.py:129、review_service.py:42。",
|
|||
|
|
"evidence": "# core/config_writer.py:50-53\nelif isinstance(child, list):\n lines.append(f\"{prefix}{key}:\")\n for item in child: # 空列表 -> 只剩 \"key:\"\n lines.append(f\"{prefix} - {_format_scalar(item)}\")\n\n# database/db_helper.py:403-406\nif value == \"\":\n child: dict[str, Any] = {} # \"key:\" 一律变成 dict\n parent[key] = child\n\n# 实测(读仓库自带 config.yaml -> dump_yaml -> 再解析):\n# roundtrip stable: False\n# DIFF .model_assignments.generation_cross_review_routes [] -> {}\n# DIFF .recognition_prompts.category_requirements [] -> {}\n# DIFF .chat_access.group_allowlist / user_allowlist / adult_* [] -> {}\n# comments in src: 15 (写回后全部丢失)",
|
|||
|
|
"impact": "任何一次保存设置/轮换集成令牌/审核写回都会触发。所有 isinstance(x, list) 判断在保存一次后失效——例如 settings_service.py:765 的 uses_plugin_adapters = isinstance(generation.get(\"adapters\"), list),用户删光 adapters 后该判断永久为 False,_runtime_override_conflicts 的兼容分支随之失效。model_assignments.*_routes 同理。运维手写的 15 条注释在用户第一次点保存后消失。",
|
|||
|
|
"fix": "dump_yaml 对空容器显式输出 \"key: []\" / \"key: {}\"(_load_yaml_subset 已能通过 _parse_scalar 正确解析 [])。更彻底的做法是引入 ruamel.yaml(round-trip 模式保注释),或至少写一个 dump_yaml(_load_yaml_subset(x)) == 归一化(x) 的属性测试锁住不变式。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "save() 把打包默认值固化进可写配置,插件升级后新默认值永久失效",
|
|||
|
|
"location": "backend/services/settings_service.py:423",
|
|||
|
|
"problem": "save() 的基线 before = self._load_current_mapping() 走的是 load_config_payload,而后者 = 打包 config.yaml ∪ 可写 config.yaml ∪ 运行时覆盖。_restore_runtime_managed_fields 只剥掉\"运行时覆盖\",打包默认值原封不动地留在 merged 里并被写进可写配置(:445)。",
|
|||
|
|
"evidence": "# database/db_helper.py:77-82\nif _should_apply_runtime_overrides(path):\n packaged_path = get_packaged_config_path()\n if not _same_path(path, packaged_path) and packaged_path.exists():\n payload = _load_yaml_subset(packaged_path.read_text(encoding=\"utf-8\"))\nif path.exists():\n payload = _merge_mappings(payload, _load_yaml_subset(path.read_text(encoding=\"utf-8\")))\n\n# backend/services/settings_service.py:423,445\nbefore = self._load_current_mapping() # 已含打包默认值\n...\nwrite_text_atomic(self.config_service.config_path, dump_yaml(merged))",
|
|||
|
|
"impact": "用户第一次点\"保存设置\"后可写配置就变成完整快照。此后插件升级修改的任何默认值(新路由、新 prompt 模板、安全默认开关)都被旧快照覆盖,用户无感知,只能手删配置文件。这与 config_service.py:68-71 刻意把 version 设为不可覆盖的意图直接矛盾——那里已踩过同一个坑,但只修了 version 一个字段。",
|
|||
|
|
"fix": "save() 的写盘基线改用只读可写文件的 loader(跳过打包合并),只把用户显式提交的 diff 落盘;打包默认值保持在读取期合并。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "write_text_atomic 的 os.replace 重置配置文件权限,临时文件短暂全局可读",
|
|||
|
|
"location": "core/config_writer.py:32",
|
|||
|
|
"problem": "每次写配置都新建临时文件再 replace。新文件权限是 0666 & ~umask(POSIX 下通常 0644),os.replace 保留的是源 inode 的权限,目标原有的 0600 被丢弃。而 config.yaml 里就有 api_key_plaintext 和数据库口令。全程也没有 fsync。",
|
|||
|
|
"evidence": "# core/config_writer.py:32-37\ntemporary = destination.with_name(f\".{destination.name}.{uuid4().hex}.tmp\")\ntry:\n temporary.write_text(content, encoding=\"utf-8\") # 默认 0644,无 os.chmod\n temporary.replace(destination) # 目标权限位被覆盖为 0644",
|
|||
|
|
"impact": "运维执行 chmod 600 config.yaml 加固后,用户第一次在 WebUI 点保存就被静默回退到 0644——同宿主机上任何本地用户/其他容器进程可读明文 API Key 与 Postgres 密码。写入窗口内临时文件同样全局可读(临时文件竞态)。无 fsync 导致崩溃后可能留下截断的 config.yaml(无备份)。",
|
|||
|
|
"fix": "用 os.open(path, O_WRONLY|O_CREAT|O_EXCL, 0o600) 创建临时文件并 os.fsync;replace 之前若目标已存在,用 os.chmod(temporary, stat.S_IMODE(destination.stat().st_mode)) 继承既有权限位。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "所有已登录 WebUI 用户被无差别授予 can_manage_config,无角色区分",
|
|||
|
|
"location": "main.py:2337",
|
|||
|
|
"problem": "RequestContext 的权限集合只有\"有 username → 全给\"\"没有 → 全不给\"两种状态。core/security.py 精心区分的 require_write / require_task_retry / require_config_manager 在真实运行时全部塌缩成 require_authenticated。",
|
|||
|
|
"evidence": "# main.py:104-106\nBRIDGE_PERMISSIONS = frozenset(\n {\"can_write\", \"can_manage_config\", \"can_manage_materials\", \"can_retry_tasks\"}\n)\n# main.py:2337\npermissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()",
|
|||
|
|
"impact": "这是上面两条 P1 的权限前提——\"只有管理员才能改 storage 路径\"的假设不成立。同时任何账号可轮换集成令牌(打断已接入的伴侣插件)、可把 generation.provider.base_url 指向 http://169.254.169.254/ 或内网地址(model_provider_runtime.py:217/:304 与 engines/image_provider.py:221 均以 allow_private=True 调用 validate_http_url,私网防护被主动关闭),构成经认证的 SSRF。",
|
|||
|
|
"fix": "从 AstrBot 侧读取真实角色(g 上的 role/claims)映射到权限集;在拿不到角色信息之前,至少把 can_manage_config 从默认集合中移除,改为要求显式配置的管理员用户名白名单(如 security.config_admin_users)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "群聊日配额统计全表扫描 + Python 侧 JSON 解析",
|
|||
|
|
"location": "backend/services/chat_command_guard_service.py:75",
|
|||
|
|
"problem": "_daily_group_count 把当天所有 chat_command_allowed 审计行全量拉回进程,逐行 json.loads(metadata_json) 再比对 group_id。SQL 里没有任何 group 过滤条件,metadata_json 是 TEXT 也没有可用索引。另外 _daily_count 统计的是所有 chat_command_allowed,而 record()(:41-55)对非生图命令也会记这个 action,导致\"每日生成上限\"被非生成类命令误消耗。",
|
|||
|
|
"evidence": "rows = conn.execute(\n f\"SELECT metadata_json FROM audit_logs WHERE action = {placeholder} AND created_at >= {placeholder}\",\n (\"chat_command_allowed\", start),\n).fetchall()\ncount = 0\nfor row in rows:\n try:\n metadata = json.loads(row[\"metadata_json\"] or \"{}\")\n except (TypeError, ValueError):\n continue\n if str(metadata.get(\"group_id\") or \"\") == group_id:\n count += 1",
|
|||
|
|
"impact": "每条群聊生图命令都触发一次当日全表扫描。audit_logs 是所有审计动作的公共表(配置保存、令牌轮换、聊天命令全在里面),随部署时间线性膨胀。繁忙群在高峰期会把 SQLite 拖到写锁竞争,成为聊天命令的延迟瓶颈;同时配额计数不准确。",
|
|||
|
|
"fix": "为聊天配额单独建 chat_command_quota(day, scope_type, scope_id, count) 计数表加唯一索引并用 UPSERT 累加;或至少给 audit_logs 增加独立 group_id 列 + (action, created_at, group_id) 复合索引。同时把配额计数的 action 从通用 chat_command_allowed 拆成 chat_generation_allowed。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "remote_fetch 的 SSRF 防护存在 DNS TOCTOU(rebinding)窗口",
|
|||
|
|
"location": "core/remote_fetch.py:73",
|
|||
|
|
"problem": "validate_http_url 通过 socket.getaddrinfo 解析主机名并逐一检查私网地址,随后 opener.open(request, ...) 让 urllib 重新解析一次 DNS。攻击者控制的权威 DNS 只需把 TTL 设为 0 并在两次解析间返回不同结果(先返回公网 IP 通过校验,再返回 127.0.0.1 / 169.254.169.254),即可绕过全部私网检查。_ValidatedRedirectHandler 同样只校验 URL 不校验最终连接的 IP。",
|
|||
|
|
"evidence": "# core/remote_fetch.py:73-75\nfor address in _resolve_addresses(host, port): # 第一次解析\n if _is_unsafe_address(address):\n raise RemoteFetchError(\"remote URL resolves to a private or local network address\")\n# core/remote_fetch.py:40\nwith opener.open(request, timeout=timeout_seconds) as response: # urllib 第二次解析",
|
|||
|
|
"impact": "backend/services/chat_media_import_service.py:31 用用户在聊天里发的 URL 调用本函数(且正确地未开 allow_private),这是唯一的不可信输入入口,也正是 rebinding 最容易被利用的地方 → 可探测并读取内网 HTTP 服务(云元数据、内部 API)。",
|
|||
|
|
"fix": "把校验通过的 IP 固定下来再连接——解析一次得到安全 IP,用该 IP 建连并通过 Host 头/SNI 保留原主机名(自定义 HTTPConnection/HTTPSConnection 子类覆写 connect),或在 http_response 钩子里用 response.fp.raw._sock.getpeername() 复核对端 IP,不安全则立即中断。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "未加引号的配置标量会被 os.path.expandvars 展开",
|
|||
|
|
"location": "database/db_helper.py:445",
|
|||
|
|
"problem": "_parse_scalar 的兜底分支对所有未加引号的标量执行环境变量展开。实测 _load_yaml_subset('plugin:\\n display_name: $PATH\\n') 返回了完整的 PATH 环境变量内容。",
|
|||
|
|
"evidence": "try:\n return int(value)\nexcept ValueError:\n return os.path.expandvars(value)",
|
|||
|
|
"impact": "display_name 等非敏感字段会通过 settings/current、settings/export 直接把环境变量内容回显给前端,而这些路径的 _mask_sensitive 按 key 名脱敏、认不出内容里的密钥。因 dump_yaml 写回时字符串都被 json.dumps 加引号(加引号后走 json.loads 分支不展开),实际风险局限于手写的 config.yaml;但 ${AWS_SECRET_ACCESS_KEY} 这类值一旦被手写进任意字段就会明文回显。同时含 $ 的合法值(如密码)会被静默篡改。",
|
|||
|
|
"fix": "改为仅对显式声明的引用语法展开(如整值为 env:NAME / ${NAME}),其余标量原样返回;core/secret_service.resolve_env_value 已实现这套显式语义,直接复用即可。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "媒体令牌置于 URL 查询串且 fetch 端点完全未认证",
|
|||
|
|
"location": "core/media_proxy.py:51",
|
|||
|
|
"problem": "签名 URL 形如 /api/plug/.../media/fetch?token=<payload>.<sig>,令牌即凭据,且 fetch 端点 auth_required: False(backend/route_registry.py:173)。令牌 payload 内不含签发者标识。",
|
|||
|
|
"evidence": "\"url\": f\"{MEDIA_FETCH_PATH}?token={token}\",",
|
|||
|
|
"impact": "令牌会进入反向代理/网关访问日志、浏览器历史、以及页面外链跳转时的 Referer 头。默认 TTL 300 秒限制了窗口,但在此期间任何拿到日志的人都可无认证取回媒体,且令牌无主可任意转发。",
|
|||
|
|
"fix": "payload 增加 actor_id 并在 _resolve_token 中与当前 ctx 比对(fetch_media_payload 已要求认证,可直接校验);响应加 Referrer-Policy: no-referrer;或改用短时 Cookie 承载而非 query 参数。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "敏感键匹配把 \"key\" 当子串,误伤普通配置项",
|
|||
|
|
"location": "backend/services/settings_service.py:26",
|
|||
|
|
"problem": "SENSITIVE_KEY_PARTS 含裸 \"key\",_is_sensitive_key(:678)做的是子串匹配。实测 _is_sensitive_key(\"advanced_materials.cost_budget.price_key\") 返回 True。该清单与 core/secret_service.SECRET_KEY_PARTS(未含裸 key)已经不一致。",
|
|||
|
|
"evidence": "SENSITIVE_KEY_PARTS = (\"password\", \"token\", \"secret\", \"api_key\", \"apikey\", \"key\", \"dsn\")",
|
|||
|
|
"impact": "cost_budget.price_key(计价键,schema 定义见 settings_service.py:143)在 get_current / export / diff 中一律显示为 ***MASKED***,前端永远读不到当前值,diff 也只显示 before_set/after_set 布尔量,运维无法确认自己配了什么。写入侧因 _merge_masked(:705)跳过哨兵值不会被破坏,属可见性缺陷。",
|
|||
|
|
"fix": "改为按最后一段键名做精确/后缀匹配(api_key、apikey、*_secret、*_token、password、dsn),或维护显式敏感字段路径清单,与 core/secret_service.SECRET_KEY_PARTS 统一到同一份定义。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "配置文件锁仅为进程内 threading.RLock",
|
|||
|
|
"location": "core/config_writer.py:17",
|
|||
|
|
"problem": "config_file_lock 返回的是按路径缓存的 threading.RLock,只在当前 Python 进程内生效。",
|
|||
|
|
"evidence": "def config_file_lock(path: str | Path) -> threading.RLock:\n return _lock_for_path(Path(path))",
|
|||
|
|
"impact": "AstrBot 自身通过 persist_runtime_config_values(database/db_helper.py:49)写同一份插件配置时不受此锁保护;多 worker 部署或外部脚本编辑 config.yaml 时,SettingsService.save() 的\"读—改—写\"会整份覆盖对方的修改。os.replace 保证单次写入原子,但保证不了读-改-写序列的隔离。",
|
|||
|
|
"fix": "叠加文件锁(POSIX fcntl.flock / Windows msvcrt.locking,或跨平台用同目录 .lock 文件 + O_EXCL);或在写入前比对文件 mtime/内容哈希,发现被外部修改则拒绝并要求前端重新拉取。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "main.py 插件入口 + backend/routes 路由层 + 权限矩阵",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A7-路由与插件入口.md",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\identity.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\security.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\media_proxy.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\params.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\profile_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\workflow_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\round_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\job_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\image_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\media_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\log_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\history_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\settings_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\data_backup_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\wardrobe_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\integration_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\advanced_material_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\prompt_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\event_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\job_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\workflow_round_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\profile_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\resource_access_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\image_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\media_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\prompt_governance_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\wardrobe_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\profile_document_import_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\model_provider_bootstrap.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\api\\client.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\api\\endpoints.ts"
|
|||
|
|
],
|
|||
|
|
"summary": "PERMISSION_MATRIX 只是一张声明式清单:dispatch_extension_request 全程不读 RouteMatch.permission,鉴权 100% 依赖每个 handler 手写的 require_*,而我逐条比对 169 条 RouteSpec 与 handler 实际 guard 后确认——声明与实现在权限类别上目前是对齐的,但矩阵在运行时被上游架空,因为 main.py 的桥接层给任何登录用户无条件发放全部四项权限,六级矩阵塌缩成\"登录/未登录\"两级,config/materials/retry 共 68 条路由的边界形同虚设。更严重的是归属校验覆盖不完整:profile 与 round 两条链路正确接入了 owner 校验,而 workflow/node/image/job/history 五条链路的 service 方法根本不接收 ctx,任何登录用户凭 ID 即可读写他人工作流与节点,GET jobs 更是直接返回他人完整生成请求体(含提示词与参考图)。另有三个结构性问题:客户端可自带 content_permissions/entrypoint 绕过成人内容路由闸门;所有 service 构造都发生在 require_* 之前,未授权请求即可触发 provider 初始化与 DB 写入且构造期异常逃出 handle_request 直接 500;157 个 handler 是完全同构的样板加 22 个各自为政的 service 工厂。此外 POST profiles/{id}/archive 实际执行的是软删除,与端点名和前端\"归档\"按钮语义完全不符。",
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P0-致命",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "桥接层给所有登录用户发放全部权限,PERMISSION_MATRIX 在运行时被架空",
|
|||
|
|
"location": "main.py:104, main.py:2337",
|
|||
|
|
"problem": "_request_context_from_astrbot_request() 只要能取到 username/user_id,就把 BRIDGE_PERMISSIONS(can_write/can_manage_config/can_manage_materials/can_retry_tasks 四项全部)塞进 RequestContext.permissions。而 AstrBot Plugin Page 桥接是前端真正走的通路,于是 config/materials/retry 三档权限与 read 完全等价。",
|
|||
|
|
"evidence": "BRIDGE_PERMISSIONS = frozenset(\n {\"can_write\", \"can_manage_config\", \"can_manage_materials\", \"can_retry_tasks\"}\n)\n...\npermissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()\nreturn RequestContext.from_astrbot(\n username=username, user_id=user_id,\n client_ip=_client_ip(request), source=\"astrbot_plugin_page\",\n permissions=permissions,\n)",
|
|||
|
|
"impact": "任意一个 AstrBot WebUI 普通登录用户可以 POST settings/save 改数据库后端与模型密钥环境变量名、POST settings/rollback、POST backups/{id}/restore 覆盖全量数据、POST integrations/private-companion/token/rotate 轮换外部集成令牌、POST jobs/failed/clear 清空他人失败队列。route_registry 中 config(22条)+materials(38条)+retry(8条) 共 68 条路由的权限边界形同虚设。同时因为没人拥有 can_manage_all_profiles,ResourceAccessService._can_access 的管理员分支是死代码。",
|
|||
|
|
"fix": "1) 从 AstrBot 请求上下文真正解析角色/权限并映射为四项能力,探测不到时降级为最小集 frozenset() 而非最大集;2) 在 config.yaml 增加显式管理员白名单(security.admin_usernames / security.material_managers),据此授予 can_manage_config / can_manage_materials / can_manage_all_profiles;3) 把 permission 真正接进分发器,让矩阵成为强制点。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P0-致命",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "dispatch_extension_request 从不检查 RouteMatch.permission,矩阵与实现只靠人工同步",
|
|||
|
|
"location": "adapters/astrbot_adapter.py:171",
|
|||
|
|
"problem": "分发器解析出 match.permission 后直接丢弃,立刻调用 handler。鉴权 100% 依赖 handler 内部手写 require_*。仓库里没有任何测试或启动断言校验 RouteSpec.permission 与 handler 实际 guard 一致;runtime_health() 只校验路由唯一性与权限名已知。",
|
|||
|
|
"evidence": "def dispatch_extension_request(ctx, method, path, *, query=None, payload=None) -> dict[str, Any]:\n match = resolve_route(method, path)\n if match is None:\n return error_response(...)\n return match.handler(*_handler_args(match, ctx, query or {}, payload or {}))\n # match.permission 从未被使用\n\n# route_registry.py:365-377\ndef assert_unique_routes() -> None: ... # 只查重复\ndef assert_known_permissions() -> None: ... # 只查权限名拼写",
|
|||
|
|
"impact": "新增一条路由时只要忘写 require_write(ctx),就是一个静默的未授权写接口,CI 与 runtime_health() 都不会报警。当前 169 条路由靠人工维持一致,是一次疏忽即破防的结构。另外真正生效的桥接注册 _bridge_registration_routes() 是 <path:endpoint> 通配符,不携带任何权限元数据(main.py:1753)。",
|
|||
|
|
"fix": "在 dispatch_extension_request 内做集中式强制:取 permission_requirement(match.permission),auth_required 时校验 ctx.is_authenticated 返回 UNAUTHORIZED,required_permissions 逐项 ctx.has_permission 校验返回 FORBIDDEN;handler 内的 require_* 保留为纵深防御;补一个用 AST 遍历 ROUTES 断言每个 handler 体内存在等价 guard 的测试。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P0-致命",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "workflow / node / image / job / history 五条链路无归属校验(IDOR)",
|
|||
|
|
"location": "backend/routes/workflow_routes.py:50",
|
|||
|
|
"problem": "profile 路由(ProfileService.get_profile(profile_id, ctx) → _require_owner)和 round 路由(WorkflowRoundService._node_and_workflow → ResourceAccessService.require_node)都做了归属校验;但 workflow/node/image/job/history 的 service 方法签名里根本没有 ctx,路由层也没有额外校验。",
|
|||
|
|
"evidence": "# backend/routes/workflow_routes.py:50-61 —— 只校验\"已登录\"\ndef get_workflow(ctx, workflow_id, service=None) -> dict:\n workflow_service = service or _create_workflow_service()\n def action() -> dict[str, Any]:\n require_authenticated(ctx)\n return workflow_service.get_workflow(workflow_id).to_dict() # 无 ctx\n return handle_request(ctx, action)\n\n# backend/services/workflow_service.py:146-150\ndef list_workflows(self, profile_id: str) -> list[ProfileWorkflow]: # 无 ctx\n self._ensure_profile_exists(profile_id) # 只判存在,不判归属\n return self.workflow_repository.list_workflows(profile_id)\n\ndef get_workflow(self, workflow_id: str) -> WorkflowDetail: # 无 ctx\n\n# backend/services/workflow_service.py:650\ndef get_node(self, node_id: str) -> WorkflowNode: # 无 ctx\n\n# backend/routes/job_routes.py:29-35 —— 列表参数无任何 owner 维度\nresult = queue_service.list_jobs(status=..., workflow_id=..., profile_id=..., page=..., page_size=...).to_dict()",
|
|||
|
|
"impact": "任意登录用户可枚举/读取/修改他人资源:GET profiles/{他人id}/workflows、GET|PUT workflows/{id}、POST workflows/{id}/start|cancel|archive|execute-character(直接花对方模型额度出图)、POST profiles/{他人id}/workflows 在他人档案下建工作流、GET|PUT nodes/{id}、POST nodes/{id}/generate;GET jobs 无任何过滤且 GenerationTask.to_dict() 含 request 全量生成请求(他人完整提示词、参考图 URI、模型路由);GET images 无 owner 过滤。行为还自相矛盾:GET nodes/{id} 能读到他人节点,而 GET nodes/{id}/rounds 会正确 404。",
|
|||
|
|
"fix": "把已存在的 ResourceAccessService 推广到这五条链路:1) WorkflowService.list_workflows/get_workflow/get_node/list_nodes 增加 ctx 参数并走 require_workflow/require_node;2) QueueService.list_jobs 与 ImageService.list_images 增加 owner_user_id 过滤参数,路由层传 ctx.actor_id(有 can_manage_all_profiles 时传 None);3) HistoryService 相关方法前置 require_workflow;4) 补越权矩阵测试:用户 A 建资源,用户 B 对全部路由逐条请求断言 404/403。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "客户端自带 content_permissions / entrypoint,可绕过成人内容路由闸门",
|
|||
|
|
"location": "backend/services/generation_service.py:2366",
|
|||
|
|
"problem": "POST nodes/{node_id}/generate 的 payload 原样透传给 GenerationService.generate_node,而路由裁决所依赖的\"用户是否被允许成人内容\"直接从请求体里取,从未与 ctx 校验。entrypoint 同样来自 payload。",
|
|||
|
|
"evidence": "def _permission_context(payload: dict[str, Any]) -> dict[str, Any]:\n configured = payload.get(\"content_permissions\") if isinstance(payload.get(\"content_permissions\"), dict) else {}\n result = {\n **configured,\n \"adult_user_allowed\": bool(configured.get(\"adult_user_allowed\", payload.get(\"adult_user_allowed\", False))),\n \"quota_allowed\": bool(configured.get(\"quota_allowed\", payload.get(\"quota_allowed\", True))),\n }\n\n# generation_service.py:2352\ndef _entrypoint(payload): value = _optional_text(payload.get(\"entrypoint\") or payload.get(\"platform\") or payload.get(\"source\")) or \"web\"\n\n# prompt_governance_service.py:262-267 消费点\nentry_allowed = bool((self.config.get(\"entrypoints\") or {}).get(entrypoint, False))\nuser_allowed = bool(permissions.get(\"adult_user_allowed\", False))\nquota_allowed = bool(permissions.get(\"quota_allowed\", True))",
|
|||
|
|
"impact": "只要服务端 prompt_governance.routing.adult_auto_route_enabled 为 true,任意登录用户 POST {\"content_permissions\": {\"adult_user_allowed\": true, \"adult_group_allowed\": true, \"quota_allowed\": true}, \"entrypoint\": \"<配置放行的入口>\"} 即可让 adult_allowed 级别提示词自动路由到 nsfw_route,同时绕过配额判定。聊天侧是对的(chat_workflow_service 由服务端 _content_permissions(config, chat) 生成),唯独 Web 路由缺失。",
|
|||
|
|
"fix": "在路由层剥离客户端提供的 content_permissions / adult_* / quota_allowed / entrypoint / platform / source 字段,由服务端按 ctx.actor_id 查配置白名单重建(复用 chat_workflow_service._content_permissions 的逻辑),entrypoint 固定为 \"web\"。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "所有 service 在 require_* 之前构造:未授权即触发重型初始化 + DB 写入,构造期异常逃出 handle_request",
|
|||
|
|
"location": "backend/routes/node_routes.py:140",
|
|||
|
|
"problem": "每个 handler 的固定写法是先在函数体里造 service,再把 require_* 放进闭包 action,于是鉴权发生在最贵的操作之后。_create_generation_service() 会读 load_config_payload()、建 Database、调 sync_generation_provider_from_config()(向数据库 upsert provider/route 行)、构造 provider runtime 与多个 governance service。而 handle_request 的 try 只包住 callback。",
|
|||
|
|
"evidence": "# backend/routes/node_routes.py:132-146\ndef generate_node(ctx, node_id, payload=None, service=None, generation_service=None) -> dict:\n workflow_service = service or _create_workflow_service()\n resolved_generation_service = generation_service or (None if service else _create_generation_service()) # 鉴权之前\n def action() -> dict[str, Any]:\n require_write(ctx) # 鉴权在这里,已经太晚\n\n# backend/routes/round_routes.py:28 —— 只读接口也走全量 provider 初始化\nround_service = service or _create_round_service() # → create_generation_service() → sync_..._from_config()\n\n# backend/routes/common.py:12-16\ndef handle_request(ctx, callback):\n try:\n return success_response(callback(), request_id=ctx.request_id)\n except Exception as exc:\n return response_from_exception(exc, request_id=ctx.request_id)",
|
|||
|
|
"impact": "1) 未授权资源消耗:无权限用户狂刷 GET nodes/x/rounds 或 POST nodes/x/generate,每次都触发配置解析 + provider 表写入,SQLite 写锁争用(timeout=30s)会拖垮整个插件;2) 500 泄漏:service 构造在 try 之外,配置损坏/数据库锁死时异常直接穿透 dispatch_extension_request(该函数也无 try)到 AstrBot Quart,返回带 traceback 的 500;3) 只读接口产生写副作用。",
|
|||
|
|
"fix": "1) 把 service 构造挪进 action 闭包(可脚本化批量修改);2) 或引入 @route(permission=..., factory=...) 装饰器统一\"先鉴权 → 再懒构造 → 再异常包装\";3) sync_generation_provider_from_config 从请求路径移出,改为启动时执行一次 + 配置变更事件触发;4) dispatch_extension_request 外层加兜底 try。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "POST profiles/{profile_id}/archive 实际执行的是删除",
|
|||
|
|
"location": "backend/routes/profile_routes.py:109",
|
|||
|
|
"problem": "archive_profile 直接把调用转给 delete_profile,语义与端点名、与前端 profileArchive 按钮的用户预期完全不符。对照 workflow 侧是真归档(workflow_routes.py:106 → _set_workflow_status(..., \"archived\", ...))。",
|
|||
|
|
"evidence": "def delete_profile_via_post(ctx, profile_id, payload=None, service=None) -> dict:\n return delete_profile(ctx, profile_id, service)\n\ndef archive_profile(ctx, profile_id, service=None) -> dict:\n return delete_profile(ctx, profile_id, service) # \"归档\" == \"删除\"",
|
|||
|
|
"impact": "前端 endpoints.ts:11 的 profileArchive 标注为\"归档\",用户点击后档案被 repository.soft_delete() 置为 status=\"deleted\",随后 get_profile 对 status == \"deleted\" 抛 404 —— 该档案及其下所有工作流从 UI 彻底消失且无归档列表可恢复。归档与删除两种不同用户意图被合并成一种破坏性操作。",
|
|||
|
|
"fix": "archive_profile 改为 profile_service.update_profile(profile_id, {\"status\": \"archived\"}, ctx),并确认 ProfileRepository.list 对 archived 状态可查(list_profiles 已支持 status 参数)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "所有错误一律 HTTP 200,UNAUTHORIZED/FORBIDDEN/NOT_FOUND 无法被 HTTP 层识别",
|
|||
|
|
"location": "main.py:2439",
|
|||
|
|
"problem": "error_response() 返回普通 dict,_to_bridge_response() 只是把它包成 {\"status\": \"error\", ...},全程没有任何 HTTP 状态码映射。",
|
|||
|
|
"evidence": "def _to_bridge_response(response: dict[str, Any]) -> dict[str, Any]:\n if response.get(\"ok\"):\n return {\"status\": \"ok\", \"message\": \"\", \"data\": response.get(\"data\"), \"request_id\": response.get(\"request_id\")}\n return {\"status\": \"error\", \"message\": str(response.get(\"message\") or \"request failed\"),\n \"data\": response, \"request_id\": response.get(\"request_id\")}\n # 无论 UNAUTHORIZED / FORBIDDEN / RESOURCE_NOT_FOUND / UNKNOWN_ERROR,AstrBot 都按 200 返回",
|
|||
|
|
"impact": "AstrBot 访问日志、反向代理限流、监控告警、浏览器 devtools 全部看到 200,权限拒绝与服务器内部错误在 HTTP 层不可观测;前端 client.ts:96 的 !response.ok 分支在桥接模式下永远不触发。安全审计(谁被拒绝了多少次)无法从访问日志得出。",
|
|||
|
|
"fix": "建立 ErrorCode → HTTP 状态码映射表(UNAUTHORIZED=401, FORBIDDEN=403, RESOURCE_NOT_FOUND=404, VALIDATION_FAILED=422, UPLOAD_TOO_LARGE=413, UNKNOWN_ERROR=500),在 _to_bridge_response / _to_media_http_response 层构造带状态码的 Response;前端 unwrapApiResponse 已能兼容两种包装,改动向后兼容。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "handle_request 吞掉全部异常且不写日志,未知错误无从定位",
|
|||
|
|
"location": "backend/dto/common.py:75",
|
|||
|
|
"problem": "任何非 ApiError 异常都被转成 UNKNOWN_ERROR / \"服务器内部错误\",details 只留一个异常类名,既不 logging 也不保留 traceback。157 个 handler 全部走这一条路径。",
|
|||
|
|
"evidence": "def response_from_exception(exc: Exception, request_id: str | None = None) -> dict[str, Any]:\n if isinstance(exc, ApiError):\n return exc.to_response(request_id=request_id)\n return error_response(\n error_code=ErrorCode.UNKNOWN_ERROR,\n message=\"服务器内部错误\",\n details={\"exception\": exc.__class__.__name__}, # 只有类名\n request_id=request_id,\n )",
|
|||
|
|
"impact": "生产环境出现 KeyError / sqlite3.OperationalError 时,运维只能看到\"服务器内部错误 + KeyError\",没有文件行号、没有与 request_id 关联的堆栈,排障必须靠复现。项目自带 LogService(logs/errors 端点)却没有被这条路径喂数据,该端点长期为空。",
|
|||
|
|
"fix": "handle_request 分离 ApiError 与未知异常两个 except 分支,未知异常用 logger.exception 打印含 request_id 与 actor 的完整堆栈,并落一条 LogService 错误日志,让 GET logs/errors 真正有内容。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "分页之后再过滤,total 与页大小失真",
|
|||
|
|
"location": "backend/routes/job_routes.py:36",
|
|||
|
|
"problem": "service 已按 page/page_size 从数据库分页取回,路由层再对当前页做可见性过滤,然后用当前页的隐藏数去修正全局 total。image_routes.py:45-49 有同样问题。",
|
|||
|
|
"evidence": "# job_routes.py:36-38\nif not ctx.has_permission(\"can_manage_materials\"):\n visible = [item for item in result[\"items\"] if item.get(\"metadata\", {}).get(\"source\") != \"advanced_material\"]\n result = {**result, \"items\": visible,\n \"total\": max(int(result.get(\"total\") or 0) - (len(result[\"items\"]) - len(visible)), 0)}\n\n# image_routes.py:45-49\nvisible = [item for item in payload[\"items\"] if not _is_restricted_image(item)]\npayload[\"restricted_hidden\"] = len(payload[\"items\"]) - len(visible)\npayload[\"items\"] = visible\npayload[\"total\"] = max(len(visible), int(payload.get(\"total\") or 0) - payload[\"restricted_hidden\"])",
|
|||
|
|
"impact": "total 是全表计数(来自 repository.list() 第二个返回值),减去的却只是当前页被隐藏的条数 → 分页器页数算错;每页实际条数少于 page_size 且各页不等长;极端情况某页全是受限项会显示空页却还有下一页。用户看到\"共 87 条\"但翻完只有 40 条。",
|
|||
|
|
"fix": "把可见性下推到仓储层 WHERE 条件:JobRepository.list(..., exclude_sources=(\"advanced_material\",));ImageRepository.list_images(..., include_restricted: bool) 在 SQL 里过滤 metadata 的 quarantined / gallery_visibility。这样 total 与分页天然一致,路由层只剩权限判定。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "ProfileService._require_owner 不认管理员权限,与 ResourceAccessService._can_access 语义相反",
|
|||
|
|
"location": "backend/services/profile_service.py:160",
|
|||
|
|
"problem": "同一套归属判定有两份实现且规则不同——一处认 can_manage_all_profiles / *,另一处只认 actor_id 相等。",
|
|||
|
|
"evidence": "# backend/services/profile_service.py:160-169\ndef _require_owner(self, profile, ctx) -> None:\n if ctx.actor_id and ctx.actor_id == profile.owner_user_id:\n return\n raise ApiError(ErrorCode.RESOURCE_NOT_FOUND, \"角色档案不存在\", {\"profile_id\": profile.profile_id})\n\n# backend/services/resource_access_service.py:80-84\n@staticmethod\ndef _can_access(owner_user_id, ctx) -> bool:\n if ctx.has_permission(\"can_manage_all_profiles\") or ctx.has_permission(\"*\"):\n return True\n return bool(ctx.actor_id and owner_user_id and ctx.actor_id == owner_user_id)",
|
|||
|
|
"impact": "即使按 P0 建议把 can_manage_all_profiles 正确授予管理员,GET/PUT/DELETE profiles/{id} 依然会对管理员 404,而 GET nodes/{id}/rounds 却放行——管理能力在两条链路上表现相反,运维无法完成跨用户支持工单。连 RequestContext.system()(permissions={\"*\"})在 profile 路径上也会被拒。",
|
|||
|
|
"fix": "删掉 _require_owner,把归属判定提升为 core/security.py 中唯一实现 can_access_owner(owner_user_id, ctx),ProfileService 与 ResourceAccessService 共用同一份。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "157 个 handler 是同构样板 + 22 个各自为政的 service 工厂 + 30 处 Database(load_database_settings())",
|
|||
|
|
"location": "backend/routes/settings_routes.py:42",
|
|||
|
|
"problem": "每个 handler 都是同一个五行骨架(service = service or _create_X() → def action(): require_*(ctx); return svc.foo(...) → return handle_request(ctx, action));每个路由模块自己写一份 _create_*_service(),都重复 Database + AuditService + 一堆 repository 手工装配。统计:return handle_request 157 处,Database(load_database_settings()) 30 处,_create_*_service 工厂 22 个,其中 WorkflowService 装配在 node_routes / workflow_routes / history_routes 里是逐字重复的三份。",
|
|||
|
|
"evidence": "# settings_routes.py:42-53 与 56-67 与 70-81 …… 仅函数名和一行调用不同\ndef validate_settings(ctx, payload, service=None) -> dict:\n settings_service = service or SettingsService()\n def action() -> dict[str, Any]:\n require_config_manager(ctx)\n return settings_service.validate(payload)\n return handle_request(ctx, action)\n\n# data_backup_routes.py:9-11 —— 为压行改用元组副作用\ndef list_backups(ctx:RequestContext,service:DataBackupService|None=None)->dict:\n instance=service or DataBackupService()\n return handle_request(ctx,lambda:(require_config_manager(ctx),instance.list_backups())[1])\n\n# wardrobe_routes.py:94-96 —— 同样的元组 hack\nreturn handle_request(ctx, lambda: (require_permission(ctx, \"can_manage_materials\"), wardrobe_service.save_template(profile_id, payload, ctx))[1])",
|
|||
|
|
"impact": "新增路由的成本与出错面被放大——前述越权与\"构造在鉴权前\"两类缺陷正是因为骨架靠复制粘贴而非统一封装才成片出现;改一处横切逻辑(日志、审计、限流、状态码)要动 157 个地方。三份重复的 WorkflowService 装配意味着依赖变更要改三处,漏一处就是运行时 TypeError。(guard, call())[1] 把鉴权伪装成表达式的一部分,静态检查与人眼 review 都极易漏看。",
|
|||
|
|
"fix": "1) 引入 @route(permission, factory) 装饰器把\"鉴权 → 懒构造 service → 异常包装\"收敛到 common.py 一处,handler 退化成一行业务调用,permission 参数可与 RouteSpec.permission 用测试断言强制一致;2) 建 backend/routes/container.py 提供 get_workflow_service()/get_image_service()/get_queue_service() 等进程级或按请求缓存的工厂,删掉 22 份重复装配与 30 处 Database 构造;3) 用普通 def action() 替换 data_backup_routes 与 wardrobe_routes 里的元组写法。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "public_token 权限档与实际暴露面不符,且 media/fetch 绕过分发器不留审计",
|
|||
|
|
"location": "main.py:329",
|
|||
|
|
"problem": "矩阵里 public_token 声明 auth_required: False,但真正的注册路径是通配符桥接(/{PLUGIN_NAME}/<path:endpoint>,不携带权限元数据),media/fetch 与其余路由共用 AstrBot 同一层鉴权——media_routes.py:56-63 的注释也承认 /api/plug 请求需要登录态。同时该端点在 dispatch_bridge_request 里被提前拦截,完全不经过 dispatch_extension_request。",
|
|||
|
|
"evidence": "# main.py:329-332\nif dispatch_method == \"GET\" and normalize_extension_path(dispatch_endpoint) == \"media/fetch\":\n file_response = await _to_media_file_response(dispatch_query)\n if file_response is not None:\n return file_response # 不经过路由解析、不经过 handle_request、无审计\n\n# main.py:1753-1758 通配符注册,permission 字段在这条路径上不存在\ndef _bridge_registration_routes() -> list[dict[str, str]]:\n path = f\"/{PLUGIN_NAME}/<path:endpoint>\"\n return [{\"method\": method, \"path\": path, \"endpoint\": \"\"} for method in sorted({r.method for r in ROUTES})]",
|
|||
|
|
"impact": "矩阵读起来像\"这里有一个免鉴权端点\",实际不是,误导后续维护者对暴露面的判断;反过来若 AstrBot 放开 /api/plug 的鉴权,这条提前返回路径会真的变成公开文件读取(目前靠 MediaProxy 的 HMAC 签名 + 300s TTL + 进程随机密钥兜住)。该路径不产生任何访问记录。",
|
|||
|
|
"fix": "把 public_token 的描述改为\"签名 token + 宿主鉴权\",或在 _to_media_file_response 前补一次 ctx.is_authenticated 判定使其与声明一致;同时在提前返回前记一条 LogService 媒体访问日志(含 request_id、actor、uri)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "node_routes._bool 与 params.as_bool 重复;images/{id}/moderation-review 无前端调用方",
|
|||
|
|
"location": "backend/routes/node_routes.py:399",
|
|||
|
|
"problem": "node_routes 自带一份布尔转换,与公共 params.as_bool(value, non_string_truthiness=True) 语义完全等价却各写一遍;另外 POST images/{image_id}/moderation-review(materials 权限)在 frontend/shared/api/endpoints.ts 全表里没有对应条目,前端无任何调用方。",
|
|||
|
|
"evidence": "# backend/routes/node_routes.py:399-402\ndef _bool(value: Any) -> bool:\n if isinstance(value, str):\n return value.strip().lower() in {\"1\", \"true\", \"yes\", \"on\"}\n return bool(value)\n\n# backend/routes/params.py:60-82 已有等价实现\ndef as_bool(value, *, default=False, non_string_truthiness=False) -> bool: ...",
|
|||
|
|
"impact": "两份布尔语义未来会漂移(例如公共版本新增 \"y\" token 而 node_routes 版本不跟);受限图片复核能力做了后端却在 UI 上不可达,_is_restricted_image 隔离的图片实际上没有人工放行入口。",
|
|||
|
|
"fix": "get_node_dependency_context 改用 as_bool(params.get(\"include_draft\"), non_string_truthiness=True) 并删除 _bool;确认 moderation-review 是否应在图库详情补一个\"管理员复核\"操作,否则从 ROUTES 移除以缩小暴露面。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "engines/ 模型引擎 + adapters/ 平台适配器",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A6-引擎与适配器.md",
|
|||
|
|
"summary": "图像 provider 链路(engines/image_provider.py + core/generation_errors.py)工程质量明显高于其余部分:URL 校验、响应字节上限、按图片数动态计算 JSON 上限、错误分类、密钥脱敏都做到位了。问题集中在它旁边的三条\"次要\"链路——vision provider、structured_text/structured_vision callable、聊天媒体导入——这三条都缺少图像链路已有的防护,是典型的\"主路径加固过、旁路径没跟上\"。最严重的是一条功能性死链:size 字段从聊天命令一路透传到 task.request,但 engines/single_flow.py:61 只读 aspect_ratio,全仓库无任何 size→aspect_ratio 映射,导致 `/yp 单图 --尺寸 4K` 和 `/yp 高清 <id> 4K` 全部产出 1024x1024,高清修复功能等于不生效。其次 ChatMediaImportService 的兜底分支把任意字符串当本地文件路径直接 read_bytes(),输入来自聊天消息组件的 file/path 字段,构成任意本地文件读取面。关于四个 flow 文件合并:结论是不建议合并——single_flow/review_flow 是真引擎(455/451行),另三个是纯声明式节点模板目录,实质重复仅约 10 行构造器样板且 metadata 契约各不相同;真正值得合并的是 onebot11_adapter 与 wechat_adapter(结构性重复约 90%)。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\__init__.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\image_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\vision_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\single_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\review_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\character_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\fast_character_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\quick_profile_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_card_renderer.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_command_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_command_parser.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_media_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\chat_commands.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\onebot11_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\wechat_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\webui_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_session_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_media_import_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\model_provider_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\web_hosting_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\profile_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\generation_errors.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\image_formats.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\identity.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\remote_fetch.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\chat_command_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\security.py"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "size 请求字段全链路无人消费,聊天尺寸参数与高清修复静默失效",
|
|||
|
|
"location": "engines/single_flow.py:61",
|
|||
|
|
"problem": "SingleFlowEngine.generate 只从请求里读 aspect_ratio,而聊天单图(chat_workflow_service.py:63)与高清修复(chat_workflow_service.py:929)写入的是 size(\"1920x1080\"/\"2560x1440\"/\"3840x2160\")。generation_service.py:733 的 runtime_generation_request = {**payload, ...} 原样透传 size 进 task.request,但全仓库(排除 .tmp/、tests/)对 aspect_ratio 的写入只出现在 advanced_material_service.py:912、advanced_style_set_service.py、node_injection_template_service.py、prompt_expansion_service.py,没有任何一处把 size 映射成 aspect_ratio。",
|
|||
|
|
"evidence": "# engines/single_flow.py:56-61\ndef generate(self, node: WorkflowNode, task: GenerationTask, request: dict[str, Any]) -> list[MockGeneratedAsset]:\n count = _bounded_count(request.get(\"count\", 1))\n ...\n width, height = _dimensions(str(request.get(\"aspect_ratio\") or \"1:1\"))\n\n# engines/single_flow.py:248-255\ndef _dimensions(aspect_ratio: str) -> tuple[int, int]:\n if aspect_ratio == \"3:4\": return 768, 1024\n if aspect_ratio == \"4:3\": return 1024, 768\n if aspect_ratio == \"16:9\": return 1280, 720\n return 1024, 1024\n\n# engines/single_flow.py:154-159(真正发给 provider 的尺寸)\nresult = provider.generate_image(ImageGenerationRequest(..., size=f\"{width}x{height}\", ...))",
|
|||
|
|
"impact": "/yp 单图 ... --尺寸 4K 与 /yp 高清 <image_id> 4K 都发出 size=\"1024x1024\";高清修复的唯一价值(提升分辨率)完全失效,用户只会得到一张同尺寸重绘图。astrbot_command_parser.py:49-52 精心校验的 1K/2K/4K 参数是纯装饰。数据库记录的 width/height 也随之全是 1024。",
|
|||
|
|
"fix": "在 single_flow.generate 增加显式尺寸解析优先级并让 size 成为一等输入:width, height = _resolve_dimensions(request.get(\"size\"), request.get(\"aspect_ratio\")),其中 _resolve_dimensions 先用 re.fullmatch(r\"(\\d{2,5})\\s*[x×]\\s*(\\d{2,5})\", text) 解析 size 并 clamp 到 64..4096,解析失败再回退 _dimensions(aspect_ratio)。同时补一条端到端断言测试覆盖 /yp 高清 <id> 4K 产出的 ImageGenerationRequest.size。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "聊天媒体导入把任意字符串当本地文件路径读取(任意本地文件读取面)",
|
|||
|
|
"location": "backend/services/chat_media_import_service.py:33",
|
|||
|
|
"problem": "import_image 的兜底分支在既不是 storage://、也不是 data:image/、也不是 http(s):// 时,直接把字符串当本地路径 expanduser().resolve().read_bytes(),没有任何根目录约束或白名单。输入来自聊天消息 Image 组件的 url/file/path/image_url 属性(adapters/astrbot_media_adapter.py:15-18 → main.py:1161,1167),这些字段是消息内容的一部分而非可信服务端常量。同一函数还完全没有图片魔数校验(对比 engines/single_flow.py:447 _require_image_format),扩展名只是被强制改写成 jpg;data:image/ 分支的 base64.b64decode 也在解码之后才检查大小。",
|
|||
|
|
"evidence": "# backend/services/chat_media_import_service.py:30-36\nelif text.startswith((\"http://\", \"https://\")):\n body = read_http_response_limited(text, timeout_seconds=30, max_bytes=MAX_CHAT_IMAGE_BYTES)\n extension = Path(urlsplit(text).path).suffix.lstrip(\".\").lower() or \"jpg\"\nelse:\n source = Path(text).expanduser().resolve()\n body = source.read_bytes() # <- 无任何根目录约束\n extension = source.suffix.lstrip(\".\").lower() or \"jpg\"",
|
|||
|
|
"impact": "只要消息组件的路径字段可被影响(file:///、相对路径、~/.astrbot/...),任意宿主文件即可被读入并作为素材落盘到 storage://materials/chat/...,随后通过素材列表接口读回,等价于任意文件读取 + 外带。非图片内容也会被当图片存储。超大 data URI 会先被完整解码进内存再判超限。",
|
|||
|
|
"fix": "1) 删除本地路径兜底分支,或改为对显式配置的缓存根做 Path.resolve() 前缀约束(参考 backend/services/web_hosting_service.py:58-67 的 _resolve 已写对的做法);2) 落盘前统一走 core.image_formats.detect_image_format(body),None 则拒绝,并用检出的 image_format.extension;3) data: 分支先按 len(encoded) > ((MAX_CHAT_IMAGE_BYTES + 2) // 3) * 4 预检再解码(engines/single_flow.py:398-404 已有正确写法可复用)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "内容安全审核 provider 解析缺 modality 与 provider_type 校验,与同文件其他两处不一致",
|
|||
|
|
"location": "backend/services/model_provider_runtime.py:285",
|
|||
|
|
"problem": "_structured_vision_callable_for_single_route 调用 resolve_route(route_name) 不带 modality=,且没有 provider_type == \"openai_compatible\" 的前置判断。同文件 _vision_provider_for_single_route(:131-136) 两者都有,_structured_text_callable_for_single_route(:200) 有 provider_type 判断。该 callable 承担图像内容安全审核(涉政/涉恐/未成年人识别),下面 review() 硬编码 OpenAI chat/completions 报文形状(:331-341)。",
|
|||
|
|
"evidence": "# :131-136(正确的对照组)\nresolved = self.model_provider_service.resolve_route(route_name, modality=\"vision_recognition\")\nif not resolved:\n return None\nprovider = resolved.get(\"provider\") or {}\nif str(provider.get(\"provider_type\") or \"openai_compatible\") != \"openai_compatible\":\n return None\n\n# :285-289(缺两道校验)\nresolved = self.model_provider_service.resolve_route(route_name)\nif not resolved:\n return None\nprovider = resolved.get(\"provider\") or {}\nsecret = resolve_secret_value(...) # 直接进入取密钥,没有 provider_type 判断",
|
|||
|
|
"impact": "同名路由若在 image_generation 模态下存在会被解析成审核模型;非 openai_compatible 的 provider 也会被套上 OpenAI 报文格式发送。两种情况都在审核时抛异常,而 review_with_fallback(:264-271) 对所有异常一律吞掉换下一条路由,全部失败后抛 RuntimeError——上层若把审核不可用当放行即安全审核静默绕过,当拒绝则是全量误杀。",
|
|||
|
|
"fix": "与 :131-136 对齐——补 modality=\"vision_recognition\"(或为审核新增独立 modality)与 provider_type 判断;并把三处 resolve_route + 密钥解析 + provider_type 校验抽成一个私有方法,杜绝再次漂移。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "缩略图是原图的字节级完整拷贝,存储与带宽占用直接翻倍",
|
|||
|
|
"location": "engines/single_flow.py:178",
|
|||
|
|
"problem": "provider 返回的图片字节被原样写入 image_uri 和 thumbnail_uri 两个路径。全仓库(engines/、backend/services/generation_service.py)不存在任何降采样/缩放代码,core/image_formats.py 也只做魔数识别、不解析尺寸。",
|
|||
|
|
"evidence": "# engines/single_flow.py:176-179\nimage_format = _require_image_format(image_bytes, provider_id=result.provider_id)\nimage_uri, thumbnail_uri = _output_uris(task, request, index, image_format.extension)\nself._write_image(image_uri, image_bytes)\nself._write_image(thumbnail_uri, image_bytes) # 同一份字节写两遍",
|
|||
|
|
"impact": "单张图上限 MAX_PROVIDER_IMAGE_BYTES = 25 MiB(engines/image_provider.py:21),最坏情况每张生成图占 50 MiB。候选池默认一轮 4 张、快速角色 4 节点并发,磁盘增长是设计预期的 2 倍。前端列表页加载缩略图实际拉的是全尺寸原图,网络与内存开销同样翻倍。",
|
|||
|
|
"fix": "新增 _write_thumbnail(uri, image_bytes, max_edge=512),用已作为可选依赖存在的 Pillow(adapters/astrbot_card_renderer.py:56 已 from PIL import Image)做等比缩放;Pillow 不可用时退化为写原图并在 metadata 打 thumbnail_downscaled: False,便于后续补偿任务识别。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "vision provider 链路完全没有错误分类与可重试判定,retryable 字段是死代码",
|
|||
|
|
"location": "backend/services/material_service.py:863",
|
|||
|
|
"problem": "图像链路有 core/generation_errors.provider_failure_attribution 把 401/403→auth_error、429→rate_limited、5xx→provider_unavailable 并计算 retryable(core/generation_errors.py:102-125)。vision 链路没有任何等价物:_vision_provider_error 只是把 error_type 原样搬运,retryable 依赖 exc.details.get(\"retryable\")——而 engines/vision_provider.py 全文从未往 details 里写过 retryable,因此那两行永远不会执行。",
|
|||
|
|
"evidence": "# backend/services/material_service.py:863-865\n\"response_issue\": str(exc)[:300],\n}\nif isinstance(exc.details.get(\"retryable\"), bool): # vision_provider 从不设置该键\n failure[\"retryable\"] = exc.details[\"retryable\"]\n\n# engines/vision_provider.py:304-312(唯一写 details 的地方,只有 body_preview / body_truncated)\nexcept urllib.error.HTTPError as exc:\n raise VisionProviderError(..., status_code=exc.code, error_type=\"http_error\", # 401/429/503 全都是 http_error\n details={\"body_preview\": body[:500], \"body_truncated\": len(body) > MAX_PROVIDER_ERROR_BYTES}) from exc",
|
|||
|
|
"impact": "素材识别失败时鉴权错误(需改配置)、限流(应退避重试)、上游 5xx(应重试)在 error_type 上完全无法区分,重试策略无从制定,UI 只能给通用文案。_vision_provider_error 还绕过了 core/generation_errors._safe_message 的密钥脱敏正则,reason/parse_error 未清洗即落入 ApiError.details 返回给客户端(advanced_material_service.py:787、advanced_style_set_service.py:1283)。",
|
|||
|
|
"fix": "把 core/generation_errors.py 的 _category/_is_retryable/_safe_details/_safe_message 提成与异常类型无关的公共函数(入参 error_type + status_code + details),让 provider_failure_attribution 与新增的 vision_failure_attribution 共用;material_service._vision_provider_error 改为调用后者,删掉那两行死代码。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "vision 请求体无上限,8 张 20MB 图可构造约 213MB 的单次 JSON 请求",
|
|||
|
|
"location": "engines/vision_provider.py:294",
|
|||
|
|
"problem": "vision_provider 只对响应设了 MAX_PROVIDER_JSON_BYTES = 4 MiB,对请求体没有任何限制。多图分析路径用 image_uris 一次传最多 MAX_VISION_IMAGES = 8 张(advanced_style_set_service.py:1545-1548,1654-1657,2281-2284),每张经 material_service._storage_image_data_url 转成 base64 data URI,单张原图上限 MAX_UPLOAD_BYTES = 20 MiB(material_service.py:43),base64 后约 27 MiB。",
|
|||
|
|
"evidence": "# engines/vision_provider.py:17-19\nMAX_PROVIDER_JSON_BYTES = 4 * 1024 * 1024 # 只约束响应\nMAX_PROVIDER_ERROR_BYTES = 64 * 1024\nMAX_VISION_IMAGES = 8\n\n# engines/vision_provider.py:294\ndata = json.dumps(payload, ensure_ascii=False).encode(\"utf-8\") # 完整物化,无上限",
|
|||
|
|
"impact": "单次风格集共识分析峰值内存约为 payload dict + json 字符串 + utf-8 bytes ≈ 3 份 213 MB。插件运行在 AstrBot 主进程内,OOM 会拖垮整个 Bot。且几乎所有上游会以 413 拒收,等于一条注定失败但先把宿主打爆的路径。对比 engines/image_provider.py:392-401 的 _provider_json_limit,图像链路已按图片数动态算过上限,vision 没跟上。",
|
|||
|
|
"fix": "新增 MAX_VISION_REQUEST_BYTES(如 32 MiB),在 _post_json 里 if len(data) > MAX_VISION_REQUEST_BYTES: raise VisionProviderError(..., error_type=\"invalid_request\");更好的做法是在 _storage_image_data_url 之前按张数 × 单张预算约束,超限时先降采样再编码。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "三处 fallback 对所有错误一视同仁地换 provider,包括不该重试的鉴权/配置/内容策略错误",
|
|||
|
|
"location": "engines/image_provider.py:191",
|
|||
|
|
"problem": "core/generation_errors._is_retryable(:122-125)已把可重试语义定义清楚,但 FallbackImageProvider、FallbackVisionProvider、complete_with_fallback(model_provider_runtime.py:180-189)、review_with_fallback(:264-271) 四个循环没有一个使用它。",
|
|||
|
|
"evidence": "# engines/image_provider.py:191-192\nexcept ImageProviderError as exc:\n failures.append({\"provider_id\": provider.provider_id, \"error_type\": exc.error_type, \"status_code\": exc.status_code})\n # 无论 configuration_error / 401 / 400 内容策略拒绝,一律继续下一个 provider\n\n# backend/services/model_provider_runtime.py:182-186\nfor completion in callables:\n try:\n return completion(request_payload)\n except Exception as exc: # 连 ValueError/JSONDecodeError 都算\"该换路由\"\n failures.append(str(exc)[:500])",
|
|||
|
|
"impact": "一个被内容策略拒绝的提示词(400)会依次打完所有已启用路由——而 image_provider_for_route 在 fallback_enabled 时会把全部启用的 image_generation 路由追加进候选(model_provider_runtime.py:60-71)。一次用户请求放大成 N 次必然失败的付费调用,同时 N 倍推高各家限流计数。密钥填错时用户看到的是\"所有候选生成服务均调用失败\"而不是\"鉴权失败\",排障方向被带偏。",
|
|||
|
|
"fix": "让 _is_retryable 成为 fallback 准入条件:except ImageProviderError as exc: attribution = provider_failure_attribution(exc); failures.append(attribution); if not attribution[\"retryable\"]: raise —— 鉴权/配置/请求错误立即上抛,不消耗其余 provider。structured_text_callable / structured_vision_callable 需先给 HTTP 错误分类才能做同样判断。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "structured_text_callable 用裸 json.loads 解析模型输出,缺 vision 侧已有的容错",
|
|||
|
|
"location": "backend/services/model_provider_runtime.py:246",
|
|||
|
|
"problem": "engines/vision_provider._json_from_text(:398-429)实现了一套克制的修复策略——剥离 BOM、从首个 { 到末个 } 截取(处理 markdown 代码围栏)、去尾逗号,共 4 个候选依次尝试。structured_text_callable 走同类 OpenAI chat/completions 接口、同样带 response_format: json_object,却直接 json.loads(content)。",
|
|||
|
|
"evidence": "# backend/services/model_provider_runtime.py:244-249\nresponse = json.loads(raw)\ncontent = _completion_content(response)\nparsed = json.loads(content) # 模型多输出一个 ```json 围栏就炸\nif not isinstance(parsed, dict):\n raise ValueError(\"structured completion must return an object\")",
|
|||
|
|
"impact": "structured_text_callable 是提示词扩写/结构化处理的入口。大量 OpenAI 兼容网关(尤其国产模型代理)不真正支持 response_format: json_object,会返回带围栏的 JSON——这些模型在 vision 路径能用、在 text 路径必失败,且失败形态是裸 json.JSONDecodeError 冒泡,complete_with_fallback 把它当作换路由信号继续烧下一条路由(与上一条叠加)。",
|
|||
|
|
"fix": "把 vision_provider._json_from_text 提到 core/(例如 core/llm_json.py),vision、structured_text、structured_vision 三处共用。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "单条聊天消息内的图片数量无上限,放大为 N 次下载 + N 次视觉大模型调用",
|
|||
|
|
"location": "adapters/astrbot_media_adapter.py:12",
|
|||
|
|
"problem": "image_references_from_event 遍历整个消息链、对每个 image 组件收集一个引用,没有任何数量截断;main.py:1167 随后对每个引用串行执行完整的导入+识别流程,每次识别都是一次视觉大模型调用。ChatCommandGuardService 的配额校验在 submit_single_image 里有(chat_workflow_service.py:38),import_material 路径上完全没有。",
|
|||
|
|
"evidence": "# main.py:1161-1167\nreferences = image_references_from_event(event)\nif not category.strip() or not references:\n ...\nresults = [await asyncio.to_thread(self.chat_workflow_service.import_material, reference, category.strip(), context) for reference in references]\n\n# backend/services/chat_workflow_service.py:948-954(每个引用的代价)\nuri, filename, size_bytes = self.media_importer.import_image(image_reference) # 最多 20MB 下载\nmaterial = self._data(self.dispatcher(ctx, \"POST\", \"materials/upload\", payload={...}))\nrecognized = self._data(self.dispatcher(ctx, \"POST\", f\"materials/{material['material_id']}/recognize\", payload={})) # 视觉大模型调用",
|
|||
|
|
"impact": "QQ 合并转发或一条多图消息可轻易携带数十张图。一条命令即触发数十次 20MB 下载与数十次付费视觉调用,全部在 asyncio.to_thread 里串行执行,期间该用户的其它聊天命令排队等待。",
|
|||
|
|
"fix": "在 image_references_from_event 内部截断(MAX_CHAT_IMAGE_REFERENCES = 4,与 MAX_VISION_IMAGES 语义对齐),超出部分在回复里明示\"仅处理前 N 张\";并让 import_material 也走 self.guard.check_generation / self.guard.record,与单图命令一致。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "onebot11_adapter 与 wechat_adapter 结构性重复约 90%",
|
|||
|
|
"location": "adapters/wechat_adapter.py:49",
|
|||
|
|
"problem": "两个文件各定义一个仅字段名不同的消息 dataclass,然后逐字复制 match_workbench_command、workbench_reply、handle_message_event、quick_profile_reply、_optional_text 五个函数。除 platform 常量、群标识字段名(group_id vs room_id)、回复字典的键(\"message\" vs \"text\")外没有任何差异。",
|
|||
|
|
"evidence": "# adapters/onebot11_adapter.py:53-71\ndef handle_message_event(event, *, api_base=API_BASE):\n message = normalize_message_event(event)\n command = parse_chat_command(message.text)\n if command is None: return None\n if command.action == \"quick_profile\":\n return {\"platform\": \"onebot11\", \"action\": command.action, \"payload\": command.payload,\n \"source\": message.to_dict(), \"reply\": quick_profile_reply(command.payload[\"questionnaire\"][\"name\"])}\n return {\"platform\": \"onebot11\", \"action\": \"open_workbench\", \"source\": message.to_dict(), \"reply\": workbench_reply(api_base)}\n\n# adapters/wechat_adapter.py:49-67 —— 逐行相同,仅 \"onebot11\" -> \"wechat\"",
|
|||
|
|
"impact": "chat_commands.py 每次新增命令都必须记得同步改两个文件;两者已开始漂移(onebot11 的 workbench_reply 用 \"message\" 键、wechat 用 \"text\" 键)。新增第三个平台(Telegram/Discord,_chat_entrypoint 已为它们预留分支)会变成第三份拷贝。",
|
|||
|
|
"fix": "抽 adapters/chat_platform_adapter.py,定义 PlatformSpec(platform, group_field_names, user_field_names, reply_text_key) 与通用 handle_message_event(event, spec, *, api_base);两个现有模块退化为 spec 常量加薄封装,保持导出符号不变以免破坏 adapters/__init__.py。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "聊天身份未按平台命名空间隔离,跨平台同 ID 可互相访问档案",
|
|||
|
|
"location": "backend/services/chat_workflow_service.py:1157",
|
|||
|
|
"problem": "RequestContext.from_astrbot(user_id=chat.user_id) 直接用平台原始用户 ID;ProfileService 的归属判定基于 ctx.actor_id(profile_service.py:36,92,161),即 user_id or username。source 字段虽带了平台(f\"astrbot_chat:{chat.platform}\"),但它不参与任何归属判定。WebUI 侧的 actor_id 是 AstrBot 登录名(main.py:2325-2332),与聊天 ID 共用同一命名空间。",
|
|||
|
|
"evidence": "# backend/services/chat_workflow_service.py:1157-1164\nreturn RequestContext.from_astrbot(\n username=chat.user_id or \"chat-user\",\n user_id=chat.user_id or \"chat-user\", # 无平台前缀\n permissions={\"can_write\", \"can_retry_tasks\", \"can_manage_materials\"},\n source=f\"astrbot_chat:{chat.platform}\", # 平台信息只进了这里,不参与鉴权\n\n# backend/services/profile_service.py:160-163\ndef _require_owner(self, profile: CharacterProfile, ctx: RequestContext) -> None:\n if ctx.actor_id and ctx.actor_id == profile.owner_user_id:\n return",
|
|||
|
|
"impact": "QQ 号 12345 与 Telegram 用户 ID 12345(两者都是纯数字)会落进同一个 owner_user_id,互相看到并操作对方的全部角色档案、工作流、图片。一个用户名恰好是纯数字的 WebUI 账号同样会与聊天用户撞车。越界不需要任何攻击技巧,是配置巧合即触发。",
|
|||
|
|
"fix": "_request_context 里改用带平台前缀的复合标识:user_id=f\"{chat.platform or 'astrbot'}:{chat.user_id}\",WebUI 侧对应用 webui: 前缀。需配套一次性迁移脚本(把现存无前缀记录按 source 字段回填),对无法判定来源的记录保守归入 webui:。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "聊天卡片 PNG 缓存目录无淘汰机制,用户可任意撑爆磁盘",
|
|||
|
|
"location": "adapters/astrbot_card_renderer.py:59",
|
|||
|
|
"problem": "_render_card 以内容 sha256 前 16 位为文件名缓存到 data/chat_cards/,命中即复用。缓存永不过期、永不清理,而 key 的输入(角色名/基础/外观/风格)完全由用户在 /yp 快速角色 命令里提供(main.py:526-529)。",
|
|||
|
|
"evidence": "# adapters/astrbot_card_renderer.py:58-61\nCARD_DIR.mkdir(parents=True, exist_ok=True)\ntarget = CARD_DIR / f\"{hashlib.sha256(key.encode('utf-8')).hexdigest()[:16]}.png\"\nif target.exists():\n return target\n\n# main.py:526-529 —— key 的四个组成部分全部来自用户输入的角色描述\nlambda: render_profile_card(\n name=str(preview.get(\"name\") or \"\"), base=str(preview.get(\"base_prompt\") or \"\"),\n appearance=str(preview.get(\"appearance_prompt\") or \"\"), style=str(preview.get(\"style_prompt\") or \"\"),\n),",
|
|||
|
|
"impact": "每次 /yp 快速角色 <任意不同描述> 都产生新 sha256 → 一张新的 1000×560 PNG(约 50-200 KB),反复调用可无限增长。height = 300 + max(len(lines), 1) * 66(:63)在 render_status_card 的 details 变长时会线性放大画布,进一步放大单文件体积。撑爆的是 AstrBot 宿主的数据盘。",
|
|||
|
|
"fix": "给 CARD_DIR 加简单 LRU 上限——渲染前统计目录文件数,超过阈值(如 500)时按 st_mtime 删除最旧一批;或改为渲染到内存 BytesIO 直接投递、完全不落盘。另外把 lines 的长度与条数做截断,避免画布尺寸被输入放大。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "dispatch_extension_request 解析出了路由声明的 permission 却从不校验",
|
|||
|
|
"location": "adapters/astrbot_adapter.py:179",
|
|||
|
|
"problem": "RouteMatch 携带 permission 字段(来自 backend/route_registry.ROUTES,实测分布 read×62 / write×37 / materials×36 / config×22 / retry×8 / public_token×1),但 dispatch 时只用 handler 和 path_params,permission 从未参与判断。全部权限校验依赖每个 handler 自己记得调 core.security.require_*。",
|
|||
|
|
"evidence": "# adapters/astrbot_adapter.py:179-187\nmatch = resolve_route(method, path)\nif match is None:\n return error_response(...)\nreturn match.handler(*_handler_args(match, ctx, query or {}, payload or {}))\n# match.permission 在此之后再无引用",
|
|||
|
|
"impact": "路由清单里声明的权限是纯文档,不构成任何强制力。抽查的 10 个\"未直接调用 require_*\"的 handler 目前都通过内层委托做了校验(profile_routes.archive_profile→delete_profile、image_routes.review_moderated_image 内联 ctx.has_permission),当下没有实际漏洞——但这是纯靠约定维系的状态,任何新增 handler 漏掉一行 require_permission 就是静默的越权入口,且没有任何测试会发现。",
|
|||
|
|
"fix": "在 dispatch 层做统一前置校验,把 RouteSpec.permission 映射到 core.security 的检查函数(read→require_authenticated、write→require_write、config→require_permission(ctx, \"can_manage_config\")、public_token→跳过),handler 内的细粒度检查保留为纵深防御。这样清单即契约,新增路由无法绕过。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "review_flow.review_node 内联复制了 _candidate_warnings 的告警逻辑",
|
|||
|
|
"location": "engines/review_flow.py:150",
|
|||
|
|
"problem": "MODEL_ROUTE_MISMATCH 与 PROMPT_MISMATCH 两条告警的判定条件、消息文案、附加字段在同一文件里写了两遍,一遍针对 selected_image(:141-159)、一遍针对候选图(:362-380)。",
|
|||
|
|
"evidence": "# engines/review_flow.py:150-159\nif selected_image and node.prompt and selected_image.positive_prompt:\n node_prompt = _normalize_prompt(node.prompt)\n image_prompt = _normalize_prompt(selected_image.positive_prompt)\n if node_prompt and image_prompt and node_prompt not in image_prompt and image_prompt not in node_prompt:\n warnings.append({\"code\": \"PROMPT_MISMATCH\", \"message\": \"selected image prompt differs from node prompt\"})\n\n# engines/review_flow.py:375-379 —— 判定与文案逐字相同,仅变量名不同\nif node.prompt and image.positive_prompt:\n node_prompt = _normalize_prompt(node.prompt)\n image_prompt = _normalize_prompt(image.positive_prompt)\n if node_prompt and image_prompt and node_prompt not in image_prompt and image_prompt not in node_prompt:\n warnings.append({\"code\": \"PROMPT_MISMATCH\", \"message\": \"candidate prompt differs from node prompt\"})",
|
|||
|
|
"impact": "review_node 返回的 warnings 与其内部 stage_review.candidates[*].warnings 存在漂移风险——一致性规则改了一处忘了另一处,同一张图在两个字段里会给出矛盾结论,而 _candidate_score(:392) 直接用 len(warnings) 打分,漂移会体现在分数上。",
|
|||
|
|
"fix": "review_node 里对 selected_image 直接复用 _candidate_warnings(node, selected_image),删除 141-159 的内联副本(_candidate_warnings 已带 node_model_route/image_model_route 字段,无信息损失)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "_match_endpoint 未对路由字面量做 re.escape",
|
|||
|
|
"location": "adapters/astrbot_adapter.py:226",
|
|||
|
|
"problem": "把路由模板直接拼进正则,只替换了 {param} 占位符,字面量部分未转义。已用脚本枚举验证当前 ROUTES 里没有任何 endpoint 含正则元字符,所以现在是安全的,但这是靠\"恰好没人写带点号的路由\"维持的隐性约束。",
|
|||
|
|
"evidence": "# adapters/astrbot_adapter.py:225-227\nparam_names = re.findall(r\"\\{([^{}]+)\\}\", pattern)\nregex = \"^\" + re.sub(r\"\\{([^{}]+)\\}\", r\"([^/]+)\", pattern) + \"$\"\nmatch = re.match(regex, endpoint)",
|
|||
|
|
"impact": "将来出现 files/{name}.json 这类端点时,. 会变成任意字符匹配,产生静默的路由串台(例如 filesXajson 也会命中),且不会有任何报错提示。",
|
|||
|
|
"fix": "先 re.escape(pattern),再把被转义后的 \\{name\\} 换回捕获组;或改用逐段比较(split \"/\" 后按段匹配),既更快也不需要正则。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "后端内部耦合与架构分层",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\B2-后端内部耦合.md",
|
|||
|
|
"summary": "分层意图清晰(routes→services→repositories,core 做横切),但组合根缺失是所有耦合问题的根因:35 个 `_create_*_service()` 工厂散落在 route 模块里,34 处 `Database(load_database_settings())` 各自无缓存重读 config.yaml。由此产生三类硬伤——后台 worker 必须反向 import route 才能拿到装配好的 GenerationService(service→route 违规);同一 service 在不同 route 被装配成行为不一致的实例,node_material_routes 构造的 WorkflowService 漏传 event_bus 导致素材绑定事件被静默丢弃(P1 可复现);装配代码在 node_routes 与 advanced_material_service 中逐行重复。core/event_bus.py 形同虚设——全仓含 tests 的 subscribe/unsubscribe 调用方为 0,只有 6 处 publish,它只是 GET events/stream 的进程内环形缓冲,没承担任何解耦职责。另有 9 条明确方向违规:database/repositories 反向 import backend.services.branch_lineage_service 与 engines,core/* 反向 import backend.dto.common 与 database.db_helper。模块级无循环导入(SCC 仅在 adapters 包导出间),但 job_routes.py:311 函数内 import 跨 route 取私有 `_create_generation_service` 是循环压力痕迹。main.py 2580 行中真正的生命周期初始化仅 94 行(3.6%),聊天业务与呈现占 1301 行(50.4%,其中 165 行是纯转发的短命令别名),AstrBot 配置与 HTTP 桥接占 825 行(32%)。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_material_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\job_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\round_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\event_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\branch_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\history_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\review_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\profile_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\workflow_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\prompt_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_node_material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_style_set_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\event_stream_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\branch_lineage_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\quick_profile_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\event_bus.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\config_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\security.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\audit_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\astrbot_provider_bridge.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\media_proxy.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\storage.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_round_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_cross_review_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\profile_card_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_adapter.py"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "装配分散导致 WorkflowService 实例行为不一致:素材绑定事件被静默丢弃",
|
|||
|
|
"location": "backend/routes/node_material_routes.py:69",
|
|||
|
|
"problem": "8 处 `WorkflowService(...)` 构造中只有 4 处传入 `event_bus`。`node_material_routes._create_service()` 是素材绑定两个路由的唯一装配路径,构造 WorkflowService 时未传 event_bus;而 WorkflowNodeMaterialService 又通过 `self.workflow_service.event_bus` 穿透邻居属性取总线,于是判空短路,事件永不发出。同样漏传的还有 branch_routes.py:64、history_routes.py:72、review_routes.py:90。",
|
|||
|
|
"evidence": "# backend/routes/node_material_routes.py:64-73\ndef _create_service() -> WorkflowNodeMaterialService:\n database = Database(load_database_settings())\n profiles = ProfileRepository(database)\n workflows = WorkflowRepository(database)\n workflow_service = WorkflowService(\n workflow_repository=workflows,\n profile_repository=profiles,\n audit_service=AuditService(database),\n ) # ← 没有 event_bus=\n\n# backend/services/workflow_node_material_service.py:208-218\nif self.workflow_service.event_bus: # ← 恒为 None,整块被跳过\n self.workflow_service.event_bus.publish(\n \"workflow.node_material_bindings_updated\", {...})\n\n# 对照 backend/routes/profile_routes.py:176-181 是传了的\nreturn WorkflowService(..., event_bus=get_default_event_bus())",
|
|||
|
|
"impact": "前端 GET events/stream 永远看不到素材绑定变更事件;基于该事件做增量刷新/审计的下游功能静默失效。因为是 if 短路,无日志无异常,排查成本极高。同类风险覆盖 4 个漏传点。",
|
|||
|
|
"fix": "1) 立即把 `event_bus=get_default_event_bus()` 补进 4 处工厂;2) 让 WorkflowService.event_bus 默认取 get_default_event_bus() 而非 None,并把 publish 收敛成服务自身的 self._publish(...)(仿 QueueService._publish),禁止 self.<other_service>.event_bus 穿透写法;3) 加单测断言 route_registry.ROUTES 中每个 handler 默认装配路径产出的 WorkflowService.event_bus is not None。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "service → route 反向依赖:后台 worker 必须 import route 模块才能启动",
|
|||
|
|
"location": "backend/services/workflow_round_worker_runtime.py:10",
|
|||
|
|
"problem": "整个应用的组合根(DI 装配逻辑)住在 route 模块里。`_create_generation_service()` 在 node_routes.py:315-390 完成 Database、9 个 repository、ModelProviderRuntime、SingleFlowEngine、4 个 governance 服务的装配。后台 round worker 是 service 层组件却必须反向 import route 层。另有 job_routes.py:311 函数内跨 route 取私有符号。",
|
|||
|
|
"evidence": "# backend/services/workflow_round_worker_runtime.py:9-16\nfrom backend.dto.job_dto import GenerationTask\nfrom backend.routes.node_routes import create_generation_service # ← service 依赖 route\nfrom backend.services.generation_service import GenerationService\n\n# backend/routes/job_routes.py:309-311\ndef _create_generation_job_retry_service() -> GenerationJobRetryService:\n # Import lazily so route discovery does not construct provider runtimes.\n from backend.routes.node_routes import _create_generation_service # ← 取的是私有名",
|
|||
|
|
"impact": "backend/routes 变成不可卸载的运行时依赖,无法把 HTTP 层与 worker 层分开打包/测试;main.py:390 也必须在 initialize() 里函数内 import worker runtime 规避启动期副作用;单测想构造 GenerationService 就得连带 import node_routes 的 31 个依赖(全仓扇出第一)。",
|
|||
|
|
"fix": "新建 backend/composition.py 作为唯一组合根,把 node_routes.py:282-390、job_routes.py:300-318、round_routes.py:137+、advanced_material_service.py:1700-1764 的装配代码全部搬入,导出 create_generation_service()/create_workflow_service()/create_round_service()。route 与 worker 都从 backend.composition 取,依赖方向恢复为 routes → composition → services。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "database 层反向依赖 backend.services 与 engines",
|
|||
|
|
"location": "database/repositories/workflow_round_repository.py:17",
|
|||
|
|
"problem": "仓储层(最底层)反向 import 业务服务层与引擎层。branch_state_for/with_branch_state 是 backend/services/branch_lineage_service.py 的分支血缘业务规则,被两个 repository 直接调用;profile_card_repository.py:11 还 import 了 engines.character_flow。而 branch_lineage_service.py:17 自己又 import database.repositories.workflow_repository,形成包级双向依赖。",
|
|||
|
|
"evidence": "# database/repositories/workflow_round_repository.py:17\nfrom backend.services.branch_lineage_service import branch_state_for, with_branch_state\n\n# database/repositories/workflow_cross_review_repository.py:17\nfrom backend.services.branch_lineage_service import branch_state_for, with_branch_state\n\n# database/repositories/profile_card_repository.py:8-11\nfrom backend.dto.common import ApiError, ErrorCode\nfrom database.db_helper import Database\nfrom engines.character_flow import canonical_gradual_node_key\n\n# backend/services/branch_lineage_service.py:17(反方向)\nfrom database.repositories.workflow_repository import WorkflowRepository",
|
|||
|
|
"impact": "database ↔ backend.services 形成包级双向依赖,模块粒度目前恰好未成环,但只要 branch_lineage_service 再多 import 一个 round/cross-review repository 就立即真成环。分支血缘规则被钉死在 SQL 层,血缘策略调整要改仓储;仓储层无法脱离服务层单独测试。",
|
|||
|
|
"fix": "branch_state_for/with_branch_state 是纯函数(对 dict 做字段级投影,BRANCH_STATE_FIELDS 是 frozenset),连同 BRANCH_STATE_FIELDS 下沉到 core/branch_state.py,database 与 backend.services 都从下层取;branch_lineage_service 保留需要 WorkflowRepository 的部分;canonical_gradual_node_key 同理下沉或由 service 算好再传给 repository。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "core 层反向依赖 backend.dto 与 database,横切层不再是最底层",
|
|||
|
|
"location": "core/security.py:5",
|
|||
|
|
"problem": "core 本应是无依赖的横切基础设施(core.identity 扇入 65 是正确样板),但 5 个模块反向依赖上层包:core/security.py、core/storage.py:11、core/media_proxy.py:17 import backend.dto.common;core/config_service.py:10、core/audit_service.py:12 import database.db_helper。",
|
|||
|
|
"evidence": "# core/security.py:5-6\nfrom backend.dto.common import ApiError, ErrorCode\nfrom core.identity import RequestContext\n\n# core/config_service.py:9-10\nfrom core.runtime_paths import resolve_config_path\nfrom database.db_helper import load_config_payload\n\n# core/audit_service.py:11-12\nfrom core.identity import RequestContext\nfrom database.db_helper import Database, DatabaseSettings",
|
|||
|
|
"impact": "core 无法独立复用或测试;backend.dto.common(扇入 53,全仓第三)实际是全局错误契约却放在 backend 包下,命名与真实分层错位。core.config_service 依赖 database.db_helper 把「读配置」和「连数据库」绑死,而 db_helper 里的 load_config_payload 本质是 YAML 加载,与数据库无关。",
|
|||
|
|
"fix": "1) 把 ApiError/ErrorCode/success_response/response_from_exception 从 backend/dto/common.py 移到 core/errors.py,原位置保留 re-export 过渡;2) 把 load_config_payload/load_config_file_payload/_RUNTIME_CONFIG_OVERRIDES 这套从 database/db_helper.py:26-83 抽到 core/config_loader.py,db_helper 反过来 import 它。core 恢复零上行依赖。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "core/event_bus.py 零订阅者:事件总线不承担任何解耦职责",
|
|||
|
|
"location": "core/event_bus.py:36",
|
|||
|
|
"problem": "全仓(含 tests/ 46 个测试文件)对 EventBus.subscribe()/unsubscribe() 的调用方为 0;publish() 只有 6 处(generation_service.py:1417/1427/1703、queue_service.py:702、workflow_node_material_service.py:209)。总线实际只是 GET events/stream 用的 500 条进程内环形缓冲,跨服务协作仍全靠硬 import。",
|
|||
|
|
"evidence": "# core/event_bus.py:36-50\ndef subscribe(self, event_type: str, handler: EventHandler) -> None: # ← 无任何调用方\n normalized = _normalize_event_type(event_type)\n if handler not in self._handlers[normalized]:\n self._handlers[normalized].append(handler)\n...\ndef publish(self, event_type, payload=None) -> Event:\n event = Event(...)\n self._history.append(event)\n for handler in [*self._handlers.get(event.event_type, []), *self._handlers.get(\"*\", [])]:\n handler(event) # ← 永远是空列表\n return event\n\n# backend/services/event_stream_service.py:11\nDEFAULT_EVENT_BUS = EventBus(history_limit=500) # 进程级全局单例",
|
|||
|
|
"impact": "已有基础设施未兑现解耦价值,workflow_round_service→generation_service、generation_job_retry_service→generation_service、deletion_safety_service→image/media 这类通知型调用本可完全走总线。同步 publish 中 handler 抛异常会打断业务事务;_handlers 是无锁 defaultdict,DEFAULT_EVENT_BUS 被 QueueWorkerPool 工作线程与请求线程共享,一旦开始订阅即为竞态温床。",
|
|||
|
|
"fix": "1) publish 加异常隔离(try/except 包每个 handler,失败只记日志)并给 subscribe/unsubscribe/_handlers 加 threading.Lock;2) 用总线替换 3 组通知型直连(generation_service→workflow_service 节点状态回写、queue_service 终态通知目前靠 terminal_notification_callback 手动穿到 main.py:1672、deletion_safety_service 级联清理);3) 若确认不做订阅,删掉 subscribe/unsubscribe 并改名为 EventRecorder。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "ChatWorkflowService 通过字符串路径反向调用 route 层(69 次进程内自我 HTTP)",
|
|||
|
|
"location": "backend/services/chat_workflow_service.py:26",
|
|||
|
|
"problem": "1315 行的 ChatWorkflowService 不直接调用任何业务 service,而是把 dispatch_web_request 当 HTTP 客户端,用字符串 method+path 重新进入路由层,共 69 处 self.dispatcher(...)。数据流变成 service → adapter → route → service。",
|
|||
|
|
"evidence": "# main.py:372\nself.chat_workflow_service = ChatWorkflowService(dispatch_web_request)\n\n# backend/services/chat_workflow_service.py:33-47\nactive = self._data(self.dispatcher(ctx, \"GET\", \"jobs\", query={\"status\": \"running\", ...}))\nworkflow = self._data(self.dispatcher(ctx, \"POST\", f\"profiles/{profile['profile_id']}/workflows\", payload={...}))\n\n# adapters/astrbot_adapter.py:179-187 + 240\nmatch = resolve_route(method, path)\nif match is None:\n return error_response(error_code=ErrorCode.RESOURCE_NOT_FOUND, ...)\nreturn match.handler(*_handler_args(match, ctx, query or {}, payload or {}))\n...\nfor index, parameter in enumerate(inspect.signature(match.handler).parameters.values()):",
|
|||
|
|
"impact": "类型安全全失——路径/参数名拼错只在运行时返回 RESOURCE_NOT_FOUND,IDE 与静态检查看不到这 69 条调用边,重命名任一路由会静默打断聊天流程;性能上每次 dispatch 走 inspect.signature 反射 + route 内 _create_*_service() 全量重建(含 config.yaml 磁盘读取);业务异常被 handle_request 包成字典再由 self._data() 拆包,异常栈丢失。",
|
|||
|
|
"fix": "让 ChatWorkflowService 依赖 backend/composition.py 暴露的具体 service 对象,复用同一批实例。若要保留「聊天走公开契约」的设计意图,至少把 69 个字符串端点收敛成类型化门面(class ImageStudioApi + 每端点一方法,内部仍走 dispatcher),使重命名有静态检查保护。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "装配代码大段复制:governance/provider 装配在 route 与 service 中各写一遍",
|
|||
|
|
"location": "backend/routes/node_routes.py:331",
|
|||
|
|
"problem": "node_routes.py:315-390 与 advanced_material_service.py:1700-1764 独立实现了同一套「读 config → sync provider → 解析 prompt_governance → 造 callable → 造 SingleFlowEngine」流程,逐行近似。prompt_routes.py:63-64 又独立造了第三遍 PromptModerationService 与 GenerationRouteService。",
|
|||
|
|
"evidence": "# backend/routes/node_routes.py:331-339\ngovernance_config = config_payload.get(\"prompt_governance\") if isinstance(config_payload.get(\"prompt_governance\"), dict) else {}\nmoderation_config = governance_config.get(\"moderation\") if isinstance(governance_config.get(\"moderation\"), dict) else {}\nmoderation_route = str(moderation_config.get(\"model_route\") or model_assignments.get(\"cross_review_primary\") or \"\").strip()\nmoderation_callable = (model_provider_runtime.structured_text_callable(moderation_route) or astrbot_structured_text_callable(moderation_route)) if moderation_route else None\n\n# backend/services/advanced_material_service.py:1731-1741(同构复制)\nmoderation_config = governance_config.get(\"moderation\") if isinstance(governance_config.get(\"moderation\"), dict) else {}\nmoderation_route = str(... or model_assignments.get(\"cross_review_primary\") or \"\").strip()\nmoderation_callable = (\n runtime.structured_text_callable(moderation_route)\n or astrbot_structured_text_callable(moderation_route)\n) if moderation_route else None",
|
|||
|
|
"impact": "governance 配置结构变化(新增字段、改默认回退键)必须同步改 3 处;漏改会导致「网页生成走新策略、素材导入/聊天走旧策略」的行为分叉,且不会报错。",
|
|||
|
|
"fix": "在 backend/services/prompt_governance_service.py 新增 build_governance_bundle(config_payload, runtime) -> GovernanceBundle(含 translation/moderation/routing/image_moderation 四个服务),三处调用方统一使用。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "全局可变状态:AstrBot 桥接与配置覆盖是进程级模块变量,热重载与测试隔离困难",
|
|||
|
|
"location": "core/astrbot_provider_bridge.py:22",
|
|||
|
|
"problem": "跨请求、跨线程共享的运行时状态用模块级 global 承载,无锁、无作用域、无重置入口。core/astrbot_provider_bridge.py:22-31 的 _context/_runtime_loop 被工作线程用于 asyncio.run_coroutine_threadsafe;database/db_helper.py:26-46 的 _RUNTIME_CONFIG_OVERRIDES/_RUNTIME_CONFIG_PERSISTER 被 34 处 load_database_settings() 隐式读取。",
|
|||
|
|
"evidence": "# core/astrbot_provider_bridge.py:22-31\n_context: Any | None = None\n_runtime_loop: asyncio.AbstractEventLoop | None = None\n\ndef configure_astrbot_provider_bridge(context, loop) -> None:\n global _context, _runtime_loop\n _context = context\n _runtime_loop = loop\n\n# core/astrbot_provider_bridge.py:70-86(工作线程侧读取)\nloop = _runtime_loop\nif provider is None or loop is None or loop.is_closed():\n return None\nfuture = asyncio.run_coroutine_threadsafe(invoke(), loop)\ntext = future.result(timeout=180)\n\n# database/db_helper.py:26-33\n_RUNTIME_CONFIG_OVERRIDES: dict[str, Any] = {}\ndef set_runtime_config_overrides(payload) -> None:\n global _RUNTIME_CONFIG_OVERRIDES\n _RUNTIME_CONFIG_OVERRIDES = _copy_plain_mapping(payload or {})",
|
|||
|
|
"impact": "插件热重载(terminate()→initialize())期间 _runtime_loop 可能指向已关闭 loop,工作线程 run_coroutine_threadsafe 会拿到 RuntimeError,而 main.py:433 的 terminate() 未复位这两个 global;测试必须靠 test_astrbot_provider_bridge.py 等 3 个文件手动 set/复位,测试间存在顺序耦合无法并行;配置改动的影响面在 34 个调用点完全不可见。",
|
|||
|
|
"fix": "1) terminate() 中调用 configure_astrbot_provider_bridge(None, None) 清场;2) 把两组 global 收进 AstrBotRuntimeBridge / RuntimeConfigStore 小对象并暴露 reset(),测试用 fixture 统一 setup/teardown;3) run_coroutine_threadsafe 外包 try/except RuntimeError。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "无组合根导致每次路由调用都重读 config.yaml 并重建整棵对象树",
|
|||
|
|
"location": "database/db_helper.py:297",
|
|||
|
|
"problem": "34 处 Database(load_database_settings())(23 个 route 模块 + 4 个 service)。load_database_settings() → load_config_payload() → path.read_text(),全链路无缓存(db_helper.py 与 core/config_service.py 中 lru_cache 出现次数为 0)。node_routes._create_generation_service() 内还额外再调一次 load_config_payload()。",
|
|||
|
|
"evidence": "# database/db_helper.py:297-298\ndef load_database_settings(config_path: str | Path = \"config.yaml\") -> DatabaseSettings:\n return DatabaseSettings.from_mapping(load_config_payload(config_path))\n\n# database/db_helper.py:64-69\ndef load_config_payload(config_path: str | Path = \"config.yaml\") -> dict[str, Any]:\n path = resolve_config_path(config_path)\n payload = load_config_file_payload(path) # ← 每次 read_text + 自研 YAML 子集解析\n if _should_apply_runtime_overrides(path):\n return _merge_mappings(payload, _RUNTIME_CONFIG_OVERRIDES)\n return payload\n\n# backend/routes/node_routes.py:316, 327\ndatabase = Database(load_database_settings()) # 第 1 次读盘\nconfig_payload = load_config_payload() # 第 2 次读盘",
|
|||
|
|
"impact": "耦合上:每个 route 都必须知道整棵依赖树怎么装,35 个工厂各自维护同一份知识。性能上:配置文件 IO 进入请求热路径,一次生成请求至少两次 YAML 解析 + 9 个 repository/1 个 provider runtime/4 个 governance 服务重建。行为上:同一次请求内两次读取甚至可能读到不同内容。",
|
|||
|
|
"fix": "load_config_payload 加进程级缓存(按 (path, mtime) 失效),set_runtime_config_overrides 时主动 invalidate;组合根持有单例 service 图,route 只从容器取。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "服务穿透邻居对象的私有属性链获取 Database(迪米特法则违规)",
|
|||
|
|
"location": "backend/services/workflow_round_worker_runtime.py:193",
|
|||
|
|
"problem": "拿不到 Database 就顺着别的对象的内部属性链摸过去,共 5 处:workflow_round_service.py:67、workflow_round_worker_runtime.py:193、round_routes.py:139、workflow_routes.py:94、profile_routes.py:156。",
|
|||
|
|
"evidence": "# backend/services/workflow_round_worker_runtime.py:192-193\ngeneration_service = create_generation_service()\ndatabase = generation_service.queue_service.repository.database # ← 三级穿透\n\n# backend/services/workflow_round_service.py:66-71\nself.resource_access_service = resource_access_service or ResourceAccessService(\n ProfileRepository(round_repository.database), # ← 穿透 repository 的内部 database\n workflow_repository, image_repository, job_repository,\n)\n\n# backend/routes/round_routes.py:138-139\ngeneration_service = create_generation_service()\ndatabase = generation_service.queue_service.repository.database",
|
|||
|
|
"impact": "Repository.database 从实现细节变成事实上的公开 API,仓储重构(如改为持有连接池)会同时打断 5 个调用点;对象图的真实依赖在构造函数签名上完全看不见。",
|
|||
|
|
"fix": "组合根统一持有 Database 实例并显式注入;过渡期至少把 ProfileRepository 作为 WorkflowRoundService 的构造参数传入,而不是就地用穿透来的 database 现造。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "generation_service 内部 event_bus 取用方式自相矛盾",
|
|||
|
|
"location": "backend/services/generation_service.py:1416",
|
|||
|
|
"problem": "同一个类里有两套取总线的写法:self.event_bus(构造注入,_publish() 用,5 处:786/859/1066/1766/1889)和 self.workflow_service.event_bus(穿透邻居,2 处:1417/1427)。",
|
|||
|
|
"evidence": "# backend/services/generation_service.py:104\nself.event_bus = event_bus\n# :1701-1703\ndef _publish(self, event_type: str, payload: dict[str, Any]) -> None:\n if self.event_bus:\n self.event_bus.publish(event_type, payload)\n\n# :1416-1427(同一个类,换了个总线来源)\nif self.workflow_service.event_bus:\n self.workflow_service.event_bus.publish(\"workflow.node_status_changed\", {...})\n self.workflow_service.event_bus.publish(\"workflow.node_updated\", {...})",
|
|||
|
|
"impact": "在 _create_generation_service() 里两者恰好是同一个 DEFAULT_EVENT_BUS 所以现在未暴露;但只要有人用未注入 event_bus 的 WorkflowService(node_material_routes / branch_routes / history_routes / review_routes 构造的正是这种)来构造 GenerationService,workflow.node_status_changed 就会静默丢失,与前述 P1 同源。",
|
|||
|
|
"fix": "把 1416-1427 改用 self._publish(...),删除 self.workflow_service.event_bus 引用。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "冗余的函数内 import(模块顶部已导入同一模块),制造虚假的循环依赖信号",
|
|||
|
|
"location": "backend/services/generation_service.py:1493",
|
|||
|
|
"problem": "模块顶部 generation_service.py:17-24 已 from backend.services.prompt_governance_service import 了 6 个名字,函数体内又对同一模块做了两次局部 import。该模块无反向依赖,不是循环依赖规避,纯属遗漏。",
|
|||
|
|
"evidence": "# 顶部 :17-24 已有\nfrom backend.services.prompt_governance_service import (\n GenerationFailureClassifier, GenerationRouteService, ImageModerationService,\n PromptModerationService, PromptSemanticNormalizer, PromptTranslationService,\n)\n# :1492-1494\nif moderation is None:\n from backend.services.prompt_governance_service import PromptModerationDecision\n moderation = PromptModerationDecision(\"safe\", 1.0, [], [\"moderation_not_configured\"], \"allow\", None, \"none\", None, \"\")\n# :1505-1507\nelse:\n from backend.services.prompt_governance_service import GenerationRouteDecision",
|
|||
|
|
"impact": "制造「这里有循环依赖」的误导性信号,干扰后续重构判断(本次审计正是靠函数内 import 定位循环依赖痕迹,全仓仅 12 处,其中这 2 处是噪音)。",
|
|||
|
|
"fix": "把 PromptModerationDecision、GenerationRouteDecision 并入顶部 import 块。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "main.py 内 66 条配置映射被正反两个方向各写一遍(相隔 1800 行)",
|
|||
|
|
"location": "main.py:1929",
|
|||
|
|
"problem": "_RUNTIME_PATH_TO_ASTRBOT_KEY(main.py:108-172)用声明式 dict 表达「ImageStudio 配置路径 → AstrBot 配置键」共 66 条;_runtime_config_patch_from_mapping()(main.py:1929-2205)用 279 行命令式代码表达完全相同的 66 条关系的反方向。脚本比对确认两侧覆盖的 66 个 AstrBot 键当前完全一致——即两份数据靠人工维持同步。",
|
|||
|
|
"evidence": "# main.py:108-113(声明式,共 66 条)\n_RUNTIME_PATH_TO_ASTRBOT_KEY = {\n \"plugin.display_name\": \"display_name\",\n \"generation.default_model\": \"default_generation_model\",\n \"generation.provider.provider_id\": \"generation_provider_id\",\n ...\n}\n\n# main.py:1946-1953(命令式,同一批关系的反方向)\ndisplay_name = _runtime_text(_runtime_value(mapping, \"display_name\", fallback=plugin.get(\"display_name\", _MISSING)))\nif display_name:\n _set_nested(patch, (\"plugin\", \"display_name\"), display_name)\ndefault_model = _runtime_text(_runtime_value(mapping, \"default_generation_model\"))\nif default_model:\n _set_nested(patch, (\"generation\", \"default_model\"), default_model)",
|
|||
|
|
"impact": "新增一个配置项需要在两个相隔 1800 行的位置各改一次;只改一边会造成「能读不能写」或「能写不能读」的单向配置,且不会有任何报错。",
|
|||
|
|
"fix": "保留 dict 作为唯一真相,_runtime_config_patch_from_mapping 改为遍历 _RUNTIME_PATH_TO_ASTRBOT_KEY.items() + _set_nested(patch, tuple(path.split('.')), value),只把少数需要特殊类型转换的键(postgresql.port 转 int、*_enabled 转 bool、allowlist 拆列表)保留为显式分支。整块可从 279 行压到约 50 行,并整体搬到 core/astrbot_config_mapping.py。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "main.py 职责膨胀:2580 行中真正的生命周期初始化仅 94 行(3.6%)",
|
|||
|
|
"location": "main.py:1299",
|
|||
|
|
"problem": "main.py 按职责分解:环境 shim 192 行(28-98 包路径 hack、174-221 AstrBot fallback stub);配置映射 350 行(108-172 正向表 + 1929-2205 反向实现);对外门面 130 行(223-352);生命周期 94 行(354-447);聊天命令 849 行(449-1297);短命令别名 165 行纯转发(1299-1463);格式化/后台任务/推送 287 行(1465-1751);AstrBot HTTP 桥接 548 行(1753-1927 + 2208-2580)。其中 1299-1463 的 40+ 个方法全部是 3 行样板转发。imagestudio_confirm(549-612)还内嵌了 session action 分派业务逻辑。",
|
|||
|
|
"evidence": "# main.py:1304-1312(165 行里每一个都长这样)\n@imagestudio_short_command.command(\"帮助\")\nasync def imagestudio_short_help(self, event: AstrMessageEvent):\n async for result in self.imagestudio_help(event):\n yield result\n\n@imagestudio_short_command.command(\"状态\")\nasync def imagestudio_short_status(self, event: AstrMessageEvent, workflow_id: str = \"\"):\n async for result in self.imagestudio_status(event, workflow_id):\n yield result\n\n# main.py:555-575(业务分派逻辑埋在命令处理器里)\nif session.action == \"confirm_quick_profile\":\n try:\n result = await asyncio.to_thread(self.chat_workflow_service.save_quick_profile_draft, session.payload[\"draft\"], context)\n workflow_id = str(result.get(\"workflow\", {}).get(\"workflow_id\") or \"\")\n ...",
|
|||
|
|
"impact": "main.py 扇出 22(全仓第三),任何聊天命令改动都要碰插件入口文件;40+ 个短别名的转发样板每新增一个命令就要写两遍;配置桥接与 HTTP 适配共 898 行与插件生命周期混在一起,导致 main.py 无法被单测独立加载(tests/test_plugin_bootstrap.py 必须 mock 全套 AstrBot API)。",
|
|||
|
|
"fix": "按职责拆为 6 个模块:core/plugin_bootstrap.py(包路径 shim)、adapters/astrbot_stubs.py(fallback)、core/astrbot_config_mapping.py(双向配置映射)、adapters/astrbot_http.py(2208-2580 的 20 个请求响应适配函数)、adapters/astrbot_bridge.py(1753-1927)、backend/chat/commands/*.py + backend/chat/formatters.py + backend/chat/notifier.py(449-1751)。短别名块用注册表替代:_SHORT_ALIASES = {\"帮助\": \"imagestudio_help\", ...} + 循环绑定,165 行降到约 45 行。拆分后 main.py 目标约 250 行。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "数据库层(模型 / 仓储 / 迁移 / 事务)",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A4-数据层.md",
|
|||
|
|
"summary": "参数化做得干净,全部 SQL 走绑定占位符,唯一的动态表名有白名单,未发现可利用的 SQL 注入;迁移脚本全部 CREATE ... IF NOT EXISTS、可重入可自愈;仓储内部的 *_atomic 方法(BEGIN IMMEDIATE + FOR UPDATE + rollback + 幂等键)写得相当认真。真正的问题集中在三处:PostgreSQL 分支实际是坏的——profile_repository 的 CASE 列里有字面量 % ,在 psycopg 的 %s paramstyle 下角色列表/详情会直接抛异常,而 config.yaml 把 PG 作为正式可选后端;SQLite 未启用 WAL,却同时存在 ThreadPoolExecutor(4) 并发写和大量 BEGIN IMMEDIATE 长事务(连纯读的 get_consistent_round_snapshot 都去拿写锁),写事务会阻塞全库读;metadata_json 是整列覆盖语义而服务层普遍跨两个连接做读-改-写,generation_service 专门加了一把进程内 threading.Lock 打补丁,说明丢更新已是已知痛点但只堵住了一条路径。此外跨仓储组合写完全没有事务边界(删图与清 workflow_nodes.selected_image_id 分属两次提交,而该列还没有外键),陈旧任务恢复的时间戳比较绑死了 SQLite 的 UTC 语义、PG 下会静默失效。工程上最突出的是零抽象:14 个仓储手写了 131 处 connect/try/finally、13 份 JSON 助手(语义还不一致)、4 组 _page/_page_size(分页上限 100 vs 200 不一致)、3 份 _rollback,缺失的 rollback 与遗漏的写法正是重复的直接产物,应抽取 BaseRepository + connection() 上下文管理器。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\models.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\seed.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0001_init_core.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0002_character_profiles.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0003_profile_workflows_and_nodes.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0004_generation_images.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0005_generation_tasks.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0006_workflow_history.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0007_workflow_branches.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0008_materials.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0009_wardrobe_items.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0010_integration_logs.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0011_capability_profiles.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0012_config_snapshots.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0013_model_providers.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0014_workflow_generation_rounds.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0015_character_profile_cards.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0016_advanced_materials.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0017_advanced_wardrobe_items.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0018_advanced_style_sets.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\job_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\profile_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\image_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_round_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\profile_card_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\advanced_material_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_cross_review_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\wardrobe_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\image_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\cross_review_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_analytics_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\orphan_media_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\worker_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\config.yaml"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "profile_repository 的 SQL 含字面量 %,PostgreSQL 后端下角色列表/详情 100% 报错",
|
|||
|
|
"location": "database/repositories/profile_repository.py:46",
|
|||
|
|
"problem": "PROFILE_GENERATION_KIND_COLUMN / PROFILE_CATEGORY_COLUMN / DETAILED_WORKFLOW_STATE_COLUMN 三个 SQL 常量里写死了 LIKE '%\"final_archive_history_id\"%'(分别在 :46、:73、:99,逐字重复三份)。当 backend == \"postgresql\" 时 Database._placeholder 返回 %s(db_helper.py:288-290),psycopg/psycopg2 会把整条 query 当成 %-format 模板去展开绑定参数,SQL 文本中的 %\" 是非法格式符。",
|
|||
|
|
"evidence": "# database/repositories/profile_repository.py:40-47\nWHEN EXISTS (\n SELECT 1\n FROM profile_workflows AS legacy_card_workflow\n WHERE legacy_card_workflow.profile_id = character_profiles.profile_id\n AND legacy_card_workflow.workflow_type = 'character_gradual'\n AND LOWER(COALESCE(legacy_card_workflow.status, '')) IN ('completed', 'archived')\n AND COALESCE(legacy_card_workflow.metadata_json, '') LIKE '%\"final_archive_history_id\"%'\n) THEN 'detailed'\n\n# database/repositories/profile_repository.py:186-194 —— 带参调用,触发 %-展开\nrow = conn.execute(\n f\"\"\"\n SELECT {PROFILE_COLUMNS}, {PROFILE_GENERATION_KIND_COLUMN},\n {PROFILE_CATEGORY_COLUMN}, {DETAILED_WORKFLOW_STATE_COLUMN}\n FROM character_profiles\n WHERE profile_id = {placeholder}\n \"\"\",\n (profile_id,),\n).fetchone()",
|
|||
|
|
"impact": "config.yaml:153-167 明确把 backend: auto|sqlite|postgresql 作为受支持配置对外暴露(postgresql.enabled: true 即自动切换)。一旦用户切到 PG,ProfileRepository.get() 和 .list() 这两个最核心的读接口立刻抛 ValueError: unsupported format character '\"'(psycopg2)或 IndexError(psycopg3),角色详情与角色列表完全不可用,且错误来自驱动层格式化,完全指向不了根因。",
|
|||
|
|
"fix": "1) 立刻把三处字面量改成参数绑定:定义 LEGACY_ARCHIVE_MARKER = '%\"final_archive_history_id\"%',SQL 里写 LIKE {placeholder},调用处按 CASE 列出现次数补齐参数。2) 更彻底:把 final_archive_history_id 提升为 profile_workflows 的真实列(新增 0019 迁移 + 回填),同时解决转义和前缀通配符全表扫描。3) 加静态断言测试:正则扫描 database/repositories/*.py 中未转义的裸 %(模式 %(?!s))。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "SQLite 未启用 WAL,却同时存在 4 线程并发写与 BEGIN IMMEDIATE 长事务",
|
|||
|
|
"location": "database/db_helper.py:159",
|
|||
|
|
"problem": "Database.connect() 只设置了 foreign_keys 和 busy_timeout,没有 PRAGMA journal_mode = WAL。默认 rollback journal 模式下写事务持有 RESERVED/EXCLUSIVE 锁会阻塞全库所有读。而代码里长写事务遍地:confirm_candidate_atomic 在一个 BEGIN IMMEDIATE 里跑 12 条语句(workflow_round_repository.py:815-1120),get_consistent_round_snapshot 甚至为纯读也去拿 BEGIN IMMEDIATE 写锁(:312-313)。",
|
|||
|
|
"evidence": "# database/db_helper.py:159-171\ndef connect(self):\n if self.backend == \"sqlite\":\n db_path = self._resolve_sqlite_path()\n db_path.parent.mkdir(parents=True, exist_ok=True)\n # Parallel image tasks briefly overlap while recording queue,\n # node, image, and audit rows. Give SQLite enough time to hand the\n # writer lock to the next task instead of failing immediately.\n conn = sqlite3.connect(db_path, timeout=30.0)\n conn.row_factory = sqlite3.Row\n conn.execute(\"PRAGMA foreign_keys = ON\")\n conn.execute(\"PRAGMA busy_timeout = 30000\")\n return conn\n return self._connect_postgresql()\n\n# backend/services/chat_workflow_service.py:216-220 —— 4 线程并发跑整条生成链路\nwith ThreadPoolExecutor(max_workers=4, thread_name_prefix=\"imagestudio-chat-fast\") as executor:\n futures = {\n executor.submit(self.generate_node, str(node[\"node_id\"]), chat): node\n for node in nodes\n }",
|
|||
|
|
"impact": "源码注释里 \"Give SQLite enough time\" 本身就是症状——作者观察到了锁竞争,选择把超时调到 30s 而非消除竞争。快速角色四图并发生成时,任一线程进入 confirm_candidate_atomic,其余 3 个线程连读节点状态都被阻塞;provider 调用稍慢就会让其他线程 30s 后抛 sqlite3.OperationalError: database is locked,而该异常被 generation_service.py:1866、:1890 的 except Exception: pass 静默吞掉,表现为节点状态莫名其妙不更新。",
|
|||
|
|
"fix": "1) connect() 中补 conn.execute(\"PRAGMA journal_mode = WAL\") 与 conn.execute(\"PRAGMA synchronous = NORMAL\"),让读写互不阻塞。2) get_consistent_round_snapshot(workflow_round_repository.py:312)改用 BEGIN DEFERRED——它只读不写,WAL 下天然获得一致性快照,不需要写锁。3) 复查 generate_node 链路,确保 provider 网络调用不落在任何事务边界内。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "metadata_json 整列覆盖 + 跨连接读改写导致丢更新,仅一条路径靠进程内锁打补丁",
|
|||
|
|
"location": "database/repositories/workflow_repository.py:298",
|
|||
|
|
"problem": "仓储的 update 语义是整列覆盖 metadata_json = ?,没有 merge、版本号或 CAS。服务层普遍是\"读 DTO → {**node.metadata, ...} → 写回\",两步跨两个独立连接与事务,任意两个并发写入者互相覆盖对方写入的键。workflow_repository.update_workflow(:154-178)同理。",
|
|||
|
|
"evidence": "# database/repositories/workflow_repository.py:298-318 —— 整列覆盖\nfor key in allowed:\n if key not in payload:\n continue\n column = \"metadata_json\" if key == \"metadata\" else key\n value = _json_or_none(payload[key]) if key == \"metadata\" else payload[key]\n assignments.append(f\"{column} = {placeholder}\")\n params.append(value)\n...\nconn.execute(f\"UPDATE workflow_nodes SET {', '.join(assignments)} WHERE node_id = {placeholder}\", tuple(params))\n\n# backend/services/generation_service.py:1870-1886 —— 跨连接读-改-写\nnode = self.workflow_service.get_node(node_id) # 连接 A:读\nnode_metadata = node.metadata if isinstance(node.metadata, dict) else {}\nif (_optional_text(node_metadata.get(\"last_task_id\")) == task.task_id\n and node.status in {\"queued\", \"running\"}):\n self.workflow_service.update_node( # 连接 B:整列覆盖\n node_id,\n {\"status\": \"failed\",\n \"metadata\": {**node_metadata, \"generation_failure\": failure,\n \"last_task_id\": task.task_id}},\n ctx,\n )\n\n# backend/services/generation_service.py:1145-1149 —— 作者已知问题,但只用进程内锁堵了一条路径\nlock_entry = _retain_quick_character_workflow_lock(workflow.workflow_id)\nlock_entry.lock.acquire()\ntry:\n current_workflow = self.workflow_service.get_workflow(workflow.workflow_id).workflow\n metadata = dict(current_workflow.metadata or {})",
|
|||
|
|
"impact": "profile_workflows.metadata_json 里装着 branch_state / stage_inputs / character_execution / completed_node_ids 这些工作流状态机的真身(workflow_round_repository.py:1097-1119)。confirm_candidate_atomic 在自己的 DB 事务里正确写入 stage_inputs,但只要 generation_service._prepare_quick_character 用更早读到的 metadata 快照调用 update_workflow,stage_inputs 就被整个抹掉——表现为\"确认了候选图但下一节点拿不到上游图\"。threading.Lock 只覆盖同进程同一条代码路径,confirm_candidate_atomic 完全不参与这把锁。",
|
|||
|
|
"fix": "1) 短期给 update_workflow/update_node 加乐观并发控制,复用已有 updated_at:WHERE workflow_id = ? AND updated_at = ?,rowcount == 0 时由调用方重读重试而非静默覆盖。2) 中期给仓储加 patch_metadata(id, patch) 语义,在同一个 BEGIN IMMEDIATE 事务内 SELECT → merge → UPDATE,把读改写收进仓储,服务层不再手工展开 {**node.metadata}。3) 长期把 stage_inputs / completed_node_ids 从 JSON blob 拆成真表(workflow_branch_state、workflow_stage_input),行级写入天然无冲突。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "跨仓储组合写无事务边界:删图与清 node 指针分属两次提交,且 selected_image_id 无外键",
|
|||
|
|
"location": "backend/services/image_service.py:216",
|
|||
|
|
"problem": "delete_image_metadata 先在一个连接里 DELETE 并 commit(),再在另一个连接里把 workflow_nodes.selected_image_id 置空。两步之间进程崩溃或第二步抛异常就留下指向不存在图片的悬挂指针,而 selected_image_id 在建表时没有声明外键,数据库层不兜底。同一仓库里 job_repository.delete_if_status(:382-431)已经做对了(同连接内删任务 + 修复 node 引用 + 一次 commit),说明这是遗漏而非有意。",
|
|||
|
|
"evidence": "-- database/migrations/0003_profile_workflows_and_nodes.sql:36 —— 无 FOREIGN KEY\nselected_image_id TEXT,\n(该表末尾 :41 只有一条外键 FOREIGN KEY(workflow_id) REFERENCES profile_workflows(workflow_id) ON DELETE CASCADE)\n\n# database/repositories/image_repository.py:260-270 —— 独立事务提交\nconn = self.database.connect()\ntry:\n conn.execute(f\"DELETE FROM generation_images WHERE image_id = {placeholder}\", (image_id,))\n conn.commit()\nfinally:\n conn.close()\nreturn before\n\n# backend/services/image_service.py:219-225 —— 第二次独立提交\nbefore = self.image_repository.delete_metadata(image_id)\nif before is None:\n raise ApiError(ErrorCode.RESOURCE_NOT_FOUND, \"图片不存在\", {\"image_id\": image_id})\nif before.node_id:\n node = self.workflow_repository.get_node(before.node_id)\n if node and node.selected_image_id == image_id:\n self.workflow_repository.update_node(before.node_id, {\"selected_image_id\": None})",
|
|||
|
|
"impact": "selected_image_id 悬挂后,node_dependency_context_service.py:81-83 去 get_image() 拿到 None,下游节点的参考图静默丢失;workflow_service 判断\"节点是否已完成\"也会得出错误结论。批量清理路径 generation_service.py:1789-1804 循环调用 delete_metadata,每张图一个独立事务,中途失败只把错误塞进 errors 列表继续走,放大了不一致。另外 character_profile_card_assets.image_id 是 ON DELETE RESTRICT(0015:32),图片被角色卡引用时 delete_metadata 会抛 IntegrityError 且无任何捕获,直接 500。",
|
|||
|
|
"fix": "1) 下沉成仓储方法 ImageRepository.delete_with_node_release(image_id),同一连接内 DELETE + UPDATE workflow_nodes SET selected_image_id = NULL WHERE selected_image_id = ?,一次 commit(),照抄 job_repository.delete_if_status 的做法。2) 补迁移 0019 给 workflow_nodes.selected_image_id 加 REFERENCES generation_images(image_id) ON DELETE SET NULL(SQLite 需重建表),并附一次性孤儿清理 UPDATE。3) delete_metadata 捕获 IntegrityError,翻译成\"该图片已被角色卡引用,无法删除\"的业务错误。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "陈旧任务恢复的时间戳比较绑死 SQLite 的 UTC 语义,PostgreSQL 下静默失效",
|
|||
|
|
"location": "backend/services/queue_service.py:457",
|
|||
|
|
"problem": "cutoff_value 在 Python 侧按 UTC 格式化成 \"%Y-%m-%d %H:%M:%S\" 字符串,直接与列 updated_at 比较。SQLite 的 CURRENT_TIMESTAMP 恰好是 UTC 且恰好是这个格式所以能工作;PostgreSQL 里 updated_at TIMESTAMP(无时区)的 CURRENT_TIMESTAMP 返回会话本地时间,两者相差一个时区偏移。",
|
|||
|
|
"evidence": "# backend/services/queue_service.py:456-468\nobserved_at = now or datetime.now(timezone.utc)\n...\ncutoff = observed_at.astimezone(timezone.utc) - timedelta(\n seconds=max(int(stale_after_seconds), 1)\n)\ncutoff_value = cutoff.strftime(\"%Y-%m-%d %H:%M:%S\")\ncandidates = self.repository.list_stale(\n statuses=(\"running\", \"retrying\"),\n updated_before=cutoff_value,\n limit=min(max(int(limit), 1), 500),\n)\n\n# database/repositories/job_repository.py:182, 195-197\nage_clause = f\"AND updated_at <= {placeholder}\" if updated_before else \"\"\n...\nWHERE status IN ({status_placeholders})\n {age_clause}\nORDER BY updated_at ASC, created_at ASC, task_id ASC",
|
|||
|
|
"impact": "PG 服务器在 UTC+8 时库里的 updated_at 永远比 UTC 的 cutoff 大 8 小时,updated_at <= cutoff 恒不成立,recover_stale_tasks 永远返回空,上个进程遗留的 running 任务永久卡死且不报错,日志只显示 recovered: 0。反向时区(UTC-x)则会把刚提交 1 分钟的任务当成陈旧任务抢走,造成任务重复执行。",
|
|||
|
|
"fix": "1) 时间差在 SQL 侧计算,不在 Python 侧格式化字符串:SQLite 用 datetime('now', ?),PG 用 NOW() - (?::text)::interval,由 Database.backend 分派。2) 或把所有 TIMESTAMP 列迁移成 TIMESTAMPTZ(PG)并在建连时 SET TIME ZONE 'UTC',让两个后端的 CURRENT_TIMESTAMP 语义对齐。3) 补测试断言 _row_to_task 中 str(row[\"created_at\"]) 在两个后端下的格式一致性——目前是巧合而非契约。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "MaterialAnalyticsService.summarize 无条件全表扫描 9 张表,含 generation_tasks 的 JSON blob",
|
|||
|
|
"location": "backend/services/material_analytics_service.py:16",
|
|||
|
|
"problem": "一次汇总调用发出 9 条 SELECT,全部没有 WHERE 时间窗、没有 LIMIT、没有聚合下推,把整张表拉进 Python 内存后用 Counter 统计。其中 generation_tasks 还带上了 request_json / metadata_json 两个大 JSON 列。",
|
|||
|
|
"evidence": "# backend/services/material_analytics_service.py:41-45\nstyle_validation_cases = conn.execute(\n \"SELECT status, fidelity_score FROM advanced_style_validation_cases\"\n).fetchall()\naudits = conn.execute(\"SELECT action, after_snapshot, metadata_json FROM audit_logs WHERE action IN ('wardrobe_match','apply_wardrobe_template','recognize_material','renormalize_material')\").fetchall()\ntasks = conn.execute(\"SELECT model_route, request_json, metadata_json FROM generation_tasks\").fetchall()",
|
|||
|
|
"impact": "generation_tasks 和 audit_logs 是只增不删的表(job_repository.delete_if_status 只删终态任务且需显式调用)。运行几个月后单次调用即拉数百 MB 进内存;在没有 WAL 的前提下这个长读事务还会把整库读锁住阻塞所有写。audit_logs 那条虽有 action IN (...) 可走 idx_audit_logs_action(action, created_at),但仍要回表读 after_snapshot 大字段。",
|
|||
|
|
"fix": "1) 加时间窗参数(默认最近 30 天)走索引 WHERE created_at >= ?,并补 idx_generation_tasks_created ON generation_tasks(created_at)(现有三个索引都以 status/workflow_id/profile_id 打头)。2) 能下推的统计全部下推:recognition_status 计数、fidelity_score 的 AVG/COUNT 改成 GROUP BY + 聚合函数,只有必须解析 JSON 的部分才拉行。3) 结果加秒级过期缓存,这是只读 dashboard 指标。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "profile_repository 每行 10 个相关子查询,且同一组谓词被复制三份并已开始漂移",
|
|||
|
|
"location": "database/repositories/profile_repository.py:32",
|
|||
|
|
"problem": "PROFILE_GENERATION_KIND_COLUMN(3 个 EXISTS)、PROFILE_CATEGORY_COLUMN(3 个 EXISTS,谓词与前者几乎逐字相同)、DETAILED_WORKFLOW_STATE_COLUMN(4 个 EXISTS,前两个又是同样谓词)被拼进同一条 SELECT,结果集每行跑 10 个相关子查询,其中\"sealed 角色卡\"和\"legacy 归档工作流\"各被重复计算 3 次。",
|
|||
|
|
"evidence": "# database/repositories/profile_repository.py:33-39 与 :60-66 —— 逐字重复,仅 THEN 分支不同\nWHEN EXISTS (\n SELECT 1\n FROM character_profile_cards AS profile_card\n WHERE profile_card.profile_id = character_profiles.profile_id\n AND LOWER(COALESCE(profile_card.status, '')) = 'sealed'\n) THEN 'detailed' # 第二处:THEN 'profile_card'\n\n# database/repositories/profile_repository.py:222-232 —— 三份常量同时拼入\nrows = conn.execute(\n f\"\"\"\n SELECT {PROFILE_COLUMNS}, {PROFILE_GENERATION_KIND_COLUMN},\n {PROFILE_CATEGORY_COLUMN}, {DETAILED_WORKFLOW_STATE_COLUMN}\n FROM character_profiles\n {where}\n ORDER BY updated_at DESC, created_at DESC\n LIMIT {normalized_page_size} OFFSET {offset}\n \"\"\",\n tuple(params),\n).fetchall()",
|
|||
|
|
"impact": "_page_size 上限 100(:362-367),最坏一次列表请求触发 1000 次相关子查询,其中 3 次带前缀通配符 LIKE 无法走索引,必须读出 profile_workflows.metadata_json 全文逐行匹配。叠加无 WAL,列表接口成为写路径的主要阻塞源。三份重复的 CASE 已经开始漂移:'archived' 状态出现在 PROFILE_CATEGORY_COLUMN:80 的 quick 分支里,却不在 PROFILE_GENERATION_KIND_COLUMN:53 的对应分支——同一语义两处判定已不一致。",
|
|||
|
|
"fix": "1) 用 LEFT JOIN 派生表把三个 EXISTS 结果算一遍再在外层 CASE 复用,例如 LEFT JOIN (SELECT profile_id, 1 AS has_sealed_card FROM character_profile_cards WHERE LOWER(COALESCE(status,''))='sealed' GROUP BY profile_id) sc ON sc.profile_id = cp.profile_id,子查询数从 10 降到 3。2) LIKE '%\"final_archive_history_id\"%' 提升为真实列(与 P1-1 的修复合并做)。3) 补 idx_profile_cards_profile_status ON character_profile_cards(profile_id, status)——现有 idx_profile_cards_profile_version(profile_id, version) 对该谓词只能用到前缀。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "14 个仓储零抽象:131 处手写连接样板、13 份 JSON 助手语义不一致、分页上限 100 vs 200 不统一",
|
|||
|
|
"location": "database/repositories/job_repository.py:617",
|
|||
|
|
"problem": "没有基类、没有 context manager、没有共享 helper 模块。实测重复量:conn = self.database.connect() + try/finally close() 共 131 处(advanced_material 37、workflow_repository 16、workflow_round 14、job 12、model_provider 10 等);placeholder property 13 份,实现全部是 return self.database._placeholder;_json_or_none 12 份,实现全部相同;JSON 解码助手 13 份且语义不一致;_page/_page_size 4 组且上限不一致;_rollback 3 份。此外 for key in allowed 中的 allowed 是 set(workflow_repository.py:281、job_repository.py:569、profile_repository.py:243),字符串哈希随机化让 SET 子句顺序每进程不同。",
|
|||
|
|
"evidence": "# database/repositories/job_repository.py:617-622 —— 上限 100,非法输入回落 30\ndef _page_size(value: Any) -> int:\n try:\n page_size = int(value)\n except (TypeError, ValueError):\n page_size = 30\n return min(max(page_size, 1), 100)\n\n# database/repositories/advanced_material_repository.py:2115-2116 —— 上限 200,非法输入直接抛\ndef _page_size(value: int) -> int:\n return min(max(int(value or 30), 1), 200)\n\n# database/repositories/job_repository.py:625-628 —— 脏数据直接抛\ndef _json_loads(value: str | None) -> dict[str, Any]:\n if not value:\n return {}\n return json.loads(value)\n\n# database/repositories/workflow_round_repository.py:1324-1328 —— 脏数据降级为 {}\ndef _json_object(value: str | None) -> dict[str, Any]:\n if not value:\n return {}\n parsed = json.loads(value)\n return parsed if isinstance(parsed, dict) else {}",
|
|||
|
|
"impact": "_json_loads 遇到非 dict 的 JSON(历史遗留数组)会把 list 塞进 GenerationTask.metadata,后续 task.metadata.get() 抛 AttributeError;同一份脏数据在 round 仓储里则安全降级为 {}——行为取决于走哪个仓储。分页上限不一致导致 /advanced-materials 单页 200 条而其他接口 100 条,前端分页组件无法统一。131 处样板意味着任何连接层改动(加 WAL、加 retry、加慢查询日志)要改 131 处,本次审计已发现 image_repository.delete_metadata 缺 rollback、profile_repository 全文无 _rollback 而 advanced_material_repository 有——遗漏正是重复的直接产物。SET 子句顺序不稳定还会打散 SQLite statement cache。",
|
|||
|
|
"fix": "1) 新建 database/repositories/base.py 定义 BaseRepository,提供 placeholder property、@contextmanager connection(write: bool)(内含 BEGIN IMMEDIATE / commit / rollback / close)与 lock_clause(),131 处 try/finally 塌缩成 with self.connection(write=True) as conn:。2) 新建 database/repositories/_sql.py 唯一实现 json_or_none / json_object / page / page_size / optional_text,分页上限统一成一个常量,JSON 解码统一采用防御式(_json_object 那版)。3) 所有 allowed 集合改成 tuple 保证 SQL 文本稳定。4) 抽出 build_update(table, pk, allowed, payload) 通用构造器——workflow_repository 里 update_workflow / update_workflow_if_status / update_node / update_node_if_status 四份 assignment 循环(:154-160、:203-209、:298-304、:357-363)几乎逐字相同。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "关键词过滤未转义 LIKE 通配符",
|
|||
|
|
"location": "database/repositories/profile_repository.py:305",
|
|||
|
|
"problem": "用户输入直接包进 %...%,输入中的 % 和 _ 被当作通配符。advanced_material_repository.py:493-498 同样问题。",
|
|||
|
|
"evidence": "# database/repositories/profile_repository.py:305-310\nif keyword:\n clauses.append(\n f\"(name LIKE {placeholder} OR description LIKE {placeholder})\"\n )\n like_value = f\"%{keyword}%\"\n params.extend([like_value, like_value])",
|
|||
|
|
"impact": "不是注入(参数是绑定的),但搜索 50% 会匹配所有以 50 开头的记录,搜索 _ 匹配任意单字符,用户可用单个 % 拉出全表。结合前缀通配符导致的全表扫描,构成一个小的 DoS 面。",
|
|||
|
|
"fix": "转义后再拼接:keyword.replace(\"\\\\\", \"\\\\\\\\\").replace(\"%\", \"\\\\%\").replace(\"_\", \"\\\\_\"),SQL 侧加 ESCAPE '\\'(SQLite 与 PG 均支持)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "PostgreSQL 迁移执行器按分号裸切分;psycopg2 适配器不关闭游标",
|
|||
|
|
"location": "database/db_helper.py:280",
|
|||
|
|
"problem": "_execute_script 在 PG 分支按字面 ; 切分脚本,任何出现在字符串字面量、$$ ... $$ 函数体或注释里的分号都会把语句劈成两半。当前 18 个迁移恰好未触发(唯一带引号的是 WHERE status IN ('queued','running'))。另外 _Psycopg2ConnectionAdapter.execute 每次新建游标且从不关闭。",
|
|||
|
|
"evidence": "# database/db_helper.py:280-286\ndef _execute_script(self, conn, script: str) -> None:\n if self.backend == \"sqlite\":\n conn.executescript(script)\n return\n statements = [statement.strip() for statement in script.split(\";\") if statement.strip()]\n for statement in statements:\n conn.execute(statement)\n\n# database/db_helper.py:305-308 —— 每次 execute 新建游标,从不关闭\ndef execute(self, query: str, params: tuple[Any, ...] | list[Any] | None = None):\n cursor = self.connection.cursor()\n cursor.execute(query, tuple(params or ()))\n return cursor",
|
|||
|
|
"impact": "迁移切分是定时炸弹:现在无害,下一条含触发器、函数或默认值字符串的迁移即爆,且失败点在半条 SQL 上、报错信息毫无意义。游标不关闭因连接即建即关而泄漏窗口很短,但 advanced_material_repository 单个方法内可能连发 5-6 条语句,psycopg2 路径下这些游标要等 close() 才释放。",
|
|||
|
|
"fix": "1) 迁移执行改用 sqlparse.split(),或约定迁移脚本用 -- statement-breakpoint 显式分隔。2) 适配器改成 with self.connection.cursor() as cursor:,或返回持有游标并在上下文退出时关闭的包装。更彻底的做法是删掉 psycopg2 回退分支——psycopg3 已是标准,维护两条驱动路径的收益不抵成本。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "list_stale 与图库列表的排序键与现有索引不匹配",
|
|||
|
|
"location": "database/repositories/job_repository.py:191",
|
|||
|
|
"problem": "list_stale 是 WHERE status IN (...) AND updated_at <= ? ORDER BY updated_at ASC,但 generation_tasks 上只有 idx_generation_tasks_status(status, created_at)(0005:22-23),updated_at 完全没索引,过滤后必须 filesort。image_repository.list_images(:171-180)的 ORDER BY is_archived ASC, created_at DESC, image_id DESC 首排序键 is_archived 不在 idx_images_profile(profile_id, created_at)(0004:30-31)中,索引有序性被浪费。",
|
|||
|
|
"evidence": "# database/repositories/job_repository.py:191-201\nrows = conn.execute(\n f\"\"\"\n SELECT {TASK_COLUMNS}\n FROM generation_tasks\n WHERE status IN ({status_placeholders})\n {age_clause}\n ORDER BY updated_at ASC, created_at ASC, task_id ASC\n {limit_clause}\n \"\"\",\n tuple(params),\n).fetchall()",
|
|||
|
|
"impact": "任务表只增不删,且 advanced_material_worker_runtime.py:82-84 在启动时以 updated_before=None, limit=None 全量扫描,随数据量线性劣化;图库分页在归档图较多时同理。",
|
|||
|
|
"fix": "补迁移 0019:CREATE INDEX IF NOT EXISTS idx_generation_tasks_status_updated ON generation_tasks(status, updated_at); CREATE INDEX IF NOT EXISTS idx_images_profile_archived ON generation_images(profile_id, is_archived, created_at DESC);"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "前端功能模块与共享层(App.vue 之外)",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A9-前端功能组件.md",
|
|||
|
|
"files_read": [
|
|||
|
|
"frontend/shared/api/client.ts",
|
|||
|
|
"frontend/shared/api/endpoints.ts",
|
|||
|
|
"frontend/shared/types/workflow-round.ts",
|
|||
|
|
"frontend/shared/types/workflow.ts",
|
|||
|
|
"frontend/shared/types/profile.ts",
|
|||
|
|
"frontend/shared/types/material.ts",
|
|||
|
|
"frontend/shared/types/image.ts",
|
|||
|
|
"frontend/shared/utils/presentation.ts",
|
|||
|
|
"frontend/shared/utils/collections.ts",
|
|||
|
|
"frontend/plugin-page/src/App.vue",
|
|||
|
|
"frontend/plugin-page/src/main.ts",
|
|||
|
|
"frontend/plugin-page/vite.config.ts",
|
|||
|
|
"frontend/standalone-web/vite.config.ts",
|
|||
|
|
"frontend/standalone-web/src/main.ts",
|
|||
|
|
"frontend/package.json",
|
|||
|
|
"frontend/standalone-web/src/features/materials/AdvancedMaterialImportPage.vue",
|
|||
|
|
"frontend/standalone-web/src/features/materials/advanced-material-state.ts",
|
|||
|
|
"frontend/standalone-web/src/features/gallery/GalleryPage.vue",
|
|||
|
|
"frontend/standalone-web/src/features/gallery/types.ts",
|
|||
|
|
"frontend/standalone-web/src/features/profiles/ProfilesPage.vue",
|
|||
|
|
"frontend/standalone-web/src/features/profiles/ProfileDetailPane.vue",
|
|||
|
|
"frontend/standalone-web/src/features/profiles/ProfileWardrobePane.vue",
|
|||
|
|
"frontend/standalone-web/src/features/profiles/profile-ui.ts",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowCandidatePool.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowRoundPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowGenerationReviewDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowMaterialDrawer.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowCompareDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowBranchPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowComposer.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/WorkflowDependencyContextCard.vue",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/types.ts",
|
|||
|
|
"frontend/standalone-web/src/features/workflow/workflow-lock-state.ts",
|
|||
|
|
"frontend/standalone-web/src/features/settings/SettingsPage.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/BackupPage.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/CapabilityProfilesPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/ReviewPolicyPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/SystemManagementPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/StorageMaintenance.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/PromptGovernancePanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/RecognitionPromptsPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/settings/ModelAssignmentsPanel.vue",
|
|||
|
|
"frontend/standalone-web/src/features/quick-profile/profile-import.ts",
|
|||
|
|
"frontend/standalone-web/src/components/MaterialOperationDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/components/ConfirmDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/components/ImagePreviewDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/components/MaterialDetailDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/components/MaterialCategoryCreateDialog.vue",
|
|||
|
|
"frontend/standalone-web/src/material-labels.ts",
|
|||
|
|
"frontend/standalone-web/src/safe-backdrop-close.ts",
|
|||
|
|
"frontend/tests/production-data-boundary.test.mjs",
|
|||
|
|
"backend/dto/workflow_round_dto.py",
|
|||
|
|
"backend/dto/profile_dto.py",
|
|||
|
|
"backend/services/wardrobe_service.py",
|
|||
|
|
"backend/services/settings_service.py",
|
|||
|
|
"backend/routes/wardrobe_routes.py",
|
|||
|
|
"core/secret_service.py",
|
|||
|
|
"database/repositories/profile_repository.py"
|
|||
|
|
],
|
|||
|
|
"summary": "plugin-page 与 standalone-web 之间没有代码重复(plugin-page/src/App.vue 仅 7 行转发 standalone-web 的 App.vue),双产物单源码架构是干净的;真正的问题是共享层缺失导致同一逻辑在功能组件里反复手写——ApiClientError 提供了结构化 code/details,但唯一的中文化函数 readableApiError 被关在 App.vue 内无法导出,结果 13 个直接调 API 的组件各写各的 catch(18+ 处),把英文超时原文直抛给中文用户。查到 4 条确凿的功能性缺陷:AdvancedMaterialImportPage.vue:1314-1316 存在编码损坏的 \"????????????????\" 字面量且已被模板渲染;profile-import.ts 把「性格关键词待从原文继续整理」等占位提示当真实值写入问卷并最终进入 base_prompt;SettingsPage.vue:98 把 UI 明确承诺「不写入配置或日志」的 API Key 以 api_key_plaintext 塞进 settings/save,后端会原样落盘到 config.yaml;快速角色问卷默认值是一个填满的虚构示例角色,且绕过了 production-data-boundary 测试的扫描范围。架构层面:ProfilesPage→ProfileDetailPane→ProfileWardrobePane 是三层 prop drilling(24 props/13 emits,16 个纯透传),ProfileWardrobePane 状态归属分裂(衣橱走 props、模板自己调 API),焦点陷阱被复制 6 份且行为各异、另有 2 个 aria-modal 对话框零键盘管理,生成状态中文标签在 4 处重复且译法冲突。类型层与后端 DTO 目前逐字段对齐无实际漂移,但全靠人工维护、无契约测试兜底。另发现死组件 WorkflowDependencyContextCard.vue(279 行零引用)与零使用的 pinia/vue-router 依赖。",
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "styleValidationReviewTrust 返回编码损坏的问号串,且已渲染到界面",
|
|||
|
|
"location": "frontend/standalone-web/src/features/materials/AdvancedMaterialImportPage.vue:1314",
|
|||
|
|
"problem": "函数三个返回分支全部是编码损坏后残留的 ASCII 问号,中文字符在某次保存/转码中被替换成 ?。这不是注释或占位常量,而是直接返回给模板的显示文本。模板 :1656 处 v-if=\"styleValidationFidelityPercent(validationCase) != null\" 会渲染它。",
|
|||
|
|
"evidence": "if (inconsistent || risks) return \"????????????????\"\nif (Number.isFinite(coverage)) return `????? ? ???? ${Math.round(coverage * 100)}%`\nreturn \"?????\"\n\n<!-- :1656 -->\n<small v-if=\"styleValidationFidelityPercent(validationCase) != null\" class=\"advanced-style-review-trust\">{{ styleValidationReviewTrust(validationCase) }}</small>",
|
|||
|
|
"impact": "只要画风验证轮返回了 fidelity_score,进阶素材审核页就会在每个验证场景卡片下显示一串问号。审核人员本应据此判断「评分可信度/维度覆盖率」,现在这条关键信息完全丢失,等于在缺少可信度提示的情况下做封存决策。全仓库仅此一处出现 ???? 字面量,可确定是单点编码事故。",
|
|||
|
|
"fix": "恢复三条中文文案,例如 \"评分可信度存疑:维度不一致或存在语义风险\" / `评分可信 · 维度覆盖 ${pct}%` / \"评分可信度未知\";同时在 frontend/tests/ 增加断言:features/ 下所有 .vue 不得出现连续 3 个以上 ? 的字符串字面量,防止编码回退再次发生。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "文本角色卡导入把「待整理」占位提示当成真实设定写入档案,最终污染 base_prompt",
|
|||
|
|
"location": "frontend/standalone-web/src/features/quick-profile/profile-import.ts:149",
|
|||
|
|
"problem": "resolveImportedProfileField 的第 4 个参数 fallback 在「没识别到」时会被写进 .value(见 :336 return { value: fallback, recognized: false })。personality/appearance/outfit/scene/avoid 五个字段的 fallback 不是空串而是给人看的提示语。调用方 App.vue:9721 只用 recognizedKeys 控制 UI 高亮,却把 .value 无条件复制进问卷。",
|
|||
|
|
"evidence": "// profile-import.ts:149\nconst personality = resolveImportedProfileField(source, [\"性格\", ...], [...], \"性格关键词待从原文继续整理\")\nconst appearance = resolveImportedProfileField(source, [\"外观\", ...], [...], \"外观特征待从原文继续整理\")\nconst outfit = resolveImportedProfileField(source, [\"服装\", ...], [...], \"服装与配饰待从原文继续整理\")\n\n// :336\nreturn { value: fallback, recognized: false }\n\n// App.vue:9721 —— 未按 recognized 过滤\npersonality: draft.personality,\nappearance: draft.appearance,\n\n// App.vue:9568 —— 进入角色实体,再经 :9367 拼进 base_prompt\npersonality: [questionnaire.personality, questionnaire.likes, questionnaire.dislikes].map(i=>i.trim()).filter(Boolean).join(\";\") || \"性格待细化\",",
|
|||
|
|
"impact": "导入一份没有「性格:」小节的角色卡后,角色档案的性格字段会变成中文句子「性格关键词待从原文继续整理」,并作为 personality: 片段进入 base_prompt 送给生图/文本模型。模型会把这句元指令当作角色描述处理,生成结果不可控;用户看到问卷里已经\"填好了\"也不会去清理。",
|
|||
|
|
"fix": "把这 5 个 fallback 改为空串,提示语移到 UI 的 placeholder 属性(ExternalProfileImportPreviewRow 已有 recognized 标志可用于渲染灰色提示)。若要保留提示,则在 App.vue:9721 处按 draft.recognizedKeys.includes(key) 过滤后再赋值。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "「仅本次会话使用、不写入配置」的 API Key 被前端主动写进 settings/save 请求体",
|
|||
|
|
"location": "frontend/standalone-web/src/features/settings/SettingsPage.vue:98",
|
|||
|
|
"problem": "输入框标签(:133)明确写着「仅本次页面会话使用」「不写入配置或日志」,但 saveAndProbeAdapter 把该值作为 api_key_plaintext 注入 config.generation.adapters[] / config.generation.provider,然后调用 endpoints.settingsSave。后端 SettingsService.save() 会把合并后的 config 通过 write_text_atomic(config_path, dump_yaml(merged)) 原样落盘(backend/services/settings_service.py:445),_mask_sensitive 只作用于返回值和审计快照,不作用于写入磁盘的内容。",
|
|||
|
|
"evidence": "// SettingsPage.vue:98\nconst key = probeApiKeyValue(a.id).trim()\nif (Array.isArray(generation.adapters)) {\n generation.adapters = generation.adapters.map(item => {\n const adapter = obj(item)\n if (String(adapter.id||adapter.provider_id) !== a.id) return item\n return { ...adapter, ...(key ? { api_key_plaintext: key } : {}) }\n })\n}\nconst saved = await postJson<SaveSettingsResult>(endpoints.settingsSave, { config })\n\n<!-- :133 -->\n<small>点击测试时会直接读取输入框当前值,兼容浏览器自动填充;不写入配置或日志。</small>",
|
|||
|
|
"impact": "用户基于「不落盘」的明确承诺输入生产环境 API Key,实际会以明文形式持久化到 config.yaml(该文件在仓库中被 git 跟踪)。承诺与行为直接背离,属于凭据泄漏风险。此外 duplicateAdapter(:89)会把该会话密钥复制到新适配器 probeApiKeys[duplicateId],扩大扩散面。",
|
|||
|
|
"fix": "二选一:(a) 保持承诺——改用不落盘的探测链路,把明文 key 只放在 apiDiscoveryProbeProvider 请求体里(backend/services/api_discovery_service.py:229 已支持读 api_key_plaintext),去掉先 save 再探测;(b) 修正承诺——把 <small> 文案改成「保存后会写入插件配置文件」并加二次确认。无论哪种,duplicateAdapter 都不应复制会话密钥。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "快速角色问卷默认值是完整虚构示例角色,且绕过了 production-data-boundary 测试的防线",
|
|||
|
|
"location": "frontend/standalone-web/src/features/quick-profile/profile-import.ts:32",
|
|||
|
|
"problem": "quickProfileQuestionnaire 的初始值(App.vue:759)不是空表单,而是一个填满的示例角色。仓库已有 frontend/tests/production-data-boundary.test.mjs 专门禁止生产 UI 内联演示数据,但该测试只读取 App.vue 和 MaterialDetailDialog.vue,profile-import.ts 不在扫描范围内,这批示例数据从防线缝隙里漏了过去。",
|
|||
|
|
"evidence": "export function createDefaultQuickProfileQuestionnaire(): QuickProfileQuestionnaire {\n return {\n roleCore: \"近未来港区巡查员,负责夜间巡逻和异常记录。\",\n distinctiveFeatures: \"发光臂章,青绿色眼睛\",\n personality: \"冷静、可靠、行动快,外冷内热。\",\n appearance: \"黑色短发,青绿色眼睛,轻机能外套。\",\n scene: \"星港、雨夜街区、巡查终端、航灯背景。\",\n\n// App.vue:759\nconst quickProfileQuestionnaire = ref<QuickProfileQuestionnaire>(createDefaultQuickProfileQuestionnaire())\n\n// tests/production-data-boundary.test.mjs:5-8 —— 扫描范围不含 profile-import.ts\nconst appPath = new URL(\"../standalone-web/src/App.vue\", import.meta.url)\nconst materialDialogPath = new URL(\"../standalone-web/src/components/MaterialDetailDialog.vue\", import.meta.url)",
|
|||
|
|
"impact": "任何用户第一次打开快速角色问卷,看到的是别人的角色设定。直接点「生成」就会消耗真实 API 额度生成这个虚构角色的四张图并落库。同时与 ProfilesPage 空态文案「新建或导入角色后即可管理生成设定」自相矛盾。",
|
|||
|
|
"fix": "把 App.vue:759 改为 createEmptyQuickProfileQuestionnaire()(同文件 :61 已存在),示例值降级为各输入框的 placeholder。同时把 production-data-boundary.test.mjs 的扫描目标改为 standalone-web/src 下全部 .vue/.ts,让防线真正覆盖 features 层。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "client.ts 的结构化错误在功能层完全失效:18+ 处各写各的 catch,英文超时文案直抛给中文用户",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:4356",
|
|||
|
|
"problem": "ApiClientError 携带 code 和 details,客户端有 8 种明确错误码。但唯一做码→中文映射的 readableApiError 定义在 App.vue 的 <script setup> 内部(局部函数,无法被其它组件 import),所以每个直接调 API 的组件都退化成裸 err.message。全项目 13 个组件直接调用 shared/api/client,其中 12 个没有任何错误码分支。readableApiError 本身也只覆盖 2/8 个错误码。",
|
|||
|
|
"evidence": "// shared/api/client.ts:143 —— 面向用户的英文文案\n`AstrBot Plugin Page POST request timed out after ${Math.round(timeoutMs / 1000)} seconds`\n\n// AdvancedMaterialImportPage.vue:1355 —— 自己重造一份,丢弃 code/details\nfunction errorMessage(err: unknown) {\n return err instanceof Error ? err.message || \"请求失败\" : String(err || \"请求失败\")\n}\n\n// ProfileWardrobePane.vue:250 / SettingsPage.vue:109,111 / ReviewPolicyPanel.vue:8,9,10 / SystemManagementPanel.vue:8,9,10 / BackupPage.vue:15 / StorageMaintenance.vue:9 ...\ncatch(e){ error.value = e instanceof Error ? e.message : String(e) }\n\n// App.vue:4356 —— bridge_timeout / http_error / bridge_method_unavailable 都落到英文原文\nreturn err.message || \"ImageStudio 后端接口请求失败。\"",
|
|||
|
|
"impact": "插件桥超时(30 秒 POST / 15 秒 GET,慢速生图场景很常见)时,设置页、进阶素材页、衣橱模板等处会弹出纯英文的 AstrBot Plugin Page POST request timed out after 30 seconds。details 里的 endpoint/status/url 等排查信息全部被丢弃,用户和维护者都拿不到可操作线索。",
|
|||
|
|
"fix": "把 readableApiError 提升到 frontend/shared/api/client.ts 导出(如 export function describeApiError(err): { message: string; hint?: string }),补齐 bridge_timeout/bridge_method_unavailable/http_error 的中文分支并把 details.endpoint 作为副文本返回;然后用它替换全部 18 处内联 catch 和 AdvancedMaterialImportPage.vue:1355 的私有副本。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "ProfilesPage 三层 prop drilling:24 个 props / 13 个 emits,其中 16 个 props 纯透传",
|
|||
|
|
"location": "frontend/standalone-web/src/features/profiles/ProfilesPage.vue:10",
|
|||
|
|
"problem": "ProfilesPage 自身只使用 profiles/selectedId/loading/error/canWrite/backendConnected 六个 props,其余 18 个在 <script> 中零引用,只在模板里原样往下传。wardrobe* 这一组穿过两层组件才到达真正的消费者 ProfileWardrobePane。emits 同理:forwardWardrobeAdd 这个纯转发函数在 ProfilesPage.vue:78 和 ProfileDetailPane.vue:137 各写了一遍,代码逐字符相同。withDefaults 的默认值也在两个组件里重复声明。",
|
|||
|
|
"evidence": "// ProfilesPage.vue:78 —— 与 ProfileDetailPane.vue:137 逐字符相同\nfunction forwardWardrobeAdd(materialId: string, wardrobeCategory: string) {\n emit(\"wardrobeAdd\", materialId, wardrobeCategory)\n}\n\n<!-- ProfilesPage.vue:121 —— 21 行里 16 行是透传 -->\n<ProfileDetailPane\n :wardrobe-items=\"wardrobeItems\" :wardrobe-materials=\"wardrobeMaterials\"\n :material-categories=\"materialCategories\" :wardrobe-permissions=\"wardrobePermissions\"\n :wardrobe-loading=\"wardrobeLoading\" :wardrobe-busy=\"wardrobeBusy\" :wardrobe-error=\"wardrobeError\"\n ... @wardrobe-add=\"forwardWardrobeAdd\" @wardrobe-remove=\"emit('wardrobeRemove', $event)\" />",
|
|||
|
|
"impact": "衣橱相关的任何新字段(如「常用标记」「排序」)都要在 App.vue、ProfilesPage、ProfileDetailPane、ProfileWardrobePane 四处同步修改 props/defaults/emits/模板,共 8 个改动点。两处 withDefaults 默认值一旦不一致就会产生难以定位的行为差异。",
|
|||
|
|
"fix": "把衣橱这一族状态收敛为一个对象 prop(:wardrobe=\"{ items, materials, categories, permissions, loading, busy, error }\"),props 从 24 降到约 12;或直接让 ProfileWardrobePane 自己按 profileId 拉取衣橱数据(模板部分它已经在这么做),彻底断开中间两层的透传链。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "ProfileWardrobePane 状态归属分裂:衣橱条目走 props/emits,套装模板自己调 API",
|
|||
|
|
"location": "frontend/standalone-web/src/features/profiles/ProfileWardrobePane.vue:249",
|
|||
|
|
"problem": "同一组件里存在两套互相矛盾的数据所有权模型。衣橱条目(wardrobeItems)、加载态(loading/busy)、错误(error)由父组件持有,增删走 emit;而套装模板(outfitTemplates)、其加载态(templateBusy)、错误(templateMessage)由组件自己持有,增删改直接 getJson/postJson/deleteJson。loadTemplates 的 catch 还完全吞掉错误。",
|
|||
|
|
"evidence": "// :59 —— 自持状态,压缩成一行\nconst outfitTemplates=ref<Array<Record<string,unknown>>>([]),templateName=ref(\"\"),templateBusy=ref(false),templateMessage=ref(\"\"),compositionPreview=ref<Record<string,unknown>|null>(null)\n\n// :151 —— 自己监听 profileId 拉数据\nwatch(()=>props.profileId,()=>{void loadTemplates()},{immediate:true})\n\n// :249 —— 错误被吞\nasync function loadTemplates(){if(!props.profileId){outfitTemplates.value=[];return}try{outfitTemplates.value=await getJson(endpoints.wardrobeTemplates(props.profileId))}catch{outfitTemplates.value=[]}}\n\n// :250 —— 自己写库,父组件完全不知情\nasync function saveCurrentTemplate(){...await postJson(endpoints.wardrobeTemplates(props.profileId),{name:...,material_ids:props.wardrobeItems.map(item=>item.material_id)})...}",
|
|||
|
|
"impact": "busy(父级)和 templateBusy(自持)互不感知,用户可以在「添加衣橱素材」进行中同时点击「保存当前衣橱」,两个写操作并发命中同一个 profile 的 metadata——而后端 WardrobeService.save_template 是读-改-写整个 profile.metadata(backend/services/wardrobe_service.py:167),存在后写覆盖先写的丢更新风险。模板保存后父组件的衣橱数据也不会刷新。权限不足或后端 500 时,用户看到的是「暂无模板」而不是失败提示。",
|
|||
|
|
"fix": "统一到一侧。推荐让本组件全权负责衣橱域(自己按 profileId 拉取 items 和 templates),父组件只传 profileId + permissions;或反之把 templates 也提到 App.vue。同时给 loadTemplates 的 catch 加错误展示,并让 templateBusy 与 props.busy 合并成一个 anyBusy computed 用于禁用所有写按钮。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "焦点陷阱被复制 6 份且实现互不相同,另有 2 个 aria-modal 对话框完全没有键盘管理",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/WorkflowMaterialDrawer.vue:149",
|
|||
|
|
"problem": "项目已有共享的 safe-backdrop-close.ts 抽象(背景点击关闭),说明作者知道这类逻辑该提取,但焦点陷阱没做。6 份手写实现分成 3 种行为:ConfirmDialog 会检查 !dialogRef.contains(active) 从而能把逃逸的焦点拉回来;其余 5 份只比较 activeElement === first/last;WorkflowCompareDialog 在无可聚焦元素时直接 return 不做处理。另有 WorkflowMaterialDrawer 和 SettingsPage 适配器弹窗声明了 aria-modal=\"true\" 却零键盘管理(无 @keydown、无 nextTick focus、无 Tab 陷阱、Esc 无效)。",
|
|||
|
|
"evidence": "// ConfirmDialog.vue:61 —— 唯一正确处理焦点逃逸的版本\nif (event.shiftKey && (active === first || !dialogRef.value?.contains(active))) {\n\n// WorkflowGenerationReviewDialog.vue:168 / ImagePreviewDialog.vue:134 / MaterialOperationDialog.vue:153 / ProfileWardrobePane.vue:230 / WorkflowCompareDialog.vue:53 —— 焦点落到 body 就锁不住\nif (event.shiftKey && document.activeElement === first) {\n\n<!-- WorkflowMaterialDrawer.vue:149 —— 无任何键盘处理 -->\n<aside class=\"workflow-material-drawer\" role=\"dialog\" aria-modal=\"true\" aria-label=\"节点素材抽屉\" :aria-busy=\"busy\">\n\n<!-- SettingsPage.vue:131 —— 同样只有 aria 属性 -->\n<article ... class=\"adapter-dialog\" role=\"dialog\" aria-modal=\"true\" :aria-label=\"`${a.name} 配置`\">",
|
|||
|
|
"impact": "aria-modal=\"true\" 向屏幕阅读器承诺「焦点被限制在此对话框内」。素材抽屉和适配器配置弹窗违背了这个承诺,键盘用户 Tab 几下就会走到被 backdrop 遮住的背景内容上,且无法用 Esc 关闭。另外 5 个对话框在 busy 态下按钮被 :disabled 后当前焦点元素会失焦到 <body>,此时 Tab 同样会逃出弹窗。",
|
|||
|
|
"fix": "把 ConfirmDialog.vue:41-67 那份(最完整的)提取为 standalone-web/src/focus-trap.ts,导出 createFocusTrap(dialogRef) 返回 { onKeydown, focusFirst, restore },与已有的 createSafeBackdropClose 并列。8 处全部改用它,WorkflowMaterialDrawer 和 SettingsPage 适配器弹窗补上 @keydown 绑定。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "生成状态中文标签在 4 处重复定义且译法互相冲突",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/WorkflowCandidatePool.vue:132",
|
|||
|
|
"problem": "同一批后端状态码在 App.vue:6927、WorkflowCandidatePool.vue:132、WorkflowRoundPanel.vue:113、shared/utils/presentation.ts:11 四个地方各建一张映射表,译法不一致。App.vue 那份还缺 retrying/dead_letter,会落到 || status 显示英文原文。",
|
|||
|
|
"evidence": "// App.vue:6927\nconst labels = { planned: \"已规划\", succeeded: \"已成功\", success: \"已成功\", completed: \"已成功\", failed: \"失败\" } // 缺 retrying / dead_letter\n\n// WorkflowCandidatePool.vue:133 —— 同样的 candidate.generation_status,不同译法\nconst labels = { planned: \"待入队\", retrying: \"重试中\", succeeded: \"已完成\", failed: \"生成失败\", dead_letter: \"重试耗尽\" }\n\n// presentation.ts:11 —— 第四份\nsucceeded: \"已成功\", dead_letter: \"死信\", retrying: \"重试执行中\", superseded: \"已被重试替代\",",
|
|||
|
|
"impact": "同一个候选在候选池里显示「已完成」,在 App.vue 的其它视图(工作流概览、对比弹窗标题)里显示「已成功」,在队列页显示「已成功」;planned 更是「待入队」和「已规划」两种说法。用户会误以为是不同状态。重试中的候选在 App.vue 侧会直接显示英文 retrying。",
|
|||
|
|
"fix": "在 frontend/shared/utils/presentation.ts 增加 export function generationStatusLabel(status: string) 作为唯一真源,覆盖 planned|queued|running|retrying|succeeded|success|completed|failed|dead_letter|cancelled|superseded,删掉三份局部表(WorkflowRoundPanel 的轮次级状态 partial_success|all_failed|confirmed 可另建 roundStatusLabel 一并放入)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "WorkflowDependencyContextCard.vue(279 行)已成为无引用的死组件",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/WorkflowDependencyContextCard.vue:1",
|
|||
|
|
"problem": "全仓库(排除 node_modules、构建产物)搜索 WorkflowDependencyContextCard,除文件自身外零命中。而 App.vue 仍在维护对应数据链路:workflowDependencyContextByNodeId(:792)、workflowDependencyContextLoadingId(:793)、workflowDependencyContextError(:794)、loadWorkflowDependencyContext()(:6894),并在 :4129/:7214/:8324 三处触发加载——数据被拉取了但展示组件不再挂载,改由 App.vue 内联渲染。",
|
|||
|
|
"evidence": "// WorkflowDependencyContextCard.vue:9-26 —— 完整 props/emits 定义,无人调用\nconst props = withDefaults(defineProps<{\n context?: WorkflowNodeDependencyContext | null\n loading?: boolean; error?: string; locked?: boolean\n}>(), { context: null, loading: false, error: \"\", locked: false })\nconst emit = defineEmits<{ refresh: [] }>()\n\n// 旁证:git status 显示构建产物 web/assets/WorkflowDependencyContextCard-CVx_4GIk.js 与\n// pages/imagestudio/assets/WorkflowDependencyContextCard-Da-Ez1_X.js 已被删除",
|
|||
|
|
"impact": "279 行代码 + 对应 scoped CSS 会误导后续维护者以为依赖上下文卡片有独立组件可复用,实际改它不产生任何效果。vue-tsc --noEmit 仍会检查它,拖慢类型检查。",
|
|||
|
|
"fix": "确认 App.vue 内联版本已完全覆盖其功能后删除该文件;若打算复用,则把 App.vue 里内联的依赖上下文渲染换回该组件,并把三个 workflowDependencyContext* ref 通过 props 传入。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "GalleryPage 批量操作串行执行且完全吞掉错误,用户只看到一个无解释的数字",
|
|||
|
|
"location": "frontend/standalone-web/src/features/gallery/GalleryPage.vue:121",
|
|||
|
|
"problem": "runBatch 用 for...of 串行 await 每一张图的请求,且 catch{} 里没有任何记录,最终只把成功计数拼进文案。checkedIds 结束后被无条件清空。",
|
|||
|
|
"evidence": "for(const id of ids){try{if(action===\"select\")await postJson(endpoints.imageSelect(id));...else await postJson(endpoints.imageSafeDelete(id),{dry_run:false,delete_files:false});success++}catch{/* keep remaining items running */}}\nactionMessage.value=`已处理 ${success}/${ids.length} 张图片`;checkedIds.value=[];actionBusy.value=false;emit(\"refresh\")",
|
|||
|
|
"impact": "选中 50 张执行「安全删除」,若因引用检查全部失败,用户看到的是「已处理 0/50 张图片」,没有任何原因(后端安全删除返回的引用冲突详情全被丢弃)。串行 50 次请求在插件桥模式下每次最多 30 秒超时,最坏情况阻塞 25 分钟且无进度反馈。checkedIds 被清空导致用户无法只重试失败项。",
|
|||
|
|
"fix": "收集失败项 const failures: Array<{id: string; reason: string}> = [],用 Promise.allSettled 配合并发上限(如 5)替代串行循环,结束后把失败 id 保留在 checkedIds 里并在 actionMessage 展示前几条原因。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "browserImageUrl 校验被复制 4 份,衣橱分类标签被复制 2 份且函数名与共享导出冲突",
|
|||
|
|
"location": "frontend/standalone-web/src/features/profiles/ProfileWardrobePane.vue:40",
|
|||
|
|
"problem": "完全相同的正则 /^(?:https?:|data:|blob:|\\/)/i 出现在 profile-ui.ts:155、WorkflowCandidatePool.vue:165、WorkflowGenerationReviewDialog.vue:97、ProfileWardrobePane.vue:189 四个文件。top/bottom/dress/shoes/accessory 五条中文标签在 ProfileWardrobePane 的 WARDROBE_CATEGORY_LABELS 和 material-labels.ts 的 MATERIAL_CATEGORY_LABELS 各存一份。更糟的是 ProfileWardrobePane.vue:158 定义了局部函数 materialCategoryLabel(material: Material),与 material-labels.ts 导出的 materialCategoryLabel(value: unknown) 同名但签名和语义都不同。",
|
|||
|
|
"evidence": "// ProfileWardrobePane.vue:41 —— 与 material-labels.ts:31-35 五条完全重合\nconst WARDROBE_CATEGORY_LABELS: Readonly<Record<string, string>> = {\n top: \"上装\", bottom: \"下装\", dress: \"连衣裙\", shoes: \"鞋履\", accessory: \"配饰\",\n}\n\n// ProfileWardrobePane.vue:158 —— 与共享导出同名不同义\nfunction materialCategoryLabel(material: Material | undefined) { ... }\n// material-labels.ts:42\nexport function materialCategoryLabel(value: unknown) { ... }",
|
|||
|
|
"impact": "新增衣橱分类(如 outerwear)需要改两个文件;material-labels.ts 已有 outfit、style、composition 三条而 ProfileWardrobePane 没有,导致同一分类在素材库显示中文、在衣橱显示英文 key。同名函数会让 IDE 的「跳转到定义」和后续 import 产生歧义。",
|
|||
|
|
"fix": "把 URL 白名单校验提到 frontend/shared/utils/presentation.ts(或新建 shared/utils/media-url.ts)导出 isBrowserImageUrl,4 处统一引用;ProfileWardrobePane 删除 WARDROBE_CATEGORY_LABELS 改用 materialCategoryLabel,并把局部的 materialCategoryLabel(material) 改名为 materialCategoryName(material) 消除冲突。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "package.json 声明了零使用的 pinia / vue-router,且构建工具全在 dependencies、devDependencies 为空",
|
|||
|
|
"location": "frontend/package.json:16",
|
|||
|
|
"problem": "全仓库搜索 pinia、vue-router、createRouter、defineStore,在 frontend/ 源码(排除 node_modules)中零命中——路由由 App.vue 的 activeNav + navigation-config.ts 自己实现,状态管理靠 ref/computed。同时 typescript、vite、vue-tsc、@vitejs/plugin-vue 这些纯构建期依赖被放进 dependencies,而 devDependencies 是空对象。",
|
|||
|
|
"evidence": "\"dependencies\": {\n \"@vitejs/plugin-vue\": \"^5.2.0\",\n \"pinia\": \"^2.3.0\",\n \"typescript\": \"^5.7.0\",\n \"vite\": \"^6.0.0\",\n \"vue\": \"^3.5.0\",\n \"vue-router\": \"^4.5.0\",\n \"vue-tsc\": \"^2.2.0\"\n},\n\"devDependencies\": {}",
|
|||
|
|
"impact": "npm ci --omit=dev 在 CI/发布环境下不会减少任何体积;pinia/vue-router 会误导新维护者以为项目有 store 和路由体系,从而按错误假设写代码(这也间接解释了 App.vue 为何膨胀到 22563 行——没有 store 层可放业务状态)。",
|
|||
|
|
"fix": "删除 pinia 和 vue-router;把 @vitejs/plugin-vue、typescript、vite、vue-tsc 移到 devDependencies,dependencies 只保留 vue。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "shared/types/ 与后端 DTO 全靠人工同步,无契约测试兜底",
|
|||
|
|
"location": "frontend/shared/types/workflow-round.ts:106",
|
|||
|
|
"problem": "逐字段核对后当前没有发现实际漂移——WorkflowGenerationCandidate 的 15 个字段、WorkflowRoundCandidatePoolEntry 的 11 个字段、CharacterProfile 的 detailed_workflow_state 字面量联合与 database/repositories/profile_repository.py:85-115 的 SQL CASE 取值都精确对应。问题在于这种对齐纯靠人工纪律:endpoints.contract 端点存在但前端没有任何测试消费它,8 个前端测试全是字符串包含断言,没有一个校验类型定义。",
|
|||
|
|
"evidence": "# backend/dto/workflow_round_dto.py:102 —— 后端真源\n\"display_index\": self.display_index, \"display_code\": self.display_code,\n\"pool_state\": self.pool_state, \"preference_status\": self.preference_status,\n\"preference_reason\": self.preference_reason, \"is_final\": self.is_final,\n\n// shared/types/workflow-round.ts:106 —— 手工镜像,无生成、无校验\nexport interface WorkflowRoundCandidateView {\n round_id: string; candidate_id: string; display_index: number; display_code: string\n pool_state: string; preference_status: WorkflowCandidatePreference | string\n preference_reason?: string | null; is_final: boolean",
|
|||
|
|
"impact": "后端 DTO 增删字段时 TypeScript 不会报错(多出的字段被静默忽略,删除的字段在运行时变 undefined 而类型上仍是 string)。WorkflowCandidatePool.vue:129 的 String(candidate.candidate.generation_status || \"planned\") 已经在用防御性写法兜底,说明团队对此有隐忧但没有系统性解法。风险随 backend/dto/ 下模块数量增长而累积。",
|
|||
|
|
"fix": "加一个 node 测试,读取 backend/dto/*.py 中每个 to_dict 的字面量 key 集合,与 frontend/shared/types/*.ts 对应 interface 的字段名集合做双向差集断言(正则提取即可,不需真正解析 AST)。放进 frontend/tests/dto-contract.test.mjs,纳入 npm run test:contracts。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "多个功能文件被压成单行,等同于把源码当构建产物维护",
|
|||
|
|
"location": "frontend/standalone-web/src/features/settings/SettingsPage.vue:98",
|
|||
|
|
"problem": "GalleryPage.vue:113-122/:214、SettingsPage.vue:41-42/:79-111/:126-153、ProfileWardrobePane.vue:59/:249-252 这些不是构建产物(在 src/ 源码树内、有 <script setup lang=\"ts\">),而是被手工去掉换行和空格的源码。SettingsPage.vue:98 的 saveAndProbeAdapter 单行超过 2000 字符,内含 3 次 await、2 个循环、1 个 try/catch 和 6 个条件分支。同一文件内两种排版风格并存。",
|
|||
|
|
"evidence": "// GalleryPage.vue:113 —— 与同文件 :99-111 的正常缩进风格并存\nfunction toggleChecked(imageId:string){checkedIds.value=checkedIds.value.includes(imageId)?checkedIds.value.filter(id=>id!==imageId):[...checkedIds.value,imageId]}\n\n// ProfileWardrobePane.vue:59 —— 与同文件 :48-57 的每行一个 ref 风格并存\nconst outfitTemplates=ref<Array<Record<string,unknown>>>([]),templateName=ref(\"\"),templateBusy=ref(false),templateMessage=ref(\"\"),compositionPreview=ref<Record<string,unknown>|null>(null)",
|
|||
|
|
"impact": "git blame/git diff 在这些行上完全失效——改一个字符就是整行重写,code review 无法定位实际变更。SettingsPage.vue:98 这样的巨型单行里藏着 API Key 明文落盘的缺陷,正是因为被压成一行才长期未被发现。",
|
|||
|
|
"fix": "对这 3 个文件跑一次 Prettier/ESLint --fix 展开(项目当前无 formatter 配置,可先加 .prettierrc 约定 printWidth: 120),单独提一个纯格式化 commit 以免污染后续 blame。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "素材库 / 高级风格集 / 衣柜 / 素材绑定",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A3-素材与风格集.md",
|
|||
|
|
"summary": "这块代码的安全底线基本守住了:存储 URI 有 split_uri + resolve_uri 双重防越权,上传走魔数嗅探而非信任 content_type,生成时解析素材引用会二次校验 confirmed,MaterialCatalogResolver 用不可序列化的属性标记做信任边界,SQLite 已开 foreign_keys 所以级联删除真实生效,本域内不存在任何缓存因而不存在缓存失效问题。但有三处真实可触发的缺陷:normalize_material_bindings 对携带正确 schema_version 的载荷完全跳过归一化、随后 _index_by_node_key 直接 rule[\"target_node_keys\"] 取键,一条空规则即可让 material_bind 节点被永久毒化(节点元数据先落库、同步后抛 KeyError,此后每次更新必 500);资产包清单把 material_id 与 recognition_status 当可信输入,能静默覆盖已有素材并把未识别素材直接写成 confirmed;normalize_category 用裸子串匹配推断类别,\"landscape\" 判成 headwear、\"rooftop\" 判成 top。另有统计口径完全失效(workflow_kind 常量 advanced_style_set 与统计里写死的 style_set 对不上,风格集导入数恒为 0)与两处全表加载(list_catalog、带高级筛选的 list_materials)以及列表接口的队列对账 N+1。精简空间很大:advanced_style_set_service.py 4491 行是典型 god object(类体 2091 行 32 方法、_execute_consensus 单体 467 行、模块级 81 个纯函数 2092 行,且与 AdvancedMaterialService 双向循环引用 owner: Any),可按提示词编译/文本清洗/评审打分/证据校验拆成 4 个纯函数模块 + 3 个阶段协作者;仓库层 12 组实体的 CRUD 连接-提交-回滚模板逐字复制 6 次以上,_optional_text 在全仓被定义 28 次、_string_list 17 次、_json_or_none 13 次。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_style_set_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_normalization_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_binding_schema_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_catalog_resolver.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_analytics_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\wardrobe_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_node_material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\asset_import_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_reference_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\asset_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\advanced_material_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\models.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0009_wardrobe_items.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0016_advanced_materials.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0018_advanced_style_sets.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\storage.py"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "素材绑定 schema 校验可被 schema_version 直接绕过,并导致 material_bind 节点被永久毒化",
|
|||
|
|
"location": "backend/services/material_binding_schema_service.py:47",
|
|||
|
|
"problem": "normalize_material_bindings 的第一分支只要客户端载荷里 schema_version 等于常量值,就 deepcopy 原样返回,完全跳过 _normalize_rule 归一化;随后仅重建 by_node_id,而重建路径 _index_by_node_id -> _index_by_node_key 直接对未归一化的 rule 做 rule[\"target_node_keys\"] 下标取键。提交 {\"schema_version\": \"2026-07-09.material-binding-schema.v1\", \"active_rules\": [{}]} 即触发 KeyError。更严重的是 WorkflowService.update_node 在第 729 行已经把节点元数据写库成功,第 732 行才调用 _sync_material_bindings_from_node,异常发生时脏数据已持久化。",
|
|||
|
|
"evidence": "# material_binding_schema_service.py:47-58\nif isinstance(payload, dict) and payload.get(\"schema_version\") == MATERIAL_BINDING_SCHEMA_VERSION:\n normalized = deepcopy(payload)\n if workflow_nodes:\n rules = normalized.get(\"active_rules\")\n if not isinstance(rules, list):\n rules = [rule for rule in normalized.get(\"rules\", []) if isinstance(rule, dict) and rule.get(\"use_reference\")]\n normalized[\"by_node_id\"] = _index_by_node_id(rules, workflow_nodes)\n return normalized # rules 从未被 _normalize_rule 处理\n\n# material_binding_schema_service.py:326-331\ndef _index_by_node_key(rules: list[dict[str, Any]]) -> dict[str, list[dict[str, Any]]]:\n result: dict[str, list[dict[str, Any]]] = {}\n for rule in rules:\n for node_key in rule[\"target_node_keys\"]: # KeyError,非 .get()\n result.setdefault(canonical_gradual_node_key(node_key), []).append(rule)\n\n# workflow_service.py:729-732\nafter = self.workflow_repository.update_node(node_id, update_payload) # 已提交\nif after is None:\n raise ApiError(ErrorCode.RESOURCE_NOT_FOUND, \"node not found\", {\"node_id\": node_id})\nself._sync_material_bindings_from_node(after) # 才在这里炸",
|
|||
|
|
"impact": "任意登录用户可用一次 PUT 让某个 material_bind 节点变成只能读不能改的死节点(每次 update_node 都 500,脏数据已持久化无法自愈)。同时该分支意味着绑定规则里的 material_id / prompt / reference_strength 都是未经校验的原样存储;虽然 GenerationReferenceService._resolve_material_references 在生成时补校验了素材存在性与 confirmed 堵住了引用图泄露,但 prompt 字段仍会经 prompt_expansion_service._material_prompt_parts 拼进最终提示词,形成一条未经素材库校验的任意文本注入通道。",
|
|||
|
|
"fix": "1) 删掉整个 fast-path,或降级为仍然逐条 _normalize_rule、只跳过 by_stage/by_node_key 的重复计算;2) _index_by_node_key / _node_view 改用 rule.get(\"target_node_keys\") or [] 与 (rule.get(\"material\") or {}).get(\"prompt\");3) 把 _sync_material_bindings_from_node 的归一化提到 update_node 写库之前,归一化失败抛 ApiError(VALIDATION_FAILED) 而不是让 KeyError 逃逸成 500。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "资产包清单可静默覆盖已有素材,并把未识别素材直接写成 confirmed",
|
|||
|
|
"location": "backend/services/asset_import_service.py:141",
|
|||
|
|
"problem": "_apply_manifest 用清单里显式给出的 material_id(_material_id 优先取 item[\"material_id\"] or item[\"id\"])去查库,命中已存在的素材就走 update_material 整体覆盖(display_name/image_uri/prompt/tags/metadata/recognition_status 全被替换)。同时 recognition_status 直接取自清单,MaterialService._validate_status 只检查它属于 MATERIAL_STATUSES 集合,因此清单写 \"recognition_status\": \"confirmed\" 就能造出一条从未经过任何 VLM 识别、prompt 完全由清单作者指定的已确认素材。_asset_uri 还允许清单写 storage://<任意根>/... 绕过包目录范围。",
|
|||
|
|
"evidence": "# asset_import_service.py:141-152\nexisting = self.material_service.repository.get_material(material_payload[\"material_id\"])\nif existing is not None:\n self.material_service.get_material(material_payload[\"material_id\"])\nself.material_service.validate_material_payload(material_payload, creating=existing is None)\n...\nelse:\n material = self.material_service.update_material(material_payload[\"material_id\"], material_payload, ctx)\n\n# asset_import_service.py:340\n\"recognition_status\": _optional_text(item.get(\"recognition_status\")) or \"pending_recognition\",\n\n# asset_import_service.py:361-366\ndef _material_id(item, package_uri, manifest_uri, index, image_uri) -> str:\n explicit = _optional_text(item.get(\"material_id\") or item.get(\"id\"))\n if explicit:\n return explicit # 清单完全掌控目标 ID",
|
|||
|
|
"impact": "confirmed 是整条生成链路的信任标记(generation_reference_service.py:95 与 workflow_node_material_service.py:293 都只认这一个状态)。一个从外部下载的资产包(清单内容对操作者不透明)导入后,既能悄悄改写用户手工确认过的素材(连 normalized_asset_profile 一起被清掉),又能把任意 prompt 文本 + 任意存储根下的图片直接注入受信任素材目录并被生成阶段当作可信参考使用。",
|
|||
|
|
"fix": "导入产生的素材 ID 一律服务端生成(走 _material_id 的 sha1 分支),清单里的 material_id 只用作包内去重键;覆盖已有素材必须显式确认(payload[\"allow_overwrite\"] 且只允许覆盖同一 metadata.asset_pack.package_uri 来源的素材);recognition_status 从清单里移除,导入一律落 pending_recognition;_asset_uri 对 storage:// 前缀的值校验其 root 必须等于包所在 root 且 relative 必须在 base_relative 之下。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "normalize_category 用子串匹配推断类别:landscape 判成 headwear、rooftop 判成 top",
|
|||
|
|
"location": "backend/services/material_normalization_service.py:314",
|
|||
|
|
"problem": "normalize_category 遍历 CATEGORY_ALIASES 做 if alias in text 的裸子串包含判断,而 _build_profile 在既无 value[\"category\"] 又无 category_hint 时会把 tags + prompt 拼成一整段自由文本喂进来(_category_from_text 就是 \" \".join(...))。别名表里有 top、cap、ring、hat、boot 这类高频子串。",
|
|||
|
|
"evidence": "# material_normalization_service.py:190\ncategory = normalize_category(value.get(\"category\") or category_hint or _category_from_text([*tags, prompt]))\n\n# material_normalization_service.py:314-321\ndef normalize_category(value: Any) -> str:\n text = _text(value).lower().replace(\"_\", \" \").strip()\n if text in ASSET_CATEGORIES:\n return text\n for alias, category in CATEGORY_ALIASES.items():\n if alias in text: # 子串包含,无词边界\n return category\n return \"other\"\n\n# material_normalization_service.py:342\ndef _category_from_text(values: list[str]) -> str:\n return \" \".join(_text(item) for item in values)",
|
|||
|
|
"impact": "按 dict 声明顺序实测:\"wide landscape background\" 命中 cap(在 landscape 中)判成 headwear;\"a girl standing on a rooftop\" 命中 top 判成 top;\"spring flowers\" 命中 ring 判成 jewelry。CATEGORY_SEMANTICS[category] 直接决定 role / body_slots / wear_layer,一个背景素材会被打成头部配饰 accessory 层,进而污染 _matches_material_filters 的 body_slot/wear_layer 筛选与 CompositionBuilder 的槽位冲突判定,用户按部位筛选查不到自己的素材且看不出原因。",
|
|||
|
|
"fix": "把别名匹配改成词边界匹配并按别名长度降序尝试:_ALIAS_PATTERNS = sorted(CATEGORY_ALIASES.items(), key=lambda kv: -len(kv[0])),ASCII 别名用 re.search(rf\"(?<![A-Za-z0-9]){re.escape(alias)}(?![A-Za-z0-9])\", text),CJK 别名保持原子串匹配。另外建议 _build_profile 不再拿整段 prompt 去猜类别,只在 tags 上猜。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "list_catalog 每次分页请求都把两张素材表整表读入内存再切片",
|
|||
|
|
"location": "backend/services/advanced_material_service.py:605",
|
|||
|
|
"problem": "接口签名带 page/page_size 且上限 200,看上去是分页的;实现却先 _all_advanced_materials() + _all_standard_materials() 把两个表全部行取出、全部 to_dict()(含 metadata_json/capsule_json 反序列化),拼成 merged 后再内存切片。同样模式出现在 material_service.py:188-216(带高级筛选就整表加载再过滤)与 material_service.py:423-434(renormalize_batch 无参时整表加载)。",
|
|||
|
|
"evidence": "# advanced_material_service.py:608-628\nadvanced = self._all_advanced_materials()\nstandard = self._all_standard_materials()\nadvanced_items = [{**item.to_dict(), \"library_kind\": \"advanced\", \"priority\": ADVANCED_MATERIAL_PRIORITY} for item in advanced if item.status == \"active\"]\nstandard_items = [{**item.to_dict(), \"library_kind\": \"standard\", \"priority\": 0} for item in standard if item.recognition_status == \"confirmed\"]\nmerged = [*advanced_items, *standard_items]\noffset = (normalized_page - 1) * normalized_size\nreturn {\"items\": merged[offset : offset + normalized_size], ...}\n\n# advanced_material_service.py:646-654\ndef _all_standard_materials(self) -> list[Material]:\n result: list[Material] = []\n page = 1\n while True:\n items, total = self.material_repository.list_materials(page=page, page_size=100)\n result.extend(items)\n if len(result) >= total or not items:\n return result\n page += 1",
|
|||
|
|
"impact": "素材库到几千条时每次翻页都是 N/100 次 SQL + N 次 JSON 反序列化 + 全量 dict 构造。进阶素材的 metadata 里还嵌着 style_set.validation_cases(三条完整 case 的 to_dict(),见 advanced_style_set_service.py:1160),单条就是几十 KB。前端每翻一页重跑一次,容易把与 AstrBot 同进程的插件内存和 SQLite 打满。",
|
|||
|
|
"fix": "把过滤和排序下推到 SQL:两库合并用 UNION ALL + 统一 (priority DESC, accepted_at DESC) 排序在数据库层做 LIMIT/OFFSET。若暂不动 SQL,至少改成先各取 offset+page_size 条再归并取前 page_size 条,避免全表物化。_matches_material_filters 里的 body_slot/wear_layer/confidence 都在 metadata_json 里,可用 SQLite json_extract 建生成列 + 索引。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "素材统计里的风格集口径完全失效:workflow_kind 常量对不上,导入数恒为 0",
|
|||
|
|
"location": "backend/services/material_analytics_service.py:168",
|
|||
|
|
"problem": "_is_style_set_metadata 判断 workflow_kind == \"style_set\",而实际写入的常量是 STYLE_SET_KIND = \"advanced_style_set\"(advanced_style_set_service.py:55)。它的第二、三个兜底条件(metadata[\"style_set\"] 是 dict / 是 True)只有已封存的进阶素材满足(accept() 在 :1153 写了 style_set 字典);advanced_material_imports 行的 metadata 只有 workflow_kind、style_source_count、generation_model_route、governance_context、material_tier、upload(见 :388-395),没有任何 style_set 键。",
|
|||
|
|
"evidence": "# advanced_style_set_service.py:55\nSTYLE_SET_KIND = \"advanced_style_set\"\n\n# advanced_style_set_service.py:388-395(导入行 metadata,无 style_set 键)\n\"metadata\": {\n \"workflow_kind\": STYLE_SET_KIND,\n \"style_source_count\": len(persisted),\n \"generation_model_route\": generation_route,\n ...\n},\n\n# material_analytics_service.py:168-173\ndef _is_style_set_metadata(metadata: dict[str, Any]) -> bool:\n return (\n str(metadata.get(\"workflow_kind\") or \"\").strip().casefold() == \"style_set\" # 永远 False\n or isinstance(metadata.get(\"style_set\"), dict)\n or metadata.get(\"style_set\") is True\n )",
|
|||
|
|
"impact": "summarize() 返回的 advanced_style_sets.imports 与 advanced_style_sets.accepted_imports 恒为 0(style_set_import_rows 永远是空列表,见 :95-97、:127-128),而同一份响应里 sources / consensus_runs / validation_cases 却是真实数量。看板上出现「0 个风格集导入,却有 24 个源图识别」的自相矛盾数据,且没有任何报错提示。",
|
|||
|
|
"fix": "material_analytics_service.py 直接从 advanced_style_set_service 导入 STYLE_SET_KIND,把第一个条件改成 == STYLE_SET_KIND.casefold()(保留旧值 \"style_set\" 作兼容分支)。更根本的做法是把是否风格集下沉成 advanced_material_imports 的真实列(如 workflow_kind TEXT),别再用 JSON 里的约定字符串在三个模块间对暗号。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "导入列表接口的队列对账是 N+1,且单条详情会重复对账两次",
|
|||
|
|
"location": "backend/services/advanced_material_service.py:212",
|
|||
|
|
"problem": "list_imports 对本页每一条 import 调 _reconcile_import_task;风格集分支进入 reconcile_missing_tasks,它对每个 source、每个 consensus run、每个 attempt 下的每个 case 各调一次 _task_missing -> queue_service.get_job(task_id)(一次独立 SQL + 独立连接)。此外 get_import 在第 221 行已经对账过一次,第 224 行的 style_sets.detail(record) 在 advanced_style_set_service.py:456 又完整重跑一遍。get_import 第 259-262 行对 accepted_material_id 也查了两次。",
|
|||
|
|
"evidence": "# advanced_material_service.py:212 / 220-224\nreconciled = [self._reconcile_import_task(item) for item in items]\n...\nrecord = self._reconcile_import_task(self._require_import(import_id)) # 第 1 次\nattempts = self.repository.list_attempts(import_id)\nis_style_set = self.style_sets.is_style_set(record)\nstyle_detail = self.style_sets.detail(record) if is_style_set else {} # detail() 内部第 2 次\n\n# advanced_style_set_service.py:551-558\ndef _task_missing(self, task_id: str | None) -> bool:\n if not str(task_id or \"\").strip():\n return True\n try:\n self.queue_service.get_job(str(task_id)) # 每个子任务一次独立查询\n except ApiError:\n return True\n return False",
|
|||
|
|
"impact": "一个 8 源图的风格集单条记录就是 1+8+1+K+1+M×(1+3) 次查询,page_size 默认 30、上限 100,几十个风格集时列表页产生上千次 SQLite 查询。SQLite 在 busy_timeout=30000 下与生成 worker 的写事务竞争,表现为素材页面转圈数秒。而且这是读接口带写副作用(会把 source/case 状态改成 failed 并改 import 状态),并发列表请求之间存在重复写。",
|
|||
|
|
"fix": "1) detail() 增加 reconcile: bool = True 参数,get_import 传 False 消除重复对账;2) 给 QueueService 加 get_jobs(task_ids: set[str]) -> dict[str, Job],reconcile_missing_tasks 先收集所有 task_id 再一次 WHERE task_id IN (...) 查全;3) 列表接口不做对账,改由后台定时任务或显式刷新动作触发,避免读接口写库。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "衣柜导入把所有 sqlite3.IntegrityError 一律当成重复条目",
|
|||
|
|
"location": "backend/services/wardrobe_service.py:379",
|
|||
|
|
"problem": "_is_unique_constraint_error 对 SQLite 分支只判断 isinstance(exc, sqlite3.IntegrityError),而这个异常类型同时覆盖 UNIQUE、FOREIGN KEY、NOT NULL、CHECK 全部约束违规(PG 分支反倒精确匹配了 23505)。wardrobe_items.material_id 有 FOREIGN KEY REFERENCES materials(material_id)(migrations/0009),而 _repository_for_material_id 只有在 advanced_wardrobe_repository 与 material_catalog 都不为 None 时才把 advmat_* 路由到进阶表,否则进阶素材 ID 会被写进标准 wardrobe_items 表,必然触发外键违规。",
|
|||
|
|
"evidence": "# wardrobe_service.py:94-97\nexcept Exception as exc:\n if not _is_unique_constraint_error(exc):\n raise\n raise _duplicate_wardrobe_item_error(profile_id, material_id, wardrobe_category) from None\n\n# wardrobe_service.py:379-382\ndef _is_unique_constraint_error(exc: Exception) -> bool:\n if isinstance(exc, sqlite3.IntegrityError): # UNIQUE / FK / NOT NULL 全中\n return True\n return getattr(exc, \"sqlstate\", None) == \"23505\" or getattr(exc, \"pgcode\", None) == \"23505\"\n\n# wardrobe_service.py:328-335\ndef _repository_for_material_id(self, material_id: str) -> WardrobeRepository:\n if (self.advanced_wardrobe_repository is not None\n and self.material_catalog is not None\n and str(material_id or \"\").startswith(\"advmat_\")):\n return self.advanced_wardrobe_repository\n return self.wardrobe_repository",
|
|||
|
|
"impact": "外键违规被伪装成 VALIDATION_FAILED \"material already exists in wardrobe category\"。用户看到「素材已在该分类中」,去衣柜里找却找不到,反复重试永远失败,日志里也不留原始异常(from None 抹掉了 cause)。以字符串前缀 advmat_ 决定写哪张表本身也是脆弱耦合。",
|
|||
|
|
"fix": "_is_unique_constraint_error 改为检查 sqlite3.IntegrityError 且 getattr(exc, \"sqlite_errorcode\", None) == sqlite3.SQLITE_CONSTRAINT_UNIQUE(Py3.11+),或退化为 \"UNIQUE constraint failed\" in str(exc);非唯一约束一律 raise。仓库路由改用 MaterialCatalogResolver 返回的 metadata[\"library_kind\"](由 material_catalog_resolver.py:113-188 在服务端赋值、不可被用户元数据伪造)而不是 ID 前缀。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "update_material 中的结构校验会把 MaterialNormalizationError 泄漏成 500",
|
|||
|
|
"location": "backend/services/material_service.py:246",
|
|||
|
|
"problem": "validate_material_payload(:276-280)已经把 MaterialNormalizationError 转成 ApiError(VALIDATION_FAILED),但紧接着 update_material 里的第二处 validate_material_structures(metadata) 没有 try/except。该分支在 payload 顶层直接带 raw_analysis / normalized_asset_profile / generation_capsules 时触发(不是放在 metadata 里)。",
|
|||
|
|
"evidence": "# material_service.py:240-247\nstructured_keys = {\"raw_analysis\", \"normalized_asset_profile\", \"generation_capsules\"}\nif structured_keys & update_payload.keys():\n metadata = {**before.metadata}\n for key in structured_keys:\n if key in update_payload:\n metadata[key] = update_payload.pop(key)\n validate_material_structures(metadata) # 抛 MaterialNormalizationError(ValueError),无捕获\n update_payload[\"metadata\"] = metadata",
|
|||
|
|
"impact": "PATCH /materials/{id} 带一个 {\"normalized_asset_profile\": {\"category\": \"bogus\"}} 就返回 500 而不是 400,前端拿不到可读的字段级错误,服务端日志被无意义堆栈刷屏。同类问题在仓库层:advanced_material_repository.py:2078-2089 的 _json_object / _json_list 对损坏的 JSON 列直接 json.loads 抛 JSONDecodeError,一条脏行会让整个列表接口 500。",
|
|||
|
|
"fix": "把这里也包成 try: validate_material_structures(metadata) except MaterialNormalizationError as exc: raise ApiError(ErrorCode.VALIDATION_FAILED, str(exc), {\"field\": \"metadata\"}) from None;并把提取结构键 -> 合并 -> 校验的逻辑与 validate_material_payload 合并成一个函数,避免两处各写一遍校验。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "Pillow 缺失时风格集共识会把多张原图以 base64 全量载入单个请求",
|
|||
|
|
"location": "backend/services/material_service.py:786",
|
|||
|
|
"problem": "_optimized_vision_input 在模块顶部 except ImportError: Image = None(material_service.py:13-16,注释明说 standalone 用户可能没有 Pillow),此时原样返回最多 20 MB 的原始字节,_storage_image_data_url 再对其做 base64(约 ×1.34)。_execute_consensus 的无对比分支会对全部 succeeded 源图逐张调用它并放进 image_urls 列表。",
|
|||
|
|
"evidence": "# material_service.py:786-787\nif Image is None:\n return content, content_type # 20 MB 原样返回\n\n# advanced_style_set_service.py:1446-1449\nelse:\n for source in sources:\n image_url, _, _ = _storage_image_data_url(source.source_image_uri, {}, self.storage)\n image_urls.append(image_url)",
|
|||
|
|
"impact": "风格集允许 2-8 张源图(MIN_STYLE_SOURCES=2 / MAX_STYLE_SOURCES=8),无 Pillow 时单个 worker 任务峰值常驻约 8 × (20 MB 原始 + 27 MB base64) ≈ 375 MB,还要整体 json.dumps 进 VisionRecognitionRequest.prompt。作为与 AstrBot 同进程运行的插件,这足以 OOM 掉整个 bot。",
|
|||
|
|
"fix": "为送进模型的图设一个独立于存储上限的硬上限(如 VISION_INPUT_MAX_BYTES = 3 MB):Pillow 可用时缩放后仍超限则继续降质;Pillow 不可用且原图超限时直接抛 ApiError(UPLOAD_TOO_LARGE) 并提示安装 Pillow。共识阶段改成流式逐张编码、用完即弃,而不是先把 8 张全部物化成列表。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "advanced_style_set_service.py 是典型 god object:4491 行 = 2091 行类体 + 2092 行模块函数,且与 AdvancedMaterialService 双向循环引用",
|
|||
|
|
"location": "backend/services/advanced_style_set_service.py:309",
|
|||
|
|
"problem": "实测结构为 32 个方法 / 81 个模块级函数;最长方法 _execute_consensus 单体 467 行,_execute_validation_case 220 行,_queue_validation_batch 161 行。类通过 self.owner 反向持有 AdvancedMaterialService 的 8 个协作者,而 AdvancedMaterialService.__init__ 又持有 AdvancedStyleSetService(advanced_material_service.py:118),形成双向循环引用,两边互相调私有方法(self.owner._require_import / _govern_validation_prompt / _generation_route)。",
|
|||
|
|
"evidence": "# advanced_style_set_service.py:309-318\ndef __init__(self, owner: Any):\n self.owner = owner # 无类型标注的反向引用\n self.repository = owner.repository\n self.material_repository = owner.material_repository\n self.queue_service = owner.queue_service\n self.audit_service = owner.audit_service\n self.storage = owner.storage\n self.normalization_service = owner.normalization_service\n self.model_provider_runtime = owner.model_provider_runtime\n self.single_flow_engine = owner.single_flow_engine\n\n# advanced_material_service.py:118\nself.style_sets = AdvancedStyleSetService(self)",
|
|||
|
|
"impact": "_execute_consensus 里 text_only / 单次多图 / 分组两阶段三条 provider 调用路径与状态机、错误分类、图片编码、提示词组装全部纠缠在一个 467 行函数里,任何一条分支的改动都要重读全文;单元测试只能通过完整 mock owner 来构造,实际覆盖只能走端到端。owner: Any 让 IDE 与类型检查完全失效。",
|
|||
|
|
"fix": "按依赖方向做纯搬运式拆分(无行为变更):1) style_prompt_compiler.py 收 _detailed_style_prompt / _detailed_style_negative_prompt / _scale_specific_* / _style_kernel_parts / _compact_style_prompt_parts / _artist_signature_* / _join_prompt(约 600 行纯函数);2) style_text_sanitizer.py 收 _sanitize_minor_age_text / _strip_mature_body_locks / _normalize_style_prompt_conflicts / _clean_prompt_text / _apply_semantic_risk_rewrites 及其正则常量(约 400 行);3) style_review_scoring.py 收 _normalize_style_dimension_scores / _weighted_style_fidelity_score / _cap_inconsistent_dimension_scores / _aggregate_style_review_feedback / STYLE_REVIEW_* 常量(约 400 行);4) style_evidence.py 收 _normalized_feature_evidence / _style_trait_is_supported / _prompt_traits_overlap / _extract_head_ratio_range 系列(约 300 行);5) 剩余的类拆成 StyleSourceStage / StyleConsensusStage / StyleValidationStage 三个协作者,把 owner: Any 换成明确的 StyleSetRuntime Protocol(只暴露 6 个方法)以消除双向引用;_execute_consensus 内部按三种 analysis_strategy 拆三个私有方法,共用一个 _persist_consensus_result。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "仓库层 12 组实体的 CRUD 模板逐字复制,_optional_text 在项目里被定义了 28 次",
|
|||
|
|
"location": "database/repositories/advanced_material_repository.py:683",
|
|||
|
|
"problem": "create_style_source / create_style_consensus_run / create_attempt / create_material 是同一段 connect-execute-commit-rollback-close 的复制;update_style_source / update_style_consensus_run / update_style_validation_case / update_attempt / update_import 是同一段 _update_assignments -> 追加 updated_at -> 连接执行 -> 回读 的复制;update_*_if_status 又是 update 版本加一段 AND status IN (...) 与 rowcount 判断的复制。全仓 def _optional_text( 出现 28 次、def _string_list( 17 次、def _json_or_none( 13 次,逻辑完全一致。",
|
|||
|
|
"evidence": "# advanced_material_repository.py:683-696(update_style_source)\nconn = self.database.connect()\ntry:\n conn.execute(f\"UPDATE advanced_style_sources SET {', '.join(assignments)} WHERE source_id = {self.placeholder}\", tuple(params))\n conn.commit()\nexcept Exception:\n _rollback(conn)\n raise\nfinally:\n conn.close()\nreturn self.get_style_source(source_id)\n# 上面 14 行在本文件里出现 6 次,仅表名/主键列/回读函数不同",
|
|||
|
|
"impact": "维护成本与一致性风险:例如「status == 'succeeded' 时自动补 completed_at」的规则在 update_style_source(:677)和 update_style_source_if_status(:729)里各写了一遍,将来只改一处就会产生行为分叉。同时每个方法都自开自关连接,一次 reconcile_missing_tasks 会开关几十个 SQLite 连接。",
|
|||
|
|
"fix": "1) 抽 _ExecutionMixin 或模块级 _write(self, sql, params) / _write_if(self, sql, params) -> int 承载连接-提交-回滚-关闭,各 CRUD 只提供 SQL 与参数;2) 抽 _EntitySpec(table, pk, columns_map, row_factory, json_fields, int_fields),把 create/get/list/update/update_if_status 做成泛型方法,12 组实体只保留各自的 spec 常量;3) 把 _optional_text / _string_list / _json_or_none / _optional_int / _optional_float / _page / _page_size 统一收进 core/coercion.py,全仓改为 import;服务层同理(advanced_material_service.py 与 advanced_style_set_service.py 各有一份 _safe_failure / _optional_score / _optional_text / _string_list)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "advanced_style_set_service 从 material_service 导入 7 个下划线私有函数,上传校验层没有公共 API",
|
|||
|
|
"location": "backend/services/advanced_style_set_service.py:33",
|
|||
|
|
"problem": "advanced_style_set_service 与 advanced_material_service 都跨模块 import 了 material_service 的私有名。这些函数(data URL 解码、文件名/大小/内容类型归一、存储图转 data URL)实际上是整个素材域的公共上传校验契约,却因为下划线命名而对静态检查器和 IDE 表现为不应被外部使用。",
|
|||
|
|
"evidence": "# advanced_style_set_service.py:33-41\nfrom backend.services.material_service import (\n _normalize_upload_content_type,\n _normalize_upload_filename,\n _normalize_upload_metadata,\n _normalize_upload_size,\n _recognition_metadata,\n _storage_image_data_url,\n _vision_provider_error,\n)",
|
|||
|
|
"impact": "任何人重构 material_service 内部时不会预期这些名字有外部调用方。契约已经开始漂移:material_service.upload_material:314 用 UPLOAD_ALLOWED_CONTENT_TYPES[detected] 集合校验扩展名,而 advanced_material_service._persist_source:1571-1579 却自己维护了一张 {(\".jpg\",\".jpeg\")} 别名表,两套判定规则已不一致。",
|
|||
|
|
"fix": "新建 backend/services/material_upload_contract.py,把这 7 个函数 + MAX_UPLOAD_BYTES / UPLOAD_ALLOWED_* 常量迁过去并改成公共名(normalize_upload_filename 等),三个服务统一 import;同时把 _persist_source 的扩展名别名逻辑合并进 validate_upload_file_type,消除两套判定。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "衣柜分类无白名单校验,delete_item 用 assert 做并发保护",
|
|||
|
|
"location": "backend/services/wardrobe_service.py:128",
|
|||
|
|
"problem": "1) wardrobe_category 只校验非空,不校验是否属于 DEFAULT_WARDROBE_CATEGORIES(:31)。由于唯一索引是 (profile_id, material_id, wardrobe_category),同一素材可以用 top、Top、随便写 等无限多个分类反复导入,绕过重复检查。2) delete_item 用 assert deleted is not None 断言仓库删除结果;assert 在 python -O 下会被完全剥离,且并发删除时第二个请求会命中 AssertionError 变成 500。",
|
|||
|
|
"evidence": "# wardrobe_service.py:58-66\nwardrobe_category = str(payload.get(\"wardrobe_category\") or \"\").strip()\nif not wardrobe_category:\n raise ApiError(ErrorCode.VALIDATION_FAILED, \"wardrobe_category is required\", {\"field\": \"wardrobe_category\"})\n# 之后再无任何白名单校验\n\n# wardrobe_service.py:128-129\ndeleted = repository.delete(wardrobe_item_id)\nassert deleted is not None",
|
|||
|
|
"impact": "脏分类导致衣柜条目数虚高,match() 只按 DEFAULT_WARDROBE_CATEGORIES 或调用方传入的分类查,这些条目永远匹配不到,成为衣柜里的僵尸数据;assert 在并发下把预期内的竞态变成 500 而不是 404。",
|
|||
|
|
"fix": "wardrobe_category 归一化为小写并校验 in DEFAULT_WARDROBE_CATEGORIES(若需自定义分类,改为从 profile 元数据读用户自定义白名单)。assert 替换为 if deleted is None: raise ApiError(ErrorCode.RESOURCE_NOT_FOUND, \"wardrobe item not found\", {\"wardrobe_item_id\": wardrobe_item_id})。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "归档素材在标准库与进阶库产生两种不同的错误语义",
|
|||
|
|
"location": "backend/services/material_catalog_resolver.py:75",
|
|||
|
|
"problem": "MaterialCatalogResolver.get_material 对 advmat_* 显式过滤 status != \"active\" 返回 None;对标准素材则不看 recognition_status,归档素材照样返回 DTO。下游 _resolve_material_references 因此给出两种错误。",
|
|||
|
|
"evidence": "# material_catalog_resolver.py:75-81\nif normalized_id.startswith(ADVANCED_MATERIAL_ID_PREFIX):\n advanced = self.advanced_materials.get_material(normalized_id)\n if advanced is None or advanced.status != \"active\":\n return None # -> 404 \"reference material does not exist\"\nmaterial = self.materials.get_material(normalized_id)\nreturn self._normal_material(material) if material is not None else None\n# 归档的标准素材 recognition_status == \"archived\" -> 落到 generation_reference_service.py:95-104\n# -> 400 \"reference material is not confirmed\"",
|
|||
|
|
"impact": "同样是用户把素材删了,进阶素材报 404 素材不存在,标准素材报 400 素材未确认。用户会误以为标准素材只是没确认,跑去重新识别,结果 MaterialService.get_material(material_service.py:232)又把归档素材当作不存在,形成前后矛盾的提示闭环。",
|
|||
|
|
"fix": "在 _normal_material 之前补一行 if material.recognition_status == \"archived\": return None,让两库语义一致;同时给 _resolve_material_references 的 404 分支加上 \"reason\": \"material_missing_or_archived\" 细节字段,前端据此直接提示该素材已被删除、请重新绑定。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "前端主体 frontend/standalone-web/src/App.vue(22563 行)",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\A8-前端App主体.md",
|
|||
|
|
"summary": "App.vue 是全仓唯一的上帝组件:22563 行(script 11356 / template 2784 / 未 scoped 的 style 8395),塞进 9 个导航页、3 个生成子工作台、12 个受控浮层、206 个 ref、203 个 computed、365 个顶层函数和 129 个 API 调用点。值得肯定的是 55 个 v-for 全部带稳定 key、9 对事件监听在 onBeforeUnmount 全部成对摘除、定时器全部清理、profile 详情/衣橱有正经的 request-generation 竞态防护,且 183 处 API 调用全部走 shared/api/endpoints、零硬编码 URL,vue-tsc strict 也实跑通过。致命问题集中在三处:displayMediaUrl() 是被 4 个 computed 和整个模板 render 直接调用的伪纯函数,内部发网络请求并整体替换 resolvedMediaUrls ref,把整库图片拖进 O(N²) 重算且 Blob 对象只在卸载时才回收;5 秒事件轮询既无重入保护又在图库页触发全量分页重扫,且 slice(-1) 写法会在事件积压时静默丢弃最多 19 条事件导致 UI 停止刷新;前端 quickCharacterPrompt() 重写了后端 character_workflow_service._step_instruction() 的提示词模板且措辞已漂移(丢了 keep identity consistent),generateQuickCharacterImages() 更是在浏览器里 290 行编排后端工作流状态机并自行 PUT completed/running/failed。结构上给出了 30+ 个独立单元的完整拆分清单(模板 18 段 + 脚本 18 个 composable,含精确行号区间),按表拆完 App.vue 可收缩到 300-500 行。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\App.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tsconfig.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\package.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\character_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\quick_character_consistency.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\quick-profile\\profile-import.ts"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P0-致命",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "displayMediaUrl() 在 computed 与模板 render 中产生副作用:触发网络请求 + 写回响应式 ref,导致图库 O(N²) 重算与 Blob 无界泄漏",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:11146",
|
|||
|
|
"problem": "displayMediaUrl 看起来是格式化函数,实际同时做三件带副作用的事:读 resolvedMediaUrls.value(建立响应式依赖)、未命中时发起 HTTP 请求(resolveStorageMediaUrl/resolveSignedMediaUrl)、请求返回后整体替换 resolvedMediaUrls ref。它被 galleryImages computed(:997)、selectedWorkflowRoundImageUrlByCandidateId(:2341)、workflowGenerationReviewImageUrlByCandidateId(:2352)、materialLibraryCardFromBackend(:5450) 调用,也被模板经 imageUrl() 在 :11668/:11693/:12826/:12865/:13961 五处直接调用——这是 Vue 明令禁止的在计算属性/渲染中做副作用。",
|
|||
|
|
"evidence": "// App.vue:11146\nfunction displayMediaUrl(rawUrl?: string | null): string {\n const value = String(rawUrl || \"\").trim()\n if (!value) return \"\"\n if (/^storage:\\/\\//i.test(value)) {\n const resolved = resolvedMediaUrls.value[value] // ← 建立响应式依赖\n if (resolved) return resolved\n if (!resolvingMediaUrls.has(value)) {\n const resolution = resolveStorageMediaUrl(value) // ← computed 里发请求\n .finally(() => { resolvingMediaUrls.delete(value) })\n resolvingMediaUrls.set(value, resolution)\n }\n return \"\"\n }\n ...\n}\n\n// App.vue:11215 —— 每张图 resolve 完成后整体替换 ref\nconst objectUrl = URL.createObjectURL(new Blob([bytes], { type: payload.content_type || \"application/octet-stream\" }))\nownedMediaObjectUrls.add(objectUrl)\nresolvedMediaUrls.value = { ...resolvedMediaUrls.value, [rawUrl]: objectUrl, [sourceKey]: objectUrl }\n\n// App.vue:997 —— 图库 computed 对全部图片逐张调用\nconst galleryImages = computed<GalleryImage[]>(() => {\n const backendImages = uniqueImages([...]).filter(isAuthenticGenerationImage)\n return [...backendImages].sort(...).map((image) => ({\n ...image,\n image_url: displayMediaUrl(image.image_url),\n thumbnail_url: displayMediaUrl(image.thumbnail_url),\n }))\n})",
|
|||
|
|
"impact": "1) O(N²) 重算:galleryImages 依赖 resolvedMediaUrls,而它自己让 N 张图各触发一次 ref 整体替换 → 每张图解析完成都让整个 galleryImages(含排序 + N 次 displayMediaUrl)重跑。loadGalleryImages(:4511) 会拉全库所有页,1000 张图 ≈ 100 万次 displayMediaUrl 调用,且 watch(galleryImages)(:4184) 跟着抖 1000 次。2) Blob 内存无界增长:ownedMediaObjectUrls 只在 onBeforeUnmount(:4041) 统一 revoke,SPA 会话中用户浏览过的每张图完整二进制常驻内存直到关标签页。3) render 期间发请求,任何无关响应式变更(切主题、点通知)重渲染时都可能补发媒体请求。",
|
|||
|
|
"fix": "把媒体解析从渲染时惰性触发改成数据加载时显式预取。新建 composables/useMediaUrls.ts 导出纯读表的 mediaUrl(raw) 和显式的 prefetchMediaUrls(rawUrls[]);在 loadGalleryImages / loadWorkflowRoundDetail / hydrateGenerationImages 拿到数据后批量预取,一次 resolvedMediaUrls.value = {...} 写入全部结果(N 次 ref 替换压成 1 次);给 resolvedMediaUrls 加 LRU 上限(如 300 条),淘汰时 URL.revokeObjectURL 并从 ownedMediaObjectUrls 移除;galleryImages 改用纯查表的 mediaUrl()。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "事件流轮询无重入保护,且在图库页每 5 秒触发整库分页重扫",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:4561",
|
|||
|
|
"problem": "window.setInterval(poll, 5000) 中的 poll 是 async 且内部 await Promise.allSettled(refreshes),其中 loadGalleryImages() 会把整个图库逐页拉完。没有任何 in-flight 标志,上一轮没跑完下一轮就会启动。loadGalleryImages 自身也没有并发去重(只置 loadingGalleryImages 但从不检查它)。",
|
|||
|
|
"evidence": "// App.vue:4561\nfunction startEventStreamPolling() {\n if (eventStreamTimer !== undefined) window.clearInterval(eventStreamTimer)\n const poll = async () => {\n if (backendConnection.value.state !== \"connected\" || document.hidden) return\n ...\n if (activeNav.value === \"gallery\" && generationChanged) refreshes.push(loadGalleryImages())\n ...\n await Promise.allSettled(refreshes) // ← 可能远超 5s\n }\n void poll()\n eventStreamTimer = window.setInterval(poll, 5000) // ← 无 in-flight 守卫\n}\n\n// App.vue:4519 —— 每次调用把整库翻完\ndo {\n const result = await getJson<ImageListResult>(endpoints.images, { include_archived: true, page, page_size: pageSize })\n loaded.push(...result.items.filter(isAuthenticGenerationImage))\n total = result.total\n if (!result.items.length) break\n page += 1\n} while (loaded.length < total)",
|
|||
|
|
"impact": "用户停留在图库页且后台批量生图时,每 5 秒发起 ceil(total/100) 个 GET /images;1000 张图 = 每 5 秒 10 个请求,且上一批未完成时新一批叠加,请求数堆积。叠加 displayMediaUrl 的 Blob 解析,页面会在几十秒内卡死。另外 loaded.length 只统计过滤后的图而 total 是全量计数,只要库里有被 isAuthenticGenerationImage 过滤掉的占位图,循环就必然多翻一整轮直到返回空页。",
|
|||
|
|
"fix": "给 poll 加 let eventPollInFlight = false 守卫(if (eventPollInFlight) return / try...finally 复位);给 loadGalleryImages 加 if (loadingGalleryImages.value) return;轮询触发的图库刷新改成增量(只拉第 1 页并 merge),全量扫描只保留给用户显式点刷新;循环条件改为按页数 page <= Math.ceil(total / pageSize) 而非按 loaded.length。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "事件流 slice(-1) 在事件积压时静默丢弃最多 19 条事件,导致队列/图库/日志停止刷新",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:4576",
|
|||
|
|
"problem": "lastEventStreamId 若不在本次返回的 20 条事件窗口内(findIndex 返回 -1),代码写成 slice(-1)——这是取最后一条,而不是取全部。作者显然想表达整段都是新事件。",
|
|||
|
|
"evidence": "// App.vue:4572\nconst previousIndex = hadPrevious\n ? stream.events.findIndex((event) => event.event_id === lastEventStreamId)\n : -1\nconst freshEvents = hadPrevious\n ? stream.events.slice(previousIndex >= 0 ? previousIndex + 1 : -1) // ← previousIndex === -1 时只取最后 1 条\n : []\nlastEventStreamId = newest\nif (!hadPrevious) return\nconst eventTypes = freshEvents.map((event) => String(event.event_type || \"\"))\nconst generationChanged = eventTypes.some((eventType) => eventType.startsWith(\"generation.\"))\nconst workflowChanged = eventTypes.some((eventType) => eventType.startsWith(\"workflow.\"))",
|
|||
|
|
"impact": "两个常见场景会命中:① 标签页被切走(poll 在 document.hidden 时直接 return,lastEventStreamId 停止推进),回来时游标早已滑出 20 条窗口;② 批量生图 5 秒内产生 >20 条事件。此时只有最后一条事件参与判定——若它恰好是 workflow.*,generationChanged 就是 false,队列、图库、日志全部不刷新,UI 显示几分钟前的旧数据且无任何报错。reconnectBackendWhenVisible(:4471) 在已连接状态下只补刷当前轮次,补不了这些。",
|
|||
|
|
"fix": "改为 stream.events.slice(previousIndex >= 0 ? previousIndex + 1 : 0)——游标丢失时应视全窗口为新事件(宁可多刷一次也不能漏)。同时建议提高 limit: 20,或在 visibilitychange 恢复时主动做一次 loadAll() 级别的对齐。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "前端重写了后端已有的快速角色提示词模板,两边文案已经漂移",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:6126",
|
|||
|
|
"problem": "后端 character_workflow_service._step_instruction() 已为 character_fast 的四个 node_key 定义了权威指令表,quick_character_consistency.py 也已定义服务端拥有的 identity 字段别名表(_FIELD_ALIASES)与身份锁(build_identity_lock)。前端把这两张表都重抄一遍,且措辞已不一致——前端 quick_avatar 丢掉了后端明确写的 keep identity consistent。",
|
|||
|
|
"evidence": "// 前端 App.vue:6126\nconst instructions: Record<string, string> = {\n quick_avatar: \"square avatar, close-up portrait, clear face, stable hair and eye details\",\n quick_bust: \"half-body character reference, upper-body clothing, natural posture\",\n quick_stand: \"full-body standing character illustration, complete outfit, accurate body proportions\",\n quick_sheet: \"character design sheet, multiple views, outfit details, material callouts\",\n}\n\n# 后端 backend/services/character_workflow_service.py:280\nfast = {\n \"quick_avatar\": \"square avatar, face, hair, eyes, expression, keep identity consistent\",\n \"quick_bust\": \"half-body character reference, clothing, posture, keep identity consistent\",\n \"quick_stand\": \"full-body standing character illustration, complete outfit and proportions\",\n \"quick_sheet\": \"character design sheet, multiple views, outfit details and material callouts\",\n}\n\n// 前端 App.vue:6138 —— 连身份一致性约束都在前端拼\nreturn [\n \"same single character in every image\",\n \"preserve exact gender presentation, age, species, face identity, body type, hair, eyes and distinctive features\",\n identityPrefix, persistedProfile?.base_prompt, persistedProfile?.appearance_prompt,\n profile.name, profile.identity, profile.personality, quickProfileLibraryAppearance(profile),\n profile.style, instructions[nodeKey], ...materialPrompts,\n].filter(Boolean).join(\", \")",
|
|||
|
|
"impact": "同一个 quick_avatar 节点,走 Web 前端触发和走后端自身触发会产生不同的提示词,四联图身份一致性表现不可复现。后端改提示词策略时前端不会跟着改,prompt_expansion_service.py 的任何 prompt 治理都会被前端拼串绕过。",
|
|||
|
|
"fix": "删除 quickCharacterPrompt 里的 instructions 表和身份约束前缀,只把结构化字段(identityFields、materialPrompts、用户自定义 patch)作为 payload 传给 endpoints.nodeGenerate,由后端 _step_instruction() + build_identity_lock() 组装最终 prompt。若前端需要预览,新增只读的 POST /prompt/preview 端点返回后端组装结果。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "generateQuickCharacterImages() 在前端编排后端工作流状态机(290 行),前端自行判定并 PUT completed/running/failed",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:5955",
|
|||
|
|
"problem": "这个函数在浏览器里完成:创建工作流 → 启动 → 提交锚点任务 → 并行提交 3 个从属任务 → 重新拉节点 → 自行推断工作流终态 → PUT 状态与 metadata → 失败时再 PUT status:\"failed\" 做补偿。工作流状态推进是后端职责,前端只是一个可以随时被关闭标签页打断的客户端。",
|
|||
|
|
"evidence": "// App.vue:5950-5979\nconst pendingTargetKeys = QUICK_CHARACTER_TARGET_SPECS.filter((spec) => {\n const node = nodes.find((item) => item.node_key === spec.nodeKey)\n return Boolean(node && [\"created\",\"pending\",\"queued\",\"running\",\"retrying\"].includes(normalizeStatus(node.status)))\n}).map((spec) => spec.nodeKey)\nconst generationCompleted = completedTargetKeys.length === QUICK_CHARACTER_TARGET_SPECS.length\nconst nextWorkflowStatus = generationCompleted ? \"completed\" : pendingTargetKeys.length ? \"running\" : \"failed\"\n...\nif (normalizeStatus(workflow.status) !== nextWorkflowStatus) workflowUpdate.status = nextWorkflowStatus\nworkflow = await putJson<ProfileWorkflow>(endpoints.workflow(workflow.workflow_id), workflowUpdate)\n\n// App.vue:6048-6061 —— 失败补偿也在前端\nif (activeWorkflowId) {\n try {\n const latestFailedWorkflow = await getJson<WorkflowDetail>(endpoints.workflow(activeWorkflowId))\n const failedWorkflow = await putJson<ProfileWorkflow>(endpoints.workflow(activeWorkflowId), {\n status: \"failed\",\n metadata: { ...(latestFailedWorkflow.workflow.metadata ?? {}), quick_character_failure: message },\n })",
|
|||
|
|
"impact": "用户在生成过程中刷新页面/关标签页,工作流就永久停留在 draft 或 running,没有任何服务端兜底把它推到终态——数据库累积僵尸工作流。同时 metadata 里 quick_character_batch_id、quick_character_pending_target_keys、anchor_image_id 等 10+ 字段的写入契约完全由前端定义,后端被动接收,任何字段改名都是隐式破坏。",
|
|||
|
|
"fix": "把整段编排下沉为后端端点(如 POST /workflows/{id}/quick-character/run),后端负责锚点优先 + 并行提交 + 状态推进 + 失败补偿;前端只提交一次请求并轮询状态。已有的 backend/services/quick_character_consistency.py 说明后端本来就有这一层,前端这段是重复实现。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "workflowNodeActionError 一个 ref 同时承载成功与失败文案(67 处赋值),模板按错误样式渲染",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:8292",
|
|||
|
|
"problem": "命名为 ...Error 的 ref 被用作通用状态行,成功、失败、锁定提示混在一起,且状态永不按语义复位。selectWorkflowNode()(:6472) 切换节点时也不清空它。",
|
|||
|
|
"evidence": "// App.vue:8292 —— 成功路径写进 Error ref\nworkflowNodeActionError.value = `档案卡已原子封存:${shortId(card.card_id)} / ${card.branch_name},4 类必填资产均已入库。`\n// App.vue:8536\nworkflowNodeActionError.value = `轮次 ${detail.round.display_code} 已加入生成队列。`\n// App.vue:8732\nworkflowNodeActionError.value = successMessage\n// App.vue:8631 —— 同一个 ref 又装失败\nworkflowNodeActionError.value = `${options.retry ? \"节点重试\" : \"节点生成\"}失败:${errorMessage(err)}`\n\n<!-- App.vue:12952 模板按错误样式渲染 -->\n<em v-else-if=\"workflowNodeActionError || workflowBranchError\">{{ workflowNodeActionError || workflowBranchError }}</em>",
|
|||
|
|
"impact": "成功提示被渲染成错误样式的 <em>;用户在 A 节点操作失败后切到 B 节点,B 节点继续挂着 A 的错误文案;archiveSelectedWorkflowFinal 的成功文案会一直留在界面上直到下一次操作。",
|
|||
|
|
"fix": "拆成 workflowNodeActionError 与 workflowNodeActionNotice 两个 ref(或一个 { kind: \"ok\" | \"error\"; text: string } | null),模板按 kind 分别渲染;在 selectWorkflowNode() 里统一复位。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "模板中残留字面量 \\n 字符串,会在「创建角色档案问卷」弹窗标题栏渲染出可见的 \\n",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:13091",
|
|||
|
|
"problem": "疑似脚本化编辑时把转义序列直接写进了 SFC 模板。\\n(反斜杠 + n 两个字符)落在 <header> 元素的文本内容位置。",
|
|||
|
|
"evidence": "<!-- App.vue:13091 实际内容(注意 </button> 后的两个字符) -->\n<button type=\"button\" title=\"关闭\" @click=\"closeOverlay('personaQuestionnaire')\">×</button>\\n </header>",
|
|||
|
|
"impact": "HTML 中反斜杠和 n 是普通文本,会原样显示在弹窗右上角关闭按钮旁边,是用户可见的 UI 脏字符。而且这是「创建角色档案问卷」这个主入口弹窗的标题栏。",
|
|||
|
|
"fix": "把 \\n 替换成真正的换行(</button> 后直接换行再写 </header>)。并在 CI 里加一条 lint:模板段内出现 \\\\n / \\\\t 字面量即失败。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "loadCharacterWorkflowForProfile 的缓存命中分支无条件重拉分支列表,与 watch 里的缓存守卫矛盾",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:7273",
|
|||
|
|
"problem": "watch([backendConnection.state, selectedCharacterWorkflow.workflow_id])(:4111) 明确用 workflowBranchesByWorkflowId.value[workflowId] 做了缓存去重,但 loadCharacterWorkflowForProfile 的早退分支绕过了这个守卫。",
|
|||
|
|
"evidence": "// App.vue:4114 —— watch 有缓存守卫\nif (connectionState !== \"connected\" || !workflowId || workflowBranchesByWorkflowId.value[workflowId]) return\nvoid loadWorkflowBranches(workflowId)\n\n// App.vue:7271 —— 这里没有\nconst missingWorkflowIds = workflowIds.filter((workflowId) => !workflowNodesByWorkflowId.value[workflowId]?.length)\nif (!missingWorkflowIds.length) {\n if (profile.workflowId) await loadWorkflowBranches(profile.workflowId) // ← 每次都发\n return\n}\n\n// App.vue:7529 —— 触发源:map 重建数组导致 selectedQuickProfileLibraryItem 换引用\nlocalQuickProfileLibraryItems.value = localQuickProfileLibraryItems.value.map(applyWorkflowLink)\nbackendQuickProfileLibraryItems.value = backendQuickProfileLibraryItems.value.map(applyWorkflowLink)",
|
|||
|
|
"impact": "每次角色卡对象被重建(保存工作流链接、快速生成状态更新、profile 详情加载完成)都会多打一次 GET /workflows/{id}/branches,并把 workflowBranchError/workflowBranchesLoadingId 抖动一次。虽不会死循环(第二轮不再改动角色卡数组),但属于确定性的冗余请求。",
|
|||
|
|
"fix": "抽出 ensureWorkflowBranches(workflowId):if (!workflowId || workflowBranchesByWorkflowId.value[workflowId]) return cached; return loadWorkflowBranches(workflowId)。watch(:4111) 与 loadCharacterWorkflowForProfile(:7273) 都调用它。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "27 处写操作前置校验 + 40+ 处 busy/error/try/catch/finally 请求样板完全复制粘贴",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:10105",
|
|||
|
|
"problem": "同一段五步骨架(并发守卫 → backendWriteBlockMessage 校验 → 置 busy/清 error → try/await/catch 拼错误文案 → finally 复位 + 关浮层)在文件里重复 40 多次,只有 endpoint、busy ref、error ref 和文案不同。典型样本见 :10025/:10105/:10157/:10179/:10210/:8222/:8407/:8462/:8554/:8692/:8741/:8784。",
|
|||
|
|
"evidence": "// App.vue:10105 —— 与 :10179 executeMaterialAssetPack 结构逐字相同\nasync function registerMaterialUri(payload: MaterialRegistrationRequest) {\n if (materialOperationExecuting.value) return\n const blocked = backendWriteBlockMessage(\"登记素材 URI\")\n if (blocked) { materialOperationError.value = blocked; return }\n materialOperationExecuting.value = true\n materialOperationError.value = \"\"\n let completed = false\n try {\n const created = await postJson<Material>(endpoints.materialUpload, { ...payload, metadata: {...} })\n await loadMaterials()\n selectedMaterialCategoryId.value = payload.category_id\n selectedMaterialId.value = created.material_id\n materialActionMessage.value = `...`\n completed = true\n } catch (err) {\n materialOperationError.value = `登记素材 URI 失败:${errorMessage(err)}`\n } finally {\n materialOperationExecuting.value = false\n if (completed) closeOverlay(\"materialOperation\")\n }\n}",
|
|||
|
|
"impact": "37 个 loading ref + 25 个 error ref 的生命周期各自手写,任何一处漏写 finally 就是永久卡死的按钮;新增写操作时必须记住 5 个步骤,遗漏 backendWriteBlockMessage 就绕过权限/连接检查。",
|
|||
|
|
"fix": "抽 composables/useWriteAction.ts:返回 { busy, error, notice, run(label, fn, opts) },run 内部统一做并发守卫 + backendWriteBlockMessage + try/catch/finally + 错误文案拼装。40 处调用点各自收缩到 5-10 行,可删约 400 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "8395 行未 scoped 的全局样式,与全部子组件的 <style scoped> 形成两套并行体系,是拆分的直接阻力",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:14145",
|
|||
|
|
"problem": "App.vue 的 <style>(无 scoped)占全文件 37%(14145-22540,约 1100 条选择器)全部注入全局作用域;而 components/ConfirmDialog.vue:137、components/EmptyState.vue:20、components/ImagePreviewDialog.vue:209 以及 features/** 下所有 44 个子组件都用 <style scoped>。",
|
|||
|
|
"evidence": "App.vue:14145 <style> ← 无 scoped,约 1100 条选择器全局生效\nApp.vue:14212 body { ... } ← 直接改 body\n\n// 对照:features/ 与 components/ 下全部使用 scoped\ncomponents/ConfirmDialog.vue:137 <style scoped>\ncomponents/EmptyState.vue:20 <style scoped>\ncomponents/ImagePreviewDialog.vue:209 <style scoped>",
|
|||
|
|
"impact": "1) 类名污染:App.vue 定义的 .workflow-action-card、.quick-profile-material-slot 等泛用名一旦被新组件意外命中就串样式,而 scoped 子组件反过来无法覆写 App.vue 规则(特异性相同、无 :deep() 通道)。2) 拆分阻力:任何把模板块搬进新组件的动作都必须同步搬走对应几百行 CSS,否则样式立刻丢失——这正是这个文件迟迟拆不动的直接原因。3) 单文件 8395 行 CSS 无法定位归属。",
|
|||
|
|
"fix": "按模板拆分清单把 CSS 同步切成 styles/generate.css、styles/materials.css、styles/queue.css、styles/overview.css、styles/dialogs.css,在对应新组件里 <style scoped> 引入;App.vue 只保留真正全局的 shell/token 部分(已有 styles/tokens.css、styles/workbench.css 可承接)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "queueBatchActionMessage 内部局部变量 errorMessage 遮蔽同名全局工具函数",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:10504",
|
|||
|
|
"problem": "函数内声明 const errorMessage = ... 字符串,与文件级 function errorMessage(err: unknown)(:11354) 同名。",
|
|||
|
|
"evidence": "// App.vue:10493\nfunction queueBatchActionMessage(result: Record<string, unknown>, fallback: string) {\n ...\n const errorMessage = errors ? `,${errors} 条处理失败` : \"\" // ← 遮蔽 App.vue:11354 的 errorMessage()\n const remainingMessage = remaining > 0 ? `;仍有 ${remaining} 条,可继续清理` : \"\"\n return `${fallback} ${count} 条${errorMessage}${remainingMessage}。`\n}",
|
|||
|
|
"impact": "当前逻辑正确,但这个函数体内再也无法调用错误格式化工具;后续维护者在此加 catch 分支时会拿到一个字符串而不是函数,触发 errorMessage is not a function。",
|
|||
|
|
"fix": "把局部变量改名为 failureSuffix。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "deleteSelectedProfile 遗漏清理 6 个 per-profile 缓存 Map",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:6316",
|
|||
|
|
"problem": "归档角色后清理了 8 个缓存(backendProfiles、两个 library items、profileWorkflowsByProfileId、profileAvatarUrls、两个 generation Map、workflow* 三组),但漏掉衣橱/档案卡/请求代次这几组。",
|
|||
|
|
"evidence": "// App.vue:6316-6319 —— 清理了这些\nprofileWorkflowsByProfileId.value = omitRecordKey(profileWorkflowsByProfileId.value, profile.profile_id)\nprofileAvatarUrls.value = omitRecordKey(profileAvatarUrls.value, profile.profile_id)\nprofileGenerationBusyByProfileId.value = omitRecordKey(profileGenerationBusyByProfileId.value, profile.profile_id)\nprofileGenerationErrorsByProfileId.value = omitRecordKey(profileGenerationErrorsByProfileId.value, profile.profile_id)\n// 未清理:profileWardrobeItemsByProfileId(:665)、profileWardrobeErrorsByProfileId(:666)、\n// profileWardrobeLoadingByProfileId(:667)、profileCardsByProfileId(:819)、\n// profileDetailsRequestGenerationByProfileId、profileWardrobeRequestGenerationByProfileId",
|
|||
|
|
"impact": "长会话中反复删除角色会让这几个 Map 单调增长;若 profileWardrobeLoadingByProfileId 恰好停留在 true,同 id 记录再次出现时会显示假的加载态。",
|
|||
|
|
"fix": "把 6 个 Map 的清理并进同一个 purgeProfileCaches(profileId) 辅助函数,deleteSelectedProfile 与 invalidateProfileDetailsRequests/invalidateProfileWardrobeRequests 共用。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "后端代码精简空间(重复代码 / 死代码 / 过度设计)",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\C1-后端精简空间.md",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\params.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\wardrobe_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\branch_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\job_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\material_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\profile_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\workflow_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\web_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\image_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\profile_card_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\round_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\advanced_material_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\material_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\workflow_round_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\chat_command_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\event_stream_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\web_hosting_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_material_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\settings_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_style_set_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\model_provider_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\composition_builder.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\prompt_expansion_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_normalization_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\config_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\identity.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\security.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\audit_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\event_bus.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\queue_manager.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\interop.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\webhook.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\interactions.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\fsm.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\astrbot_provider_bridge.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\profile_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\wardrobe_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\advanced_material_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_cross_review_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\job_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\workflow_round_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\image_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\vision_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\webui_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\wechat_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\onebot11_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\config.yaml",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\metadata.yaml",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\_conf_schema.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\requirements.txt"
|
|||
|
|
],
|
|||
|
|
"summary": "后端真实源码约 54,551 行,系统性扫描确认可保守削减约 1,746 行(不含高风险的路由层改造),含路由改造约 2,046 行,占比约 3.8%。最大三块冗余是:仓储层 `conn = connect() / try / finally: close()` 样板重复 131 次(≈390 行)、DTO 中 15 个纯字段镜像的 `to_dict()`(≈220 行)、以及 29 个全仓零引用的函数/类/方法加上整文件死代码 `core/queue_manager.py`(≈375 行)。复制粘贴型工具函数尤其严重:`_optional_text` 独立定义 30 次、`_string_list` 18 次且至少有 6 种互不兼容的语义、`_json_or_none` 12 次,而 `backend/routes/params.py:22` 已有的 `optional_text` 别名却从未被任何模块 import。更关键的是,复制粘贴已造成两个真实缺陷:`backend/routes/branch_routes.py:62` 的 `_create_workflow_service()` 漏配 `event_bus`,导致分支创建/切换事件被静默吞掉永不进 SSE;`engines/vision_provider.py:40` 的 `endpoint_url()` 缺少 image 版本里的 `/v1` 去重,用户填 `base_url` 以 `/v1` 结尾时会拼出 `/v1/v1/chat/completions` 而 404。另发现 `EventBus.subscribe/unsubscribe` 零调用导致 `publish` 的分发循环永远空转——它实际只是个环形缓冲,却按发布订阅总线的形态注入到三个 service,这正是 branch_routes 漏配一直没被发现的根因。",
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "仓储层连接样板重复 131 次,可用一个上下文管理器收敛",
|
|||
|
|
"location": "database/repositories/advanced_material_repository.py:346",
|
|||
|
|
"problem": "`conn = self.database.connect() / try / finally: conn.close()` 在 13 个 repository 与 core/audit_service.py 中出现 131 次(grep 统计 conn.close() 132、finally: 131)。写方法还各自手写 `except Exception: _rollback(conn); raise`,而 `_rollback` 本身被独立定义 3 次(advanced_material_repository.py:2119、workflow_cross_review_repository.py:1464、workflow_round_repository.py:1378)。13 个仓储的 `__init__` 与 `placeholder` 属性也逐字节重复。",
|
|||
|
|
"evidence": "def get_import(self, import_id: str) -> AdvancedMaterialImport | None:\n conn = self.database.connect()\n try:\n row = conn.execute(\n f\"SELECT {IMPORT_COLUMNS} FROM advanced_material_imports WHERE import_id = {self.placeholder}\",\n (import_id,),\n ).fetchone()\n finally:\n conn.close()\n return _row_to_import(row) if row else None\n\ndef get_import_by_idempotency_key(self, idempotency_key: str) -> AdvancedMaterialImport | None:\n conn = self.database.connect()\n try:\n row = conn.execute(\n f\"SELECT {IMPORT_COLUMNS} FROM advanced_material_imports WHERE idempotency_key = {self.placeholder}\",\n (idempotency_key,),\n ).fetchone()\n finally:\n conn.close()\n return _row_to_import(row) if row else None",
|
|||
|
|
"impact": "约 520 行纯噪音;「忘记 rollback」只能靠人工 review 保证(profile_repository.update 没有 rollback 分支而 advanced_material_repository.update_import 有);任何连接策略变更(连接池 / WAL / 超时)需要改 131 处。",
|
|||
|
|
"fix": "在 database/repositories/__init__.py 建 BaseRepository,提供 `__init__`、`placeholder` 属性、`@contextmanager _connect(*, write=False)`(自动 commit/rollback/close)以及 `_fetch_one(sql, params, mapper)` / `_fetch_all(...)`。上例的 get_import 塌缩为 3 行。预计削减约 390 行,并消除 13 份重复的 __init__ + placeholder。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "branch_routes 的 _create_workflow_service 漏配 event_bus,分支事件永不推送",
|
|||
|
|
"location": "backend/routes/branch_routes.py:62",
|
|||
|
|
"problem": "`_create_workflow_service()` 在 4 个 route 模块各复制一份(branch_routes.py:62、node_routes.py:305、workflow_routes.py:143、profile_routes.py:174),其中 branch_routes 那份漏掉 `event_bus=get_default_event_bus()`。WorkflowService._publish 是空守卫(workflow_service.py:1012-1014,`if self.event_bus:`),而 create_branch(workflow_service.py:514)与 switch_branch(workflow_service.py:640)都调用了 _publish。",
|
|||
|
|
"evidence": "# backend/routes/branch_routes.py:62-68 ← 缺 event_bus\ndef _create_workflow_service() -> WorkflowService:\n database = Database(load_database_settings())\n return WorkflowService(\n workflow_repository=WorkflowRepository(database),\n profile_repository=ProfileRepository(database),\n audit_service=AuditService(database),\n )\n\n# backend/routes/workflow_routes.py:143-150 ← 有 event_bus\ndef _create_workflow_service() -> WorkflowService:\n database = Database(load_database_settings())\n return WorkflowService(\n workflow_repository=WorkflowRepository(database),\n profile_repository=ProfileRepository(database),\n audit_service=AuditService(database),\n event_bus=get_default_event_bus(),\n )",
|
|||
|
|
"impact": "走 HTTP POST /workflows/{id}/branches 与 POST /workflows/{id}/switch-branch 时,workflow.branch_created / branch_switched 事件被静默吞掉且不报错,前端 SSE(backend/routes/event_routes.py)永远看不到分支变更。",
|
|||
|
|
"fix": "把 22 个 _create_*_service() 工厂集中到 backend/services/factory.py,每种 service 只保留一份构造。同时消除层级倒置:backend/services/workflow_round_worker_runtime.py:10 反向 import backend.routes.node_routes.create_generation_service,而 node_routes.py:393-396 的 create_generation_service() 只是 `return _create_generation_service()` 的纯转发壳,应一并删除。削减约 20 行并修复该功能缺陷。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "_optional_text 被独立定义 30 次,且已有的共享别名从未被 import",
|
|||
|
|
"location": "backend/routes/params.py:22",
|
|||
|
|
"problem": "同一个 5 行函数在 backend/services(19 处)、database/repositories(4 处)、engines(4 处)、adapters(2 处)、core(1 处)共 30 处独立定义。而 backend/routes/params.py:22 早已提供 `optional_text = optional_str` 别名,注释宣称「两个名字在后端都在用」,实际 grep 显示 optional_text 从未被任何模块 import——注释是错的、别名是死的。此外 3 份变体(core/astrbot_provider_bridge.py:349、adapters/wechat_adapter.py:77、adapters/onebot11_adapter.py:96)用 `str(value or \"\")` 会把 0 / False / [] 当成 None,与主流版本语义不同。",
|
|||
|
|
"evidence": "# backend/routes/params.py:20-22 —— 死别名 + 错误注释\n# Both names are used for this coercion across the backend. Keep the alias so\n# route and service code can use the vocabulary that best fits the field.\noptional_text = optional_str\n\n# backend/services/queue_service.py:713 —— 与另外 26 份逐字节相同\ndef _optional_text(value: Any) -> str | None:\n if value is None:\n return None\n text = str(value).strip()\n return text or None\n\n# core/astrbot_provider_bridge.py:349 —— 语义不同的变体\ndef _optional_text(value: Any) -> str | None:\n text = str(value or \"\").strip() # 0 / False 被当成 None\n return text or None",
|
|||
|
|
"impact": "约 150 行冗余;`str(value or \"\")` 变体对 0 / False 的处理与主流版本不一致,跨模块搬运代码时会静默改变行为。",
|
|||
|
|
"fix": "新建 core/coerce.py(不要复用 backend/routes/params.py——database/repositories 与 engines 依赖 routes 包属层级倒置),导出 optional_text / optional_int / optional_float / as_bool / mapping;30 处改为 `from core.coerce import optional_text as _optional_text`。删掉 params.py:22 的死别名及错误注释。预计削减约 145 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "_string_list 定义 18 次且至少 6 种不同语义——同名不同义的语义地雷",
|
|||
|
|
"location": "backend/services/generation_service.py:2417",
|
|||
|
|
"problem": "同名函数在 18 个模块中行为完全不同:是否对逗号字符串切分、是否去重、是否排序、是否接受 tuple/set,四个维度自由组合出至少 6 种实现。tags / prompt_parts 这类字段在 service 间流转时会得到顺序不同、元素数不同的列表。",
|
|||
|
|
"evidence": "# core/astrbot_provider_bridge.py:344 —— 逗号切分 + 去重 + 排序\ndef _string_list(value: Any) -> list[str]:\n values = value.split(\",\") if isinstance(value, str) else value if isinstance(value, list) else []\n return sorted({str(item).strip() for item in values if str(item).strip()})\n\n# backend/services/generation_service.py:2417 —— 逗号切分 + 不去重 + 不排序\ndef _string_list(value: Any) -> list[str]:\n if isinstance(value, str):\n values = value.split(\",\")\n elif isinstance(value, (list, tuple, set)):\n values = value\n else:\n return []\n return [str(item).strip() for item in values if str(item).strip()]\n\n# backend/services/workflow_round_service.py:1544 —— 字符串直接返回 [] + 去重保序\ndef _string_list(value: Any) -> list[str]:\n values = value if isinstance(value, list) else []\n return list(dict.fromkeys(str(item).strip() for item in values if str(item).strip()))\n\n# backend/services/prompt_expansion_service.py:257 —— 字符串整体作为单元素,不切分\ndef _string_list(value: Any) -> list[str]:\n if isinstance(value, str):\n return [value] if value.strip() else []",
|
|||
|
|
"impact": "prompt 片段顺序在 advanced_style_set_service(保序去重)与 astrbot_provider_bridge(排序)之间不可预测;把逗号分隔的 tags 传给 workflow_round_service 版本会直接得到空列表而不报错;任何人跨模块复制一行调用都会踩坑。",
|
|||
|
|
"fix": "在 core/coerce.py 定义语义显式命名的三个函数替代:csv_tags(value)(切分+去重+保序)、prompt_parts(value)(不切分+去重+保序)、sorted_tags(value)(切分+去重+排序),18 处按各自实际语义替换。预计削减约 90 行,并把语义分叉从隐式变为显式。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "core/queue_manager.py 整文件死代码(112 行,零引用)",
|
|||
|
|
"location": "core/queue_manager.py:1",
|
|||
|
|
"problem": "QueueManager / QueueItem 是一套完整的内存队列实现(enqueue / dequeue / complete / fail / retry / summary + FSM 状态校验),但全仓(含 tests/)没有任何模块 import core.queue_manager,也没有任何地方引用 QueueManager 或 QueueItem。真实队列走的是 backend/services/queue_service.py + database/repositories/job_repository.py 的持久化实现。",
|
|||
|
|
"evidence": "class QueueManager:\n def __init__(self):\n self._pending: deque[str] = deque()\n self._items: dict[str, QueueItem] = {}\n self._idempotency_index: dict[str, str] = {}\n\n def enqueue(self, payload: dict[str, Any], *, idempotency_key: str | None = None) -> QueueItem:\n ...\n\n# 验证:grep -rn \"queue_manager|QueueItem\" --include=*.py . 的全部命中都落在 core/queue_manager.py 自身",
|
|||
|
|
"impact": "112 行需要维护的「影子队列实现」,与真实队列的状态机语义并行存在,新人极易误用;它单独依赖 core.fsm.task_transition,让 fsm 看起来比实际有更多消费者。",
|
|||
|
|
"fix": "整文件删除。削减 112 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "VisionProviderConfig.endpoint_url 缺 /v1 去重,base_url 以 /v1 结尾时视觉请求 404",
|
|||
|
|
"location": "engines/vision_provider.py:40",
|
|||
|
|
"problem": "endpoint_url() 与 engines/image_provider.py:51 是明显的复制关系,但 image 版本有 `/v1` 去重规则,vision 版本没有。vision_endpoint 默认值就是 \"/v1/chat/completions\"(vision_provider.py:35)。backend/services/model_provider_runtime.py:210-216 与 :297-303 拼 URL 时同样没有去重。",
|
|||
|
|
"evidence": "# engines/image_provider.py:51-56\ndef endpoint_url(self) -> str:\n base = self.base_url.rstrip(\"/\")\n endpoint = self.image_endpoint if self.image_endpoint.startswith(\"/\") else f\"/{self.image_endpoint}\"\n if base.lower().endswith(\"/v1\") and endpoint.lower().startswith(\"/v1/\"):\n endpoint = endpoint[3:] # ← 去重\n return f\"{base}{endpoint}\"\n\n# engines/vision_provider.py:40-43 ← 没有去重\ndef endpoint_url(self) -> str:\n base = self.base_url.rstrip(\"/\")\n endpoint = self.vision_endpoint if self.vision_endpoint.startswith(\"/\") else f\"/{self.vision_endpoint}\"\n return f\"{base}{endpoint}\"",
|
|||
|
|
"impact": "用户填 `base_url: https://api.example.com/v1`(OpenAI 兼容服务最常见写法,也是 _conf_schema.json 里 provider base_url 的自然填法)时,视觉识别请求打到 https://api.example.com/v1/v1/chat/completions → 404,而生图请求正常。用户极难定位。",
|
|||
|
|
"fix": "把 endpoint_url 提到 engines/_provider_base.py 的共享函数 join_endpoint(base_url, endpoint),image / vision / model_provider_runtime 三处共用。削减约 10 行并修复该配置陷阱。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "DTO 中 15 个 to_dict() 是纯字段镜像,可由一个 mixin 替代",
|
|||
|
|
"location": "backend/dto/workflow_round_dto.py:29",
|
|||
|
|
"problem": "AST 扫描确认 15 个 to_dict() 的 key 列表与 value 表达式逐项等于 dataclass 字段声明顺序,无任何重命名或转换,合计 234 行。分布在 workflow_round_dto(3)、workflow_dto(4)、profile_dto、material_dto(2)、image_dto、job_dto、wardrobe_dto、profile_card_dto、workflow_feedback_dto。",
|
|||
|
|
"evidence": "@dataclass(frozen=True)\nclass WorkflowGenerationRound:\n round_id: str\n workflow_id: str\n node_id: str\n ...\n def to_dict(self) -> dict[str, Any]:\n return {\n \"round_id\": self.round_id,\n \"workflow_id\": self.workflow_id,\n \"node_id\": self.node_id,\n ...\n }",
|
|||
|
|
"impact": "新增字段必须改两处,漏改会静默导致 API 少字段——这类 bug 没有任何类型检查能拦住。234 行零信息量代码。",
|
|||
|
|
"fix": "在 backend/dto/common.py 加 `class DictMixin: def to_dict(self): return {f.name: getattr(self, f.name) for f in dataclasses.fields(self)}`,15 个类改为继承并删除各自的 to_dict(另 18 个有转换逻辑的保持手写,可覆写)。预计削减约 220 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "main.py 中 50 个 /yp 短别名命令是纯转发壳(164 行)",
|
|||
|
|
"location": "main.py:1300",
|
|||
|
|
"problem": "imagestudio_short_command 命令组下的 49 个方法,每一个都只是把参数原样转给同名长命令。代码风格还不统一——前 8 个用两行 async for / yield,从 main.py:1345 起的 41 个压成一行,说明是分批粘贴出来的。",
|
|||
|
|
"evidence": "@imagestudio_short_command.command(\"帮助\")\nasync def imagestudio_short_help(self, event: AstrMessageEvent):\n async for result in self.imagestudio_help(event):\n yield result\n\n@imagestudio_short_command.command(\"状态\")\nasync def imagestudio_short_status(self, event: AstrMessageEvent, workflow_id: str = \"\"):\n async for result in self.imagestudio_status(event, workflow_id):\n yield result\n\n@imagestudio_short_command.command(\"确认\")\nasync def imagestudio_short_confirm(self, event: AstrMessageEvent):\n async for result in self.imagestudio_confirm(event): yield result",
|
|||
|
|
"impact": "164 行样板;新增一条命令必须在两处注册,漏掉短别名不报错只表现为「短命令不存在」;命令名映射散落在 100 个装饰器里,无法一眼审计。",
|
|||
|
|
"fix": "把命令名 → 处理方法的映射抽成 _SHORT_ALIASES 常量表,在类体末尾用循环 + setattr 生成短别名(若 AstrBot 装饰器要求签名内省,则退化为保留装饰器但由代码生成脚本维护该表)。预计削减约 140 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "EventBus 的订阅分发链路全空转——subscribe/unsubscribe 零调用",
|
|||
|
|
"location": "core/event_bus.py:36",
|
|||
|
|
"problem": "EventBus 实现了 handler 注册 + 分发,但 subscribe / unsubscribe 在全仓(含 tests/、前端)零调用,self._handlers 永远是空 defaultdict,publish 里的分发循环永远迭代 0 次。唯一消费者 EventStreamService.recent_events(event_stream_service.py:22)只读 history()。",
|
|||
|
|
"evidence": "class EventBus:\n def __init__(self, *, history_limit: int = 200):\n self._handlers: DefaultDict[str, list[EventHandler]] = defaultdict(list)\n self._history: deque[Event] = deque(maxlen=max(int(history_limit), 1))\n\n def subscribe(self, event_type, handler) -> None: # 零调用\n ...\n def unsubscribe(self, event_type, handler) -> None: # 零调用\n ...\n def publish(self, event_type, payload=None) -> Event:\n event = Event(...)\n self._history.append(event)\n for handler in [*self._handlers.get(event.event_type, []), *self._handlers.get(\"*\", [])]:\n handler(event) # 永远不执行\n return event",
|
|||
|
|
"impact": "EventBus 实际只是一个有界环形缓冲,却按发布/订阅总线的形态被注入到 WorkflowService、GenerationService、QueueService,误导读者以为存在事件驱动链路——这正是 branch_routes 漏配 event_bus 一直没被发现的原因。",
|
|||
|
|
"fix": "删除 subscribe / unsubscribe、publish 中的分发循环、_handlers 字段与 EventHandler 类型别名(约 14 行),并把类改名为 EventLog 或在 docstring 明确「仅记录,不分发」。若确实计划做订阅,应先接一个真实订阅方。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "29 个函数/类/方法全仓零引用(确证死代码,约 263 行)",
|
|||
|
|
"location": "database/repositories/workflow_round_repository.py:488",
|
|||
|
|
"problem": "AST 定义扫描 + 全范围 grep(main.py + backend + core + database + engines + adapters + tests + frontend/src)交叉验证,引用数为 1(仅定义处)的符号共 29 个。其中体量最大的是 update_pool_preferences(60 行)、RawAnalysis / NormalizedAssetProfile / GenerationCapsule 三个 DTO(51 行)、QueueWorker.run_batch(15 行)、normalize_action(14 行)、requeue_dead_letters(10 行)、next_version(10 行)、get_import_by_idempotency_key(10 行)。core/interop.py 更是整条调用链死亡:只有 PRIVATE_COMPANION 常量被外部引用,PluginRequest / normalize_plugin_request / is_allowed_plugin / daily_outfit_request / ALLOWED_PLUGINS 组成一条无人调用的链(唯一入口 daily_outfit_request 自身零引用)。",
|
|||
|
|
"evidence": "# backend/services/settings_service.py:882-884 —— 注释说「为已有测试保留」,但没有任何测试引用\n# Compatibility alias for existing tests and services that imported the\n# previous private serializer. New code should import core.config_writer.\n_dump_yaml = dump_yaml\n\n# backend/services/queue_service.py:539 —— 死信重入队能力已实现但无路由、无 CLI、无测试触发\ndef requeue_dead_letters(self, ctx, *, limit: int = 50) -> dict[str, Any]:\n\n# database/repositories/workflow_round_repository.py:488 —— 60 行候选池偏好写入,零调用\ndef update_pool_preferences(self, ...):",
|
|||
|
|
"impact": "263 行需要被 review、被 grep 命中、被误认为「有人在用」的代码。requeue_dead_letters 和 update_pool_preferences 尤其危险——看起来是已交付的产品功能(死信恢复、候选池偏好),实际从未接线。",
|
|||
|
|
"fix": "全部删除。core/interop.py 保留 PRIVATE_COMPANION 常量(或把它挪到 integration_token_service.py 后整文件删除,可再减约 45 行)。若 requeue_dead_letters / update_pool_preferences 是计划中的功能,转成 issue 并删代码,不要留半成品。预计削减约 263 行(含 interop 整清理约 300 行)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "_json_or_none / _json_loads / _json_object 在仓储层重复 20 次,且两个变体保护级别不一致",
|
|||
|
|
"location": "database/repositories/capability_repository.py:156",
|
|||
|
|
"problem": "_json_or_none 有 13 份逐字节相同的定义(12 个 repository + core/audit_service.py:113)。_json_loads(6 处)与 _json_object(2 处)是同一功能的两个名字,后者多了 `isinstance(parsed, dict)` 保护而前者没有——JSON 列被写成数组时前者会返回 list,其 `-> dict[str, Any]` 类型标注是假的。",
|
|||
|
|
"evidence": "# 13 份完全相同\ndef _json_or_none(value: Any | None) -> str | None:\n if value is None:\n return None\n return json.dumps(value, ensure_ascii=False, sort_keys=True)\n\n# capability_repository.py:156 —— 无 dict 保护,类型标注撒谎\ndef _json_loads(value: str | None) -> dict[str, Any]:\n if not value:\n return {}\n return json.loads(value)\n\n# advanced_material_repository.py:2078 —— 有 dict 保护\ndef _json_object(value: str | None) -> dict[str, Any]:\n if not value:\n return {}\n parsed = json.loads(value)\n return parsed if isinstance(parsed, dict) else {}",
|
|||
|
|
"impact": "约 100 行冗余;_json_loads 的 6 个使用点在遇到非 dict JSON 时会把 list 当 dict 返回,下游 .get() 直接 AttributeError。",
|
|||
|
|
"fix": "建 database/repositories/_json.py 导出 dump_json 与 load_object(带 dict 保护),删除全部 20 个本地定义,_json_loads 的调用点统一切到有保护的版本。预计削减约 75 行并消除一类运行时 AttributeError。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "update 赋值构建块在 workflow_repository 内复制 4 次,allowed 集合各写两遍",
|
|||
|
|
"location": "database/repositories/workflow_repository.py:280",
|
|||
|
|
"problem": "update_workflow(141-176) 与 update_workflow_if_status(180-227) 用同一份 7 键 allowed 集合 + 同一段赋值循环;update_node(280-321) 与 update_node_if_status(324-379) 用同一份 13 键 allowed 集合 + 同一段循环。四段循环体逐字节相同,allowed 集合本身也各复制两遍(:281-294 与 :333-346)。同型代码另见 profile_repository.py:242、image_repository.py:185。",
|
|||
|
|
"evidence": "# workflow_repository.py:295-303 与 :348-356 完全相同\nassignments: list[str] = []\nparams: list[Any] = []\nplaceholder = self.placeholder\nfor key in allowed:\n if key not in payload:\n continue\n column = \"metadata_json\" if key == \"metadata\" else key\n value = _json_or_none(payload[key]) if key == \"metadata\" else payload[key]\n assignments.append(f\"{column} = {placeholder}\")\n params.append(value)",
|
|||
|
|
"impact": "给 node 加一个可更新字段要改 2 个 allowed 集合,漏改会导致「普通更新生效、CAS 更新静默丢字段」,且无任何测试能覆盖这种不对称。",
|
|||
|
|
"fix": "抽出模块级常量 WORKFLOW_UPDATABLE / NODE_UPDATABLE,并把 advanced_material_repository.py:2042 已有的通用 _update_assignments(payload, placeholder, columns) 提升到 database/repositories/_sql.py 供 workflow / profile / image / job 四个仓储共用(job_repository.py:568 是它的第二个分叉实现)。预计削减约 85 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "_signed_url 复制 7 份,异常范围不一致且其中一份每张图新建一次 MediaService",
|
|||
|
|
"location": "backend/routes/advanced_material_routes.py:428",
|
|||
|
|
"problem": "同一个「签名失败返回 None」的包装函数复制了 7 份(advanced_material_routes:428、image_routes:226、material_routes:317、profile_card_routes:90、round_routes:192、workflow_round_worker_runtime:257、dashboard_service:89),签名与异常范围各不相同:有的只吞 ApiError,有的吞所有异常;有的接收注入的 media_service,advanced_material_routes 版本则在函数内部 new。",
|
|||
|
|
"evidence": "# backend/routes/image_routes.py:226-232 —— 只吞 ApiError,接收注入\ndef _signed_url(media_service: MediaService, uri: str | None) -> str | None:\n if not uri:\n return None\n try:\n return media_service.sign_media(uri)[\"url\"]\n except ApiError:\n return None\n\n# backend/routes/advanced_material_routes.py:428-435 —— 吞所有异常,且每次调用新建 MediaService()\ndef _signed_url(uri: Any) -> str | None:\n text = str(uri or \"\").strip()\n if not text:\n return None\n try:\n return MediaService().sign_media(text)[\"url\"]\n except Exception:\n return None",
|
|||
|
|
"impact": "约 49 行冗余;advanced_material_routes 版本在列表序列化时每张图新建一次 MediaService(),其构造内部会 ConfigService() 读盘,列表页 N 张图 = N 次配置文件加载。",
|
|||
|
|
"fix": "在 MediaService 上加实例方法 `def signed_url(self, uri: Any) -> str | None:`,7 处全部改为 media_service.signed_url(uri)。削减约 41 行并消除 advanced_material 列表页的 N 次配置读盘。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "分页/限流工具 _page / _page_size / _limit 重复 10 次且容错语义不一致",
|
|||
|
|
"location": "database/repositories/advanced_material_repository.py:2111",
|
|||
|
|
"problem": "_page / _page_size 在 job_repository:609/617、material_repository:319/327、profile_repository:354/362、advanced_material_repository:2111/2115 共 4 组。前三份逐字节相同(try/except → max(page,1)),第四份写成 `max(int(value or 1), 1)`——传入非数字字符串时前三份返回 1、第四份抛 ValueError。_limit 在 5 处(integration_repository:109、job_repository:601、log_service:293、event_stream_service:35、settings_service:659)默认值与上限各不相同。",
|
|||
|
|
"evidence": "# job_repository.py:609 / material_repository.py:319 / profile_repository.py:354 —— 三份相同\ndef _page(value: Any) -> int:\n try:\n page = int(value)\n except (TypeError, ValueError):\n page = 1\n return max(page, 1)\n\n# advanced_material_repository.py:2111 —— 不吞异常\ndef _page(value: int) -> int:\n return max(int(value or 1), 1)",
|
|||
|
|
"impact": "约 60 行冗余;高级素材列表接口对畸形 page 参数会 500,而其他列表接口降级为第 1 页——同一套 API 的错误语义不一致。",
|
|||
|
|
"fix": "core/coerce.py 提供 page(value) / page_size(value, *, default, maximum) / clamp_limit(value, *, default, maximum),10 处替换。削减约 48 行并统一 API 参数容错语义。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "EventStreamService._event_to_webhook 造两次对象再丢弃一次",
|
|||
|
|
"location": "backend/services/event_stream_service.py:42",
|
|||
|
|
"problem": "event_from_bus() 造一个带随机 event_id 和当前时间的 WebhookEvent,紧接着 _event_to_webhook 把这两个字段全部丢掉、用 bus event 的值重建一个新对象。中间对象完全浪费。",
|
|||
|
|
"evidence": "# backend/services/event_stream_service.py:42-49\ndef _event_to_webhook(event: Event) -> WebhookEvent:\n webhook_event = event_from_bus(event.event_type, event.payload) # 造一个\n return WebhookEvent( # 立刻丢掉重造\n event_type=webhook_event.event_type,\n data=webhook_event.data,\n event_id=event.event_id,\n created_at=event.created_at,\n )\n\n# core/webhook.py:48-49\ndef event_from_bus(event_type: str, payload: dict[str, Any]) -> WebhookEvent:\n return WebhookEvent(event_type=event_type, data=dict(payload))",
|
|||
|
|
"impact": "每条 SSE 事件多一次 uuid4() + datetime.now()(WebhookEvent 的 default_factory)。recent_events 默认 limit 50、上限 200,每次拉取浪费 50~200 次 UUID 生成。可读性上更糟:读者会以为 event_from_bus 有转换逻辑。",
|
|||
|
|
"fix": "_event_to_webhook 直接构造 WebhookEvent(event_type=event.event_type, data=dict(event.payload), event_id=event.event_id, created_at=event.created_at);core/webhook.py:48-49 的 event_from_bus 随之成为死代码,一并删除。削减约 6 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "ImageProviderError 与 VisionProviderError 逐字节相同,_post_json 的 urllib 分支重复 40 行",
|
|||
|
|
"location": "engines/vision_provider.py:72",
|
|||
|
|
"problem": "两个异常类的 __init__ 与全部字段完全一致,只有类名不同(image_provider.py:96-110 vs vision_provider.py:72-86)。_post_json 的 urllib 回退分支(HTTPError / URLError / RemoteFetchError 三个 except 块,image_provider.py:272-303 vs vision_provider.py:295-337)也只差异常类和错误文案里的 image/vision 前缀。",
|
|||
|
|
"evidence": "# engines/image_provider.py:96-110 / engines/vision_provider.py:72-86 —— 除类名外完全相同\nclass ImageProviderError(RuntimeError):\n def __init__(self, message: str, *, provider_id: str, status_code: int | None = None,\n error_type: str = \"provider_error\", details: dict[str, Any] | None = None):\n super().__init__(message)\n self.provider_id = provider_id\n self.status_code = status_code\n self.error_type = error_type\n self.details = details or {}",
|
|||
|
|
"impact": "约 55 行冗余。真正的代价是分叉:image 版 _post_json 已加了 httpx 流式读取 + 动态长度上限(image_provider.py:231-271、_provider_json_limit),vision 版仍停留在纯 urllib,所以视觉调用没有 httpx 的超时/重定向控制。",
|
|||
|
|
"fix": "新建 engines/_provider_http.py 提供 `class ProviderError(RuntimeError)` 基类(两个子类只留 pass)与 `post_json(url, payload, headers, timeout_seconds, *, error_cls, label, max_bytes)`,两个 provider 各传 error_cls 与 label。削减约 55 行,并让 vision 自动获得 httpx 路径。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "core/config_service.py 存储默认值在同文件写了两遍,版本号三处硬编码",
|
|||
|
|
"location": "core/config_service.py:74",
|
|||
|
|
"problem": "7 条存储路径的默认值在 StorageConfig 的 dataclass 字段默认值(:24-31)与 load() 里的 .get(key, 默认值)(:74-83)各硬编码一次。另外 PLUGIN_VERSION = \"1.0.72\" 在 core/config_service.py:13、config.yaml:4、metadata.yaml:5 三处硬编码,而 load() 的注释(:68-70)明确说会忽略 config.yaml 里的值。",
|
|||
|
|
"evidence": "# core/config_service.py:24-31\n@dataclass(frozen=True)\nclass StorageConfig:\n images_path: str = \"data/images\"\n thumbnails_path: str = \"data/thumbnails\"\n ...\n# core/config_service.py:74-83 —— 同样的字面量再写一遍\n storage=StorageConfig(\n images_path=str(storage.get(\"images_path\", \"data/images\")),\n thumbnails_path=str(storage.get(\"thumbnails_path\", \"data/thumbnails\")),\n ...\n )",
|
|||
|
|
"impact": "改默认路径必须改两处,漏改后「未配置时」和「配置为空字符串时」会落到不同目录;config.yaml 的 plugin.version 是纯误导——用户改了它不生效。",
|
|||
|
|
"fix": "用 `_STORAGE_FIELDS = {f.name for f in dataclasses.fields(StorageConfig)}` + `StorageConfig(**{k: str(v) for k, v in storage.items() if k in _STORAGE_FIELDS and v})` 替代逐字段枚举(削减约 12 行);删除 config.yaml 中误导性的 plugin.version 字段。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "两个源文件带 UTF-8 BOM,标准 AST 工具解析失败",
|
|||
|
|
"location": "backend/routes/job_routes.py:1",
|
|||
|
|
"problem": "backend/routes/job_routes.py 与 backend/services/profile_document_import_service.py 以 \\xef\\xbb\\xbf 开头。ast.parse(path.read_text(encoding=\"utf-8\")) 直接抛 SyntaxError: invalid non-printable character U+FEFF。",
|
|||
|
|
"evidence": "$ python -c \"import pathlib; print(pathlib.Path('backend/routes/job_routes.py').read_bytes()[:6])\"\nb'\\xef\\xbb\\xbf\\\"\\\"\\\"'",
|
|||
|
|
"impact": "Python 解释器自身能处理(tokenize 会跳过 BOM),但任何自建的 AST 分析、代码生成、lint 插件、批量重构脚本都会在这两个文件上失败——本次审计的死代码扫描就漏掉了这两个文件。",
|
|||
|
|
"fix": "去掉 BOM(`p.write_bytes(p.read_bytes()[3:])`),并在 .gitattributes 或 pre-commit 加 BOM 检查。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "_now ×6、_mapping ×4,以及 291 处内联 dict 守卫写法把 key 名写两遍",
|
|||
|
|
"location": "backend/routes/node_routes.py:329",
|
|||
|
|
"problem": "_now() 在 5 处的 `return datetime.now(timezone.utc).isoformat()` 完全相同(advanced_style_set_service:4279、generation_job_retry_service:493、queue_service:709、worker_service:476、queue_manager:14)。_mapping() 已被抽出 4 次却没有共享(history_service:313、model_provider_bootstrap:353、workflow_round_service:1196、workflow_cross_review_repository:1429),而其对应的内联写法 `x.get(\"y\") if isinstance(x.get(\"y\"), dict) else {}` 在后端出现 291 次。",
|
|||
|
|
"evidence": "# backend/routes/node_routes.py:329-335 —— 一个函数里连着 5 行同款,每处 key 名写两遍\ngeneration_config = config_payload.get(\"generation\") if isinstance(config_payload.get(\"generation\"), dict) else {}\nmodel_assignments = config_payload.get(\"model_assignments\") if isinstance(config_payload.get(\"model_assignments\"), dict) else {}\ngovernance_config = config_payload.get(\"prompt_governance\") if isinstance(config_payload.get(\"prompt_governance\"), dict) else {}\ntranslation_config = governance_config.get(\"translation\") if isinstance(governance_config.get(\"translation\"), dict) else {}\nmoderation_config = governance_config.get(\"moderation\") if isinstance(governance_config.get(\"moderation\"), dict) else {}",
|
|||
|
|
"impact": "_now 约 15 行冗余;291 处内联写法虽不减行数,但每处把 key 名写两遍,是「改 key 名漏改一半」的高发点——写成 `config.get(\"a\") if isinstance(config.get(\"b\"), dict) else {}` 不会有任何工具报错。",
|
|||
|
|
"fix": "core/coerce.py 提供 utc_now_iso() 与 mapping(value);291 处内联改为 `mapping(config_payload.get(\"generation\"))`,key 名只出现一次。削减约 30 行并消除一类 key 名不一致隐患。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "148 个路由处理器的 def action() 闭包样板,权限声明与 route_registry 双真相",
|
|||
|
|
"location": "backend/routes/wardrobe_routes.py:57",
|
|||
|
|
"problem": "每个路由处理器固定四段式:`xxx_service = service or _create_xxx()` → `def action():` → `require_xxx(ctx)` → `return handle_request(ctx, action)`,共 148 处 def action() + 157 处 handle_request + 147 处 `service or _create_`。权限声明散落在闭包第一行,与 route_registry.ROUTES 的 RouteSpec.permission 是两份独立真相。同文件 :86-107 已出现绕过闭包的「元组下标」写法,说明作者本人也觉得样板烦。",
|
|||
|
|
"evidence": "# backend/routes/wardrobe_routes.py:57-70\ndef delete_wardrobe_item(ctx, profile_id, wardrobe_item_id, service=None) -> dict:\n wardrobe_service = service or _create_wardrobe_service()\n\n def action() -> dict[str, Any]:\n require_permission(ctx, \"can_manage_materials\")\n return wardrobe_service.delete_item(profile_id, wardrobe_item_id, ctx)\n\n return handle_request(ctx, action)\n\n# 同文件 :86-107 —— 作者自己的规避写法\nreturn handle_request(ctx, lambda: (require_permission(ctx, \"can_manage_materials\"),\n wardrobe_service.save_template(profile_id, payload, ctx))[1])",
|
|||
|
|
"impact": "约 450 行样板;权限矩阵无法从 route_registry.ROUTES 静态核对,这种「双真相」结构正是 branch_routes 漏配 event_bus 长期不被发现的土壤。",
|
|||
|
|
"fix": "(改动面大,建议单独立项)引入 `@route_handler(permission=\"materials\", service_factory=_create_wardrobe_service)` 装饰器,让处理器只写业务行,并在启动时断言装饰器声明的 permission 与 route_registry.ROUTES 一致。预计削减约 300 行且权限矩阵可静态校验。风险:148 个处理器的测试注入点(service= 参数)需同步调整,建议分模块灰度。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "C2-前端代码精简空间",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\C2-前端精简空间.md",
|
|||
|
|
"summary": "前端源码 40182 行,CSS 占 14698 行(36.6%),standalone-web/src/App.vue 一个文件 22563 行(脚本 11357 / 模板 2785 / 样式 8396),是典型 god component(206 个 ref、135 个 computed、264 个 function、全文仅 1 条注释)。经全量静态分析确证的净删除空间约 1970 行(占前端源码 4.9%、CSS 的 13%),主要来自:App.vue 内 413 行永不命中的死 CSS、SettingsPage.vue:165 一整段重构遗留的 .backup-* 死样式、素材选择器弹层模板被逐字复制两份(109 行且已开始功能漂移)、4 份几乎一字不差的焦点陷阱实现(其中 WorkflowCompareDialog 的选择器漏了 input/select/textarea 会导致 Tab 逃逸)、图片缩放拖拽逻辑与 ImagePreviewDialog 完全重复(脚本 60 行 + 逐字节相同的 CSS 104 行)、以及上限 2251 行的重复 CSS 声明块。pinia 与 vue-router 声明为依赖但代码库零引用,devDependencies 为空且 typescript/vite/vue-tsc 全部错放在 dependencies。好消息是 plugin-page 与 standalone-web 两套构建之间零重复(plugin-page/src/App.vue 仅 7 行 re-export),无未引用组件,仓库无跟踪垃圾文件。另有约 7200 行属于\"迁移拆分\"空间,拆完后 App.vue 可压到约 4200 行。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\package.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tsconfig.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\env.d.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\plugin-page\\src\\App.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\plugin-page\\src\\main.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\plugin-page\\vite.config.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\vite.config.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\main.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\App.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\styles\\workbench.css",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\ConfirmDialog.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\ImagePreviewDialog.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\MaterialCategoryCreateDialog.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\MaterialDetailDialog.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\EmptyState.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\StateCard.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\components\\MetricCardGrid.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\workflow\\WorkflowCompareDialog.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\workflow\\WorkflowGenerationReviewDialog.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\profiles\\ProfileDetailPane.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\profiles\\ProfilesPage.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\profiles\\ProfileWardrobePane.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\gallery\\GalleryPage.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\materials\\AdvancedMaterialImportPage.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\settings\\SettingsPage.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\settings\\BackupPage.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\utils\\presentation.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\utils\\collections.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\utils\\storage-uri.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\workflow.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\workflow-material.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\integration.ts"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "App.vue 样式块内 413 行 CSS 规则永远不会命中(死代码)",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:21129",
|
|||
|
|
"problem": "全量扫描 903 个 CSS 类(已排除 Vue Transition 派生类与动态绑定类),63 个类在任何模板/脚本中都无法被引用。逐条核实后,App.vue 内由纯死类构成的完整规则块合计 277 行,另有 136 行是子组件已有 scoped 副本的重复定义。典型死类组:generate-status-grid / generate-feature-card / generate-section / generate-card-grid / generate-card-top / feature-state(21129-21286)、workflow-node-card / workflow-node-info / workflow-node-index / workflow-node-copy(17764-17845、19007-19029,模板实际用的是 workflow-node-item)、workflow-lock-notice(19031-19059)、quick-profile-hero-caption(20075-20101)、image-card / image-grid / metric-stack / compact-panel(21484-21703)。",
|
|||
|
|
"evidence": ".generate-status-grid {\n display: grid;\n grid-template-columns: repeat(auto-fit, minmax(180px, 1fr));\n gap: 12px;\n}\n\n.generate-status-grid article,\n.generate-feature-card {\n position: relative;\n overflow: hidden;\n display: grid;\n gap: 6px;\n border: 1px solid var(--border);\n border-radius: 12px;\n padding: 14px;\n background: var(--panel);\n box-shadow: var(--shadow);\n}\n\n/* 另见 App.vue:17764 */\n.workflow-node-card {\n position: relative;\n display: grid;\n grid-template-columns: 26px minmax(0, 1fr);\n align-items: start;\n gap: 10px;\n border: 1px solid var(--border);\n border-radius: 10px;\n padding: 11px;\n background: var(--panel-muted);\n}",
|
|||
|
|
"impact": "约 413 行 CSS 每次构建都打进 web/assets/index-*.css 与 pages/imagestudio/assets/index-*.css 两份产物;.workflow-node-card 这类高度迷惑的名字会让人误以为在改生效规则,实际改了没反应。",
|
|||
|
|
"fix": "删除 App.vue:17323、17764-17845、19007-19059、20075-20101、21129-21286、21484-21499、21662-21703 等规则区间(保留 .image-viewer-backdrop,仍被 App.vue 4 个内联弹层使用)。同时新增 frontend/tests/dead-css.test.mjs(项目已有 node --test tests/*.test.mjs 契约测试习惯),断言 App.vue 样式块内的类必须能在某模板中被引用,白名单放 Transition 派生类与动态拼接类,防止回潮。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": ".image-viewer-* 的 104 行样式与 60 行缩放/拖拽逻辑在 App.vue 与 ImagePreviewDialog.vue 各存在一份",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:17213",
|
|||
|
|
"problem": "App.vue:17213-17316 与 components/ImagePreviewDialog.vue:221-324 的 CSS 逐字节相同(diff 去掉尾部一条无关规则后无差异)。App.vue 侧这 104 行的唯一使用者是 ImagePreviewDialog.vue 的模板,而该组件自带 scoped 副本,所以 App.vue 这份纯冗余。脚本侧同理:App.vue:7952-8011 与 ImagePreviewDialog.vue:58-114 是同一套 MIN/MAX 缩放 + 指针捕获拖拽逻辑,只是变量名加了 imageViewer 前缀;魔数 0.5/4/0.12/2.2 在 App.vue 里硬编码,在组件里是具名常量。",
|
|||
|
|
"evidence": "// App.vue:7982-7995\nfunction startImageViewerDrag(event: PointerEvent) {\n if (imageViewerScale.value <= 1) return\n imageViewerDragging.value = true\n imageViewerDragStart.value = {\n pointerId: event.pointerId,\n x: event.clientX,\n y: event.clientY,\n offsetX: imageViewerOffset.value.x,\n offsetY: imageViewerOffset.value.y,\n }\n if (event.currentTarget instanceof HTMLElement) {\n event.currentTarget.setPointerCapture(event.pointerId)\n }\n}\n\n// ImagePreviewDialog.vue:85-98 —— 同构\nfunction startDrag(event: PointerEvent) {\n if (scale.value <= 1) return\n dragging.value = true\n dragStart.value = {\n pointerId: event.pointerId,\n x: event.clientX,\n y: event.clientY,\n offsetX: offset.value.x,\n offsetY: offset.value.y,\n }\n if (event.currentTarget instanceof HTMLElement) {\n event.currentTarget.setPointerCapture(event.pointerId)\n }\n}",
|
|||
|
|
"impact": "约 164 行冗余(CSS 104 + 脚本 60)。更严重的是缩放手感会漂移——改 ImagePreviewDialog 的 WHEEL_STEP/DOUBLE_CLICK_SCALE 不会同步到 quick-profile 查看器,两个查看器行为逐渐不一致。",
|
|||
|
|
"fix": "抽 standalone-web/src/composables/useImageZoomPan.ts,返回 { scale, offset, dragging, transform, reset, zoom, zoomFromWheel, toggleZoom, startDrag, drag, endDrag };ImagePreviewDialog.vue 与 App.vue 的 quick-profile 查看器都用它。进一步让 quick-profile 查看器直接复用 <ImagePreviewDialog>,可连带删掉 App.vue:12218-12266 的 49 行模板与 17213-17316 的 104 行 CSS,合计净省约 195 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "素材选择器弹层模板被逐字复制两份(109 行),且已开始功能漂移",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:12009",
|
|||
|
|
"problem": "App.vue:12009-12117(quickProfile 模式)与 App.vue:12658-12769(workflow 模式)是同一个素材选择弹层。去掉缩进后 diff,97 行中只有 5 行不同:material.category → binding.category(3 处)、文案「取消这个素材参考」→「默认不注入这个参考」、workflow 侧多一条 <small v-if=\"reference.advanced\">进阶素材 · 高优先级</small>。分组 tab、TransitionGroup、「不使用」选项、缩略图渲染全部一致。",
|
|||
|
|
"evidence": "<!-- App.vue:12021-12042 -->\n<header class=\"quick-profile-material-picker-header\">\n <div>\n <span>{{ material.category }}参考</span>\n <strong>选择素材</strong>\n </div>\n <small>{{ activeQuickProfileMaterialPickerCountLabel }}</small>\n</header>\n<div\n v-if=\"material.category === QUICK_PROFILE_ACCESSORY_CATEGORY\"\n class=\"quick-profile-material-picker-groups\"\n aria-label=\"配饰分类\"\n>\n\n<!-- App.vue:12672-12693 —— 仅 material→binding 之差 -->\n<header class=\"quick-profile-material-picker-header\">\n <div>\n <span>{{ binding.category }}参考</span>\n <strong>选择素材</strong>\n </div>\n <small>{{ activeQuickProfileMaterialPickerCountLabel }}</small>\n</header>\n<div\n v-if=\"binding.category === QUICK_PROFILE_ACCESSORY_CATEGORY\"\n class=\"quick-profile-material-picker-groups\"\n aria-label=\"配饰分类\"\n>",
|
|||
|
|
"impact": "约 109 行模板重复。功能已漂移:reference.advanced 的「进阶素材 · 高优先级」标记只有 workflow 侧有,用户在 quickProfile 模式下看不到哪些是进阶素材。任何后续改动都要记得改两处。",
|
|||
|
|
"fix": "抽 standalone-web/src/features/materials/MaterialPickerPopover.vue,props 为 { category: string; placement?: 'up' | 'down'; extraClass?: string },emits 为 select / clear。两处调用点各缩到 8 行左右,净省约 100 行并消除漂移。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "4 份几乎一字不差的焦点陷阱实现,且行为不一致导致 Tab 逃逸",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/WorkflowCompareDialog.vue:34",
|
|||
|
|
"problem": "ConfirmDialog.vue:30,42-83、ImagePreviewDialog.vue:27,116-151、WorkflowCompareDialog.vue:20,34-78、WorkflowGenerationReviewDialog.vue:53,149-205 各自实现了 previousFocus / focusableElements() / trapFocus() / restoreFocus() / watch(() => props.open, ...) 这套约 45 行的样板。四份实现不一致:WorkflowCompareDialog 的可聚焦选择器缺少 input/select/textarea,弹窗内若有输入框 Tab 会跳出弹窗;ConfirmDialog 多了 !dialogRef.value?.contains(active) 兜底,另三份没有。",
|
|||
|
|
"evidence": "// WorkflowCompareDialog.vue:34-40 —— 选择器缺 input/select/textarea\nfunction focusableElements() {\n return Array.from(\n dialogRef.value?.querySelectorAll<HTMLElement>(\n 'button:not([disabled]), [href], [tabindex]:not([tabindex=\"-1\"])',\n ) ?? [],\n )\n}\n\n// ConfirmDialog.vue:42-48 —— 完整版\nfunction focusableElements() {\n return Array.from(\n dialogRef.value?.querySelectorAll<HTMLElement>(\n 'button:not([disabled]), [href], input:not([disabled]), select:not([disabled]), textarea:not([disabled]), [tabindex]:not([tabindex=\"-1\"])',\n ) ?? [],\n ).filter((element) => !element.hasAttribute(\"hidden\"))\n}",
|
|||
|
|
"impact": "约 180 行样板代码,抽成 composable 后剩约 50 行,净省 130 行。同时修复 WorkflowCompareDialog 的无障碍缺陷:对比弹窗里出现输入控件时键盘焦点会逃出模态框。",
|
|||
|
|
"fix": "新增 standalone-web/src/composables/useDialogFocusTrap.ts:export function useDialogFocusTrap(open: Ref<boolean>, opts: { dialogRef: Ref<HTMLElement|null>; initialFocus?: () => HTMLElement|null|undefined; onEscape?: () => void }) { ... return { onKeydown } },内部统一使用完整可聚焦选择器 + contains 兜底。项目已有 safe-backdrop-close.ts 这种单一职责工具模块的先例,放同级目录即可。四个弹窗组件各删 45 行、改为 3 行调用。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "SettingsPage.vue 里遗留一整段 1796 字符的 .backup-* 死样式",
|
|||
|
|
"location": "frontend/standalone-web/src/features/settings/SettingsPage.vue:165",
|
|||
|
|
"problem": "备份面板早已拆成独立的 features/settings/BackupPage.vue,且该组件用的是自己的 scoped 类名(.page / .hero / .cards / .full / .snapshots / .file / .list)。SettingsPage.vue 的 scoped 样式却仍保留旧类名全套定义(.backup-page / .backup-hero / .backup-grid / .file-picker / .import-preview / .backup-scope / .backup-notice),而 SettingsPage.vue 自己的模板里这些类零出现——因为是 scoped,它们绝无可能作用到 BackupPage.vue 上。SettingsPage.vue:167 的媒体查询里还引用了 .backup-grid 和 .backup-scope>div,同样死。",
|
|||
|
|
"evidence": "/* SettingsPage.vue:165(单物理行 1796 字符,此处仅节选) */\n.backup-page{display:grid;gap:15px}.backup-hero{display:flex;align-items:center;gap:13px;border:1px solid var(--border);border-radius:10px;padding:15px;background:linear-gradient(105deg,var(--primary-soft),var(--panel-muted))}.backup-hero>div:last-child{display:grid;gap:4px}.backup-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:13px}.file-picker{display:block!important}.import-preview{display:grid;gap:4px;border-radius:7px;padding:9px;background:var(--panel)}.backup-scope{...}.backup-notice{...}\n\n<!-- BackupPage.vue 模板实际用的类,与上面完全不相干 -->\n<div class=\"page\">\n <section class=\"hero\"><b>↕</b>...</section>\n <div class=\"cards\"><article>...</article></div>",
|
|||
|
|
"impact": "约 1.8 KB / 45 逻辑行死样式打进两份构建产物;后来者改备份页样式时可能误改这段完全不生效的规则。",
|
|||
|
|
"fix": "删除 SettingsPage.vue:165 整行,并从 :167 的媒体查询选择器列表中移除 .backup-grid 与 .backup-scope>div 两项。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "App.vue 全局样式里躺着 5 个已抽离组件的样式,691 行属于隐式全局依赖",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:14145",
|
|||
|
|
"problem": "静态分析发现 31 个类在 App.vue 全局 <style> 中定义但只被子组件模板使用。其中 AppNavigationRail.vue(整个文件 97 行,无 <style> 块)依赖 App.vue 里的 rail/brand/brand-mark/brand-copy/icon-nav/nav-scroll/nav-group/nav-group-label/nav-group-items/nav-glyph/nav-copy/nav-label/rail-pin/rail-pin-label 共 347 行;MaterialCategoryCreateDialog.vue(无 <style> 块)依赖 173 行;MaterialDetailDialog.vue(scoped 块只有 4 行、不含 .materials-detail-card*)依赖 171 行。另有 136 行是子组件已有等价 scoped 副本的纯重复(image-viewer-card/-actions/-stage 104 行 + state-card/empty-state-actions/workflow-chat-panel/workflow-composer-panel 32 行)。",
|
|||
|
|
"evidence": "<!-- components/AppNavigationRail.vue 全文 97 行,无 <style> 块 -->\n<aside :class=\"['rail', { pinned, visible }]\" ...>\n <div class=\"brand\"><span class=\"brand-mark\">IS</span><div class=\"brand-copy\">...</div></div>\n <nav class=\"icon-nav\">...\n\n/* 而 .rail / .brand / .brand-mark / .icon-nav 的 347 行定义全在 App.vue 的全局 <style> 里 */",
|
|||
|
|
"impact": "组件不自洽:AppNavigationRail 和 MaterialCategoryCreateDialog 离开 App.vue 就完全没有样式,无法单独测试或复用到 plugin-page 之外的场景;App.vue 样式块因此虚胖 691 行;另 136 行是纯冗余可直接删。",
|
|||
|
|
"fix": "把 347 / 173 / 171 三组规则分别剪切进 AppNavigationRail.vue、MaterialCategoryCreateDialog.vue、MaterialDetailDialog.vue 的 <style scoped>(需要父级配合定位的 .rail 类规则用 :deep() 或保留在 workbench.css);136 行重复的直接删。App.vue 样式块净减 827 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "全局重复 CSS 声明块:上限 2251 行冗余,占 CSS 总量 15%",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:15423",
|
|||
|
|
"problem": "扫描所有 .vue/.css 的声明块(≥2 条声明),完全相同的声明体按 (出现次数-1)×(声明数+2) 计,冗余上限 2251 行,占全前端 CSS 14698 行的 15%。高频样本:overflow:hidden;text-overflow:ellipsis;white-space:nowrap 出现 21 次;color:var(--text-soft);font-size:12px;font-weight:800 出现 21 次(其中 14 次在 App.vue 内且完全一致,分别位于 15423/15659/15810/16233/16358/16612/16712/16899/17238/19236/19427/19925/20725/20959);display:grid;gap:4px;min-width:0 出现 18 次;height:100%;object-fit:cover;width:100% 出现 14 次;主按钮三件套 background/border-color/color 出现 10 次。6 个组件各自重复定义 button.primary-button(GalleryPage.vue:473、AdvancedMaterialImportPage.vue:1791、ProfileDetailPane.vue:614、ProfilesPage.vue:236、ProfileWardrobePane.vue:501、WorkflowMaterialDrawer.vue:528)。workbench.css 整个 458 行文件就是一层事后覆盖层,用超长选择器列表去覆盖 App.vue 既有规则,其中 431-451 是同一个媒体查询写了两遍(第二遍只加 .navigation-shell 前缀提优先级)。",
|
|||
|
|
"evidence": "/* App.vue 内 14 处逐字相同的\"眼眉标签\"规则,各 5 行 */\n/* 15423 / 15659 / 15810 / 16233 / 16358 / 16612 / 16712 / 16899 / 17238 / 19236 / 19427 / 19925 / 20725 / 20959 */\n{\n color: var(--text-soft);\n font-size: 12px;\n font-weight: 800;\n}\n\n/* workbench.css:431-451 —— 同一媒体查询写两遍 */\n@media (max-width: 1360px) and (min-width: 901px) {\n .overview-strip { grid-template-columns: minmax(220px, 1.2fr) repeat(3, minmax(120px, 0.8fr)) minmax(200px, 0.9fr); gap: 10px; }\n .overview-strip > * { min-width: 0; }\n}\n\n@media (max-width: 1360px) and (min-width: 901px) {\n .navigation-shell .overview-strip { grid-template-columns: minmax(220px, 1.2fr) repeat(3, minmax(120px, 0.8fr)) minmax(200px, 0.9fr); gap: 10px; }\n .navigation-shell .overview-strip > * { min-width: 0; }\n}",
|
|||
|
|
"impact": "约 10-15% 的 CSS 是复读;主色/字号改一次要改 20 处。workbench.css 这层覆盖层的存在本身就是证据——因为基础规则改不动,只能再叠一层高优先级选择器去盖,导致每条规则实际有 2-3 份定义。",
|
|||
|
|
"fix": "1) 把 styles/tokens.css(当前仅 24 行)扩成真正的 token/utility 层:.u-truncate、.u-eyebrow、.u-stack-4、.u-cover-img、.u-btn-primary;2) App.vue 内那 14 处合并成一条选择器列表规则(70 行 → 20 行);3) 把 workbench.css 的覆盖规则回填进各自基础规则,删掉这一层(458 → 150 行以内);4) 立刻可删 workbench.css:442-451 的 10 行重复媒体查询。保守可回收 700-1100 行。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "25 个 loader 函数复读同一套 loading/error/try/catch/finally 样板,其中 loadModelOptions 缺 finally 会导致 loading 永久卡死",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:4673",
|
|||
|
|
"problem": "App.vue 声明了 10 个 loadingXxx ref(:869-878)、25 个 xxxError ref(:666-892 分散)、5 个 xxxBusy ref,25 个 loader 函数(4488/4511/4611/4624/4638/4673/4700/4787/4864/5010/6872/6894/7028/7115 等)各自手写同一套开关;全文件 82 处 try {、76 处 catch (err)。样板本身还不一致:loadModelOptions(:4673-4684)没有 try/finally,直接在函数末尾赋 loadingModelOptions.value = false —— 由于用的是 Promise.allSettled 目前不会抛,但 endpoints 取值或 errorMessage 抛错时 loading 会永久为 true。",
|
|||
|
|
"evidence": "// App.vue:4673-4684 —— 缺 try/finally\nasync function loadModelOptions() {\n loadingModelOptions.value = true\n modelOptionsError.value = \"\"\n const [capabilities, runtimeModels] = await Promise.allSettled([...])\n capabilityProfiles.value = capabilities.status === \"fulfilled\" ? capabilities.value : []\n runtimeModelOptions.value = runtimeModels.status === \"fulfilled\" ? runtimeModels.value.items ?? [] : []\n if (capabilities.status === \"rejected\") modelOptionsError.value = errorMessage(capabilities.reason)\n loadingModelOptions.value = false\n}\n\n// App.vue:4611-4622 —— 标准样板,与 4488/4624/4700/4787 结构完全一致\nasync function loadDailyCostUsage() {\n loadingDailyCostUsage.value = true\n dailyCostUsageError.value = \"\"\n try {\n dailyCostUsage.value = await getJson<CostUsageSummary>(endpoints.costUsage, { days: 1 })\n } catch (err) {\n dailyCostUsage.value = null\n dailyCostUsageError.value = `今日用量读取失败:${errorMessage(err)}`\n } finally {\n loadingDailyCostUsage.value = false\n }\n}",
|
|||
|
|
"impact": "约 125 行纯样板 + 40 个平行 ref。除行数外还有正确性成本:loadModelOptions 的 loading 泄漏路径无人覆盖,一旦触发模型选择区会永久显示骨架屏。",
|
|||
|
|
"fix": "抽 standalone-web/src/composables/useAsyncResource.ts:export function useAsyncResource<T>(loader: () => Promise<T>, opts?: { initial?: T; errorPrefix?: string; onError?: (e: unknown) => void }) { const data = ref(opts?.initial), loading = ref(false), error = ref(\"\"); async function reload() { loading.value = true; error.value = \"\"; try { data.value = await loader() } catch (e) { error.value = `${opts?.errorPrefix ?? \"\"}${errorMessage(e)}`; opts?.onError?.(e) } finally { loading.value = false } } return { data, loading, error, reload } }。25 个 loader 各缩到 1-3 行,40 个 ref 缩到 25 个解构,净省 90-120 行,并统一修掉 loading 泄漏。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "pinia 与 vue-router 声明为依赖但代码库零引用;构建工具全部错放在 dependencies,devDependencies 为空",
|
|||
|
|
"location": "frontend/package.json:15",
|
|||
|
|
"problem": "dependencies 里的 pinia ^2.3.0 与 vue-router ^4.5.0 在 shared/、standalone-web/、plugin-page/、tests/ 下 grep 均为 0 命中(node_modules 里占 1.1 MB:pinia 407 K + vue-router 691 K)。standalone-web/src/main.ts 只有 4 行,从未 createPinia()。同时 devDependencies 为空对象,typescript(23 MB)、vite(2.7 MB)、vue-tsc、@vitejs/plugin-vue 这些纯构建期工具全在 dependencies —— 而插件发布只带预构建产物(web/、pages/imagestudio/),运行期完全不需要 node_modules,严格说 7 个依赖全部应是 devDependencies。",
|
|||
|
|
"evidence": "\"dependencies\": {\n \"@vitejs/plugin-vue\": \"^5.2.0\",\n \"pinia\": \"^2.3.0\",\n \"typescript\": \"^5.7.0\",\n \"vite\": \"^6.0.0\",\n \"vue\": \"^3.5.0\",\n \"vue-router\": \"^4.5.0\",\n \"vue-tsc\": \"^2.2.0\"\n},\n\"devDependencies\": {}\n\n// standalone-web/src/main.ts 全文 4 行\nimport { createApp } from \"vue\"\nimport App from \"./App.vue\"\n\ncreateApp(App).mount(\"#app\")",
|
|||
|
|
"impact": "npm ci --omit=dev 得不到任何东西,无法只装运行期依赖;基于 dependencies 的 SCA/许可证扫描会把 23 MB 的 TypeScript 编译器算作运行期依赖;pinia/vue-router 的存在会误导人以为项目有路由和 store,实际导航是 App.vue 的 activeNav ref(:5102 setActiveNav)+ navigation-config.ts。",
|
|||
|
|
"fix": "删除 pinia 与 vue-router;把剩余 5 个包全部移到 devDependencies,dependencies 置空;重新生成 package-lock.json。若未来确需状态管理,建议走 composable 路线(见下一条)而非引入 pinia。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "App.vue 用 206 个 ref + 135 个 computed 的巨型局部状态树代替 store,overlay 状态机的 fallback 分支埋雷",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:9896",
|
|||
|
|
"problem": "App.vue 顶层有 206 个 ref、135 个 computed、264 个 function、101 个 async function,全文只有 1 条区块注释(:903)。跨模块状态(bootstrap/permissions、素材库、角色档案、工作流轮次、队列)全部平铺在同一 setup 作用域,子组件靠 props 逐层透传 + emit 回传。最典型症状是 overlay 状态机被写成两条 11-12 分支的 if 链(setOverlayOpen :9846-9894 共 49 行、isOverlayOpen :9896-9931 共 36 行),且 isOverlayOpen 的兜底是 return showLogPanel.value —— 新增一个 ControlledOverlay 而忘记加分支时不报错、静默返回日志面板的开关状态。",
|
|||
|
|
"evidence": "// App.vue:9896-9931\nfunction isOverlayOpen(key: ControlledOverlay) {\n if (key === \"search\") { return searchPanelOpen.value }\n if (key === \"palette\") { return showPalette.value }\n if (key === \"notifications\") { return showNotificationPanel.value }\n if (key === \"imageViewer\") { return showImageViewer.value }\n if (key === \"materialPicker\") { return Boolean(activeQuickProfileMaterialPickerCategory.value) }\n if (key === \"materialDetail\") { return showMaterialDetailCard.value }\n if (key === \"materialCategoryCreate\") { return showMaterialCategoryCreateCard.value }\n if (key === \"materialOperation\") { return Boolean(materialOperationMode.value) }\n if (key === \"personaQuestionnaire\") { return showPersonaQuestionnaire.value }\n if (key === \"quickProfileDetail\") { return showQuickProfileDetailCard.value }\n if (key === \"externalProfileImport\") { return showExternalProfileImport.value }\n return showLogPanel.value // ← 兜底返回 logs,漏加分支会静默错判\n}",
|
|||
|
|
"impact": "85 行可被 25 行的注册表替代;isOverlayOpen 的 fallback 使得新增 overlay 时漏加分支不会被类型系统或运行时发现,表现为「点击新弹层的遮罩不关闭」或「Esc 关错弹层」这类难查的 bug。整体上任何一个功能改动都要打开 22563 行的文件。",
|
|||
|
|
"fix": "1) 低成本立刻做:改注册表驱动 —— const overlayRegistry: Record<ControlledOverlay, { isOpen(): boolean; setOpen(v: boolean): void }> = { search: { isOpen: () => searchPanelOpen.value, setOpen: (v) => { searchFocused.value = v; searchPanelOpen.value = v } }, ... };const isOverlayOpen = (k) => overlayRegistry[k].isOpen()。Record 类型强制穷尽所有 ControlledOverlay,漏写编译期报错,省 40 行。2) 中成本:不引入 pinia,按 feature 抽 composable(useWorkflowWorkspace / useMaterialLibrary / useQuickProfile / useQueue),与现有 features/*/xxx-state.ts(如 advanced-material-state.ts)的组织方式一致,迁移成本低于引入 store。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "shared/types/workflow.ts 里 35 行类型是 workflow-material.ts 的过期副本",
|
|||
|
|
"location": "frontend/shared/types/workflow.ts:96",
|
|||
|
|
"problem": "WorkflowMaterialNavigationContext 在 workflow.ts:88-94 与 workflow-material.ts:1-7 里逐字段完全相同地定义了两遍。workflow.ts 的 WorkflowNodeMaterialBindingDetail(:108-123)与 WorkflowNodeMaterialBindingSaveRequest(:125-130)是 workflow-material.ts 的 WorkflowNodeMaterialBindingsResult / WorkflowNodeMaterialBindingsSaveRequest 的旧版本。全代码库检索显示 WorkflowNodeMaterialBindingDetail、WorkflowNodeMaterialBindingSaveRequest、WorkflowTemporaryReference 只在 workflow.ts 内部互相引用,无任何外部使用者(App.vue:129-132 导入的是 workflow-material.ts 那一版)。",
|
|||
|
|
"evidence": "// shared/types/workflow.ts:88-94\nexport interface WorkflowMaterialNavigationContext {\n workflow_id: string\n branch_name: string\n node_id: string\n round_id?: string | null\n return_view: string\n}\n\n// shared/types/workflow-material.ts:1-7 —— 一模一样\nexport interface WorkflowMaterialNavigationContext {\n workflow_id: string\n branch_name: string\n node_id: string\n round_id?: string | null\n return_view: string\n}",
|
|||
|
|
"impact": "35 行死类型。两份定义在后端字段变更时若只改一份,TypeScript 会给出互相矛盾的结论——一处报错一处放行。",
|
|||
|
|
"fix": "删除 workflow.ts:96-130(WorkflowTemporaryReference + WorkflowNodeMaterialBindingDetail + WorkflowNodeMaterialBindingSaveRequest);workflow.ts:88-94 改成 export type { WorkflowMaterialNavigationContext } from \"./workflow-material\"(workflow.ts:82 仍需引用它)。改后跑 npm run typecheck 验证。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "19 个导出符号在整个代码库无任何使用者",
|
|||
|
|
"location": "frontend/shared/types/integration.ts:1",
|
|||
|
|
"problem": "全量扫描 93 个 .ts/.vue/.mjs 源文件,19 个导出标识符在整个代码库中只出现 1 次(即声明自身):integration.ts 的 DailyOutfitResult(整个文件 7 行只有它)、client.ts 的 patchJson、advanced-material.ts 的 6 个(AdvancedMaterialImportListResult / AdvancedMaterialValidationAttemptListResult / AdvancedMaterialListResult / MaterialCatalogResult / AdvancedMaterialValidationAttemptRequest / AdvancedStyleConsensusRequest)、logs.ts 的 EventStreamResult 与 LogBundle、storage.ts 的 StorageStatus 与 StorageMigrationPlan、ui.ts 的 UiAction、wardrobe.ts 的 WardrobeMatchResult、workflow-round.ts 的 WorkflowRoundNextRequest 与 WorkflowRoundConfirmRequest、features/workflow/types.ts 的 WorkflowRoundKnownStatus。其中 EventStreamResult 已定义却被调用方手写内联类型绕过。",
|
|||
|
|
"evidence": "// shared/types/integration.ts 全文\nexport interface DailyOutfitResult {\n matched: boolean\n selected_materials: Array<Record<string, unknown>>\n final_prompt: string\n job_id?: string | null\n fallback_used: boolean\n}\n\n// App.vue:4566-4568 —— 手写内联类型,没用已有的 EventStreamResult\nconst stream = await getJson<{\n events: Array<{ event_id?: string; event_type?: string; data?: Record<string, unknown> }>\n}>(endpoints.eventStream, { limit: 20 })",
|
|||
|
|
"impact": "约 45 行死类型。EventStreamResult 这种「定义了但调用方手写内联」的情况更糟:后端事件流字段变更时类型完全失去保护作用。",
|
|||
|
|
"fix": "删除 integration.ts 整个文件与其余 18 个未使用导出;例外是 EventStreamResult ——把 App.vue:4566 改成 getJson<EventStreamResult>(endpoints.eventStream, { limit: 20 }) 而不是删类型。删除前跑 npm run typecheck(package.json:12 已有该脚本)确认。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "errorMessage 在两处独立实现且行为不一致,normalizeStatus 也被重复实现",
|
|||
|
|
"location": "frontend/standalone-web/src/features/materials/AdvancedMaterialImportPage.vue:1354",
|
|||
|
|
"problem": "App.vue:11354 的 errorMessage 委托给 readableApiError(会识别 ApiClientError、区分预览模式、给中文可读文案),AdvancedMaterialImportPage.vue:1354 却自己实现了一个只取 err.message 的版本。同一个 ApiClientError 在主界面显示「后端未连接,当前为预览模式」之类可读提示,在进阶素材页只显示原始 message。此外 AdvancedMaterialImportPage.vue:1350 独立重实现了 normalizeStatus,而 shared/utils/presentation.ts:7 已有同名导出。",
|
|||
|
|
"evidence": "// App.vue:11354-11356\nfunction errorMessage(err: unknown) {\n return readableApiError(err)\n}\n\n// AdvancedMaterialImportPage.vue:1354-1356\nfunction errorMessage(err: unknown) {\n return err instanceof Error ? err.message || \"请求失败\" : String(err || \"请求失败\")\n}\n\n// AdvancedMaterialImportPage.vue:1350-1352 —— shared/utils/presentation.ts:7 已有\nfunction normalizeStatus(value: string) {\n return String(value || \"\").trim().toLowerCase()\n}",
|
|||
|
|
"impact": "行数很小(约 8 行),但错误提示体验不一致:用户在进阶素材页遇到后端断连时看到的是晦涩的原始错误而非可操作提示。属于「共享工具没被共享」的信号。",
|
|||
|
|
"fix": "把 readableApiError 与 backendWriteBlockMessage(App.vue:4356-4387)下沉到 shared/utils/api-error.ts,App.vue 与 AdvancedMaterialImportPage.vue 都从那里导入;AdvancedMaterialImportPage.vue:1350 的 normalizeStatus 改为 import { normalizeStatus } from \"../../../../shared/utils/presentation\"。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "tsconfig.json 的 @shared/* 路径别名无人使用,且 Vite 侧没有对应 alias(类型检查过、构建会炸)",
|
|||
|
|
"location": "frontend/tsconfig.json:15",
|
|||
|
|
"problem": "tsconfig.json 配了 baseUrl \".\" 与 paths { \"@shared/*\": [\"shared/*\"] },但全代码库检索 @shared/ 为 0 命中——所有引用走相对路径,App.vue 是 \"../../shared/api/endpoints\",BackupPage.vue:3 甚至写到 \"../../../../shared/api/client\"。而两个 vite.config.ts(plugin-page 与 standalone-web)都没有配 resolve.alias,即便有人开始用 @shared/,vue-tsc --noEmit 能过、vite build 会失败。",
|
|||
|
|
"evidence": "// tsconfig.json:14-18\n\"baseUrl\": \".\",\n\"paths\": {\n \"@shared/*\": [\"shared/*\"]\n}\n\n// features/settings/BackupPage.vue:3-5 —— 实际写法\nimport { getJson, postJson } from \"../../../../shared/api/client\"\nimport { endpoints } from \"../../../../shared/api/endpoints\"\nimport type { DiffSettingsResult, ... } from \"../../../../shared/types/settings\"",
|
|||
|
|
"impact": "配置死代码 + 潜在的「类型检查通过但构建失败」陷阱;../../../../ 这种路径也让文件移动变得昂贵,是阻碍 App.vue 拆分的实际摩擦之一。",
|
|||
|
|
"fix": "推荐补齐而非删除:在 plugin-page/vite.config.ts 与 standalone-web/vite.config.ts 各加 resolve: { alias: { \"@shared\": resolve(rootDir, \"../../shared\") } },然后把全部 ../../shared 与 ../../../../shared 批量替换为 @shared(约 60 处 import)。这也为后续按 feature 拆分 App.vue 扫清路径障碍。若不想动构建配置,则删掉 tsconfig.json 的 baseUrl + paths 两项。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "同一文件内混用压缩风格与展开风格代码,使行数统计失真且无法 diff 审查",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:4549",
|
|||
|
|
"problem": "App.vue 绝大部分是 Prettier 风格的展开代码,但 4549-4559 的 expandImageGenerationPrompt 是手工压行;MaterialDetailDialog.vue:34-58、BackupPage.vue:14-25 整个 script 都是压行;ReviewPolicyPanel.vue:14 的整个 <style scoped> 是一行 1000+ 字符;SettingsPage.vue:160-167 的样式也是每行 1000-1800 字符。",
|
|||
|
|
"evidence": "// App.vue:4549-4559\nasync function expandImageGenerationPrompt() {\n const prompt=imageGenerationPrompt.value.trim()\n if(!prompt||promptExpansionBusy.value)return\n promptExpansionBusy.value=true\n try{\n const result=await postJson<{selected_prompt?:string;translation?:{translated_text?:string;status?:string;error?:string}}>(endpoints.promptGovernancePreview,{prompt,entrypoint:\"web\",translation_consent:true,selection:\"translated\",model_route:selectedImageModelRoute()})\n const expanded=String(result.selected_prompt||result.translation?.translated_text||\"\").trim()\n if(!expanded||result.translation?.status===\"unavailable_fallback\"||result.translation?.status===\"failed_fallback\")throw new Error(result.translation?.error||\"未找到可用的提示词扩写模型\")\n imageGenerationPrompt.value=expanded\n }catch(err){imageGenerationError.value=`提示词扩写失败:${errorMessage(err)}`}finally{promptExpansionBusy.value=false}\n}",
|
|||
|
|
"impact": "行数统计失真(压行掩盖了真实复杂度,也让本报告里的「削减行数」低估);git diff 一行全红/全绿完全无法 code review;无法用 lint 自动格式化或做 CSS 重复检测。",
|
|||
|
|
"fix": "引入 Prettier(printWidth: 110,与现有多数代码一致),加 npm run format:check 到 CI。展开后 CSS/TS 总行数会上升约 800 行,但这是把隐藏复杂度显性化——后续所有重复检测与删除操作都会更可靠。建议在执行本报告其他修复项之前先做这一步。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "测试质量 + 仓库卫生 + 配置",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\C3-测试与仓库卫生.md",
|
|||
|
|
"summary": "测试层面:263 个后端用例可全量收集通过,核心工作流覆盖扎实,但 166 个 endpoint 只有 33 个(19.9%)被测试触及,34 个路由模块有 24 个零覆盖;68 个 service 有 23 个(4579 行)零测试,含 integration_token_service、chat_command_guard_service、deletion_safety_service 三个安全模块。8 个前端\"测试\"中 7 个只是对 .vue/.css 源文本做正则匹配,35638 行 Vue 代码实质零行为测试。仓库无 conftest.py 与 pytest 配置,导致 `pytest tests/` 直接 ModuleNotFoundError 崩溃收集,21 个测试文件各自重复内联同一套 SQLite 夹具。仓库卫生方面:工作树有 8228 个未跟踪垃圾文件 / 2.57 GB(.tmp 2129 MB + 20 个 .tmp_pytest_* 507 MB),.gitignore 挡住了它们但没挡住 release/(20 个 zip 37 MB)与 web//pages/ 构建产物(38 个哈希文件),230 次提交里 86 次(37%)在改产物,历史累计 439 个死 blob,.git 已膨胀到 51 MB。最致命的是 .env.imagestudio_codex_handoff.yaml 含 2 个真实 API key、PG 明文密码、AstrBot 管理员 JWT 与明文密码,虽未入库但已被打进 19 个部署包并 scp 到 NAS。配置方面:config.yaml/metadata.yaml/CHANGELOG.md 存在 GBK-UTF8 编码损坏(私有区字符 54/1/649 个),_conf_schema.json 完全缺失 prompt_governance(NSFW 分流、内容审核、日志脱敏)与 chat_access.adult_*_allowlist,导致内容安全开关在 AstrBot 面板不可配置且升级即丢失。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\.gitignore",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\.env.imagestudio_codex_handoff.yaml",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\config.yaml",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\_conf_schema.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\metadata.yaml",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\requirements.txt",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\DEVELOPMENT.md",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\config_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\package.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_profile_delete_route.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_generation_task_display.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_advanced_style_consensus_route.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_resource_access_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_material_vision_input.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_image_provider_timeout.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_worker_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_advanced_style_set_workflow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tests\\ui-theme-contract.test.mjs",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tests\\workflow-model-route.test.mjs",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tests\\workflow-cross-review-wiring.test.mjs",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tests\\production-data-boundary.test.mjs",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\image_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\api_discovery_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\settings_service.py"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P0-致命",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "真实 API Key / DB 密码 / 管理员 JWT 被打进 19 个部署包并推送到 NAS",
|
|||
|
|
"location": ".env.imagestudio_codex_handoff.yaml:26",
|
|||
|
|
"problem": "该文件保存了可直接使用的凭据:2 个 OpenAI 兼容站点的明文 API key(weilai.chat / botcf.com)、PostgreSQL 主机端口用户与明文密码(<PG_HOST:内网IP:已脱敏>:25008 / <PG_USER:已脱敏>)、AstrBot 管理员 HS256 JWT 与明文管理员账号密码。.gitignore 的 .env* 规则确实生效(git ls-files 与 git log --all --diff-filter=A 均确认从未入库),但打包脚本用的是整目录复制而不是 git archive,于是把它一起带进了产物:8 个 deploy 暂存目录(deploy-1.0.35、deploy-clean-1.0.35/36/37、deploy-stage-1.0.38~41)各含一份 3506 字节的副本;tar -tf 逐个验证确认 19 个归档内含该文件。文件末尾还写着 rules.redact_secrets: true。",
|
|||
|
|
"evidence": "$ git ls-files | grep -i env\n(空)\n\n$ tar -tf .tmp/astrbot_plugin_imagestudio-1.0.35.tar.gz | grep env\nastrbot_plugin_imagestudio/.env.imagestudio_codex_handoff.yaml\n\n$ tar -tf .tmp/astrbot_plugin_imagestudio-1.0.41-20260720-101338.tar.gz | grep env\n./.env.imagestudio_codex_handoff.yaml\n\n# 19 个泄漏归档:1.0.35/36/37、1.0.38~1.0.41-*、1.0.64/65/67-*/68-*、\n# imagestudio-1.0.66-deploy-*、imagestudio-material-*(5)、nas-deploy-worktree-20260724(.clean).tar\n\n# 文件内容(类型,不复述完整值)\nmodel_sites.site_a.api_key_plaintext: \"sk-3af1f9dd7ffd25ff...(64 hex)\"\nmodel_sites.site_b.api_key_plaintext: \"sk-B6KqCtlccQ2MU5O3z2X2...(48)\"\npostgresql.password: 明文\nastrbot.admin_token_plaintext: \"eyJhbGciOiJIUzI1NiI...\" (HS256 JWT, exp=1783872299)\nastrbot.admin_password_plaintext: 明文",
|
|||
|
|
"impact": "任何拿到部署包的人(NAS 上其他用户、备份、误传附件)可直接盗刷两个模型站点额度、连上 <PG_HOST:内网IP:已脱敏>:25008 的 PostgreSQL 读写业务库、用 JWT 与管理员密码接管 AstrBot 实例(该实例托管本插件全部 166 个 API)。JWT 约 2026-07-12 过期但 admin_password_plaintext 永不过期,可随时重签。deploy_remote.sh / upload_via_ssh.py 表明这些包已上传 NAS。",
|
|||
|
|
"fix": "1) 立刻轮换全部 5 类凭据(两个 API key 作废重发、PG 改密、AstrBot 改密并更换 JWT secret);2) 删除全部含泄漏的归档与 staging 目录并到 NAS 删除同名包;3) 打包改用 `git archive --format=tar.gz --prefix=astrbot_plugin_imagestudio/ -o dist/plugin-$(git describe --always).tar.gz HEAD`,并在脚本末尾加断言 `tar -tf \"$OUT\" | grep -qE '(^|/)\\.env' && { echo \"SECRET LEAK\"; exit 1; }`;4) 凭据改为环境变量注入(_conf_schema.json 已有 *_api_key_env_var 字段,设计上本就支持),文件里只留 env var 名。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "pytest tests/ 直接崩溃:无 conftest.py、无 pytest 配置,测试只能用 python -m pytest 跑",
|
|||
|
|
"location": "仓库根(缺 conftest.py / pytest.ini / pyproject.toml)",
|
|||
|
|
"problem": "45 个测试文件全部使用以仓库根为起点的绝对导入(from backend... / from adapters...),但仓库不存在任何 conftest.py(find . -name conftest.py 全仓库 0 命中),也没有 pytest.ini / pyproject.toml / setup.cfg / tox.ini。只有 python -m pytest(-m 会把 CWD 插到 sys.path[0])能工作,裸 pytest 在收集阶段就 ModuleNotFoundError。",
|
|||
|
|
"evidence": "$ python -m pytest tests/test_profile_delete_route.py -q\n1 passed in 0.14s\n\n$ pytest tests/test_profile_delete_route.py -q\ntests\\test_profile_delete_route.py:3: in <module>\n from adapters.astrbot_adapter import resolve_route\nE ModuleNotFoundError: No module named 'adapters'\n=========================== short test summary ===========================\nERROR tests/test_profile_delete_route.py\n!!!!!!!!!!!!! Interrupted: 1 error during collection !!!!!!!!!!!!!\n1 error in 0.46s",
|
|||
|
|
"impact": "任何 CI 模板、IDE 测试运行器(PyCharm/VSCode 默认调 pytest)、tox、pre-commit 都会在收集阶段全军覆没,报错还指向不明。新贡献者按直觉敲 pytest 会认为仓库测试是坏的。同时缺配置也意味着没有 --strict-markers、无超时保护、无覆盖率门槛。",
|
|||
|
|
"fix": "新建 pyproject.toml:[tool.pytest.ini_options] pythonpath = [\".\"] / testpaths = [\"tests\"] / addopts = \"-q --strict-markers\"。pythonpath 需 pytest>=7;若要兼容更旧版本再加仓库根 conftest.py:sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "166 个 endpoint 只有 33 个被测试触及,24/31 路由模块零覆盖(含鉴权与文件服务)",
|
|||
|
|
"location": "backend/route_registry.py:1",
|
|||
|
|
"problem": "用 RouteSpec.endpoint 字面量与 handler 模块名对 45 个测试文件做全文匹配,发现 166 个 endpoint 中只有 33 个(19.9%)出现在测试里,31 个路由 handler 模块只有 7 个被引用。未被任何测试触及的模块包含 storage_routes(签名 URL / 文件下发,直接决定“前端不能看到本地绝对路径”这条安全约束)、integration_routes(外部插件集成入口)、settings_routes(写配置)、data_backup_routes(导出全量数据)、media_routes、round_routes。",
|
|||
|
|
"evidence": "$ python -c \"from backend.route_registry import ROUTES; ...\"\ntotal endpoints: 166 referenced in tests: 33 (19.9%)\nroute modules: 31 referenced: 7 (22.6%)\n\nuntested modules:\n api_discovery_routes asset_routes branch_routes contract_routes\n cost_routes dashboard_routes data_backup_routes event_routes\n history_routes image_routes integration_routes job_routes\n log_routes material_routes media_routes model_runtime_routes\n prompt_routes round_routes settings_routes storage_routes\n ui_routes wardrobe_routes web_routes workflow_routes",
|
|||
|
|
"impact": "高风险模块(文件下发、外部集成鉴权、写配置、全量导出)任何一次重构后的权限回归都不会被测试拦住。round_routes 虽然 service 层有 1055 行测试,但路由层的参数校验与权限装饰完全裸奔。",
|
|||
|
|
"fix": "1) 加注册表级契约测试,一次覆盖 166 条:遍历 ROUTES 断言每条 handler 可 import、可调用、permission 属于合法枚举,并调用 assert_unique_routes();2) 对 storage_routes / integration_routes / settings_routes / data_backup_routes 补越权用例,复用 tests/test_resource_access_service.py:96 已有的 owner-chain 参数化模式。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "23 个 service(4579 行)零测试,其中 3 个是安全边界模块,另有 2 个“同名测试不测同名 service”",
|
|||
|
|
"location": "backend/services/integration_token_service.py:1",
|
|||
|
|
"problem": "以下 service 名在整个 tests/ 目录全文中一次都没出现过:integration_token_service(202行,外部插件token鉴权)、chat_command_guard_service(101行,聊天访问控制)、deletion_safety_service(75行,删除安全检查)、material_binding_schema_service(408)、asset_import_service(406)、prompt_plan_schema_service(379)、model_provider_bootstrap(375)、node_injection_template_service(366)、branch_lineage_service(356)、api_discovery_service(351)、history_service(314)、log_service(300)、workflow_progression_service(240)、integration_service(171)、dashboard_service(117)、chat_session_service(86)、web_hosting_service(84)、reference_binding_mapper(69)、event_stream_service(50)、chat_media_import_service(47)、ui_service(37)、contract_service(30)、media_index_service(15),合计 4579 行 / 23 个模块(占 68 个 service 的 33.8%)。其中 branch_lineage_service 与 workflow_progression_service 存在同名测试文件(test_workflow_branch_lineage.py 332 行、test_character_workflow_progression.py 248 行)却测的是别的模块,制造“看起来有测试”的假象。",
|
|||
|
|
"evidence": "$ python -c \"svc={p.stem for p in Path('backend/services').glob('*.py')}; txt=all tests; print(sorted(s for s in svc if s not in txt))\"\ntotal services: 68\nnever mentioned in tests: 23\n - api_discovery_service - asset_import_service\n - branch_lineage_service - chat_command_guard_service\n - chat_media_import_service - chat_session_service\n - contract_service - dashboard_service\n - deletion_safety_service - event_stream_service\n - history_service - integration_service\n - integration_token_service - log_service\n - material_binding_schema_service - media_index_service\n - model_provider_bootstrap - node_injection_template_service\n - prompt_plan_schema_service - reference_binding_mapper\n - ui_service - web_hosting_service\n - workflow_progression_service",
|
|||
|
|
"impact": "integration_token_service 零测试 = 外部插件鉴权逻辑无任何回归保护;chat_command_guard_service 零测试 = chat_access 白名单/限流的绕过不会被发现;deletion_safety_service 零测试 = 误删保护失效时静默。同名测试的假象比明摆着没测试更危险。",
|
|||
|
|
"fix": "按风险补三个模块:1) integration_token_service:签发→校验→过期→篡改签名→重放,5 个用例;2) chat_command_guard_service:白名单命中/未命中、成人白名单、并发上限、日限额,4 个用例(复用 tests/test_chat_workflow_round_commands.py 的 fixture);3) deletion_safety_service:有引用拒删、级联引用、强制删除,3 个用例。同时把 test_workflow_branch_lineage.py / test_character_workflow_progression.py 与真实 service 接上,或改名以免误导。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "前端 8 个测试里 7 个是“源码正则匹配”,35638 行 Vue 零行为测试",
|
|||
|
|
"location": "frontend/tests/production-data-boundary.test.mjs:33",
|
|||
|
|
"problem": "grep -l readFile frontend/tests/*.mjs 命中 7/8 个文件。这些测试把 .vue / .css 当纯文本读进来做正则断言,从不实例化组件;唯一真正 import 被测模块的是 workflow-model-route.test.mjs(两个纯函数模块)。frontend/package.json 也证实没有任何组件测试运行器——test:contracts 就是 node --test tests/*.test.mjs,依赖树里没有 vitest / @vue/test-utils / jsdom。仓库有 36 个 .vue 共 35638 行、3616 行 .ts。",
|
|||
|
|
"evidence": "// frontend/tests/production-data-boundary.test.mjs:33\nassert.match(app, /const profilePageItems = computed<CharacterProfile\\[\\]>\\(\\(\\) => backendProfiles\\.value\\)/)\nassert.match(app, /const imageGenerationImages[\\s\\S]*localGenerationImages\\.value\\.filter\\(isAuthenticGenerationImage\\)/)\n\n// frontend/tests/workflow-cross-review-wiring.test.mjs:19\nassert.equal((app.match(/<WorkflowGenerationReviewDialog\\b/g) ?? []).length, 1)\nassert.match(app, /loadWorkflowRounds\\(result\\.round\\.node_id, true\\)/)\nassert.match(dialog, /model_routes: \\[\\.\\.\\.reviewModelRoutes\\.value\\]/)\n\n$ node --test tests/*.test.mjs\nℹ tests 26 ℹ pass 26 ℹ duration_ms 356.7 ← 只做字符串匹配所以极快",
|
|||
|
|
"impact": "假阴性:把 computed<CharacterProfile[]> 换成等价写法或格式化工具重排空格,整批测试崩但功能正常。假阳性:组件运行时抛异常、props 契约破裂、事件不触发,这些测试全部照过。35638 行 Vue 实际覆盖率为 0。",
|
|||
|
|
"fix": "1) 引入 vitest + @vue/test-utils + jsdom(放 devDependencies),把 workflow-cross-review-wiring / advanced-material-upload / production-data-boundary 改写成挂载组件、mock endpoints、断言实际请求与渲染结果;2) 把纯逻辑(isAuthenticGenerationImage、候选池筛选、时间戳归一化)从 App.vue 抽到 .ts 模块,按 workflow-model-route.test.mjs 的模式做真单测;3) ui-theme-contract.test.mjs 这类 CSS token 文本契约改名挪到 scripts/lint-tokens.mjs 接进 npm run lint,避免虚增测试数字。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "release/ 20 个 zip(37 MB)与前端构建产物已入库,.git 膨胀到 51 MB、37% 的提交在改产物",
|
|||
|
|
"location": ".gitignore:1",
|
|||
|
|
"problem": ".gitignore 挡住了 .tmp/ 与 .tmp_pytest*/,但没有 release/、web/assets/、pages/imagestudio/assets/ 规则。git ls-files 总数 455,其中 release/*.zip 20 个共 37.02 MB(单个最大 4.23 MB)、web/ 19 个、pages/imagestudio/ 19 个,二进制与构建产物占已跟踪文件 12.7%。Vite 的 content-hash 让每次 build 都生成全新文件名,历史只增不减。早期 release zip 里还装着运行期真实数据,与 DEVELOPMENT.md 自己写的“不能把真实运行数据加入发布包”直接冲突。",
|
|||
|
|
"evidence": "$ du -sh .git 51M\n$ git rev-list --all --count 230\n$ git log --oneline --all -- web pages | wc -l 86 (37.4%)\n$ git log --all --name-only --pretty=format: -- web pages | sort -u | grep -c .\n439 ← 当前只剩 38 个,约 401 个死 blob\n\n$ git ls-files release | wc -l 20\n$ python -c \"zipfile 检查 release/astrbot_plugin_imagestudio-1.0.5.zip\"\ndata/temp/provider_probe_B_image2_generation.json 1.76 MB (真实模型响应)\ndata/temp/provider_probe_B_image2_generation.png 1.32 MB\ndata/temp/s81_provider_runtime_b.png 0.76 MB\ndata/db/imagestudio.db 0.29 MB (真实 SQLite 库)",
|
|||
|
|
"impact": "每个 clone 强制拉 51 MB,其中约 45 MB 是没人会用的历史产物;37% 的提交 diff 被机器生成的 minified JS 淹没,git log -p / git bisect / code review 全部失效;早期 release zip 里的 imagestudio.db 与 provider probe 响应是真实数据,永久留在版本库历史中无法用 git rm 撤销。",
|
|||
|
|
"fix": ".gitignore 补:release/、web/assets/、pages/imagestudio/assets/(保留两个 index.html,AstrBot 靠它发现插件页面)。然后 `git rm -r --cached release web/assets pages/imagestudio/assets && git commit -m \"chore: stop tracking release archives and frontend build output\"`。要真正瘦身 .git 需 `git filter-repo --path release --path web/assets --path pages/imagestudio/assets --invert-paths`(会让所有 clone 失效,需团队同步);不改写历史时至少跑 `git gc --aggressive --prune=now`。同时删掉 DEVELOPMENT.md 里“生产构建必须同时提交 pages/imagestudio/ 产物”那句,否则与 .gitignore 打架。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "_conf_schema.json 缺失 prompt_governance 与成人内容白名单,AstrBot 面板无法配置内容安全开关",
|
|||
|
|
"location": "_conf_schema.json:1",
|
|||
|
|
"problem": "main.py:110-175 的 _CONFIG_KEY_MAP 逐条把 config.yaml 路径映射到 schema 键,schema 里没有的键在 AstrBot 面板中就不存在、也无法持久化。实测缺口:整个 prompt_governance 段(translation 6 项、moderation.llm_review_enabled/failure_policy、routing.safe_route/nsfw_route/adult_auto_route_enabled/entrypoints、image_moderation、retention 4 项含 redact_sensitive_logs)全无 schema 键;chat_access.adult_group_allowlist 与 adult_user_allowlist 缺失(普通 group/user_allowlist 有);security.integration_token_set / allowed_integration_plugins 缺失;features 只有 private_companion_integration;storage 只有 advanced_materials_path。而这些键在业务代码里是真的被读的。",
|
|||
|
|
"evidence": "# backend/services/chat_workflow_service.py:1175\nadult_users = _id_set(access.get(\"adult_user_allowlist\"))\nadult_groups = _id_set(access.get(\"adult_group_allowlist\"))\n\n# backend/services/advanced_material_service.py:1309\nadult_users = {str(value).strip() for value in access.get(\"adult_user_allowlist\", []) if str(value).strip()}\n\n# backend/services/settings_service.py:184-185 插件自己的设置页渲染了它们\n_field(\"adult_group_allowlist\", \"list\", label=\"成人内容群聊白名单\"),\n_field(\"adult_user_allowlist\", \"list\", label=\"成人内容用户白名单\"),\n\n# 但 main.py:156 附近的映射表里只有\n\"chat_access.group_allowlist\": \"chat_group_allowlist\",\n\"chat_access.user_allowlist\": \"chat_user_allowlist\",\n# 没有任何 prompt_governance.* 或 adult_*_allowlist 条目",
|
|||
|
|
"impact": "在 AstrBot 真实部署下管理员只能手改插件目录里的 config.yaml 来打开 NSFW 分流、内容审核、成人白名单和日志脱敏;而插件更新会覆盖 config.yaml,配置静默回退到默认值(adult_auto_route_enabled: false、llm_review_enabled: false、adult_*_allowlist: [])。即内容安全开关在生产环境实际不可用、且升级即丢失。",
|
|||
|
|
"fix": "1) 在 _conf_schema.json 补齐 prompt_governance.moderation.* / prompt_governance.routing.* / chat_access.adult_*_allowlist / prompt_governance.retention.redact_sensitive_logs 等键,并在 main.py:_CONFIG_KEY_MAP 加对应映射;2) 加防漂移测试:断言 _CONFIG_KEY_MAP.values() 与 _conf_schema.json 的键集双向相等,再加一条断言 config.yaml 中所有安全相关叶子路径都出现在映射表 key 里;3) postgresql_password 在 schema 里是 type:string,面板会明文回显,若 AstrBot 支持 password/secret 类型应改用,否则在 description 里引导填 dsn 并把密码放环境变量。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "config.yaml / metadata.yaml / CHANGELOG.md 存在 GBK-UTF8 编码损坏(私有区乱码)",
|
|||
|
|
"location": "config.yaml:15",
|
|||
|
|
"problem": "三个文件本身是合法 UTF-8,但内容是中文被 GBK 误解码后再存成 UTF-8 的产物,含 Unicode 私有使用区字符(U+E000–U+F8FF),已不可逆还原。全仓库扫描(跳过 web/ pages/ release/ frontend/)显示 config.yaml 54 个、CHANGELOG.md 649 个、metadata.yaml 1 个(正是 AstrBot 插件市场直接展示的 short_desc)。同一个 config.yaml 里新写的注释是正常的(config.yaml:57「留空时沿用进阶素材识别路由」可读),说明损坏是某次 Windows 工具链(PowerShell 重定向或 Set-Content 默认 ANSI)一次性造成的。",
|
|||
|
|
"evidence": "$ python -c \"print(repr(open('config.yaml',encoding='utf-8').read().splitlines()[14]))\"\n' # default_model 鏄?ImageStudio 璺敱鍚嶏紱route_name 鐣欑┖鏃跺鐢ㄥ畠銆?'\n# 原文应为:default_model 是 ImageStudio 路由名;route_name 留空时也用它。\n\n# metadata.yaml:5\nshort_desc: AstrBot 鍐呯殑瑙掕壊璧勪骇銆佸伐浣滄祦涓庡浘鐗囩敓鎴愮鐞嗗伐浣滃彴銆?\n\n# 扫描结果(PUA = 私有区字符数)\nconfig.yaml PUA= 54\nCHANGELOG.md PUA= 649\nmetadata.yaml PUA= 1\n受影响的已跟踪文件:3",
|
|||
|
|
"impact": "AstrBot 插件市场/管理面板会展示乱码描述;config.yaml 里 54 处注释是解释各字段语义的唯一文档(vision 段那几行解释了“Vision 模型必须单独填写,不能假定生图模型具备视觉理解能力”这种关键约束),现在完全不可读,运维只能靠猜;后续任何人用编辑器打开再保存会让损坏继续扩散。",
|
|||
|
|
"fix": "1) 手工重写这三个文件的中文内容(config.yaml 54 处、metadata.yaml 1 处优先,CHANGELOG.md 可从 git 历史或 Release notes 恢复);2) 在 DEVELOPMENT.md 的 Windows 章节明确写文件一律 Out-File -Encoding utf8 / Set-Content -Encoding utf8,禁止裸 > 重定向;3) 加测试守卫:遍历 config.yaml/metadata.yaml/_conf_schema.json/README.md,断言不含 U+E000–U+F8FF 区间字符。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "工作树 8228 个未跟踪垃圾文件 / 2.57 GB,并导致 546 MB 的部署包被推到 NAS",
|
|||
|
|
"location": ".tmp/deploy-1.0.35",
|
|||
|
|
"problem": ".gitignore 的 .tmp/ 与 .tmp_pytest*/ 规则是生效的(git ls-files 匹配 0 条),所以不是入库问题而是磁盘与工作流污染。.tmp/ 6303 文件 2129.08 MB,20 个根级 .tmp_pytest_* 合计 1925 文件 507.25 MB,总计 8228 文件 2636 MB。.tmp/ 内含 8 个整仓库快照、22 个部署归档、约 60 个一次性 patch/fix 脚本(fix_soft_asserts.py、patch_soft_anatomy_tests.py、patch_tests_age2.py 等)、40+ 个 pytest 输出目录,以及真实生成图片和一个真实 imagestudio.db。关键错误是 deploy-1.0.35 把根目录的 .tmp_pytest_* 一起复制了进去(2180 文件 496 MB)。",
|
|||
|
|
"evidence": "$ Get-ChildItem .tmp/deploy-1.0.35 | Select -First 8\n.tmp_pytest_advanced .tmp_pytest_advanced2 .tmp_pytest_config\n.tmp_pytest_full .tmp_pytest_full_1022 .tmp_pytest_full_1022_final\n.tmp_pytest_full_1030 .tmp_pytest_full_1030_x\n\n$ Get-ChildItem .tmp -File | Sort Length -Desc | Select -First 2\nnas-deploy-worktree-20260724.tar 546.20 MB\nastrbot_plugin_imagestudio-1.0.62.zip.b64 53.92 MB\n\n$ tar -tf nas-deploy-worktree-20260724.tar | awk -F/ '{print $2}' | sort | uniq -c | sort -rn | head\n 601 .tmp_pytest_full_1031 601 .tmp_pytest_full_1030_x 601 .tmp_pytest_full_1030\n 599 .tmp_pytest_full_1022_final 599 .tmp_pytest_full_1022 583 .tmp_pytest_full\n 365 backend 305 .tmp_pytest_route_refs 176 tests",
|
|||
|
|
"impact": "2.6 GB 拖慢 IDE 索引、ripgrep、防病毒扫描与备份;更实际的伤害是打包脚本整目录复制时把它们卷进部署包——546 MB 的 tar(正常插件包约 8 MB,膨胀 68 倍)已经把 3000+ 个 pytest 临时文件 scp 到了 NAS;60 多个 fix_*.py / patch_*.py 表明大量一次性修补发生在版本控制之外,无法追溯。",
|
|||
|
|
"fix": "1) 清理:Get-ChildItem . -Directory -Force -Filter '.tmp_pytest*' | Remove-Item -Recurse -Force;Remove-Item -Recurse -Force .tmp, .pytest_cache;清 __pycache__(排除 node_modules)。2) 统一 pytest 临时目录出口:pyproject.toml 的 addopts 加 --basetemp=.pytest_tmp(.gitignore 已有 .pytest_tmp*/ 规则)。3) 打包改用 git archive,从根本上不再复制工作树。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "simplify",
|
|||
|
|
"title": "21 个测试文件各自重复内联同一套 SQLite Database 夹具,无任何共享 fixture",
|
|||
|
|
"location": "tests/test_resource_access_service.py:15",
|
|||
|
|
"problem": "grep -l \"DatabaseSettings(\" tests/*.py 命中 21 个文件,每个都手写同一段 Database + DatabaseSettings + SqliteSettings + repository 组装 + 种子数据。同样地 _StaticConfigService(把 7 个 storage 路径指到 tmp 目录的桩类)在 test_advanced_material_workflow.py 与 test_advanced_style_set_workflow.py:61-75 各写一份完全相同的实现,RequestContext(...) 构造在 3 个文件里重复。仓库没有 conftest.py,没有任何共享位置。",
|
|||
|
|
"evidence": "# tests/test_resource_access_service.py:15\ndef _fixture(tmp_path):\n database = Database(\n DatabaseSettings(\n backend=\"sqlite\",\n sqlite=SqliteSettings(path=str(tmp_path / \"resource-access.sqlite3\")),\n )\n )\n database.initialize()\n profiles = ProfileRepository(database)\n workflows = WorkflowRepository(database)\n images = ImageRepository(database)\n jobs = JobRepository(database)\n ...\n\n$ grep -l \"DatabaseSettings(\" tests/*.py | wc -l\n21",
|
|||
|
|
"impact": "新增一个 repository 或改一次 DatabaseSettings 签名要动 21 个文件;各文件的种子数据略有差异,导致同一个 bug 在 A 测试能复现、B 不能这类难查的不一致;也是 11559 行测试代码里相当一部分体积的来源。",
|
|||
|
|
"fix": "新建 tests/conftest.py(顺带解决 sys.path 问题),提供 sqlite_db、storage_config、ctx 三个 fixture(sqlite_db 用 tmp_path 建库并 initialize;storage_config 用字典推导生成 7 个 *_path;ctx 返回构造 RequestContext 的工厂),然后分批把 21 个文件的 _fixture 改成注入。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "tests/test_material_vision_input.py 无条件 import Pillow,而 Pillow 被明确定义为可选依赖",
|
|||
|
|
"location": "tests/test_material_vision_input.py:6",
|
|||
|
|
"problem": "requirements.txt 全文只有注释、零条实际依赖,且第一段明写“为保持插件默认零依赖,不在这里强制安装 Pillow”。业务代码遵守了这个约定(engines/image_provider.py:231-235、backend/services/api_discovery_service.py:320-324 都是 try/except ImportError 的软导入),但测试是模块级硬导入,没有 pytest.importorskip 也没有 skipif,仓库也没有 requirements-dev.txt。",
|
|||
|
|
"evidence": "# requirements.txt(全文仅注释)\n$ grep -vE \"^\\s*#|^\\s*$\" requirements.txt | wc -l\n0\n# 内容:为保持插件默认零依赖,不在这里强制安装 Pillow。\n\n# tests/test_material_vision_input.py:6\nfrom PIL import Image ← 模块级硬导入,无保护\n\n# 对比业务代码 engines/image_provider.py:231\ntry:\n import httpx\nexcept ImportError:\n httpx = None",
|
|||
|
|
"impact": "在任何按 requirements.txt(零依赖)搭起来的干净环境里,python -m pytest tests/ 会在收集阶段报 ModuleNotFoundError: No module named 'PIL',整个 263 个用例的套件一起挂掉而不是跳过这 2 个。与 pytest 配置缺失叠加,等于测试套件在标准环境下有两处一碰就碎。",
|
|||
|
|
"fix": "1) 改成 Image = pytest.importorskip(\"PIL\").Image(或文件顶部 pytest.importorskip(\"PIL\")),缺 Pillow 时只跳过这 2 个用例;2) 新建 requirements-dev.txt 显式声明并钉版本:pytest>=8.0,<9 / Pillow>=10.0,<12;requirements.txt 保持零依赖不变。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "前端把 vite/typescript/vue-tsc 放进 dependencies,devDependencies 为空",
|
|||
|
|
"location": "frontend/package.json:15",
|
|||
|
|
"problem": "@vitejs/plugin-vue、typescript、vite、vue-tsc 是纯构建期工具,运行期打包产物里一行都不会出现,却全放在 dependencies 里;devDependencies 是空对象。",
|
|||
|
|
"evidence": "\"dependencies\": {\n \"@vitejs/plugin-vue\": \"^5.2.0\",\n \"pinia\": \"^2.3.0\",\n \"typescript\": \"^5.7.0\",\n \"vite\": \"^6.0.0\",\n \"vue\": \"^3.5.0\",\n \"vue-router\": \"^4.5.0\",\n \"vue-tsc\": \"^2.2.0\"\n},\n\"devDependencies\": {}",
|
|||
|
|
"impact": "npm ci --omit=dev 无法瘦身;依赖审计(npm audit)会把构建工具的漏洞算作生产依赖漏洞,噪音掩盖真实风险。因 private:true 且不发布 npm 包,影响有限,但会误导自动化供应链扫描。",
|
|||
|
|
"fix": "dependencies 只保留 pinia / vue / vue-router;把 @vitejs/plugin-vue、typescript、vite、vue-tsc 移到 devDependencies。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "无 CI,且 DEVELOPMENT.md 的“验证”流程根本不跑测试、还引用了不存在的目录",
|
|||
|
|
"location": "DEVELOPMENT.md:1",
|
|||
|
|
"problem": "文档给出的唯一后端验证命令是 python -m compileall(只做语法编译,不执行任何测试),263 个后端用例和 26 个前端用例都不在流程里。命令里还包含根本不存在的 scripts 目录,而 compileall 对缺失目录只打印警告、退出码仍是 0,静默通过。git ls-files 确认仓库无 .github/、无 Makefile、无任何 workflow 文件。",
|
|||
|
|
"evidence": "# DEVELOPMENT.md「## 验证」\npython -m compileall adapters backend core database engines scripts tests main.py\ncd frontend\nnpm.cmd run typecheck\nnpm.cmd run build:plugin\n\n$ python -m compileall -q adapters backend core database engines scripts tests main.py\nCan't list 'scripts'\nexit=0 ← 目录缺失但仍返回成功\n\n$ ls -a .github\nls: cannot access '.github': No such file or directory\n$ git ls-files | grep -iE '\\.github|workflow.*ya?ml|Makefile'\n(无命中)",
|
|||
|
|
"impact": "11559 行测试代码不在任何门禁上,全靠开发者自觉手跑;compileall 对缺失目录返回 0 意味着这个“验证”连自身完整性都保证不了。230 次提交里没有一次被自动检查拦截过。",
|
|||
|
|
"fix": "1) 修正 DEVELOPMENT.md:验证改为 python -m pytest -q(先落实 pytest 配置)+ npm run typecheck + npm run test:contracts + npm run build:plugin,并删掉不存在的 scripts;2) 加最小 CI(.github/workflows/ci.yml 或对应 Forgejo Actions),只跑 python -m pytest -q + npm run typecheck + npm run test:contracts,20 秒内出结果。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "4 份根目录设计文档已随版本严重过期,且 tests/test_prompt_expansion_stage_guard.py 未入库",
|
|||
|
|
"location": "imagestudio_full_plugin_dev_doc_v3_ui_adjusted.md:1",
|
|||
|
|
"problem": "根目录 4 份 v3 设计文档写于 2026-07-04~07-06(imagestudio_api_contract_v3_ui_adjusted.md、imagestudio_codex_hermes_dev_flow_v3.md、imagestudio_full_plugin_dev_doc_v3_ui_adjusted.md 31 KB、imagestudio_ui_navigation_queue_s130.md),而 metadata.yaml 当前版本已是 1.0.72,中间隔了 70 个版本、约 230 次提交,且没有任何“已归档/已作废”标记。另外 git status 显示 tests/test_prompt_expansion_stage_guard.py 处于未跟踪状态,而它守护的 backend/services/prompt_expansion_service.py 是 M(已修改)。",
|
|||
|
|
"evidence": "$ ls -la *.md\n-rw-r--r-- 9103 Jul 5 16:30 imagestudio_api_contract_v3_ui_adjusted.md\n-rw-r--r-- 9993 Jul 4 14:39 imagestudio_codex_hermes_dev_flow_v3.md\n-rw-r--r-- 31780 Jul 5 16:31 imagestudio_full_plugin_dev_doc_v3_ui_adjusted.md\n-rw-r--r-- 10505 Jul 6 11:57 imagestudio_ui_navigation_queue_s130.md\n\n$ cat metadata.yaml | grep version\nversion: 1.0.72\n\n$ git status --porcelain | grep -E 'prompt_expansion'\n M backend/services/prompt_expansion_service.py\n?? tests/test_prompt_expansion_stage_guard.py ← 60 行 2 个用例,从未提交",
|
|||
|
|
"impact": "新人从根目录看到 4 份体量最大的 .md 会误以为它们是当前契约(api_contract v3 与现在 166 个 endpoint 的实际形态大概率对不上)。未入库的测试文件意味着这次改动的回归保护随时可能被 git clean 抹掉。",
|
|||
|
|
"fix": "1) git mv imagestudio_*.md docs/archive/,并在每份开头加一行归档声明指向 docs/ 下最新日期文档与 backend/route_registry.py;2) 立刻把 tests/test_prompt_expansion_stage_guard.py 与它守护的 backend/services/prompt_expansion_service.py 改动放进同一个 commit 提交。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "前后端耦合",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\B1-前后端耦合.md",
|
|||
|
|
"summary": "URL 层的耦合治理做得很好:backend/route_registry.py 是 166 条路由的唯一真相来源,前端 143 个端点全集中在 endpoints.ts,逐条比对后前端调用零 404、几乎无硬编码 URL(仅 vite base 一处)。问题全部集中在\"语义契约\"层——URL 对上了但规则没对上:前端在 App.vue、workflow-lock-state.ts、workflow-model-route.ts 里重新实现了后端的顺序进度状态机、节点写锁、模型路由解析链、队列状态枚举和 retry 权限矩阵,每一处都有可举证的不一致(sealed_at 后端零产出、节点写锁少判 selected_image_id 与 completed_node_ids、快速角色流路由优先级后端是反的、retry 按钮漏判 can_write)。根因是后端有 CharacterGradualProgression.to_dict()、PERMISSION_MATRIX、ContractService 这些可下发的真相,却没有一条路由送到前端。GET contract 已提供完整路由+权限+错误码自检数据,但唯一消费点是 SystemManagementPanel.vue 的诊断 JSON dump,无任何自动化契约测试——我实测 166 条 handler 全部可导入,做成 CI 断言只需十几行,是本维度性价比最高的修复。另外 web/ 与 pages/imagestudio/ 两份 1.1MB 的 vite dist 被提交进 git(80 次提交),每改一行 UI 源码都带上一批内容哈希文件的增删。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\contract_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\ui_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\web_hosting_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_progression_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\node_injection_template_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\event_stream_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\image_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_style_set_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\media_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\contract_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\api_discovery_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\node_material_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\round_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\profile_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\image_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\advanced_material_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\event_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\job_routes.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\workflow_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\workflow_round_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\profile_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\job_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\image_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\advanced_material_dto.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\model_route_chain.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\config_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\security.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\event_bus.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\character_flow.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\profile_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\api\\endpoints.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\api\\client.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\api\\astrbot-page-bridge.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\config\\timeouts.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\utils\\presentation.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\utils\\storage-uri.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\api.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\bootstrap.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\profile.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\workflow.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\workflow-round.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\job.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\image.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\types\\advanced-material.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\App.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\workflow\\workflow-lock-state.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\workflow\\workflow-model-route.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\workflow\\workflow-node-catalog.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\workflow\\WorkflowCandidatePool.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\materials\\advanced-material-state.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\settings\\SystemManagementPanel.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\src\\features\\settings\\model-route-chain.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\standalone-web\\vite.config.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\plugin-page\\vite.config.ts",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\plugin-page\\src\\App.vue",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\package.json",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tests\\task-timeout-consistency.test.mjs",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\tests\\workflow-model-route.test.mjs",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\tests\\test_model_route_chain.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\.gitignore"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "前端复刻后端顺序进度状态机,后端从不下发权威结果",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:10843",
|
|||
|
|
"problem": "后端 is_character_gradual_node_completed()(workflow_progression_service.py:180-199)是节点完成判定的唯一权威,含 status 白名单、node_type 特例、profile_card 别名归一、legacy 元数据兼容四层规则。前端在 App.vue 逐条重写了这四层,连\"legacy 兼容只对无 sequential-v1 标记的工作流生效\"这种细节都手工复刻,并把版本号字面量抄进 App.vue:927。根因是 backend/routes/node_routes.py:60-62 的 list_workflow_nodes 只返回裸节点数组;全仓库 progression.to_dict() 唯一调用点是 character_workflow_service.py:339(execute-character 响应),列表接口拿不到。",
|
|||
|
|
"evidence": "后端 workflow_progression_service.py:180-199:\nif node.status in PROGRESSION_COMPLETED_NODE_STATUSES: # {\"selected\",\"skipped\"}\n return True\nif node.status == \"completed\" and (node.node_type in {\"planning\",\"reference_binding\"}\n or canonical_gradual_node_key(node.node_key) == \"profile_card\"): return True\nif allow_legacy_planning_metadata:\n if node.node_key == \"prompt_plan\": return \"prompt_plan\" in metadata and metadata.get(\"prompt_plan\") is not None\n\n前端 App.vue:10843-10859(第二份实现):\nif (status === \"selected\" || status === \"skipped\") return true\nif (status === \"completed\" && (backendNode.node_type === \"planning\" || backendNode.node_type === \"reference_binding\" || canonicalCharacterWorkflowNodeKey(backendNode.node_key) === \"profile_card\")) return true\nconst legacyProgression = isRecord(execution) && execution.progression_version !== CHARACTER_WORKFLOW_PROGRESSION_VERSION\nif (!legacyProgression) return false\nif (backendNode.node_key === \"prompt_plan\") return backendNode.metadata?.prompt_plan != null\n\nApp.vue:927: const CHARACTER_WORKFLOW_PROGRESSION_VERSION = \"character_gradual_sequential_v1\"\n\nbackend/routes/node_routes.py:60-62(无 progression):\nreturn [node.to_dict() for node in workflow_service.list_nodes(workflow_id)]",
|
|||
|
|
"impact": "后端改一条进度规则,前端不会报错也不会失效,只会静默给出与服务端相反的\"下一步点哪个节点\":UI 显示可用 → 用户点击 → 后端 ensure_character_gradual_node_access 抛 character_gradual_progression_locked;反之 UI 灰掉后端允许的节点,用户彻底卡死且无提示。前端还复制了一份后端的向后兼容降级策略,双份都要维护。",
|
|||
|
|
"fix": "在 list_workflow_nodes 返回体加 progression 字段(CharacterGradualProgression.to_dict() 已现成,含 current_node_id/completed_node_ids/all_completed/import_confirmed):return {\"items\": [n.to_dict() for n in nodes], \"progression\": progression.to_dict() if progression else None}。同步 frontend/shared/types/workflow.ts 并升 contract_version。删除 App.vue:10836-10861 与 App.vue:927,改为读 progression.completed_node_ids / current_node_id。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "节点写锁规则前后端不一致,前端判据含后端从不产出的 sealed_at 字段",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/workflow-lock-state.ts:12",
|
|||
|
|
"problem": "两条独立的\"节点是否可写\"规则,条件集合不同。前端多判一个不存在的字段 metadata.sealed_at,少判后端的两个真实条件 node.selected_image_id 与 progression.completed_node_ids。grep -rn \"sealed_at\" --include=*.py 在 main.py/backend/core/database/engines/adapters 全树零命中。",
|
|||
|
|
"evidence": "前端 workflow-lock-state.ts:8-27:\nexport function isWorkflowSealed(workflowStatus, metadata) {\n return (status === \"archived\" || status === \"cancelled\" ||\n Boolean(metadata?.final_archive_history_id || metadata?.sealed_at))\n}\nexport function isWorkflowNodeWriteLocked(nodeStatus, workflowStatus, workflowMetadata) {\n return normalizedStatus(nodeStatus) === \"selected\" || isWorkflowSealed(workflowStatus, workflowMetadata)\n}\n\n后端 workflow_round_service.py:1145-1155:\nif node.status == \"selected\" or node.selected_image_id or (\n progression is not None and node.node_id in progression.completed_node_ids):\n raise ApiError(ErrorCode.VALIDATION_FAILED,\n \"confirmed workflow node is locked; create a branch before adding another round\",\n {\"node_id\": node.node_id, \"required_action\": \"create_branch\"})",
|
|||
|
|
"impact": "三处真实分歧:(1) selected_image_id 已设但 status != selected 时(如 adopt-image 后 status 仍为 completed),前端放行\"再跑一轮\",后端返回 VALIDATION_FAILED/required_action: create_branch,用户看到无法自恢复的报错;(2) 节点通过 skipped 或 planning completed 进入 completed_node_ids 时同理;(3) metadata.sealed_at 是永远 undefined 的死判据,制造\"存在该封存机制\"的错觉,连测试都没覆盖。",
|
|||
|
|
"fix": "把锁状态改为服务端下发。在 WorkflowNode.to_dict() 增加 writable: bool 与 lock_reason: str|None——WorkflowNodeMaterialBindingDetail 已有 writable+lock 结构(frontend/shared/types/workflow.ts:108-123),直接复用同一形状。前端 workflow-lock-state.ts 退化为读 node.writable。若暂不改协议,至少立刻删掉 sealed_at 并补上 selected_image_id 判据。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "模型路由解析链前端重实现,快速角色流优先级与后端相反",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/workflow-model-route.ts:36",
|
|||
|
|
"problem": "前端为显示\"当前生效模型/来源标签\"复刻了后端整条路由回退链,包括 node_key→model_assignments 的映射表,并直接读取原始配置文件结构 config.model_assignments / config.generation.default_model。三处不一致:前端多了 round 层、少了 image_low 兜底、完全没有 quick_character 分支。",
|
|||
|
|
"evidence": "后端普通节点 generation_service.py:169-176:\nrequested_model_route = str(payload.get(\"model_route\") or node.model_route or assigned_route\n or profile_model_route or self.default_model_route or \"image_low\")\n\n后端快速角色 generation_service.py:617-624(优先级反转,assigned_route 抢在 payload 之前):\nif quick_character:\n route_candidate = str(assigned_route or payload.get(\"model_route\")\n or profile_model_route or self.default_model_route or \"image_low\")\n\n前端 workflow-model-route.ts:36-48(唯一一条链):\nconst candidates = [[\"override\", input.overrideRoute], [\"round\", input.roundRoute],\n [\"node\", input.nodeRoute], [\"assignment\", input.assignmentRoute],\n [\"profile\", input.profileRoute], [\"system\", input.systemRoute]]\n\n前端第二份 assignment 映射 workflow-model-route.ts:76-84 缺后端 generation_service.py:2033-2037 的 quick_avatar/quick_bust/quick_stand/quick_sheet→quick_generation、questionnaire、preview\n\n前端读原始配置 App.vue:2218/2231-2233:\ncurrentSettings.value.config.model_assignments / recordString(generation, \"default_model\")",
|
|||
|
|
"impact": "快速角色流的模型标签会直接说谎:用户设了本次覆盖,UI 显示\"本次覆盖 → route-X\",后端实际按 assigned_route 出图。对 quick_* 节点 workflowModelAssignmentKey(\"quick_avatar\",\"generation\") 返回 \"quick_avatar\" 查不到分配,UI 显示\"角色默认/系统默认\"而后端实际用 quick_generation 的路由。且前端消费的是配置文件原始结构,后端换配置键名前端静默失准。",
|
|||
|
|
"fix": "新增 GET nodes/{node_id}/effective-model-route,或在 nodeDependencyContext 响应里加 { effective_model_route, source },由 generation_service 抽出纯函数 resolve_model_route(node, workflow, profile, assignments, default) 同时供生成路径与该端点调用。前端 resolveWorkflowModelRoute 只保留\"用户 override 是否非空\"判断,其余读服务端字段;删除 workflow-model-route.ts:61-84 的别名/分配表。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "security",
|
|||
|
|
"title": "retry 类操作前端权限门禁漏判 can_write,后端要求两个权限",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:3076",
|
|||
|
|
"problem": "route_registry 已把 retry 定义为复合权限 [can_write, can_retry_tasks] 并通过 GET contract 对外发布,前端节点重试按钮只判了 can_retry_tasks。同文件 App.vue:3373 的队列面板反而判对了(canWrite && canRetry),说明是遗漏而非设计。",
|
|||
|
|
"evidence": "backend/route_registry.py:22-27:\n\"retry\": {\"auth_required\": True, \"required_permission\": \"can_retry_tasks\",\n \"required_permissions\": [\"can_write\", \"can_retry_tasks\"],\n \"description\": \"task retry management (write + retry capabilities)\"}\n\ncore/security.py:31-33:\ndef require_task_retry(ctx: RequestContext) -> None:\n require_write(ctx) # 先判 can_write\n require_permission(ctx, \"can_retry_tasks\")\n\nApp.vue:2786 + 3073-3077:\nconst canRetry = Boolean(permissions?.can_retry_tasks)\n...\nstate: canRetry ? \"可重试\" : \"无权限\",\nenabled: Boolean(canRetry && !selectedNodeWriteLocked && !workflowNodeActionExecuting.value),",
|
|||
|
|
"impact": "持有 can_retry_tasks 但无 can_write 的账号,在工作流面板看到\"可重试\"的可用按钮(portrait-retry 卡片与 3139 处同类卡片),点击后 nodes/{id}/retry 与 nodes/{id}/generate 一律被 require_write 拒为 FORBIDDEN。同一份权限模型在同一文件里出现两种实现,这类漏判还会再犯。",
|
|||
|
|
"fix": "不要在前端逐个手写权限组合。GET contract 返回的 routes[].required_permissions 已是完整数组,bootstrap 后拉一次 contract 构造统一的 canCall(method, endpoint):const contractRoutes = new Map(contract.routes.map(r => [`${r.method} ${r.endpoint}`, r.required_permissions])); function canCall(m, e) { return (contractRoutes.get(`${m} ${e}`) ?? []).every(p => permissions?.[p]) },所有按钮 enabled 统一走它。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "编译产物 web/ 与 pages/imagestudio/ 被提交进 git,且是两份重复 dist",
|
|||
|
|
"location": "frontend/standalone-web/vite.config.ts:12",
|
|||
|
|
"problem": "两个 vite 构建把产物直接写进受版本控制的目录且 emptyOutDir: true,每次构建整批删除重建内容哈希文件名。.gitignore 忽略了 frontend/**/dist/ 却没忽略这两个实际输出目录。git ls-files web/ pages/ = 38 文件各 1.1MB,git log -- web/assets = 80 次提交,最近 15 次改 App.vue 的提交每一次都同时改了两个 assets 目录。",
|
|||
|
|
"evidence": "frontend/standalone-web/vite.config.ts:11-15:\nbase: \"/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/\",\nbuild: { outDir: resolve(rootDir, \"../../web\"), emptyOutDir: true },\n\nfrontend/plugin-page/vite.config.ts:11-15:\nbase: \"./\",\nbuild: { outDir: resolve(rootDir, \"../../pages/imagestudio\"), emptyOutDir: true },\n\nbackend/services/web_hosting_service.py:14-15:\nPLUGIN_PAGE_DIR = REPO_ROOT / \"pages\" / \"imagestudio\"\nSTANDALONE_WEB_DIR = REPO_ROOT / \"web\"\n\n两份 dist 同源——frontend/plugin-page/src/App.vue 全文:\n<script setup lang=\"ts\">import WorkbenchApp from \"../../standalone-web/src/App.vue\"</script>\n<template><WorkbenchApp /></template>",
|
|||
|
|
"impact": "每个 UI 改动的 diff 被 2MB 噪音淹没,review 无法看清真实变更;没有 CI 校验 web/ 是当前 frontend/ 的构建结果,只改源码忘记 build 的提交会让线上跑旧 JS,而 WebHostingService.manifest() 只检查 index.html 存在就报 built: true,无法发现版本错位;内容哈希+emptyOutDir 让 git 累积大量永不复用的 blob;两份 dist 意味着每次 UI 修复必须构建两次,漏一次就出现\"插件页正常、独立页是旧版\",git 历史里 fix: recover material library after stale asset token 等修复正是这类症状。",
|
|||
|
|
"fix": ".gitignore 增加 /web/ 与 /pages/imagestudio/,git rm -r --cached web pages/imagestudio;发布改为 CI 构建 + release artifact。若插件分发形态必须保留产物在主干,则加 CI 检查 npm run build && git diff --exit-code web pages 保证产物与源码一致,并把两个 vite 配置合并为一次构建两个 base(或让 plugin-page 复用 standalone 产物)消除两份 dist;WebHostingService.manifest() 增加构建时注入的 build_hash 与 metadata.yaml 版本比对。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "events/stream 无游标:SSE 载荷塞进 JSON,5 秒轮询 20 条导致事件静默丢失",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:4575",
|
|||
|
|
"problem": "后端把事件流做成\"返回最近 N 条 + 预渲染 SSE 文本\"的普通 JSON 接口,没有 after_event_id/since 参数;前端只能轮询固定窗口,靠在返回数组里查找上次 event_id 切增量,窗口被冲掉后增量计算退化为只取最后一条。",
|
|||
|
|
"evidence": "后端 event_stream_service.py:22-32(无游标):\ndef recent_events(self, *, event_type=None, limit=50):\n events = self.event_bus.history(event_type)\n selected = events[-_limit(limit):]\n return {\"content_type\": \"text/event-stream\", \"events\": [...],\n \"body\": \"\".join(to_sse(e) for e in webhook_events), \"count\": ...}\n\n前端 App.vue:4566-4577, 4608:\nconst stream = await getJson<...>(endpoints.eventStream, { limit: 20 })\nconst previousIndex = hadPrevious ? stream.events.findIndex(e => e.event_id === lastEventStreamId) : -1\nconst freshEvents = hadPrevious ? stream.events.slice(previousIndex >= 0 ? previousIndex + 1 : -1) : []\n...\neventStreamTimer = window.setInterval(poll, 5000)",
|
|||
|
|
"impact": "一轮 4~8 张候选图的生成会在几秒内产生远超 20 条 generation.*/workflow.* 事件,此时 findIndex 返回 -1,slice(-1) 只留最后一条,eventTypes 只有一个元素,generationChanged/workflowChanged 基于 1/20 样本判定——loadJobs()/loadDashboard()/loadWorkflowRoundDetail() 该刷的不刷,UI 停在旧状态直到用户手动刷新。这是\"生成完了界面不动\"的结构性来源。同时 body 与 content_type 每次白白序列化传输后被丢弃。",
|
|||
|
|
"fix": "EventStreamService.recent_events 增加 after_event_id: str|None,在 _history 定位后返回其后全部事件(配合 limit 上限 200),并返回 truncated: bool;前端改传 { after_event_id: lastEventStreamId, limit: 200 },收到 truncated: true 时退化为一次全量刷新而非静默少刷;content_type/body 要么删掉,要么单独开一条真正的 SSE 路由。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "队列状态枚举在前后端共 4 处独立硬编码",
|
|||
|
|
"location": "frontend/standalone-web/src/App.vue:3285",
|
|||
|
|
"problem": "后端已把任务状态分类抽成具名常量,前端把这些集合的字面量抄了三份(App.vue:3285、3293、presentation.ts:5),还在 App.vue:3628 第五次内联,且 presentation.ts 的 TERMINAL_JOB_STATUSES 与后端任何一个常量都不完全相等。",
|
|||
|
|
"evidence": "后端 queue_service.py:31-34:\nRETRYABLE_TASK_STATUSES = {\"failed\", \"dead_letter\", \"cancelled\"}\nFAILED_TASK_STATUSES = (\"failed\", \"dead_letter\", \"cancelled\")\nCOMPLETED_TASK_STATUSES = (\"completed\", \"succeeded\", \"superseded\")\nDELETABLE_TASK_STATUSES = frozenset((*FAILED_TASK_STATUSES, *COMPLETED_TASK_STATUSES))\n\nApp.vue:3284-3286:\nconst failedJobs = computed(() => queueItems.value.filter((job) => [\"failed\", \"dead_letter\", \"cancelled\"].includes(normalizeStatus(job.status))))\n\nApp.vue:3292-3295:\nconst completedQueueJobs = computed(() => queueItems.value.filter((job) => [\"completed\", \"succeeded\", \"superseded\"].includes(normalizeStatus(job.status))))\n\npresentation.ts:5(第四份):\nconst TERMINAL_JOB_STATUSES = new Set([\"completed\", \"succeeded\", \"failed\", \"cancelled\", \"dead_letter\", \"superseded\"])",
|
|||
|
|
"impact": "TASK_STATUSES(queue_service.py:18-29)新增一个终态(如 expired)时,后端会正确算进 DELETABLE,前端四处判定全部漏掉:任务永远停在\"运行中\"分组、waitTime() 里 terminal 为 false 导致耗时读数无限增长、\"清理终态任务\"按钮的 enabled 计数与后端实际清理数量对不上。",
|
|||
|
|
"fix": "把四个集合加进 GET ui/bootstrap 返回体(UiService.get_bootstrap 已在下发 permissions 与 feature_flags,加 task_statuses: {retryable, failed, completed, terminal} 是同性质元数据),前端建 taskStatusSets.ts 从 bootstrap 初始化;或至少把四份合并成 frontend/shared/config/task-status.ts 一份,并加一条像 task-timeout-consistency.test.mjs 那样的正则测试断言它与 queue_service.py 字面量一致。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "images/{id}/moderation-review 与 include_restricted 是完整但前端零对接的功能链",
|
|||
|
|
"location": "backend/routes/image_routes.py:41",
|
|||
|
|
"problem": "后端为\"内容审核隔离图片\"实现了完整闭环——列表过滤、管理员可见开关、隐藏计数、管理员复核端点——前端一个都没接。grep -rn \"include_restricted|restricted_hidden|quarantined|gallery_visibility\" frontend/ 全树零命中(仅 App.vue:4526 的 include_archived 无关)。frontend/shared/types/image.ts:28-33 的 ImageListResult 也没有 restricted_hidden 字段。",
|
|||
|
|
"evidence": "backend/routes/image_routes.py:41-49:\ninclude_restricted = as_bool(params.get(\"include_restricted\"), default=False, non_string_truthiness=True)\nif include_restricted and not (ctx.has_permission(\"can_manage_materials\") or ctx.has_permission(\"*\")):\n include_restricted = False\nif not include_restricted:\n visible = [item for item in payload[\"items\"] if not _is_restricted_image(item)]\n payload[\"restricted_hidden\"] = len(payload[\"items\"]) - len(visible)\n payload[\"items\"] = visible\n\nbackend/services/image_service.py:135-143:\nupdate = {\"metadata\": {**metadata, \"quarantined\": False if action == \"approve\" else True,\n \"gallery_visibility\": \"normal\" if action == \"approve\" else \"restricted\",\n \"moderation_admin_review\": review}, **({\"is_archived\": True} if action == \"reject\" else {})}",
|
|||
|
|
"impact": "generation_service.py:491/968 会把审核不通过的图打上 gallery_visibility: \"restricted\"。这些图对所有人(含管理员)永久不可见、不可复核、不可解除——UI 既不显示\"有 N 张被隐藏\",也没有任何入口调用 approve。用户视角是\"生成成功了但图库里没有\",无任何解释。这是一条后端做完、前端未接、因此在生产中表现为数据丢失的功能链。",
|
|||
|
|
"fix": "图库页加\"显示受限图片\"开关(仅 can_manage_materials 可见)传 include_restricted=true;列表底部读 restricted_hidden 显示\"另有 N 张待复核\";受限图详情弹窗加 approve/reject 调新增的 endpoints.imageModerationReview(imageId);把 restricted_hidden?: number 补进 ImageListResult。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "AdvancedMaterialImportStatus 双向漂移,Union | string 写法使联合类型编译期完全失效",
|
|||
|
|
"location": "frontend/shared/types/advanced-material.ts:1",
|
|||
|
|
"problem": "前端缺 6 个后端实际产出的 style-set 状态,多 2 个后端根本不产出的值;而所有状态字段一律声明为 `Union | string`,TypeScript 会折叠成 string,联合类型不产生任何检查。",
|
|||
|
|
"evidence": "后端实际写入 import.status 的取值(扫描 update_import):\nadjustment_required, cancelled, ready_for_validation, recognition_failed, recognition_queued,\nrecognizing_style_sources, style_consensus_failed, style_consensus_queued, style_consensus_running,\nstyle_source_recognition_queued, style_source_review_required, validating, validation_failed, validation_queued\n\n前端 advanced-material.ts:1-18 联合类型缺 6 个 style_* 状态,多出 pending_validation 与 pending_review\n(grep -rn '\"pending_validation\"' --include=*.py backend/ database/ → 0 命中)\n\n声明处 advanced-material.ts:168:\nstatus: AdvancedMaterialImportStatus | string\n(同样写法见 AdvancedStyleSource.status、AdvancedStyleValidationCase.status、workflow-round.ts:73/112 等十余处)",
|
|||
|
|
"impact": "运行时目前靠 advanced-material-state.ts:24-67 的 switch 兜住(手工覆盖了 style_* 状态),但类型层面零护栏:后端加新状态,TS 编译通过、review 看不出,只有用户看到 advanced-material-state.ts:65 的\"未知状态\"文案才暴露。pending_validation/pending_review 两个幽灵值说明漂移已发生过至少一次且无人察觉。",
|
|||
|
|
"fix": "去掉 | string 让联合类型真正生效,需要容忍未知值的地方在解析边界显式收窄;状态集合从后端生成而非手抄——把各 *_STATUSES 常量集中到一个 Python 模块,加生成脚本产出 frontend/shared/types/generated-enums.ts,CI 校验 git diff --exit-code。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "backend/dto/advanced_material_dto.py 已不是进阶素材接口的响应真相来源",
|
|||
|
|
"location": "backend/dto/advanced_material_dto.py:141",
|
|||
|
|
"problem": "TS 类型里大量字段在 DTO 的 to_dict() 中不存在,而路由层序列化只补 *_url。响应形状实际由散落在 3600 行 advanced_style_set_service.py 各处的 dict 字面量决定。",
|
|||
|
|
"evidence": "路由层只做签名 URL 增强 advanced_material_routes.py:387-392:\ndef _serialize_style_source(payload):\n return {**payload, \"source_image_url\": _signed_url(payload.get(\"source_image_uri\")),\n \"source_thumbnail_url\": _signed_url(payload.get(\"source_thumbnail_uri\"))}\n\nTS 声明但 DTO 无:\n- AdvancedStyleSource: excluded/filename/image_uri/thumbnail_uri/retry_count(来自 advanced_style_set_service.py:409/414)\n- AdvancedStyleValidationCase: scenario/title/style_prompt/style_scores/rendering_layers/content_novelty(来自 :2330/2354/3345/3422)\n- AdvancedMaterialImport: task_id/workflow_kind/sources/consensus/validation_cases(来自 :389/1152)\n- AdvancedStyleConsensus: 整个接口无对应 DTO,AdvancedStyleConsensusRun 字段名完全不同(consensus_style/version_id/variable_style_features 都不在 DTO)\n\n反向:DTO 输出但 TS 未声明的 idempotency_key、recognition_model_route/provider_id/provider_model、width/height/completed_at",
|
|||
|
|
"impact": "backend/dto/ 这一层对进阶素材名存实亡——想知道接口返回什么只能全文搜索,改动时没有任何单点可对照。这也是状态枚举漂移的同一根因,并让未来把响应 schema 纳入 contract 变得不可能。",
|
|||
|
|
"fix": "为 style-set 流程补 DTO(AdvancedStyleSourceView、AdvancedStyleValidationCaseView、AdvancedStyleConsensusView),让 service 返回 DTO 实例而非裸 dict,路由层只调 to_dict() + 加签名 URL,使\"接口返回什么\"重新有唯一定义点。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "route-chain: 线格式编解码器逐字移植到 TS,无跨语言一致性测试",
|
|||
|
|
"location": "frontend/standalone-web/src/features/settings/model-route-chain.ts:33",
|
|||
|
|
"problem": "这不是逻辑重复而是同一序列化格式的两份独立实现——前端写出的字符串要被后端解析,正确性完全依赖手工对齐 json.dumps 与 JSON.stringify 的输出细节,无任何测试断言两语言编码结果相同。",
|
|||
|
|
"evidence": "core/model_route_chain.py:61-66:\nreturn ROUTE_CHAIN_PREFIX + json.dumps(\n {\"plugin\": plugin_route, \"astrbot\": astrbot_route},\n ensure_ascii=False, separators=(\",\", \":\"), sort_keys=True)\n\nmodel-route-chain.ts:33:\nreturn `${MODEL_ROUTE_CHAIN_PREFIX}${JSON.stringify({ astrbot, plugin })}`\n\n目前能对上靠三个巧合同时成立:Python sort_keys=True 产出 astrbot 在前 ≡ TS 对象字面量恰好写成 { astrbot, plugin } 的插入顺序;separators=(\",\", \":\") ≡ JSON.stringify 默认无空格;ensure_ascii=False ≡ JSON.stringify 默认不转义非 ASCII。",
|
|||
|
|
"impact": "把 TS 那行写成 JSON.stringify({ plugin, astrbot })(更符合\"plugin 优先\"的语义直觉)会产出 {\"plugin\":...,\"astrbot\":...},Python 端 json.loads 读 dict 解析仍成功所以不会立刻报错,但两端产出的字符串不再逐字节相等,任何做字符串比较的地方(配置 diff、快照对比、settings/diff)都会把\"没变的配置\"报成\"变了\"。tests/test_model_route_chain.py 只验 Python 内部往返,frontend/tests/workflow-model-route.test.mjs 根本没测这个模块。",
|
|||
|
|
"fix": "加黄金样本测试固定线格式:Python 断言 encode_model_route_chain(\"p\",\"astrbot:a\") == 'route-chain:{\"astrbot\":\"astrbot:a\",\"plugin\":\"p\"}',frontend/tests/ 用同一字符串断言 encodeModelRouteChain({plugin:\"p\",astrbot:\"astrbot:a\"}),并加反向 parse 用例。更彻底:settings 保存接口接受 {plugin, astrbot} 结构体由后端编码,别让前端编码。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "character_flow 节点目录与 legacy 别名表在前端重复,两份前端拷贝之间也不一致",
|
|||
|
|
"location": "frontend/standalone-web/src/features/workflow/workflow-model-route.ts:70",
|
|||
|
|
"problem": "同一张 legacy node_key→canonical node_key 映射表存在三份,其中两份在同一前端目录下且 expression_sheet 的映射结果不同。11 个节点的 node_key 与展示文案也在后端 CHARACTER_GRADUAL_STEPS 与前端 CHARACTER_WORKFLOW_PREVIEW_NODE_SPECS 各存一份,中文标题都不一致。",
|
|||
|
|
"evidence": "engines/character_flow.py:15-20:\nLEGACY_CHARACTER_GRADUAL_NODE_KEYS = {\"half_body\": \"upper_body\",\n \"expression_sheet\": \"turnaround\", # ← turnaround\n \"pose_sheet\": \"design_sheet\", \"final_pack\": \"profile_card\"}\n\nworkflow-node-catalog.ts:23-28(与后端一致):\nexport const LEGACY_CHARACTER_WORKFLOW_NODE_KEYS = { half_body: \"upper_body\",\n expression_sheet: \"turnaround\", pose_sheet: \"design_sheet\", final_pack: \"profile_card\" }\n\nworkflow-model-route.ts:61-74(同一前端内第三份,冲突):\nconst canonicalAliases = { portrait: \"headshot\", avatar: \"headshot\", half_body: \"upper_body\",\n halfbody: \"upper_body\", standing: \"full_body\", fullbody: \"full_body\",\n three_view: \"turnaround\", three_views: \"turnaround\", character_sheet: \"design_sheet\",\n expression_sheet: \"design_sheet\", // ← design_sheet,与另两处冲突\n pose_sheet: \"design_sheet\", hires_fix: \"style_lock\" }",
|
|||
|
|
"impact": "expression_sheet 的冲突目前不产生可见 bug(turnaround 和 design_sheet 恰好都映射到 sheet_generation),但这是巧合而非设计。真正风险:engines/character_flow.py 加新节点或改别名时前端两份表都不更新,workflowUiNodeIdForNodeKey 返回 undefined,该节点在工作流面板直接不显示——用户看到的步骤列表比后端实际创建的节点少一个。",
|
|||
|
|
"fix": "ui/bootstrap 或新增 GET workflows/node-catalog 下发 gradual_node_templates() + LEGACY_CHARACTER_GRADUAL_NODE_KEYS(后端 character_flow.py:131-145 已有 gradual_node_templates(),几乎可直接返回)。前端 workflow-node-catalog.ts 只保留纯展示增强字段(badge/phase/detail),键与别名从接口取;删除 workflow-model-route.ts:61-74 的私有别名表改用 canonicalCharacterWorkflowNodeKey。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "LONG_TASK_ENDPOINT 正则重复两份,其声称对齐的\"后端 900 秒预算\"在后端不存在",
|
|||
|
|
"location": "frontend/shared/api/client.ts:14",
|
|||
|
|
"problem": "一条 18 个分支、枚举所有慢端点的正则在 client.ts:14 与 astrbot-page-bridge.ts:6 逐字符重复;而 timeouts.ts 的 900 秒在后端没有对应常量(生成任务真实预算是 180 秒),号称守护一致性的测试只读前端文件。",
|
|||
|
|
"evidence": "client.ts:14 与 astrbot-page-bridge.ts:6(完全相同的两份):\nconst LONG_TASK_ENDPOINT = /(?:^|\\/)(?:generate|retry|review|execute-character|expand|governance-preview|probe-provider|probe-model|recognize|re-recognize|renormalize|renormalize-batch|import-pack|migrate|restore)(?:\\/|$)/\n\nfrontend/shared/config/timeouts.ts:\nexport const LONG_TASK_TIMEOUT_SECONDS = 900\n\n后端唯一的 900(与生成任务无关):\n# backend/services/advanced_style_set_service.py:57\nSTYLE_VALIDATION_TIMEOUT_SECONDS = 900\n\n生成任务真实预算 generation_service.py:2308:\nraw = request.get(\"timeout_seconds\") or settings.get(\"timeout_seconds\") or 180\n(config.yaml:25/42 亦为 timeout_seconds: 180)\n\nfrontend/tests/task-timeout-consistency.test.mjs:11-25 读的四个文件全在 frontend/ 下:\ntest(\"uses the backend 900 second task budget throughout the plugin bridge UI\", ...)\nassert.match(timeouts, /LONG_TASK_TIMEOUT_SECONDS = 900/)",
|
|||
|
|
"impact": "正则里 probe-model/renormalize/renormalize-batch/import-pack 对应的端点前端根本不调用,说明列表已与实际调用面脱节无人维护;测试名断言\"后端 900 秒预算\"却没读过任何后端文件,固化的是虚构契约;新增慢端点(如实际走 900 秒的 style-consensus)不在正则里,客户端 30 秒就超时报错而后端任务还在跑。",
|
|||
|
|
"fix": "正则提到 frontend/shared/config/timeouts.ts 一份供两个 transport import。更好的做法是后端在 ui/bootstrap 下发 { long_task_endpoint_patterns, long_task_timeout_seconds },取值来自 config.generation.provider.timeout_seconds 的上界;测试改为读 config.yaml / generation_service.py 与前端常量做真实比对。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "contract 端点只被用作诊断 JSON dump,无任何自动化契约测试消费",
|
|||
|
|
"location": "frontend/standalone-web/src/features/settings/SystemManagementPanel.vue:8",
|
|||
|
|
"problem": "契约自检机制已存在且数据完备(166 条路由的 method/endpoint/handler/permission/required_permissions + 权限矩阵 + 错误码),但唯一前端消费点是把它 JSON.stringify 塞进 <pre>;route_registry.py:365-377 的 assert_unique_routes()/assert_known_permissions() 也没有任何测试调用。",
|
|||
|
|
"evidence": "backend/services/contract_service.py:13-30:\nreturn {\"contract_version\": \"3.0\", \"api_base\": API_BASE,\n \"routes\": route_manifest(), \"permissions\": permission_matrix(),\n \"error_codes\": [code.value for code in ErrorCode], \"response_shape\": {...}}\n\nSystemManagementPanel.vue:8(唯一消费点):\nelse if(props.mode===\"diagnostics\"){const [contract,actions,manifest]=await Promise.all([\n getJson(endpoints.contract),getJson(endpoints.uiActions),getJson(endpoints.webManifest)]);\n result.value={contract,actions,manifest}}\n\nbackend/route_registry.py:365-377(现成自检,无人调用):\ndef assert_unique_routes() -> None: ...\ndef assert_known_permissions() -> None: ...",
|
|||
|
|
"impact": "契约漂移目前没有任何自动化拦截。本次审计的 14 条死接口、retry 权限漏判、状态枚举双向漂移全部是人工比对才发现的,任何一条都本可由 CI 自动发现。",
|
|||
|
|
"fix": "本维度性价比最高的修复。三条测试即可:(1) tests/test_route_contract.py 调 assert_unique_routes()/assert_known_permissions() 并逐条 importlib 校验 handler 可导入(我已实测 166 条全部可导入,当前会通过);(2) frontend/tests/api-contract.test.mjs 解析 route_registry.py 的 RouteSpec 与 endpoints.ts 模板(把 ${...} 与 {...} 归一为 {}),断言 endpoints.ts ⊆ ROUTES,并把 14 条死接口写成显式 allowlist,新增未使用路由时失败;(3) 前端按钮 enabled 统一走 canCall(method, endpoint),输入直接来自 contract 的 required_permissions。此外 contract 目前不发布请求/响应 schema,DTO 都是 frozen dataclass,用 dataclasses.fields() 自动生成字段清单纳入后可把字段级漂移一并自动化(前提是先修 DTO 真相来源问题)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "API_BASE 硬编码三份,bootstrap.api_base 下发了却没人用",
|
|||
|
|
"location": "frontend/shared/api/endpoints.ts:1",
|
|||
|
|
"problem": "同一字符串在 backend/services/ui_service.py:10、frontend/shared/api/endpoints.ts:1、frontend/standalone-web/vite.config.ts:11 三处独立字面量;后端已通过 bootstrap 下发它,前端却只拿它拼提示文案,实际请求路径始终来自编译期常量。",
|
|||
|
|
"evidence": "backend/services/ui_service.py:10:\nAPI_BASE = \"/api/v1/plugins/extensions/astrbot_plugin_imagestudio\"\n\nfrontend/shared/api/endpoints.ts:1:\nexport const API_BASE = \"/api/v1/plugins/extensions/astrbot_plugin_imagestudio\"\n\nfrontend/standalone-web/vite.config.ts:11:\nbase: \"/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/\",\n\n前端拿到 bootstrap.api_base 后的唯一用途 App.vue:4330:\n: `已通过同源 API ${loadedBootstrap.api_base} 调用后端,写入动作会进入工作流并落库。`,\n\n实际请求路径 client.ts:299:\nconst url = new URL(`${API_BASE}/${trimmed}`, window.location.origin)",
|
|||
|
|
"impact": "插件名变更或 AstrBot 调整挂载前缀时,后端改一处即可,前端要改两处且其中一处是 vite 构建配置——漏掉 vite base 会导致所有 chunk 404,这类\"资产 404\"症状在 git 历史里已出现过(fix: load material library page without expiring asset chunk)。而 bootstrap.api_base 这个正确的解耦通道就在那里没被使用。",
|
|||
|
|
"fix": "client.ts 的 buildUrl 改为优先使用运行期值:模块内维护 let runtimeApiBase = API_BASE,bootstrap 成功后调 setApiBase(bootstrap.api_base)。vite base 保持编译期常量(静态资源必须编译期确定),但加一条 frontend/tests/ 断言 vite.config.ts 的 base 以 endpoints.ts 的 API_BASE 开头。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "CharacterProfile TS 类型缺 owner 四字段,generation_kind 联合含后端不产出的值",
|
|||
|
|
"location": "frontend/shared/types/profile.ts:18",
|
|||
|
|
"problem": "DTO to_dict() 输出 20 个字段,TS 声明 16 个,缺 owner_platform/owner_context_id/owner_user_id/owner_session_key;且 TS 的 generation_kind 联合比后端 SQL 实际产出的多一个 profile_card。",
|
|||
|
|
"evidence": "frontend/shared/types/profile.ts:18:\ngeneration_kind?: \"ungenerated\" | \"quick\" | \"detailed\" | \"profile_card\"\n\ndatabase/repositories/profile_repository.py:32-56(只可能产出三个值):\nCASE WHEN EXISTS (... profile_card.status = 'sealed') THEN 'detailed'\n WHEN EXISTS (... final_archive_history_id ...) THEN 'detailed'\n WHEN EXISTS (... character_fast ...) THEN 'quick'\n ELSE 'ungenerated' END AS generation_kind\n\nbackend/dto/profile_dto.py:42-45(TS 未声明):\n\"owner_platform\": self.owner_platform, \"owner_context_id\": self.owner_context_id,\n\"owner_user_id\": self.owner_user_id, \"owner_session_key\": self.owner_session_key,",
|
|||
|
|
"impact": "owner 四字段是多租户归属信息,随每个 profile 响应传输但前端不认识——将来要做\"仅显示我创建的角色\"这类过滤会有人再造一份类型。\"profile_card\" 是死枚举值,为它写的 switch 分支永远不执行,读代码的人会误以为存在该状态(实际只有 profile_category 才有 profile_card)。",
|
|||
|
|
"fix": "DTO 与 TS 字段做一次对齐(不需要的字段就从 to_dict() 里删,别只在 TS 侧省略);删掉 generation_kind 里的 \"profile_card\"。长期应由 DTO schema 导出自动生成。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "WorkflowRoundNextRequest.count 类型比后端窄,掩盖 cross_review 的 1..8 区间",
|
|||
|
|
"location": "frontend/shared/types/workflow-round.ts:49",
|
|||
|
|
"problem": "TS 把候选数收窄为字面量联合 4 | 5,后端接受的区间取决于轮次类型(primary 为 1..5,cross_review 为 1..8)。",
|
|||
|
|
"evidence": "frontend/shared/types/workflow-round.ts:48-52:\nexport interface WorkflowRoundNextRequest {\n count: 4 | 5\n idempotency_key: string\n model_route?: string\n}\n\nbackend/services/workflow_round_service.py:1159-1171:\ndef _requested_count(value: Any, round_kind: str) -> int:\n count = int(value) if value is not None else 4\n maximum = 8 if round_kind == \"cross_review\" else 5\n if count < 1 or count > maximum:\n raise ApiError(ErrorCode.VALIDATION_FAILED, \"round candidate count is outside the supported range\",\n {\"field\": \"count\", \"minimum\": 1, \"maximum\": maximum})",
|
|||
|
|
"impact": "目前是安全方向的收窄不会产生 400,但把\"业务上只允许 4 或 5\"这条产品规则固化在 TS 类型里而后端并不这么认为,两边对同一规则的理解已分叉;将来要支持\"生成 8 张交叉复核候选\"时类型会成为隐性阻碍,且改动方向不明确(因为没人记得后端其实早就支持)。",
|
|||
|
|
"fix": "改为 count: number,把\"4 或 5\"的选择限制放在 UI 控件层(下拉框选项),并在 WorkflowGenerationRoundDetail 或 round.metadata 里由后端下发 allowed_count_range: [min, max],让 UI 选项跟着轮次类型走。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"dimension": "横切关注点(错误处理 / 日志 / 可观测性 / 性能 / 一致性)",
|
|||
|
|
"file_written": "D:\\CODEX\\projects\\AstrBot ImageStudio\\审计\\findings\\C4-横切关注点.md",
|
|||
|
|
"summary": "本仓库在局部纪律很好:时间处理全部是 aware UTC(`utcnow()` 与裸 `datetime.now()` 各 0 次)、无裸 `except:`、无同步 `requests`/`time.sleep`/`subprocess` 混进 async、文件句柄统一走 `Path.read_bytes` 与上下文管理器、脱敏与 SSRF 防护都已实现。真正的问题集中在\"看不见\"和\"扛不住\"两条线:178 个 Python 源文件中只有 `worker_service.py` 一个配置了 logger(全仓 6 处 `logger.` 调用),同时 166 条路由全部经由 `handle_request` 吞掉非 ApiError 异常,只回一个异常类名且不打任何日志,线上内部错误无法定位。`ApiError` 被写成继承 Exception 的 dataclass,`str(exc)` 返回 args 元组 repr(含完整 details),而 main.py 有 40 处直接把 `{exc}` 拼进聊天回复,用户看到 Python 元组且 details 泄漏到群聊。关联链路三处全断:前端 `unwrapApiResponse` 剥掉 request_id、ApiClientError 不带 request_id、聊天侧的 `chatreq_` ID 从未落库。性能上 SQLite 未开 WAL 却跑 13 个并发 worker,每次 DB 操作新建连接(全仓 151 处 `.connect()`),空转约 43 次 connect/秒;`cost_service` 对 generation_images 无 LIMIT 全表扫描后再在 Python 侧过滤日期;高级素材详情有按 attempt 循环查子表的 N+1。此外 `route_registry` 声明了 166 条路由的权限矩阵但 dispatch 从不校验,只有 5 处散落的手工检查。",
|
|||
|
|
"files_read": [
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\main.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\dto\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\routes\\common.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\route_registry.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\worker_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\queue_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\log_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\cost_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\event_stream_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\dashboard_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\generation_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_style_set_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\material_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\api_discovery_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\settings_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_session_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\chat_workflow_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\integration_token_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\workflow_round_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\backend\\services\\advanced_material_worker_runtime.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\generation_errors.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\identity.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\audit_service.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\event_bus.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\remote_fetch.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\interop.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\core\\astrbot_provider_bridge.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\db_helper.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\job_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\advanced_material_repository.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\migrations\\0005_generation_tasks.sql",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\astrbot_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\adapters\\webui_adapter.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\engines\\image_provider.py",
|
|||
|
|
"D:\\CODEX\\projects\\AstrBot ImageStudio\\frontend\\shared\\api\\client.ts"
|
|||
|
|
],
|
|||
|
|
"findings": [
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "全后端 178 个源文件只有 1 个配置了日志,生产环境完全不可观测",
|
|||
|
|
"location": "backend/services/worker_service.py:5",
|
|||
|
|
"problem": "对整个真实源码树做统计,`import logging` / `getLogger` 只出现在 worker_service.py 一处,全仓 `logger.` 调用共 6 次,全部是 logger.exception,全部在这一个文件里。main.py(121KB)、backend/routes/*(32 文件)、backend/services/*(60+ 文件)、database/、engines/、core/、adapters/ 零日志。作为对照,全仓有 180 处 `except Exception`,其中 33 处直接 pass / return None / continue。",
|
|||
|
|
"evidence": "$ grep -rn \"import logging|getLogger|logger\\.\" --include=\"*.py\" main.py backend core database engines adapters\nbackend/services/worker_service.py:5:import logging\nbackend/services/worker_service.py:17:logger = logging.getLogger(__name__)\nbackend/services/worker_service.py:198: logger.exception(\nbackend/services/worker_service.py:235: logger.exception(\nbackend/services/worker_service.py:253: logger.exception(\nbackend/services/worker_service.py:290: logger.exception(\nbackend/services/worker_service.py:337: logger.exception(\nbackend/services/worker_service.py:352: logger.exception(\n(以上 8 行为全仓全部匹配)",
|
|||
|
|
"impact": "只有队列 worker 崩溃这一条路径会在 AstrBot 日志里留痕。生成失败、素材识别失败、配置写入失败、媒体签名失败、Provider 探测失败全部无声。用户报「点了没反应」时运维手上除了 logs/errors 里那条 error_message 文本外没有任何上下文(无栈、无入参、无 provider、无 request_id)。这也是本报告多条问题无法自愈的根因。",
|
|||
|
|
"fix": "1) 新建 core/logging.py 导出 get_logger(name),内部优先桥接 astrbot.api.logger,回退 logging.getLogger(f\"imagestudio.{name}\");2) 在 backend/routes/common.py:handle_request 的 except 中强制 logger.exception;3) 至少给 generation_service、advanced_material_service、advanced_style_set_service、model_provider_runtime、astrbot_provider_bridge 这 5 个含外部 IO 的模块补 logger.warning/exception,把当前 33 处静默 pass 改成 logger.debug/warning + 原有兜底逻辑。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "handle_request 吞掉全部非 ApiError 异常且不记日志,166 条路由内部错误彻底静默",
|
|||
|
|
"location": "backend/routes/common.py:12",
|
|||
|
|
"problem": "全部 32 个 route 文件、166 条路由都通过 handle_request 包装 handler。它 except Exception 后交给 response_from_exception,后者对任何非 ApiError 异常一律返回固定文案 + 仅异常类名,栈、消息、上下文全部丢弃,且没有任何日志输出。",
|
|||
|
|
"evidence": "# backend/routes/common.py:12-16\ndef handle_request(ctx: RequestContext, callback: Callable[[], Any]) -> dict[str, Any]:\n try:\n return success_response(callback(), request_id=ctx.request_id)\n except Exception as exc:\n return response_from_exception(exc, request_id=ctx.request_id)\n\n# backend/dto/common.py:75-83\ndef response_from_exception(exc: Exception, request_id: str | None = None) -> dict[str, Any]:\n if isinstance(exc, ApiError):\n return exc.to_response(request_id=request_id)\n return error_response(\n error_code=ErrorCode.UNKNOWN_ERROR,\n message=\"服务器内部错误\",\n details={\"exception\": exc.__class__.__name__}, # 只有类名\n request_id=request_id,\n )",
|
|||
|
|
"impact": "任何 KeyError / TypeError / sqlite3.OperationalError / json.JSONDecodeError 都变成前端一句「服务器内部错误」,服务端没有任何记录。配合全仓无日志,这类 bug 在生产中不可能被诊断,只能靠复现。log_service 提供的 logs/errors 端点只覆盖 generation_tasks 和 integration_logs 两张表,完全不含 API 层错误。",
|
|||
|
|
"fix": "在 handle_request 中分层捕获:except ApiError 记 info(含 ctx.request_id 与 error_code),except Exception 记 logger.exception(\"api internal error request_id=%s actor=%s\", ctx.request_id, ctx.actor_id) 后再返回兜底响应;同时在 response_from_exception 的 details 里补 {\"request_id\": request_id},让前端展示的 ID 可直接反查日志。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P1-严重",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "ApiError 是继承 Exception 的 dataclass,str(exc) 返回含完整 details 的元组 repr,main.py 有 40 处直接回显给聊天用户",
|
|||
|
|
"location": "backend/dto/common.py:29",
|
|||
|
|
"problem": "ApiError 用 @dataclass(frozen=True) 继承 Exception。Exception.__new__ 会把构造时的位置参数塞进 self.args,而 Exception.__str__ 直接返回 args 的 repr。全仓 169 处 ApiError(ErrorCode.X, \"消息\", {...}) 全部用位置参数构造,因此 str(exc) 得到的是 (<ErrorCode.X: 'X'>, '消息', {details 全量}) 这个元组的 repr 而不是 message。main.py 有 40 处 f\"...{exc}\" 直接拼进聊天回复。",
|
|||
|
|
"evidence": "# backend/dto/common.py:29-33\n@dataclass(frozen=True)\nclass ApiError(Exception):\n error_code: ErrorCode\n message: str\n details: dict[str, Any] | None = None\n\n# 本机 Python 实测输出:\nstr(exc) => \"(<ErrorCode.VALIDATION_FAILED: 'VALIDATION_FAILED'>, '内容审核未通过,未创建生成任务', {'a': 1})\"\n\n# main.py:1591\nawait self._send_proactive(context.message_origin, f\"❌ 单图生成失败\\n请求:{request_id}\\n错误:{exc}\", [])\n# main.py:669\nyield event.plain_result(f\"❌ 节点生成失败:{exc}\")",
|
|||
|
|
"impact": "1) 可用性:QQ/微信用户收到的是 `❌ 节点生成失败:(<ErrorCode.GENERATION_FAILED: 'GENERATION_FAILED'>, '...', {...})`,中文提示被 Python 元组语法淹没。2) 信息泄漏:details 含 provider_id、model_id、route_name、body_preview(engines/image_provider.py:253,289,324 会塞入上游响应体前 500 字节)、内部 field 名等,直接推送到群聊;core/generation_errors._safe_details 的脱敏在这条路径上完全没生效。同仓 worker_service.py:514-534 已为 ApiError 写了正确的 _safe_error_message,说明作者知道这个坑但聊天路径没复用。",
|
|||
|
|
"fix": "1) 给 ApiError 加显式 __str__:`def __str__(self): return f\"{self.message} [{self.error_code.value}]\"`。2) 把 main.py 的 40 处 {exc} 统一改为调用共享的 user_facing_error(exc)(可直接把 worker_service._safe_error_message 提到 backend/dto/common.py 复用),确保聊天侧永远只看到 message 不看到 details。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "SQLite 未开启 WAL,却跑 13 个并发写线程,全靠 30 秒 busy_timeout 硬扛",
|
|||
|
|
"location": "database/db_helper.py:159",
|
|||
|
|
"problem": "默认后端是 SQLite。connect() 设了 foreign_keys 和 busy_timeout 但没设 journal_mode=WAL,全仓 grep PRAGMA 只有这两行 + 一处 PRAGMA table_info。同时系统默认启动 5 个 round worker(workflow_round_worker_runtime.py:25)+ 8 个 advanced material worker(advanced_material_worker_runtime.py:22),共 13 个后台线程持续读写同一个 db 文件。",
|
|||
|
|
"evidence": "# database/db_helper.py:159-171\ndef connect(self):\n if self.backend == \"sqlite\":\n db_path = self._resolve_sqlite_path()\n db_path.parent.mkdir(parents=True, exist_ok=True)\n # Parallel image tasks briefly overlap while recording queue,\n # node, image, and audit rows. Give SQLite enough time to hand the\n # writer lock to the next task instead of failing immediately.\n conn = sqlite3.connect(db_path, timeout=30.0)\n conn.row_factory = sqlite3.Row\n conn.execute(\"PRAGMA foreign_keys = ON\")\n conn.execute(\"PRAGMA busy_timeout = 30000\")\n return conn\n return self._connect_postgresql()",
|
|||
|
|
"impact": "在默认 rollback journal 模式下,任何一个写事务(audit_logs 插入、generation_tasks 状态更新、generation_images 落库)会阻塞所有读。13 个 worker + 前端轮询 + 每次 audit_service.record 都是独立写事务,实际表现是 UI 列表接口在批量生成期间随机卡顿数百毫秒到数秒;极端时触发 30 秒超时抛 sqlite3.OperationalError: database is locked,而该异常会被 handle_request 吞成「服务器内部错误」且无日志。",
|
|||
|
|
"fix": "在 connect() 中补 `conn.execute(\"PRAGMA journal_mode = WAL\")` 和 `conn.execute(\"PRAGMA synchronous = NORMAL\")`;并在 initialize() 里配置 PRAGMA wal_autocheckpoint 避免 WAL 文件无限增长。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "每次队列 claim 开 2 个新连接;13 个 worker 空转时约 43 次 connect + COUNT(*) / 秒",
|
|||
|
|
"location": "backend/services/worker_service.py:207",
|
|||
|
|
"problem": "全仓 151 处 self.database.connect(),没有连接池,每个仓储方法自己 connect/execute/close。worker 轮询链路放大了这个模式:_claim_next 每轮调 list_jobs,job_repository.list() 里额外跑一次 COUNT(*)(worker 只用 .items,total 被丢弃);然后每个候选调 claim_queued_task,它先 get_job(连接 1)再 update_if_status(连接 2)。多 worker 抢同一批 queued 行时失败的 claim 会继续尝试下一个候选,最坏 25 候选 × 2 = 50 连接/轮。",
|
|||
|
|
"evidence": "# backend/services/worker_service.py:207-221\ndef _claim_next(self, ctx: RequestContext) -> GenerationTask | None:\n queued = self.queue_service.list_jobs(status=\"queued\", page=1, page_size=self.claim_batch_size).items\n for candidate in queued:\n ...\n claimed = self.queue_service.claim_queued_task(candidate.task_id, ctx, worker_id=self.worker_id)\n\n# backend/services/queue_service.py:289-294\nbefore = self.get_job(task_id) # 连接 1\nif before.status != \"queued\":\n return None\nafter = self.repository.update_if_status(...) # 连接 2\n\n# database/repositories/job_repository.py:141-154\ntotal_row = conn.execute(f\"SELECT COUNT(*) AS total FROM generation_tasks {where_clause}\", tuple(params)).fetchone()\nrows = conn.execute(f\"SELECT {TASK_COLUMNS} FROM generation_tasks {where_clause} ORDER BY created_at DESC LIMIT ...\")\n\n# 轮询频率:round pool interval_seconds=0.25 × 5 = 20/s;advanced pool 0.35 × 8 ≈ 23/s,空载合计约 43 次/秒",
|
|||
|
|
"impact": "完全空闲的实例也持续消耗 CPU 和 IO;配合无 WAL 会显著抬高真实任务的写锁等待概率。在 PostgreSQL 后端下更糟——每次 connect() 都是完整 TCP + 认证握手,43 次/秒的建连会直接打满 max_connections。",
|
|||
|
|
"fix": "1) 给 JobRepository 加 list_items_only(...) 或 with_total 参数,worker 路径跳过 COUNT(*);2) 把 claim_queued_task 的 get_job + update_if_status 合成一条 UPDATE ... WHERE task_id=? AND status='queued'(PG 用 RETURNING,SQLite 用 changes()),单连接单语句完成 CAS;3) 在 Database 上加 threading.local() 连接复用或最小连接池;4) worker 连续 N 轮 idle 后指数退避到 2 秒。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "前端 direct-fetch 路径完全没有超时,与 plugin-bridge 路径行为不一致",
|
|||
|
|
"location": "frontend/shared/api/client.ts:81",
|
|||
|
|
"problem": "同一个 request<T>() 有两条传输路径。plugin-bridge 路径给每个调用都套了 withClientTimeout(GET 15s / POST 30s / 长任务 LONG_TASK_TIMEOUT_MS),而 direct-fetch(独立部署的 standalone-web)路径的 fetch() 没有传 signal,也没有任何 AbortController。",
|
|||
|
|
"evidence": "// frontend/shared/api/client.ts:80-93 —— 无超时\ntry {\n response = await fetch(url, {\n method: options.method ?? \"GET\",\n headers: options.body === undefined ? undefined : { \"Content-Type\": \"application/json\" },\n body: options.body === undefined ? undefined : JSON.stringify(options.body),\n })\n} catch (err) {\n throw new ApiClientError(\"network_unreachable\", \"无法连接 ImageStudio 后端接口\", { ... })\n}\n\n// frontend/shared/api/client.ts:131-137 —— 有超时\nif (method === \"GET\" && typeof bridge.apiGet === \"function\") {\n return withClientTimeout(\n bridge.apiGet<...>(endpoint, compactQuery(options.query)),\n PLUGIN_BRIDGE_GET_TIMEOUT_MS,\n \"AstrBot Plugin Page GET request timed out\",\n )\n}",
|
|||
|
|
"impact": "standalone-web 模式下只要后端某个请求 hang 住(结合 SQLite 写锁等待,或上游 provider 180 秒超时),对应 UI 按钮会永远停在 loading 状态,没有任何反馈也无法取消。同一份代码在 AstrBot 内嵌模式下 15 秒就报错,两种部署形态行为不一致,问题在 standalone 上更难复现和归因。",
|
|||
|
|
"fix": "在 direct-fetch 分支复用同一套超时常量:创建 AbortController,timeoutMs 取 GET 用 PLUGIN_BRIDGE_GET_TIMEOUT_MS、其余用 pluginBridgePostTimeout(endpoint),window.setTimeout 触发 controller.abort(),fetch 传 signal,finally 清 timer;并把 AbortError 映射成与 bridge 一致的 ApiClientError(\"bridge_timeout\", ...)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "request_id 三条关联链路全断:前端剥掉、错误对象不带、聊天侧的 ID 从未落库",
|
|||
|
|
"location": "frontend/shared/api/client.ts:205",
|
|||
|
|
"problem": "后端为每个请求生成 req_{uuid4().hex} 并写进 audit_logs.request_id,响应体也带了 request_id。但 (1) 前端 unwrapApiResponse 成功时直接 current = current.data 丢弃 request_id,失败时构造 ApiClientError(code, message, details) 同样不带(request_id 在响应顶层而非 details 里);(2) 聊天命令自己造了一套独立 ID(chatreq_ / chatfast_ / chathires_),只用于回显给用户,从未传入 RequestContext 也从未写入任何表。",
|
|||
|
|
"evidence": "// frontend/shared/api/client.ts:205-210 —— request_id 在这里蒸发\nif (isApiResponse(current)) {\n if (!current.ok) {\n throw new ApiClientError(current.error_code, current.message, current.details ?? {})\n }\n current = current.data\n continue\n}\n\n# main.py:1210-1218 —— 造一个只给人看的 ID\nrequest_id = f\"chatreq_{uuid4().hex}\"\ntask = asyncio.create_task(self._run_single_image_background(command, context, request_id))\nyield event.plain_result(\"✅ 单图请求已进入后台执行\\n\\n\" f\"请求:{request_id}\\n尺寸:{command.size}\\n数量:{command.count}\\n\" ...)\n\n# main.py:1579-1584 —— request_id 只参与字符串拼接,没进 submit_single_image\nasync def _run_single_image_background(self, command, context, request_id: str) -> None:\n result = await asyncio.to_thread(self.chat_workflow_service.submit_single_image, command, context)\n f\"✅ 单图生成完成\\n请求:{request_id}\\n任务:{generation.get('task_id') or '—'}\\n\"",
|
|||
|
|
"impact": "用户报障时给出的唯一凭据(Web 端错误提示 / 聊天里的「请求:chatreq_xxx」)在 audit_logs、generation_tasks、workflow_history 里一条都查不到。配合全仓无日志和异常被吞,线上问题定位手段为零。ID 前缀也不一致:core/identity.new_request_id() 用 req_,聊天侧用 chatreq_/chatfast_/chathires_,三者互不相干。",
|
|||
|
|
"fix": "1) client.ts:ApiClientError 增加 readonly requestId?: string,在 unwrapApiResponse 里从 current.request_id 取;UI 错误 toast 追加「(追踪号 req_xxx)」。2) main.py:删掉平行 ID,先构造 ctx = RequestContext(...),把 ctx.request_id 同时用于回显和下游调用。3) RequestContext 支持从入站 X-Request-Id 头继承(当前 _request_context_from_astrbot_request 只取 username/user_id/ip)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "cost_service 对 generation_images 做无 LIMIT 全表扫描,days 过滤放在 Python 侧",
|
|||
|
|
"location": "backend/services/cost_service.py:186",
|
|||
|
|
"problem": "GET costs/usage 是 read 权限的普通查询接口。_image_rows() 的 SQL 没有 LIMIT、没有时间条件,把 generation_images 全表(含 metadata_json 大字段)拉进内存,days 过滤在 Python 循环里做;_advanced_material_attempt_rows() 和 _advanced_style_validation_case_rows() 同样无 LIMIT。每行还要跑一次 json.loads。",
|
|||
|
|
"evidence": "# backend/services/cost_service.py:186-198\nwhere_sql = f\"WHERE {' AND '.join(where)}\" if where else \"\"\nconn = self.database.connect()\ntry:\n return conn.execute(\n f\"\"\"\n SELECT image_id, task_id, profile_id, workflow_id, node_id,\n model_route, metadata_json, created_at\n FROM generation_images\n {where_sql}\n ORDER BY created_at DESC\n \"\"\", # <- 无 LIMIT,无 created_at 过滤\n tuple(params),\n ).fetchall()\nfinally:\n conn.close()\n\n# backend/services/cost_service.py:40-44 —— 时间过滤在拉完全表之后\nfor row in self._image_rows(...):\n created_at = _parse_datetime(row[\"created_at\"])\n if since and created_at and created_at < since:\n continue\n metadata = _json_loads(row[\"metadata_json\"])",
|
|||
|
|
"impact": "仪表盘/成本页每次刷新都把全部历史图片记录 + 全部 metadata JSON 载入内存并逐行解析。跑几个月后 generation_images 到十万级时,单次请求就是几百 MB 内存峰值 + 秒级 CPU,且这次查询会长时间持有 SQLite 读锁(无 WAL 时阻塞写)。days=7 这类查询完全没有得到任何加速。",
|
|||
|
|
"fix": "把时间窗口下推到 SQL:`if since is not None: where.append(f\"created_at >= {placeholder}\"); params.append(since.isoformat())`,并加 `LIMIT {placeholder}`(如 20000,超出时在响应里标 truncated=True)。长期建议在写入侧把 usage 的 token 数拆成独立数值列或建 generation_cost_facts 汇总表,避免每次聚合都解析 JSON。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "高级素材详情存在 N+1:按 attempt 循环查子表,每次新开一个数据库连接",
|
|||
|
|
"location": "backend/services/advanced_material_service.py:228",
|
|||
|
|
"problem": "get_import / list_attempts 对每个 attempt 单独调 list_style_validation_cases(attempt.attempt_id),而该仓储方法自己 connect/close。attempt 有 N 条就产生 N 次独立连接 + N 次查询。advanced_style_set_service.py:1215-1216 和 509-510 还有两层嵌套。另外 advanced_material_service.py:258-263 把 get_material 调了两次(条件里一次、取值里一次),又是两个连接还引入 TOCTOU。",
|
|||
|
|
"evidence": "# backend/services/advanced_material_service.py:228-239\nattempt_payloads = [\n {\n **attempt.to_dict(),\n **(\n {\"validation_cases\": [\n case.to_dict()\n for case in self.repository.list_style_validation_cases(attempt.attempt_id) # N+1\n ]} if is_style_set else {}\n ),\n }\n for attempt in attempts\n]\n\n# database/repositories/advanced_material_repository.py:1147-1157\ndef list_style_validation_cases(self, attempt_id: str) -> list[AdvancedStyleValidationCase]:\n conn = self.database.connect() # 每次调用一条新连接\n try:\n rows = conn.execute(...).fetchall()\n finally:\n conn.close()\n\n# backend/services/advanced_material_service.py:258-263 —— get_material 调用两次\n\"accepted_material\": (\n self.repository.get_material(record.accepted_material_id).to_dict()\n if record.accepted_material_id and self.repository.get_material(record.accepted_material_id)\n else None\n),",
|
|||
|
|
"impact": "一个做过 20 轮验证的风格集导入,打开详情页就是 20+ 次建连 + 20+ 次查询(PG 后端下是 20+ 次 TCP 握手)。这是前端轮询的接口,会被反复调用。",
|
|||
|
|
"fix": "加批量方法 list_style_validation_cases_by_attempts(attempt_ids),用 `WHERE attempt_id IN (...)` 一次查完再按 attempt_id 分组返回 dict;调用方改为先拿 grouped 再取。accepted_material 处改成先 `material = self.repository.get_material(...) if record.accepted_material_id else None` 再用局部变量。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "ChatSessionStore 构造期无保护地做 naive/aware datetime 比较,脏持久化文件会让插件启动直接崩",
|
|||
|
|
"location": "backend/services/chat_session_service.py:76",
|
|||
|
|
"problem": "__init__ 无保护地调 self._load()。_load() 内部对每一行用 except (KeyError, TypeError, ValueError): continue 做了容错,但 datetime.fromisoformat(\"2026-07-27T10:00:00\")(无时区偏移)不会抛异常,它会安静地返回 naive datetime 存进 self._items。真正爆炸的是循环外的 self._prune()(第 76 行),那里拿 naive 和 aware 的 now 比较抛 TypeError,而这个位置不在任何 try 内。put() 里的 min(..., key=lambda item: item.created_at) 同理。",
|
|||
|
|
"evidence": "# backend/services/chat_session_service.py:57-76\ndef _prune(self) -> None:\n now = datetime.now(timezone.utc) # aware\n for key in [key for key, item in self._items.items() if item.expires_at <= now]: # <- 这里炸\n self._items.pop(key, None)\n\ndef _load(self) -> None:\n try:\n payload = json.loads(self.persistence_path.read_text(encoding=\"utf-8\"))\n except (OSError, ValueError, TypeError):\n return\n for item in payload if isinstance(payload, list) else []:\n try:\n session = ChatSession(\n key=str(item[\"key\"]), action=str(item[\"action\"]), payload=dict(item.get(\"payload\") or {}),\n expires_at=datetime.fromisoformat(item[\"expires_at\"]), created_at=datetime.fromisoformat(item[\"created_at\"]),\n ) # naive 字符串在这里不会报错\n self._items[session.key] = session\n except (KeyError, TypeError, ValueError):\n continue\n self._prune() # 无 try 保护\n\n# 第 32 行:self._load() 在 __init__ 中无保护调用",
|
|||
|
|
"impact": "只要 data/chat_sessions.json 里出现一个不带时区的时间戳(旧版本写入、手工编辑、外部工具生成、跨环境迁移),ChatSessionStore() 构造就抛 TypeError,整个聊天服务乃至插件加载失败。作者显然想让这个文件「坏了就忽略」(_load 开头的 except 已表明意图),但这层防护被绕过了。",
|
|||
|
|
"fix": "在解析处强制归一化时区:定义 `def _aware(value): parsed = datetime.fromisoformat(value); return parsed if parsed.tzinfo else parsed.replace(tzinfo=timezone.utc)`,然后 expires_at=_aware(...)、created_at=_aware(...)。仓库里 cost_service._parse_datetime、worker_service._retry_deferred、integration_token_service._parse_time 三处已经写对了这段逻辑,应抽成 core/ 共享工具消除漏网。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "coupling",
|
|||
|
|
"title": "路由权限矩阵声明了 166 条路由的权限但 dispatch 从不校验,只有 5 处散落的手工检查",
|
|||
|
|
"location": "adapters/astrbot_adapter.py:187",
|
|||
|
|
"problem": "route_registry.py 定义了完整的 PERMISSION_MATRIX(read/write/config/materials/retry/public_token),166 条路由每条都标了 permission,RouteSpec.to_dict() 还会把 auth_required / required_permissions 导出到 manifest。但唯一的请求入口 dispatch_extension_request 拿到 RouteMatch.permission 之后直接调 handler,从不做任何校验。",
|
|||
|
|
"evidence": "# adapters/astrbot_adapter.py:179-187\nmatch = resolve_route(method, path)\nif match is None:\n return error_response(error_code=ErrorCode.RESOURCE_NOT_FOUND, ...)\nreturn match.handler(*_handler_args(match, ctx, query or {}, payload or {}))\n# ^ match.permission 被解析出来了,但这里没有任何 ctx.has_permission 判断\n\n# 全仓 has_permission 只出现 5 次,集中在 2 个文件:\n# backend/routes/image_routes.py:42,70,90,156\n# backend/routes/job_routes.py:36",
|
|||
|
|
"impact": "当前没出安全事故是因为 main.py:2337 给所有已认证用户一次性授予了全部 4 个权限(permissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()),矩阵在运行时恒真。但这意味着:(1) 这套权限声明是纯装饰,读代码会误以为已生效;(2) 一旦将来 AstrBot 暴露细粒度角色、或有人收窄 BRIDGE_PERMISSIONS,166 条路由里只有 5 处会真正拦截,其余 161 条静默放行。典型的「策略集中声明、执行处处缺失」。",
|
|||
|
|
"fix": "在 dispatch_extension_request 中统一执行:拿 requirement = permission_requirement(match.permission),若 requirement[\"auth_required\"] 且 not ctx.is_authenticated 返回 UNAUTHORIZED;遍历 requirement[\"required_permissions\"],任一 not ctx.has_permission(needed) 则返回 FORBIDDEN(details 带 required_permission)。然后把 image_routes / job_routes 里那 5 处手工检查降级为纯业务判断或删除。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P2-中等",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": "主动通知全链路静默失败:签名失败返回 None、发送失败返回 False、terminal 回调静默 return,一行日志都没有",
|
|||
|
|
"location": "backend/services/workflow_round_worker_runtime.py:262",
|
|||
|
|
"problem": "轮次完成后主动推图到聊天的链路上有三个连续的静默吞异常点:_signed_url 签名失败返回 None(图片 URL 变 None 静默消失)、_send_proactive 发送失败返回 False(调用方也不看返回值)、worker_service._notify_terminal 静默 return。第三处尤其可疑:worker_service.py 是全仓唯一有 logger 的文件,其他 6 个 except 全部 logger.exception,唯独 _notify_terminal 是例外。",
|
|||
|
|
"evidence": "# backend/services/workflow_round_worker_runtime.py:257-263\ndef _signed_url(media_service: MediaService, uri: Any) -> str | None:\n if not isinstance(uri, str) or not uri:\n return None\n try:\n return media_service.sign_media(uri).get(\"url\")\n except Exception:\n return None # 签名失败 -> 通知里图片 URL 变 None -> 图片静默消失\n\n# main.py:1665-1670\n result = sender(message_origin, message)\n if inspect.isawaitable(result):\n await result\n return True\nexcept Exception:\n return False # 发送失败 -> 返回 False -> 调用方也不看返回值\n\n# backend/services/worker_service.py:374-380\ntry:\n self.terminal_callback(claimed_task, current_task, exc)\nexcept Exception:\n # Queue state is authoritative. ...\n return # 同文件其余 6 处失败都 logger.exception,唯独这里静默",
|
|||
|
|
"impact": "用户提交生成后一直收不到结果推送,服务端三个环节都「成功」返回(None / False / 静默 return),队列状态显示 succeeded。这类问题在没有日志的前提下只能靠加打印重现。",
|
|||
|
|
"fix": "三处都改为 logger.warning(...) / logger.exception(...) 后再返回兜底值。_send_proactive 的返回值当前在 _deliver_round_terminal_notification 中被忽略,应至少在返回 False 时记一条带 round_id 和 message_origin 的 warning。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "_Psycopg2ConnectionAdapter.execute 每次新建 cursor 且从不关闭",
|
|||
|
|
"location": "database/db_helper.py:305",
|
|||
|
|
"problem": "psycopg2 回退适配器的 execute() 每次调用都 self.connection.cursor() 新建游标,返回后没有任何地方 close() 它。适配器只暴露 commit / rollback / close,没有 cursor 生命周期管理。",
|
|||
|
|
"evidence": "# database/db_helper.py:301-317\nclass _Psycopg2ConnectionAdapter:\n def __init__(self, connection):\n self.connection = connection\n\n def execute(self, query: str, params: tuple[Any, ...] | list[Any] | None = None):\n cursor = self.connection.cursor()\n cursor.execute(query, tuple(params or ()))\n return cursor # 调用方拿去 .fetchall()/.fetchone() 后就丢弃,永不 close",
|
|||
|
|
"impact": "单个连接内每执行一条 SQL 就泄漏一个游标对象,直到 conn.close() 才随连接释放。因为本仓库是「每次操作新建连接」,泄漏窗口很短所以实际危害有限;但一旦按连接复用建议引入连接池,这个泄漏会立刻变成真实的内存和服务端资源问题。而且 psycopg2 只是 import psycopg 失败时的回退路径,问题更难被发现。",
|
|||
|
|
"fix": "返回一个包装对象,在 fetchall/fetchone 取完后自动 close;或让 execute 直接返回结果集。至少加 try/except 在 execute 抛错时 cursor.close() 后再 raise。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "hygiene",
|
|||
|
|
"title": ".part 临时文件在 replace() 失败时残留,无清理",
|
|||
|
|
"location": "backend/services/material_service.py:326",
|
|||
|
|
"problem": "两处上传落盘都用「先写 .part 再 replace」的原子写模式,但 except OSError 里只抛 ApiError,不删除已经写出的 .part 文件。如果 write_bytes 成功而 replace 失败(目标被占用、跨设备、权限、磁盘满导致部分写入后失败),临时文件永久留在存储目录。advanced_material_service.py:1584-1594 是同一段代码的复制。",
|
|||
|
|
"evidence": "# backend/services/material_service.py:326-336\ntry:\n persisted_path.parent.mkdir(parents=True, exist_ok=True)\n temporary_path = persisted_path.with_suffix(f\"{persisted_path.suffix}.part\")\n temporary_path.write_bytes(content)\n temporary_path.replace(persisted_path)\nexcept OSError as exc:\n raise ApiError(\n ErrorCode.STORAGE_WRITE_FAILED,\n \"material image could not be stored\",\n {\"uri\": image_uri, \"exception_type\": exc.__class__.__name__},\n ) from None\n # temporary_path 从未清理",
|
|||
|
|
"impact": "磁盘写失败(最典型是磁盘满)时,每次重试都留下一个和原图等大的 .part 残骸,把本就紧张的空间进一步耗尽,形成正反馈。orphan_media_service 扫描的是 storage:// URI 与 DB 的差集,识别不了 .part 后缀的孤儿。",
|
|||
|
|
"fix": "把 temporary_path 的定义提到 try 之前,在 except 分支加 temporary_path.unlink(missing_ok=True) 再 raise。建议把这段抽成 core/storage.py 的 atomic_write_bytes(path, content),两处共用。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "perf",
|
|||
|
|
"title": "每个请求重排 166 条路由 + 逐条编译正则 + inspect.signature,全程无缓存",
|
|||
|
|
"location": "adapters/astrbot_adapter.py:47",
|
|||
|
|
"problem": "唯一的请求入口每次都做纯静态的重复计算:ROUTES 是模块级常量,路由表和 handler 签名在进程生命周期内不变,但每个请求都重跑一遍 sorted() + 逐候选 re.findall/re.sub/re.match + inspect.signature。",
|
|||
|
|
"evidence": "# adapters/astrbot_adapter.py:44-55\ndef resolve_route(method: str, path: str) -> RouteMatch | None:\n ...\n candidates = sorted( # 每请求排序\n (route for route in ROUTES if route.method == normalized_method),\n key=_route_specificity,\n reverse=True,\n )\n for route in candidates:\n params = _match_endpoint(route.endpoint, endpoint) # 每候选编译一次正则\n\n# adapters/astrbot_adapter.py:224-230\ndef _match_endpoint(pattern: str, endpoint: str) -> dict[str, str] | None:\n param_names = re.findall(r\"\\{([^{}]+)\\}\", pattern)\n regex = \"^\" + re.sub(r\"\\{([^{}]+)\\}\", r\"([^/]+)\", pattern) + \"$\"\n match = re.match(regex, endpoint)\n\n# adapters/astrbot_adapter.py:240\nfor index, parameter in enumerate(inspect.signature(match.handler).parameters.values()):\n\n# 本机实测 2000 次循环:resolve_route + _handler_args 约 0.178 ms/次,路由总数 166",
|
|||
|
|
"impact": "0.18 ms 本身不致命,但这是所有 166 个端点的固定入口开销,且是纯浪费。前端存在多处轮询(队列状态、导入进度、事件流),这部分开销会持续累加;随路由表增长呈线性劣化。",
|
|||
|
|
"fix": "1) 模块加载时预编译 dict[str, list[tuple[re.Pattern, RouteSpec]]](按 method 分桶、已按 _route_specificity 排好序);2) _load_handler 加 @functools.lru_cache(maxsize=None);3) handler 的参数绑定方案在导入时算一次并缓存到 RouteSpec 上,避免每请求 inspect.signature。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "log_service / cost_service 的 _json_loads 无异常保护,单条脏数据可让整个日志端点 500",
|
|||
|
|
"location": "backend/services/log_service.py:297",
|
|||
|
|
"problem": "日志查询接口对每一行的 JSON 列做 json.loads,没有任何保护。只要 integration_logs.request_json / audit_logs.metadata_json / generation_images.metadata_json 里有一条被截断或写坏的记录,整个列表接口就抛 json.JSONDecodeError。",
|
|||
|
|
"evidence": "# backend/services/log_service.py:297-300\ndef _json_loads(value: str | None) -> Any:\n if not value:\n return None\n return json.loads(value) # 无 try\n\n# 调用点(每行一次):\n# log_service.py:162-163 \"request\": _json_loads(row[\"request_json\"]) or {}, \"response\": _json_loads(row[\"response_json\"]) or {}\n# log_service.py:284-286 before_snapshot / after_snapshot / metadata",
|
|||
|
|
"impact": "日志查询是排障工具,恰恰应该对脏数据最宽容。现在一条坏记录就能让 logs/integrations 和 security/audit-logs 两个端点永久 500,而且因为异常被 handle_request 吞成「服务器内部错误」、因为全仓没有日志,运维根本不知道是哪条记录坏了——排障工具自己坏了且无法排障。",
|
|||
|
|
"fix": "改为 try: return json.loads(value) except (ValueError, TypeError): return {\"_parse_error\": True, \"raw_preview\": str(value)[:200]}。cost_service._json_loads 同理(它返回 dict,失败时返回 {} 即可)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"severity": "P3-轻微",
|
|||
|
|
"category": "bug",
|
|||
|
|
"title": "EventBus.publish 不隔离订阅者异常,一个 handler 抛错会中断发布方并跳过后续 handler",
|
|||
|
|
"location": "core/event_bus.py:48",
|
|||
|
|
"problem": "publish 在循环里直接调 handler(event),没有 try/except。任一订阅者抛异常会向上冒泡到发布方(generation_service._publish,generation_service.py:1701-1703),同时后续 handler 全部不执行。",
|
|||
|
|
"evidence": "# core/event_bus.py:45-50\ndef publish(self, event_type: str, payload: dict[str, Any] | None = None) -> Event:\n event = Event(event_type=_normalize_event_type(event_type), payload=dict(payload or {}))\n self._history.append(event)\n for handler in [*self._handlers.get(event.event_type, []), *self._handlers.get(\"*\", [])]:\n handler(event) # 无隔离\n return event\n\n# backend/services/generation_service.py:1701-1703\ndef _publish(self, event_type: str, payload: dict[str, Any]) -> None:\n if self.event_bus:\n self.event_bus.publish(event_type, payload)",
|
|||
|
|
"impact": "当前全仓 grep 不到任何 event_bus.subscribe(...) 调用(事件总线只被当作 deque(maxlen=500) 的历史缓冲,供 event_stream_service.recent_events 轮询读取),所以这条现在不会触发。但 _publish 已被 generation_service 在生成成功/失败等关键路径调用,一旦将来接入第一个订阅者(例如 webhook 推送),一个下游 HTTP 超时就会把生成流程本身打断。",
|
|||
|
|
"fix": "把 handler(event) 包进 try/except Exception 并 logger.exception(\"event handler failed event_type=%s event_id=%s\", ...)。顺带修掉 unsubscribe(event_bus.py:41-43)对未订阅类型会因 defaultdict 凭空建键的小问题。"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
],
|
|||
|
|
"verdicts": [
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "backend/services/workflow_round_service.py:81 (_ensure_round_creation_allowed,定义在 :1129-1156)",
|
|||
|
|
"reason": "结论:误报。发现者引用的代码片段本身真实存在(workflow_round_service.py:115-117 的幂等重放分支、:928-940 无条件写 status=\"queued\"、database/repositories/workflow_round_repository.py:1355-1375 的 _ensure_writable_round_row confirmed 只读保护、:416-431 的裸 UPDATE update_round),但它漏读了 create_round 的第一行守卫,导致整条攻击链对 confirmed 轮次不可达。\n\n1) create_round 在读取 idempotency_key 之前就先跑了节点级锁:\n```python\nnode, workflow = self._node_and_workflow(node_id, ctx)\nself._ensure_round_creation_allowed(node, workflow) # line 81\n...\nidempotency_key = _optional_text(payload.get(\"idempotency_key\")) # line 90\nif idempotency_key:\n existing = self.round_repository.get_round_by_idempotency_key(idempotency_key) # line 92\n```\n而 _ensure_round_creation_allowed(:1145-1156)明确拒绝已定稿节点:\n```python\nif node.status == \"selected\" or node.selected_image_id or (\n progression is not None and node.node_id in progression.completed_node_ids\n):\n raise ApiError(\n ErrorCode.VALIDATION_FAILED,\n \"confirmed workflow node is locked; create a branch before adding another round\",\n {\"node_id\": node.node_id, \"branch_name\": node.branch_name, \"required_action\": \"create_branch\"},\n )\n```\n即:目标节点一旦定稿,请求在第 81 行就抛 VALIDATION_FAILED,根本走不到第 92 行的 idempotency 查表,更走不到 :116 的 _ensure_round_enqueued。\n\n2) 「轮次 status=confirmed」与「节点 status=selected / selected_image_id 非空」是同一原子事务里一起写的,不存在“轮次已 confirmed 但节点仍可创建轮次”的中间态。全仓只有两处把轮次写成 confirmed,两处都同时锁死节点:\n- workflow_round_repository.py:993-1000(confirm_candidate_atomic)先执行\n```sql\nUPDATE workflow_nodes SET status = 'selected', selected_image_id = ?, updated_at = CURRENT_TIMESTAMP WHERE node_id = ?\n```\n:1010-1017 才写 `UPDATE workflow_generation_rounds SET status = 'confirmed' ...`,同一 conn 同一事务(sqlite BEGIN IMMEDIATE,pg FOR UPDATE 行锁)。\n- workflow_cross_review_repository.py:580-583 同样先 `UPDATE workflow_nodes SET status = 'selected', selected_image_id = ?`,:599-635 才把评审轮次与 root 轮次写成 confirmed。\n\n3) 不存在“解锁节点但保留轮次 confirmed”的回退路径。唯一清 selected_image_id 的地方是 image_service.py:179/205/225(unselect/archive/delete_image),三处入口都先调 _ensure_image_node_is_editable(:236-255),对 LOCKED_NODE_STATUSES 内节点直接 raise locked_node_edit_error;即便 selected_image_id 被清空,node.status 仍是 'selected',第 1145 行照样拦得住。\n\n4) 因此发现列举的三项后果全部不成立:apply_feedback_atomic 的 _ensure_writable_round_row 拦截、workflow_cross_review_repository.py:1066-1076 的 confirmed_candidate_mismatch、:1108-1122 的 _ensure_review_after_confirmation_is_safe、:795 的 record_skip_restoration_handoff_atomic,都不会因为“status 被回退成 completed”而失效——status 压根回退不了。另外跨评审轮次也不走 create_round:backend/routes/generation_cross_review_routes.py 走 workflow_cross_review_repository 的原子路径(:165-169 有独立 _round_replay),与这条 idempotency 分支无关。\n\n5) 触发条件描述尤其站不住:“网络超时后的正常重试”发生在轮次刚创建、节点尚未定稿时,此时 status 是 draft/queued/running,_ensure_round_enqueued 写成 queued 再 reconcile 成实际聚合态,正是幂等补偿的预期语义,不是缺陷。\n\n残留的真实问题仅是防御性编码瑕疵:_ensure_round_enqueued 末尾的 status 写入确实没有终态白名单,依赖调用方(create_round:81)的前置守卫兜底,属于“守卫与写入分离、易被后续重构打破”的健壮性问题,最多 P3,需先破坏第 81 行节点锁才可能演变为真<EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/workflow_round_service.py:806-861(reconcile_round 主体);佐证:backend/services/workflow_round_service.py:465-648(create_next_round)、database/repositories/workflow_round_repository.py:416-431(update_round 无守卫)、database/repositories/workflow_round_repository.py:1355-1375(_ensure_writable_round_row)、database/repositories/workflow_round_repository.py:855-861(confirm 的 status 门槛)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "我尝试推翻但没能推翻,代码与描述基本吻合,只有一处细节需要修正。\n\n一、代码真实存在,行号对得上,不是幻觉。\n`backend/services/workflow_round_service.py:818-847` 逐字如证据所述:\n```\n updated_round = self.round_repository.update_round(\n round_id,\n {\n \"status\": round_status,\n \"metadata\": { **round_record.metadata, \"planned_only\": False, \"execution_summary\": {...} },\n },\n ) or round_record\n if self.generation_service is not None:\n ...\n if node.status not in {\"selected\", \"discarded\"}:\n update_payload = {\n \"status\": node_status,\n \"metadata\": { **metadata, \"active_round_id\": round_id, \"last_round_status\": round_status, ... },\n }\n```\n写入前只有 `round_record is None` 的存在性检查(808-813)和 `node.status not in {\"selected\",\"discarded\"}`(837),**没有任何对 `round_record.status` 是否为 `superseded` 的判断**。\n\n`_ensure_writable_round_row` 确实在 `database/repositories/workflow_round_repository.py:1355-1375`(报告只写了文件名、行号 1370-1375 完全命中),不可写集合是 `{\"cancelled\", \"all_failed\"}` + 单独的 `confirmed` 分支,**不含 `superseded`**。\n\n二、不存在使问题不可达的上游守卫。\n1. `create_next_round`(473-514)先取 `get_consistent_round_snapshot`——我读了仓储实现(repository:300-354),它只做 `SELECT ... FOR UPDATE` 取 round/candidates/pool,**没有任何 status 校验**;服务层随后唯一的门槛是 `if not 1 <= len(liked) <= 3`(504)以及对每个 liked 候选要求 `generation_status == \"succeeded\"`(518)。**未 liked 的候选处于 queued/running 完全不影响创建子轮次**。\n2. `update_feedback`(313-342)只校验 `node.status == \"selected\" or node.selected_image_id`,不校验轮次是否终态;`apply_feedback_atomic` 里的 `_ensure_writable_round_row`(repository:584)对 running/superseded 均放行。\n3. `_ensure_round_creation_allowed`(1129-1156)只校验 workflow_type / node_type / 节点是否已定稿,与父轮次执行状态无关。\n4. 我特意 grep 了 `cancel`,`create_next_round` 把父轮次置 `superseded`(621-633)时**不会取消父轮次残留的 queued/running 候选**,因此这些任务必然会走完 `_on_terminal`(workflow_round_worker_runtime.py:93)→ `handle_task_terminal` → `reconcile_round(updated.origin_round_id)`(service:790)。\n5. `_reconcile_lock`(803)只解决同一轮次内多线程覆盖,对\"父轮次已被 superseded\"这一维度无效。\n6. `update_round`(repository:416-431)是裸 UPDATE,不调用 `_ensure_writable_round_row`,所以 `superseded` 会被 `_aggregate_execution_status` 的聚合值直接覆盖。\n\n三、框架层没有消解它。这是应用层状态机问题,参数化/鉴权与之无关。\n\n四、影响需要按实读修正一处,但结论不变,反而更坐实:\n`confirm_candidate_atomic`(repository:855-861)在 `_ensure_writable_round_row` 之后还有\n```\n if str(round_row[\"status\"] or \"\") not in {\"completed\", \"partial_success\"}:\n raise WorkflowRoundWriteError(\"round_not_ready_for_confirmation\", ...)\n```\n也就是说 **`superseded` 状态下本身是无法定稿的**——报告第 4 条\"superseded 轮次……可再次 create_next_round / 可写反馈\"成立(这两条路径确实放行),但\"可定稿\"并非直接成立。然而这恰恰说明危害链条是成立的:正是 `reconcile_round` 把 `superseded` 覆写回 `completed/partial_success`,才把这道定稿门槛重新打开,让用户能在被封存的旧轮次上定稿、绕过刚创建的优化子轮次。\n\n节点侧影响也属实:`active_round_id` 被改回父轮次(843),而前端 `frontend/standalone-web/src/App.vue` 的 `selectedWorkflowActiveRound
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"corrected_location": "backend/services/worker_service.py:207-221(_claim_next 消费路径)+ database/repositories/job_repository.py:145-154(ORDER BY created_at DESC)",
|
|||
|
|
"reason": "经逐处 Read 核实,该发现成立,非幻觉、无上游守卫消解。\\n\\n1) 代码真实存在且行号吻合。database/repositories/job_repository.py:145-154 的 JobRepository.list 确为 `SELECT {TASK_COLUMNS} FROM generation_tasks {where_clause} ORDER BY created_at DESC LIMIT {placeholder} OFFSET {placeholder}`;backend/services/worker_service.py:207-208 确为 `queued = self.queue_service.list_jobs(status=\\\"queued\\\", page=1, page_size=self.claim_batch_size).items`。claim_batch_size 在 QueueWorkerPool.__init__(worker_service.py:394)默认 25,advanced_material_worker_runtime.py:34 显式传 25,QueueWorker.__init__(worker_service.py:92)clamp 到 [1,50]。所以消费者确实只看「最新的 25 条」。\\n\\n2) 无上游校验/守卫使其不可达。backend/services/queue_service.py:91-114 的 list_jobs 只做 `if status and status not in TASK_STATUSES: raise ApiError(...)` 与 page/page_size 归一,随后原样调用 self.repository.list(...),并未为消费路径替换排序。全仓唯一的 FIFO 查询是 job_repository.py:197 的 `ORDER BY updated_at ASC, created_at ASC, task_id ASC`(list_stale),但它只在 WorkflowRoundWorkerRuntime.start() 里经 recover_stale_tasks 调一次做启动恢复(workflow_round_worker_runtime.py:55-62),不参与常态认领——恰恰说明作者会写有序查询,claim 路径是遗漏。\\n\\n3) 无框架层机制消解。参数化只解决注入,不影响排序;两个 pool 共用同一张 generation_tasks,task_filter 是在取回 25 条之后于 Python 侧过滤(worker_service.py:212 `if self.task_filter is not None and not self.task_filter(candidate): continue`,_is_workflow_round_task 判 metadata.source == \\\"workflow_round\\\",_is_advanced_material_task 判 \\\"advanced_material\\\")。因此若最新 25 条 queued 全是另一类任务,本 pool 取到空集,形成跨池队头阻塞,有效窗口比 25 更小——实际比原描述更糟。_retry_deferred(worker_service.py:210、480-493)同样在窗口内白占名额,属实。\\n\\n4) 同秒乱序属实:database/migrations/0005_generation_tasks.sql:15 为 `created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP`(SQLite 秒级),ORDER BY 无 task_id 等 tiebreaker,一轮 4~8 个候选同秒入队顺序不确定。\\n\\n5) 严重度需下调为 P1 而非 P0:不造成数据损坏或进程崩溃,_run_loop(worker_service.py:191-205)有全量 try/except 保活;「最早入队永久饿死」须待认领队列稳定 >25 才成立,而一轮候选仅 4~8 个(workflow_round_service.py:878 逐候选 enqueue_prepared_generation),需 4 轮以上同时积压。常态表现为 LIFO 不公平 + 同秒乱序(后创建轮次先出图),深度积压时才升级为老任务卡死。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "backend/services/queue_service.py:406-411(配合 database/repositories/job_repository.py:206-226)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "代码属实,但严重度被严重夸大,声称的核心影响(扣费/出图/取消语义失效)已被上游 CAS 消解。\n\n一、代码真实存在,行号对得上。queue_service.py:406-411 确为「读→校验→无条件写」:\n```\ndef update_task(self, task_id, payload, ctx) -> GenerationTask:\n before = self.get_job(task_id)\n status = payload.get(\"status\")\n if status is not None:\n self._validate_task_status_transition(before.status, str(status))\n after = self.repository.update(task_id, payload)\n```\njob_repository.py:215-222 的 UPDATE 也确实只有 `WHERE task_id = {placeholder}`,与同文件 228 行 `update_if_status`(`AND status IN (...)`,并用 `cursor.rowcount` 判定)形成对比。worker_service.py:116/272/308 三处调用也属实。\n\n二、但存在两层守卫,把可利用窗口压到「两条相邻语句之间」:\n1) worker_service.py:106-115 在写入前重读并早退:`latest = self.queue_service.get_job(...)`,`if latest.status == \"cancelled\": ... return`;`_handle_failure` 在 243 行有同样的 cancelled 早退。\n2) 更关键的是 update_task 自身 407 行会重读 before 并在 410 行做 FSM 校验,而 core/fsm.py:38 中 `\"cancelled\": {\"queued\", \"retrying\"}` —— cancelled→succeeded / cancelled→failed 均非法,会抛 ApiError(run_once 的 except 接住后再进 _handle_failure,仍被 243 行 cancelled 早退拦下)。因此只有当取消的 COMMIT 恰好落在 407 行 get_job 之后、411 行 UPDATE 之前的亚毫秒缝隙里,覆写才会发生。\n\n三、证据里最重的那条影响「用户已明确取消却仍被扣费并产生图片记录」是错的。真正出图的 workflow_round 池(workflow_round_worker_runtime.py:39 的 QueueWorkerPool → _execute → execute_claimed_generation)终态并不是由 update_task 写的,而是由 generation_service.py:1039 写的:\n```\ncompleted_task = self.queue_service.update_task_if_status(\n task.task_id, {\"running\"}, update_payload, ctx,\n require_workflow_not_cancelled=True,\n)\nif completed_task is None:\n ... self._cancel_generation_if_requested(...) -> self._finalize_cancelled_generation(...)\n ... self._cleanup_generation_outputs(assets=assets, images=[*images, *restricted_images])\n```\n取消发生 → 该 CAS 返回 None → 走取消收尾并清理产物;取消未发生 → 状态已是 succeeded,而 cancel_job(queue_service.py:145-150)对 succeeded 直接抛 TASK_NOT_CANCELABLE,取消根本插不进来,随后 worker 的 update_task 只是 succeeded→succeeded 的同态覆写(can_transition 中 target==current 允许)。同理 cancel_workflow_jobs(queue_service.py:230-275)内部逐个调 cancel_job,后者本身是 update_if_status + while 重试的 CAS 循环,不会被「同样方式回滚」。\n\n四、_handle_failure 的 queued 重排队路径(308 行)在生产上基本不可达:两个生产池都配置 RetryPolicy(max_attempts=1)(workflow_round_worker_runtime.py:45、advanced_material_worker_runtime.py:35),should_retry 不成立。\n\n五、结论:仅剩 advanced_material 链存在真实但极窄的裸露面——advanced_material_service.py:656 的 execute_task 全程不做终态 CAS,终态唯一写者就是 worker 的 update_task。这条链是识别/校验任务,覆写后果是「已取消的行被改成 succeeded/failed」的终态污染,没有额外扣费或图片记录后果。属于值得顺手把 update_task 换成 update_task_if_status 的健壮性问题,而非 P0/P1。",
|
|||
|
|
"refuted": false
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_location": "backend/services/worker_service.py:191-206 (_run_loop) 与 backend/services/queue_service.py:439+ (recover_stale_tasks)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "代码引用属实,但核心机理与影响链被证伪,标题所述的\"并发抢占同一队列 / worker 线性膨胀\"不成立。\n\n1) 引用的代码确实存在、行号对得上:\n- worker_service.py:180-185 的 stop() 逐字一致(set() 后 join(timeout),返回 running=self.is_running());\n- main.py:437-439 逐字一致:`self.round_worker_result = self.round_worker_runtime.stop()` / `self.round_worker_runtime = None`,确实不看 running;\n- engines/image_provider.py:308-313 的 `_generation_request_timeout_seconds` 返回 `min(max(timeout, 10), 3600)` 也属实。\n\n2) 决定性反证——旧线程停止后不会再认领任何任务。worker_service.py:191-206:\n```\ndef _run_loop(self, ctx):\n while not self._stop_event.is_set():\n result = self.run_once(ctx)\n ...\n```\n认领动作只发生在 run_once → _claim_next 内,而 run_once 每次只认领一条任务(_claim_next 返回单个 task,见 207-221 行)。stop_event 被 set 后,阻塞中的 run_once 一旦返回,while 条件立即为假,线程退出。也就是说 join 超时后残留的旧线程**最多把手上那一条已认领任务跑完就自然死亡,绝不会与新 pool 竞争认领新任务**。因此\"新旧 worker 池并发抢占同一队列\"\"worker 数量随每次重载线性膨胀\"两条结论都不成立——每次重载最多留下 5 个短命线程,各自跑完一条即退出,不累积。\n\n3) 残余风险存在但被默认配置基本关闭。唯一真实的窄窗口是:旧线程手上那条任务若在新 runtime 启动时已 running 超过 stale_after_seconds,会被 recover_stale_tasks(workflow_round_worker_runtime.py:55-59,默认 stale_after_seconds=300)标 failed 再重新入队,造成同一候选被生成两次。但要触发必须让单条任务的 running 时长超过 300 秒,而 config.yaml:25 / config.yaml:42 的默认 `timeout_seconds: 180`,加上 terminate() 里 5 个 worker 顺序 join(最多 25 秒)才轮到 initialize() 做 recovery,默认配置下窗口实际关闭;只有把 provider 超时手动配到 >300 秒(上限 3600)才可能命中。这属于配置相关的边缘竞态,而非 P0/P1 级的必现重复计费。\n\n4) 另外 update_task 并非\"无条件写回\":queue_service.py:405-408 的 update_task 会先走 `_validate_task_status_transition`(673-686 行,依赖 core/fsm.py 的 TASK_TRANSITIONS);run_once 116 行之前还会先 get_job 复查 cancelled(106-115 行)。真正的 CAS 保护(update_task_if_status / claim_queued_task,184-228、280-312 行)确实保证了同一条任务不会被两个线程同时认领。\n\n结论:机理描述有真实成分(stop() 返回值被忽略、join 可能超时),但被断言的影响链(队列抢占、线程线性膨胀、必然重复计费)经代码验证不成立,严重度严重夸大。按误报处理,若要保留最多记为 P3 的健壮性改进(stop() 返回 running=True 时记日志/告警,或让 recovery 跳过本进程仍存活的 worker_id)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/workflow_round_worker_runtime.py:45 + backend/routes/job_routes.py:195-220(真正的缺口在这里:`_retry_task` 只为 advanced_material 做了分派,没有 workflow_round 分支)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "我逐条 Read 验证,核心事实全部属实,行号无幻觉,因此不予推翻;但严重度和部分推论被夸大,下调为 P2。\n\n一、代码确实存在且行号精确\n1) `backend/services/workflow_round_worker_runtime.py:45`:\n```python\nretry_policy=RetryPolicy(max_attempts=1),\n```\n2) `backend/services/worker_service.py:41-43`:\n```python\ndef should_retry(self, attempt: int, exc: Exception) -> bool:\n if attempt >= max(int(self.max_attempts), 1):\n return False\n```\n3) `backend/services/queue_service.py:293`(claim 时先自增):\n```python\nattempt = _int_or_default(before.metadata.get(\"worker_attempt\"), 0) + 1\n```\nworker_service.py:259-262 读回的正是这个 `worker_attempt`,首次失败时 attempt=1,`1 >= 1` → False,worker 侧自动重试确实恒不触发。\n4) `backend/services/generation_service.py:361` 确为 `\"task_type\": \"workflow_round_candidate\"`;`generation_job_retry_service.py:181` 确为 `if task.task_type != \"generation\" or not task.node_id ...: raise ApiError(ErrorCode.TASK_NOT_RETRYABLE, ...)`。\n\n二、不存在上游守卫或框架层消解\n- `_retry_task`(job_routes.py:202-220)只对 `task.metadata.get(\"source\") == \"advanced_material\"` 做了专门分派(style_sets.retry_source / retry_validation_case / retry_import),`source == \"workflow_round\"` 直接落到 `resolved_retry_service.retry_job(...)` → `_validate_executable` 抛 TASK_NOT_RETRYABLE。同类非 generation 任务有专属重试通道、唯独 round 候选没有,说明这是遗漏而非有意收口。\n- provider 层无重试:对 `model_provider_runtime.py / model_provider_service.py / model_provider_bootstrap.py` grep \"retry|Retry|429|backoff\" 零命中;`generation_service.py:441-454` 里唯一的兜底是 `_content_safety_fallback`,只针对内容安全,不覆盖 429/5xx/连接重置。\n- `round_routes.py` 全部端点为 list/create/get/update_feedback/confirm_candidate/create_next_round,`workflow_round_service.py` 全文 grep \"retry|regenerate\" 只有 715 的 audit action 名和 754 的 `retry_scheduled` 字段,确无候选级重试接口。\n- 失败任务终态是 `\"status\": \"failed\"`(worker_service.py:272-288),不是 dead_letter,所以连 `retry_dead_letters` 都扫不到。\n\n三、被夸大之处(故降级)\n1) “显式传 RetryPolicy(max_attempts=1) ← 等于关闭重试”暗示是 round 特有的错误决策。实际上 `RetryPolicy` 的 dataclass 默认值本身就是 `max_attempts: int = 1`(worker_service.py:36),全仓库仅有的两个 pool(advanced_material_worker_runtime.py:35 与本处)都是 1。这是全局既定的“不自动重试”姿态,不是这一行把某个原本工作的通道关掉了。\n2) “workflow_round_service.py:750 的 retry_scheduled 分支全部成为死代码”不成立。`recover_stale_tasks`(queue_service.py:439-484)会把僵死的 running 任务先记 failed 再回写 queued,`WorkflowRoundWorkerRuntime.start()`(第 60-62 行)随后对每个 recovered task 调 `reconcile_task` → `handle_task_terminal`,正好命中 `elif status == \"queued\": candidate_status = \"queued\"; failure = {\"retry_scheduled\": True...}`。即进程崩溃/僵死这一类中断是有恢复重跑路径的,只是覆盖不到 provider 瞬时错误。(`_retry_deferred` / `next_retry_at` 确实当前不可达。)\n3) “用户唯一出路是重建整轮”偏重。失败候选会带 failure message 落库,round 归约为 partial_success,成功候选仍可 confirm,也可 `create_next_round` 续轮;这是可用的降级形态,不是全盘作废。\n\n结论:缺陷成立——详细工作流候选遇到 429/连接重置/provider 5xx 会一次性变终态 failed,且 Web 侧无任何候选级重试入口,用户体验和成本确有损失;但它属于健壮性缺失/功能缺口,不造成数据损坏、状态错乱或崩溃,且与全局默认策略一致,定为 P2-中等而非 P1。建议修法:在 `job_routes._retry_task` 增加 `so
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/generation_service.py:119(同步入口)+ :760/:831-836/:882(同栈内 create_task→CAS running→provider 调用);main.py:333(共享默认 executor);backend/services/chat_workflow_service.py:216(每用户 4 线程)",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "我以\"默认是误报\"的立场逐条核对,四个验证点全部证实,无法推翻。\n\n1) 行号与代码真实存在,非幻觉。`backend/services/generation_service.py:119` 精确匹配:\n```python\n119: def generate_node(self, node_id, payload, ctx) -> dict[str, Any]:\n120: lock_entry = _retain_node_generation_lock(node_id)\n121: lock_entry.lock.acquire()\n123: return self._generate_node_locked(node_id, payload, ctx)\n```\n`chat_workflow_service.py:216` 也逐字匹配 `with ThreadPoolExecutor(max_workers=4, thread_name_prefix=\"imagestudio-chat-fast\") as executor:`,`main.py:333` 确为 `await asyncio.to_thread(dispatch_web_request, ...)`。\n\n2) \"同栈内发起 provider 阻塞调用\"属实。`_generate_node_locked`(:556)里::760 `task = self.queue_service.create_task({... \"status\": \"queued\" ...})`,紧接 :831-836 `running_task = self.queue_service.update_task_if_status(task.task_id, {\"queued\"}, {\"status\": \"running\", \"progress\": 20}, ...)`,然后 :882 `assets = self.single_flow_engine.generate(running_node, task, runtime_generation_request)` —— 确实是\"入队即转 running 并原地同步打 provider\",不走后台 worker。\n\n3) 上游没有守卫让它不可达。`backend/routes/node_routes.py:146` 在 `resolved_generation_service` 存在时直接调 `generate_node`;生产 dispatch 中 `service`/`generation_service` 均为 None,因而走 `_create_generation_service()` 真实分支,那个 `queue_placeholder` 兜底分支根本进不去。`route_registry.py:116` 确认 `RouteSpec(\"POST\", \"nodes/{node_id}/generate\", ...)` 是对外 Web/bridge 路由。\n\n4) 框架层没有消解它。全仓(排除 .tmp 快照与 tests)只有三处并发控制:`worker_service.py:446` 的 `_BoundedWorkerHandler.global_semaphore` 只包裹后台 worker 的 `handler(task)`,与本条的内联同步路径无关;`cross_review_service.py:62` 无关;`chat_command_guard_service.py:22-23` 的 `user_concurrency_limit=2 / group_concurrency_limit=5` 我特意查了调用点,全仓仅 `chat_workflow_service.py:38`(`submit_single_image`)一处使用——`generate_quick_character`(:190 起,含 :216 的 4 线程池)和 chat `generate_node`(:678)都**没有**调用 `guard.check_generation`,Web 路由更是完全无此守卫。也没有任何 `set_default_executor`,所以 `to_thread` 确实落在共享默认 executor(min(32, cpu+4))上。per-node RLock 只按 node_id 串行,不同 node/工作流/用户之间毫无节流。\n\n5) 影响量级也核实无误。`generation_service.py:2308-2313`:`raw = request.get(\"timeout_seconds\") or settings.get(\"timeout_seconds\") or 180` / `return min(max(timeout, 10), 3600)`,即默认 180s、上限 3600s,与描述一致。`route_registry.py:139-142` 显示 `GET jobs`、`POST jobs/{task_id}/cancel` 与生成接口共用同一条 `to_thread` 派发链,因此\"连取消接口一起挂起\"的自锁死结论成立。\n\n严重度判定:定为 P1 而非 P0。它是可用性/成本失控问题,需要写权限(`require_write(ctx)`)且需要并发量堆积,不涉及数据损坏或凭据泄露;但触发门槛极低——3 个用户各跑一次快速角色四图即 12 路并发,属正常业务负载而非攻击,且会拖垮整个 AstrBot 进程的默认线程池(跨插件影响)并使自救用的取消接口一并失效,故不宜降为 P2。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/settings_service.py:614(写入点 _record_snapshot),调用方 backend/services/settings_service.py:444(save)、:556(rollback);生产装配点 backend/routes/settings_routes.py:167-169",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "我按\"默认误报\"的立场逐条找反驳点,但四项检查全部未能推翻,该发现成立(仅严重度需下调)。\n\n1. 代码真实存在,行号完全对得上,不是幻觉。\n`backend/services/settings_service.py:595-620`:\n```python\ndef _record_snapshot(self, config: Mapping[str, Any], ctx: RequestContext, reason: str) -> str | None:\n if not self.audit_service:\n return None\n ...\n (\n snapshot_id,\n reason,\n ctx.actor_id,\n json.dumps(dict(config), ensure_ascii=False, sort_keys=True), # 第 614 行,无 _mask_sensitive\n ),\n```\n调用侧 `:444` `self._record_snapshot(before_file, ctx, \"before_save\")`、`:556` `self._record_snapshot(before_file, ctx, \"before_rollback\")` 均传入 `before_file = load_config_file_payload(self.config_service.config_path)`(`:424`、`:542`)。而 `database/db_helper.py:72-83` 的 `load_config_file_payload` 就是 `_load_yaml_subset(path.read_text(...))` 的原始合并结果,不做任何脱敏——所以写入的确是明文原文。对比读取侧 `list_snapshots:499`、`get_snapshot:523` 都写了 `_mask_sensitive(config)`,写入侧确实是唯一漏网的一环,非对称是真实的。\n\n2. 不存在让它不可达的上游守卫。`save()` 里在 `_record_snapshot` 之前只有 `validate()`、`_merge_masked`、`_runtime_override_changes`、`persist_runtime_config_values` 等业务校验,没有任何跳过快照的分支;`_record_snapshot` 唯一的早退条件是 `audit_service is None`,而生产装配路径 `backend/routes/settings_routes.py:167-169` 恰恰是 `return SettingsService(audit_service=AuditService(database))`,且 `save_settings/import_settings/rollback_settings`(:75、:115、:158)全部走这个工厂。也就是说线上每一次保存都必然落一份明文。仅测试里 `tests/test_astrbot_runtime_config_bridge.py:146` 构造的实例没有 audit_service。\n\n3. 框架层机制解决不了它。SQL 确实是参数化的(`{placeholder}` 绑定),路由确实有 `require_config_manager(ctx)` 鉴权——但这两者只防注入和防越权调用,不改变\"明文落库\"这个事实;威胁模型是 DB 文件/备份/只读账号,鉴权装饰器管不到。\n\n4. 明文字段确实存在,不是臆造:`backend/services/model_provider_runtime.py:366` `value = str(candidate.get(\"api_key_plaintext\") or \"\").strip()`、`backend/services/api_discovery_service.py:229` 均读该字段,`frontend/.../SettingsPage.vue:86` 新建适配器时就带 `api_key_plaintext:\"\"`,说明这是前端可填、会落盘的明文槽位;`config.yaml:162,167` 有 `dsn: \"\"` / `password: \"\"`(默认空,实际部署会填);`backend/services/integration_token_service.py:64-65` 会把 `integration_token_hash` / `integration_token_salt` 写入 security 段,`:115,:133` 还兼容遗留的明文 `integration_token` 槽位。这些都会被 `before_file` 原样带进 config_json。全仓 grep `config_snapshots` 只有 `models.py:402`、迁移 `0012_config_snapshots.sql`、以及本文件的三处 SQL,确无 DELETE/保留期,\"无上限累积\"属实。\n\n需要修正的两点(因此下调严重度到 P2):\n- 影响被适度夸大。同样的明文本来就以 config.yaml 明文躺在同一台主机上,默认后端是 SQLite(`config.yaml:157-159` backend: auto,路径 `data/db/imagestudio.db`),与 config.yaml 同处一个数据目录、同一信任边界,边际新增风险主要是\"历史版本永久累积\"(含已轮换的旧 Key)。只有在启用 PostgreSQL(可能是远程库、有独立只读账号与备份链路)时才真正发生跨信任边界的凭据外溢,那种部署下可算 P1。\n- \"令牌盐+哈希外泄即可离线爆破\"只在用户自定义弱令牌时成立:`integration_token_service.py:152` 允许用户自填、下限仅 16 字符,且 `_hash_token`(:166-167)是单轮 HMAC-SHA256 无 KDF;若走默认的 `secrets.token_urlsafe(32)` 则不
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "core/storage.py:45-63(migration_plan)与 core/storage.py:182-200(_copy_path);入口 backend/routes/storage_routes.py:36-47",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "代码属实,行号完全对得上,非幻觉。core/storage.py:51-53 为 `will_overwrite = target_path.exists()` / `if not dry_run: _copy_path(source_path, target_path)`,`will_overwrite` 仅在 61 行放进返回体,全文件无任何分支用它阻断覆盖;core/storage.py:182-200 的 `_copy_path` 用 `shutil.copy2` 静默覆盖,唯一守卫是 188-191 的 `if target_root == source_root or source_root in target_root.parents: raise ApiError(FORBIDDEN, \"target path cannot be inside source directory\")`,与覆盖保护无关。backend/routes/storage_routes.py:36-46 确实只有 `require_config_manager(ctx)` 且 `dry_run=as_bool(payload.get(\"dry_run\"), default=True)` 可由请求体传 false;该文件未 import AuditService,\"无审计\"属实。main.py:104-106 的 `BRIDGE_PERMISSIONS = frozenset({\"can_write\",\"can_manage_config\",\"can_manage_materials\",\"can_retry_tasks\"})` 配合 2337 行 `permissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()`,确认 require_config_manager 在真实运行时等价于\"已登录\"。backend/services/settings_service.py:195-205 将 7 个 storage root 暴露为可写 path 字段,save()(415-456)无绝对路径校验,core/runtime_paths.py:43-45 对绝对路径原样返回,故\"root 可为任意绝对路径\"前提成立。\\n\\n但\"任意文件写入\"被夸大一档:source/target 均经 resolve_uri(core/storage.py:100-109),split_uri 在 127-128 拒绝绝对路径与 `..`,resolve_uri 还做 `if resolved_path != resolved_root and resolved_root not in resolved_path.parents: raise ApiError(FORBIDDEN, \"storage uri escapes its root\")`。默认配置下写入被严格限制在 7 个 data/ 子目录内,本端点自身不存在路径穿越;要写到 AstrBot 配置/插件/启动目录,必须先经 settings/save 改 root——那是另一条独立缺陷,且实施者已持有 can_manage_config(唯一真实 HTTP 入口是 AstrBot WebUI 已认证用户,本就能改配置、装插件),所以\"提升到宿主机代码执行\"是链式结论而非该端点单独能力。聊天入口 backend/services/chat_workflow_service.py:1160 只授予 {\"can_write\",\"can_retry_tasks\",\"can_manage_materials\"},无法触达该端点。backend/route_registry.py:12-16 的 PERMISSION_MATRIX[\"config\"] 只是声明式清单,不构成额外拦截,故问题确实可达。\\n\\n真实且未被消解的危害:即便 root 全部保持默认,也能把 storage://images/x.png 覆盖到 storage://archives/x.png 等已存在文件,无二次确认、无审计、will_overwrite 事后才返回。判定为 P2:真实的数据完整性/缺失审计缺陷,需要 will_overwrite 阻断(或强制 confirm 参数)+ audit_service.record 修复,但不是无认证的任意写或直接 RCE。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "main.py:104-106(BRIDGE_PERMISSIONS 定义)、main.py:2324-2344(_request_context_from_astrbot_request,赋权语句在 2337);受影响的执行点 core/security.py:17-37、backend/route_registry.py:9-29",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "尝试推翻失败,该发现属实(但影响描述有轻微夸大,严重度应为 P1 而非致命级)。\n\n一、代码真实存在,行号完全对得上,不是幻觉。\nmain.py:104-106:\n```python\nBRIDGE_PERMISSIONS = frozenset(\n {\"can_write\", \"can_manage_config\", \"can_manage_materials\", \"can_retry_tasks\"}\n)\n```\nmain.py:2324-2344:\n```python\ndef _request_context_from_astrbot_request(request: Any | None) -> RequestContext:\n username = _first_text(_attr(request, \"username\"), _current_quart_username()) or None\n user_id = _first_text(_attr(request, \"user_id\"), username) or None\n # AstrBot authenticates extension routes before invoking the plugin, but\n # PluginRequest currently exposes only the authenticated username. ...\n permissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()\n```\n代码注释本身就承认这是\"上游只暴露 username,无法读取角色\"的妥协实现。\n\n二、不存在让问题不可达的上游守卫或框架机制。\n1. 这是插件页唯一的 HTTP 上下文构造点。全仓(排除 .tmp 与 tests)只有三处构造 RequestContext:main.py:2338(bridge)、chat_command_guard_service.py:42(聊天入口,permissions=set())、chat_workflow_service.py:1157(聊天工作流,显式三项)。前端走的 `/api/plug/{plugin}` 通路必然经过 main.py:323 `ctx = _request_context_from_astrbot_request(request)` → dispatch_web_request。\n2. dispatcher 没有集中执行权限矩阵。adapters/astrbot_adapter.py:171-187 的 dispatch_extension_request 只做 `resolve_route` 然后 `match.handler(...)`,完全没有读取 `match.permission` 做 fail-closed 校验。\n3. handler 层的校验确实存在,但对该 ctx 全部放行。已核实:settings_routes.py:78 `require_config_manager(ctx)`、data_backup_routes.py:26 `require_config_manager(ctx)`、job_routes.py:275 `require_task_retry(ctx)`。而 core/security.py:19 的实现是 `if not ctx.has_permission(permission)`,core/identity.py:36 `return permission in self.permissions or \"*\" in self.permissions` —— 这四项恰好就在 BRIDGE_PERMISSIONS 里,检查必然通过。因此 route_registry.py:9-29 中 read/write/config/materials/retry 的分级对 bridge 通路等价于 read,属实。\n4. 声称的敏感路由确实存在且确实标为 config/retry:route_registry.py:156 `settings/save` → config;:166 `backups/{backup_id}/restore` → config;:327 `integrations/private-companion/token/rotate` → config;:145 `jobs/failed/clear` → retry。\n\n三、需要修正/降级的部分(这是我唯一能站住的反驳点)。\n1. 不是\"任意人\",仍需通过 AstrBot WebUI 认证:`permissions = ... if username or user_id else frozenset()`,未登录时权限为空集,require_authenticated 会拒绝。所以这是\"已认证用户之间缺乏权限分级\",不是未授权访问,不构成 P0。真实危害取决于部署是否存在非管理员 WebUI 账号或被反代给多用户;在单管理员 dashboard 部署下影响退化为无。仓库自身的 docs/final-release-audit-2026-07-22.md:36-42 已把同一问题记为 P1-1,并写明\"当前可接受边界:仅限明确声明的可信 AstrBot 管理员工作台\"。\n2. \"管理员分支是死代码\"这一句轻微不准确。resource_access_service.py:82 是 `if ctx.has_permission(\"can_manage_all_profiles\") or ctx.has_permission(\"*\")`,后半支对 RequestContext.system()(core/identity.py:45 `permissions=frozenset({\"*\"})`)仍然生效,并非整段死代码;准确说法是\"没有任何 HTTP 用户能获得 can_manage_all_profiles,该子条件在 Web 通路上恒假\"。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "adapters/astrbot_adapter.py:171-187(分发器)+ backend/route_registry.py:365-377(断言)+ main.py:1753(桥接通配符注册)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"refuted": false,
|
|||
|
|
"reason": "技术事实全部核实成立,但影响被夸大,严重度应下调为 P2。\n\n【1. 代码真实存在,行号准确】adapters/astrbot_adapter.py:171 起:\n```\ndef dispatch_extension_request(ctx, method, path, *, query=None, payload=None) -> dict[str, Any]:\n match = resolve_route(method, path)\n if match is None:\n return error_response(error_code=ErrorCode.RESOURCE_NOT_FOUND, ...)\n return match.handler(*_handler_args(match, ctx, query or {}, payload or {}))\n```\nRouteMatch(21-28 行)确实声明了 `permission: str`,resolve_route(56-63 行)确实填充 `permission=route.permission`,分发器全程未读取。不是幻觉。\n\n【2. 无上游守卫可消解】生产链路已跟到底:register_bridge_api()(main.py:283)→ _bridge_registration_routes()(main.py:1753)确认只注册 `f\"/{PLUGIN_NAME}/<path:endpoint>\"` 通配符、不带权限元数据 → dispatch_bridge_request → dispatch_web_request(main.py:272)→ dispatch_extension_request。全链无权限判定。_register_route()(190-207 行)虽然会向宿主传 `permission=`,但属于 register_astrbot_runtime 分支,桥接实际不走这条路。\n\n【3. 无框架层鉴权机制】assert_unique_routes() 只查 (method, endpoint) 重复;assert_known_permissions() 只查权限名是否在 PERMISSION_MATRIX 中,二者均不触及权限语义。runtime_health()(113-142)比描述多做一件事——遍历 _load_handler 校验 handler 可导入——但同样不校验 guard。grep tests/ 中 ROUTES 仅两处(test_profile_card_finalize.py:292、test_workflow_generation_cross_review.py:255),只断言 (method, endpoint, handler) 三元组,无任何 permission↔require_* 一致性测试。鉴权 100% 依赖 handler 内手写 core/security.py 的 require_write/require_config_manager 等。\n\n【4. 影响被夸大——实测当前 0 条越权】我脚本遍历全部路由静态提取 handler 内 require_*,初筛 9 条可疑,逐条 Read 后全部是委托调用、守卫在下游:archive_profile/delete_profile_via_post → delete_profile 内 `require_write(ctx)`(profile_routes.py:95);start_workflow/archive_workflow → _set_workflow_status 内 `require_write(ctx)`(workflow_routes.py:123);四条 logs 路由 → _read_logs 内 `require_authenticated(ctx)`(log_routes.py:98);review_moderated_image 用内联 `ctx.has_permission(\"can_manage_materials\")`(image_routes.py:156)。另路由总数为 166 而非声称的 169。\n\n结论:这是真实的\"缺少纵深防御 + 缺少一致性断言\"的架构/可维护性问题,属于未来回归风险(新增路由漏写 guard 时 CI 与 runtime_health 均不报警),而非现存可利用漏洞。因此不予推翻,但严重度从\"静默未授权写接口\"的暗示下调至 P2-中等。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "core/runtime_paths.py:43-45 + core/storage.py:105-108 + backend/services/settings_service.py:240-370(validate 无 storage 段校验)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "机制部分属实,不是幻觉;但攻击链中\"未认证第三方外传\"这一最致命的环节站不住脚,因此严重度需要下调。\n\n一、代码确实存在,行号完全对得上\ncore/runtime_paths.py:43-45:\n resolved = Path(path).expanduser()\n if resolved.is_absolute():\n return resolved\n绝对路径原样返回,无任何白名单/前缀约束。\n\ncore/storage.py:105-108:\n resolved_root = resolve_data_path(root_path).resolve()\n resolved_path = (resolved_root / relative).resolve()\n if resolved_path != resolved_root and resolved_root not in resolved_path.parents:\n raise ApiError(ErrorCode.FORBIDDEN, \"storage uri escapes its root\", {\"uri\": uri})\n越界检查只以 resolved_root 为锚,root 本身被改掉后该检查无意义。split_uri(storage.py:127)也只挡 \"..\" 和绝对相对路径,同样只解决\"root 内越界\"。\n\ncore/config_service.py:74-84:storage.*_path 直接 str(storage.get(...)) 落库,无 sanitize;load() 每次重读文件、无缓存,改完立刻生效。\n\n二、确无上游守卫\n我通读了 backend/services/settings_service.py 的 validate()(240-370 行):只处理 generation / vision / database / prompt_governance / advanced_materials,storage 段只在 schema 描述里出现过一次(195 行 \"key\": \"storage\"),validate 里一次都没出现(grep -i storage 在该文件仅命中第 195 行)。save()(415 行)只调用 validate 后就 _merge_masked 写盘。storage.temp_path 也不在 main.py 的 _RUNTIME_PATH_TO_ASTRBOT_KEY 白名单里(只有 storage.advanced_materials_path 在),因此不会被 _restore_runtime_managed_fields 还原。所以\"配置任意绝对根 → media/index 枚举 → media/sign 签名\"这条链,对已持有 can_manage_config 的调用者是真实可达的。\n\n三、被夸大之处(严重度下调的依据)\n1)\"未认证第三方 GET media/fetch 取走内容\"缺乏依据。backend/route_registry.py 的 PERMISSION_MATRIX 只是一份**说明性清单**:全仓 grep \"auth_required\" 只命中 route_registry.py 自身,消费者仅有 adapters/webui_adapter.py:19 和 backend/services/contract_service.py:17-18,都是把它塞进 manifest/contract 返回给前端,没有任何代码依据 auth_required 去放行或拦截请求。真正的鉴权只有两层:处理函数内的 require_authenticated(core/security.py:9),以及宿主 AstrBot 的中间件。而插件自身的注释明确说明后者是生效的——main.py:2333-2337:\"AstrBot authenticates extension routes before invoking the plugin, but PluginRequest currently exposes only the authenticated username\";backend/routes/media_routes.py:56-62 更进一步解释了 fetch_media_payload 存在的原因:\"a plain <img> request to /api/plug does not [carry the dashboard login state]... lets browser clients create a local Blob URL without making the signed streaming endpoint less strict at the AstrBot middleware layer\"。即:media/fetch 走的是宿主已鉴权的扩展路由面,浏览器 <img> 拿不到 JWT 才需要单开 payload 接口。所以\"把 URL 交给任何未认证第三方\"这一步没有证据支撑,属于把清单里的声明性字段误当成实际的鉴权开关。\n\n2)前置条件被弱化了。settings/save 在 route_registry.py:156 标注为 \"config\" → require_permission(ctx, \"can_manage_config\")。main.py:2337 的 permissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()(BRIDGE_PERMISSIONS 定义在 main.py:104-106,含 can_manage_config)确实让每个已认证桥接调用者都拿到该权限,所以\"任一已登录用户\"在插件内部成立;但这里的\"已登录\"就是 AstrBot 面板管理员登录态,该主体本来就能改 AstrBot 自身配置、装插件,读宿主文件对他并非实质越权。\n\n结论:漏洞成立但不是 P0。它是一条真实的加固缺口——配置面缺少\"存储根必须落在插件数据目录内\"的约束,使 VfsStorage 的沙箱边界完全<EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/generation_service.py:2352 与 :2366(_entrypoint / _permission_context 定义),真实触发点在同文件 :1461、:1503(_govern_prompt)与 :1578、:1581(_content_safety_fallback);入口为 backend/routes/node_routes.py:132-146",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "我尝试按\"默认误报\"立场推翻,但读代码后无法推翻,该缺陷成立(仅影响描述有夸大,已修正)。\n\n一、代码真实存在,行号对得上,非幻觉。backend/services/generation_service.py:2366:\n def _permission_context(payload: dict[str, Any]) -> dict[str, Any]:\n configured = payload.get(\"content_permissions\") if isinstance(payload.get(\"content_permissions\"), dict) else {}\n result = {\n **configured,\n \"adult_user_allowed\": bool(configured.get(\"adult_user_allowed\", payload.get(\"adult_user_allowed\", False))),\n \"quota_allowed\": bool(configured.get(\"quota_allowed\", payload.get(\"quota_allowed\", True))),\n }\n同文件 :2352:\n def _entrypoint(payload: dict[str, Any]) -> str:\n value = _optional_text(payload.get(\"entrypoint\") or payload.get(\"platform\") or payload.get(\"source\")) or \"web\"\n两者的唯一数据源都是请求体,全程未引用 ctx。消费点确认在 :1461/:1503(_govern_prompt 内 self.generation_route_service.decide(..., entrypoint=entrypoint, user_id=ctx.actor_id, permission_context=_permission_context(payload)))与 :1578/:1581(_content_safety_fallback,此处 payload = runtime_request,即持久化下来的用户请求)。\n\n二、不存在让问题不可达的上游守卫。backend/routes/node_routes.py:132 起的 generate_node 中,action() 只有 require_write(ctx) 一行鉴权,随后 return resolved_generation_service.generate_node(node_id, request_payload, ctx),request_payload = payload or {} 原样透传。generation_service.py:556 的 _generate_node_locked 只做幂等键、节点状态、工作流状态、模型路由校验;:1636 的 _enrich_generation_payload 仅 dict(payload) 后补 dependency_context 并编排素材绑定,全程未剔除或覆写 content_permissions / entrypoint。我逐条查了 payload 的写入点(grep 'payload\\['),没有任何一处重置这两个键。\n\n三、无框架层机制消解。core/security.py:27 require_write -> require_permission(ctx, \"can_write\"),core/identity.py:35 has_permission 仅做集合成员判断,与\"是否被允许成人内容\"完全正交,不构成任何间接约束。\n\n四、对照组证明这是 Web 路径的遗漏而非有意设计。同仓库 backend/services/advanced_material_service.py:1760 起的 _governance_context(config, ctx) 才是正确模式:\n \"entrypoint\": _entrypoint_from_context(ctx),\n \"content_permissions\": {\"adult_user_allowed\": bool(actor_id and actor_id in adult_users), \"quota_allowed\": True}\n即由 ctx.actor_id 比对 config chat_access.adult_user_allowlist 服务端推导;chat_workflow_service.py:1173 _content_permissions(config, chat) 同理(adult_user_allowed 由 chat.user_id 是否在白名单决定)。唯独 generation_service 的 Web 生成路径读客户端输入。原报告\"聊天侧是对的、Web 侧缺失\"的判断准确。\n\n五、消费端确实按这些值放行。prompt_governance_service.py:261-267:\n enabled = bool(self.config.get(\"adult_auto_route_enabled\", False))\n entry_allowed = bool((self.config.get(\"entrypoints\") or {}).get(entrypoint, False))\n user_allowed = bool(permissions.get(\"adult_user_allowed\", False))\n group_allowed = bool(permissions.get(\"adult_group_allowed\", group_id is None))\n quota_allowed = bool(permissions.get(\"quota_allowed\", True))\n if not all((enabled, entry_allowed, user_allowed, group_allowed, quota_allowed)):\n return self._decision(..., \"adult_route_permission_denied\", ...)\n五个条件里有三个(user/group/quota)直接来自请求体,第四个 entry_allowed 的 key 也由请求体决定。额外一点原报告未提:_entrypoint 会把任意含 \"onebot\"/\"qq\" 的字符串归一为 \"qq\",因此 Web 用户即使配置里 entrypoints.web=false,也能提交 entrypoint=\"onebot\" 冒充 QQ 入口,绕过入口维度的闸门。\n\n六、影响需要下调两处(故<EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/routes/workflow_routes.py:50(同类问题散布于 node_routes.py:79 get_node、image_routes.py:22 list_images、job_routes.py:19 list_jobs / :44 get_job、history_routes.py:35 list_workflow_history;根因在 backend/services/workflow_service.py:146-164、:960 与 backend/services/queue_service.py:91)",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "核验结论:无法推翻,该发现成立,行号与代码完全对得上。\n\n1) 代码真实存在,行号精确。backend/routes/workflow_routes.py:50-61 与证据片段逐字一致:\n```\ndef get_workflow(ctx, workflow_id, service=None) -> dict:\n def action() -> dict[str, Any]:\n require_authenticated(ctx)\n return workflow_service.get_workflow(workflow_id).to_dict()\n```\nbackend/services/workflow_service.py:146-150 亦一致:`def list_workflows(self, profile_id: str)` / `def get_workflow(self, workflow_id: str)` 签名里确实没有 ctx。\n\n2) 不存在上游守卫使其不可达。`core/security.py:9` 的 require_authenticated 只判 `ctx.is_authenticated`;`require_write` → `require_permission(ctx, \"can_write\")` 只判权限位,两者都不涉及资源归属。唯一的\"存在性\"检查 workflow_service.py:960 `_ensure_profile_exists` 只判 `profile is None or profile.status == \"deleted\"`,明确不判 owner_user_id,所以 `GET/POST profiles/{他人id}/workflows` 会正常返回/创建。get_workflow / get_node 连 profile 都不查,直接按 id 取库。\n\n3) 框架层未消解,反而证明这是\"漏改\"而非\"设计如此\"。仓库里明确存在归属链服务 backend/services/resource_access_service.py,其模块 docstring 写着\"Regular AstrBot chat callers must own the profile at the end of the resource chain\",并实现了 require_profile/require_workflow/require_node/require_image/require_task 五个方法,_can_access 比对 `ctx.actor_id == owner_user_id`。round 链路(workflow_round_service.py:1108-1114 `_node_and_workflow` → `resource_access_service.require_node`)和 profile 链路(profile_service.py:70-80 → :160 `_require_owner`)都接入了它,唯独 workflow/node/image/job/history 五条链路没有调用。所有权模型是真实存在的:profile_service.py:33-35 `owner_user_id = ctx.actor_id`(注释:\"Accepting an arbitrary owner_user_id would allow cross-tenant profile creation\"),聊天侧调用方经 `RequestContext.from_astrbot` 拿到的是聊天 user_id。因此\"GET nodes/{id} 可读他人节点、GET nodes/{id}/rounds 返回 404\"的自相矛盾属实。\n\n4) jobs/images 无 owner 维度也属实。job_routes.py:29-35 传给 queue_service.list_jobs 的只有 status/workflow_id/profile_id/page/page_size;queue_service.py:91-114 同样只有这些参数,直接透传 repository.list,无任何 owner 过滤;job_routes.py:52-56 `get_job(task_id)` 仅在 metadata.source == \"advanced_material\" 时才追加权限判断。backend/dto/job_dto.py 中 GenerationTask 字段含 `request: dict[str, Any]` 且 to_dict 输出 `\"request\": self.request`,泄露面确实是完整生成请求。image_routes.py:33-39 list_images 也只按 profile_id/workflow_id 过滤,restricted 过滤是内容分级不是归属。\n\n唯一需要下修的是严重度口径:攻击面不是匿名,必须先持有一个通过 AstrBot WebUI 网关认证的 bridge 会话(main.py:2337 只有在拿到 username/user_id 时才授予 BRIDGE_PERMISSIONS),且 BRIDGE_PERMISSIONS = {can_write, can_manage_config, can_manage_materials, can_retry_tasks} 不含 can_manage_all_profiles,所以这不是权限提升到 system,而是同层已认证主体之间的横向越权。在只有单一 WebUI 管理员、且全部数据都由该管理员创建的部署里实际影响为零;但只要存在聊天侧创建的档案(owner_user_id 为聊天用户 id,与 WebUI actor_id 不同)或多个 WebUI 账号,跨租户读写/耗费他人模型额度即可复现。定级 P1-严重(而非 P0-致命)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "backend/routes/profile_routes.py:109-114(仅命名问题);frontend/shared/api/endpoints.ts:10-11(profileDelete 定义后未使用)",
|
|||
|
|
"reason": "代码本身存在,但这条发现的\"影响\"是编造的,P1-严重定级完全站不住脚。\n\n一、代码事实核对(这部分属实)\n`backend/routes/profile_routes.py:109-114` 确实是:\n```python\ndef archive_profile(\n ctx: RequestContext,\n profile_id: str,\n service: ProfileService | None = None,\n) -> dict:\n return delete_profile(ctx, profile_id, service)\n```\n`backend/route_registry.py:69` 也确实注册了 `RouteSpec(\"POST\", \"profiles/{profile_id}/archive\", \"backend.routes.profile_routes.archive_profile\", \"write\")`。行号无误,非幻觉。\n\n二、致命反驳:前端根本没有\"归档\"按钮,唯一调用方是一个红色\"删除\"按钮\n发现声称\"前端 profileArchive 按钮标注为'归档',用户点击后档案被删除……归档与删除两种不同用户意图被合并成一种破坏性操作\"。我把 profileArchive 的全部引用查了一遍,全仓库只有一处调用(`frontend/standalone-web/src/App.vue:6300`),而它所在的函数是:\n```js\nasync function deleteSelectedProfile() {\n const profile = selectedProfilePage.value\n if (!profile || profileDeletingId.value) return\n const blocked = backendWriteBlockMessage(\"删除角色档案\")\n ...\n profileDeletingId.value = profile.profile_id\n try {\n await postJson<CharacterProfile>(endpoints.profileArchive(profile.profile_id))\n```\n失败文案是 `` `删除角色失败:${errorMessage(err)}` ``(App.vue:6334),状态变量叫 `profileDeletingId`,模板绑定是 `@delete=\"deleteSelectedProfile\"`(App.vue:13369)。\n\n再看用户实际看到的控件(`frontend/standalone-web/src/features/profiles/ProfileDetailPane.vue:190-198`):\n```html\n<button\n type=\"button\"\n class=\"danger-button\"\n :disabled=\"!canWrite || !backendConnected || deleting || Boolean(generationBusy)\"\n title=\"删除角色档案\"\n @click=\"emit('delete')\"\n>\n {{ deleting ? \"删除中\" : \"删除\" }}\n</button>\n```\n按钮文案是\"删除\"、hover 提示是\"删除角色档案\"、样式是 `danger-button`。用户意图 = 删除,后端行为 = `repository.soft_delete()` 置 `status=\"deleted\"`,两者**完全一致**。不存在任何\"用户以为在归档、结果被删除\"的场景,也就不存在所声称的\"两种不同用户意图被合并\"和\"破坏性操作\"。发现里\"该档案及其下所有工作流从 UI 彻底消失\"这一后果是删除按钮的**预期行为**,不是缺陷。\n\n三、被引用的\"证据\"反而说明这是有意为之\n发现自己引用的 `delete_profile_via_post`(profile_routes.py:101-107)同样转发 `delete_profile`,对应 `route_registry.py:68` 的 `POST profiles/{profile_id}/delete`。也就是说仓库为软删除提供了 DELETE、POST /delete、POST /archive 三条等价入口(`endpoints.ts:10` 的 `profileDelete` 已定义但全仓库无人调用),前端只是挑了 `/archive` 这条当删除用。这是路径命名/冗余入口的整洁性问题,不是语义错误导致的数据事故。\n\n四、与 workflow 侧的对照不成立\n`workflow_routes.py:106-111` 的 `archive_workflow` 走 `_set_workflow_status(ctx, workflow_id, \"archived\", service)`,确实是真归档,但 workflow 有 `core/fsm.py:9-15` 定义的 `archived` 终态,而 profile 侧的状态机里压根没有 `archived` 这个态——`ProfileService.get_profile`(profile_service.py:70-77)只对 `status == \"deleted\"` 做 404,`list_profiles` 默认 `status=\"active\"`。profile 从未设计过归档态,拿 workflow 的 FSM 去要求 profile 属于类比失当。\n\n五、结论\n唯一站得住的残留问题是:REST 路径叫 `/archive` 却执行删除,命名有误导性,且与 `imagestudio_full_plugin_dev_doc_v3_ui_adjusted.md:657` 的契约文档字面不符,未来若真要加归档功能会撞名。这是 P3 命名/契约整洁度问题。发现给出的\"修复建议\"(把 archive 改成写入 `status=\"archived\"`)如果照做,反而会让当前这个\"删除\"按钮变<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "backend/routes/node_routes.py:140(构造顺序);真正残留问题在 adapters/astrbot_adapter.py:171-187 与 main.py:315-344 缺少兜底 try",
|
|||
|
|
"reason": "代码本身没有幻觉,但这条发现的**核心指控(\"未授权即触发重型初始化\")在本仓库不存在可达路径**,属于误报级别的过度定级。\n\n一、代码确实如描述(这部分成立)\n`backend/routes/node_routes.py:132-146` 逐字对得上:\n```python\ndef generate_node(ctx, node_id, payload=None, service=None, generation_service=None) -> dict:\n workflow_service = service or _create_workflow_service()\n resolved_generation_service = generation_service or (None if service else _create_generation_service())\n def action() -> dict[str, Any]:\n require_write(ctx)\n```\n`_create_generation_service()`(node_routes.py:315-390)确实调 `load_config_payload()`、`Database(...)`、`sync_generation_provider_from_config(...)`;后者在 `backend/services/model_provider_bootstrap.py:68,79` 里确实执行 `service.upsert_provider(...)` 与 `service.upsert_route(...)` 写库。`round_routes.py:137-138` 的只读 `list_node_rounds` 也确实经 `_create_round_service()` → `create_generation_service()` 走同一条重型链路。`backend/routes/common.py:12-16` 的 try 也确实只包住 callback。\n\n二、致命反驳:不存在\"有权限差、但能打到 handler\"的调用者\n生产环境只有两个 ingress,两个都在 service 构造**之前**就完成了准入判定:\n\n1. Web 桥接(前端唯一通路):`main.py:2324-2344`\n```python\n # AstrBot authenticates extension routes before invoking the plugin, but\n # PluginRequest currently exposes only the authenticated username.\n permissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()\n```\n`BRIDGE_PERMISSIONS`(main.py:104-106)= `{\"can_write\",\"can_manage_config\",\"can_manage_materials\",\"can_retry_tasks\"}` 全给。也就是说:**凡是能走到 handler 的请求,`require_write` / `require_task_retry` 必然通过**;而未登录请求根本进不来——`docs/project-migration-handoff-2026-07-18.md:263` 是实测证据:\"插件 Web 路径直接无认证访问会返回 `Missing API key`,这是正常鉴权行为\"。攻击者拿不到 401 之后的执行权,谈不上\"未授权用户狂刷 GET nodes/x/rounds\"。\n\n2. 聊天命令通路:`backend/services/chat_workflow_service.py:1149-1164`,白名单校验 `chat_action_allowed(...)` 在**构造 ctx 和 dispatch 之前**,不通过直接 `raise ValueError(\"当前群聊或用户不满足聊天白名单规则\")`;通过者被显式授予 `{\"can_write\",\"can_retry_tasks\",\"can_manage_materials\"}`。同样不存在\"权限不足但已进入 handler\"的状态。\n\n结论:把 `require_*` 提到构造前,唯一能省下的是一个**在生产中永远不会发生**的分支的开销。原文声称的影响 1(未授权资源消耗 / SQLite 写锁被未授权流量拖垮)与影响 3 的\"未授权\"定性,均不可达。\n\n三、影响 2(500 泄漏)被夸大\n`adapters/astrbot_adapter.py:171-187` 的 `dispatch_extension_request` 与 `main.py:315-344` 的 `dispatch_bridge_request` 确实都没有 try,构造期异常会逃逸到 Quart —— 这一条**技术上成立**。但:(a) 只影响响应体格式(丢失 `{\"ok\": false, ...}` 信封),不是安全问题;(b) \"带 traceback 的 500\"是臆测,Quart 非 debug 模式只返回通用 Internal Server Error,仓库中没有任何证据显示 AstrBot 以 debug 模式运行;(c) 即便包进 try,`response_from_exception` 也只是把同一个错误换个外壳返回,实际收益是格式一致性而非防泄漏。\n\n四、残留的真问题与原文定性不同\n真正值得修的是\"**只读 GET 也跑一遍 config bootstrap 并 upsert provider/route**\"(node_routes.py:328 + model_provider_bootstrap.py:68,79),这是**已授权用户**的每请求性能/副作用问题,与鉴权顺序无关——把 `require_write` 前移一行也解决不了它。注意默认 `config.yaml`(generation.provider 的 base_url/model_id/route_name 均为空)会在 bootstrap.py:49-51 因 `missing` 直接 continu
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/workflow_round_worker_runtime.py:10(附带 backend/routes/job_routes.py:311)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "代码事实核对全部属实,非幻觉:backend/services/workflow_round_worker_runtime.py:10 顶层确为 `from backend.routes.node_routes import create_generation_service`;node_routes.py:315-390 的 `_create_generation_service()` 确实是组合根(Database、7+ 个 repository、ModelProviderRuntime、SingleFlowEngine 及 PromptTranslationService/PromptModerationService/GenerationRouteService/ImageModerationService 四个治理服务);job_routes.py:311 确为 `from backend.routes.node_routes import _create_generation_service`(私有名)。因此 service → route 的层次反向依赖客观存在,findings 的定位成立。\n\n但严重度被显著夸大,且两条核心\"影响\"经代码验证为虚构:\n\n(1) 作者已主动消解大半。node_routes.py:393-396 存在显式公共出口:`def create_generation_service() -> GenerationService: \"\"\"Public runtime factory shared by HTTP routes and background workers.\"\"\" return _create_generation_service()`。worker 走的正是这个公共名,属于有意设计的公开缝;真正不规范的只有 job_routes.py:311 舍近求远取私有名(公共封装就在同文件 80 行外)。\n\n(2) 无循环导入。grep 确认 node_routes.py 中 `worker_runtime`/`round_worker` 零命中,方向单一,运行时不会失败。\n\n(3) \"单测想构造 GenerationService 就得连带 import node_routes 的 31 个依赖\" —— 与真实代码矛盾。tests/test_workflow_round_queue_runtime.py:12 直接 `from backend.services.generation_service import GenerationService`,并在测试内手工装配 QueueService/ImageService/WorkflowService,完全绕开 node_routes,测试从未被这条依赖绑架。\n\n(4) \"main.py:390 必须在 initialize() 里函数内 import 以规避启动期副作用\" —— 无据。main.py:385-388 的注释讲的是\"数据库迁移必须先于 AstrBot 注册/检视路由处理器完成\",而 389-409 的函数内导入被包在 try/except 中,目的是 worker 启动失败时插件与恢复 UI 仍可用(`self.round_worker_result = {\"running\": False, \"error_type\": ...}`),与 route 层副作用无关。且 node_routes 导入期本身没有副作用——全部装配都在函数体内,import 不会构造 provider runtime。\n\n综上:这是纯架构整洁度/可打包性问题,无任何可触发的失败路径、正确性或安全影响,仅代价为一次模块导入。合理处置是把 job_routes.py:311 改用公共 `create_generation_service`,并可选地把组合根下沉到独立 factory 模块。严重度应降至 P3-轻微。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/routes/node_material_routes.py:69(同类漏传:branch_routes.py:64、history_routes.py:72、review_routes.py:90)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "我尝试推翻但没能推翻——事实层面全部核实为真,只是\"影响\"被明显夸大。\n\n【1. 代码真实存在,行号精确对得上,非幻觉】\nbackend/routes/node_material_routes.py:65-73 原文:\n def _create_service() -> WorkflowNodeMaterialService:\n database = Database(load_database_settings())\n profiles = ProfileRepository(database)\n workflows = WorkflowRepository(database)\n workflow_service = WorkflowService(\n workflow_repository=workflows,\n profile_repository=profiles,\n audit_service=AuditService(database),\n )\n确实没有 event_bus=。另三处也逐一 Read 确认行号精确:branch_routes.py:64、history_routes.py:72、review_routes.py:90,构造参数同样只有三个。对照组也属实:profile_routes.py:180、workflow_routes.py:149、job_routes.py:305、node_routes.py:311 都传了 event_bus=get_default_event_bus()。\n\n【2. 没有上游守卫/兜底让它不可达】\nbackend/services/workflow_service.py:78-86:\n event_bus: EventBus | None = None,\n ):\n ...\n self.event_bus = event_bus\n构造函数是纯赋值,没有 `event_bus or get_default_event_bus()` 的惰性兜底。因此 workflow_node_material_service.py:208 的 `if self.workflow_service.event_bus:` 在该装配路径下恒为 None,209-218 的 publish 整块被跳过,无日志无异常,属实。\n\n【3. 没有框架层机制消解它,且是唯一生产装配路径】\nbackend/route_registry.py:89-90 以点号字符串挂载 handler:\n RouteSpec(\"GET\", \"nodes/{node_id}/material-bindings\", \"backend.routes.node_material_routes.get_node_material_bindings\", \"read\")\n RouteSpec(\"PUT\", ... .save_node_material_bindings\", \"write\")\n调用时不注入 service 参数,故 `service or _create_service()` 必走 _create_service()。tests/test_workflow_node_material_bindings.py:413-421 则显式传 service=fixture[\"binding_service\"],所以测试完全绕开了这条路径,也没有任何用例断言事件发布——这解释了缺陷为何能长期存在。event_stream_service.py:11-20 的 DEFAULT_EVENT_BUS 是模块级单例,core/webhook.py:50-51 的 event_from_bus 不做事件类型白名单,所以\"若发布则可见、不发布则永不可见\"的推断成立。\n\n【4. 严重度被夸大,应降级】\n(a) 描述称\"下游的增量刷新/审计静默失效\",其中\"审计\"是错的:workflow_node_material_service.py:200-207 的 self.workflow_service.audit_service.record(action=\"update_workflow_node_material_bindings\", ...) 走的是另一条正确装配的依赖,审计记录照常落库。\n(b) 业务数据不丢:189-197 行 update_node 正常写入 metadata,198-199 行还有 None 校验;219-223 行 save 后直接 return self.get_node_bindings(...),发起操作的前端同步拿到了最新完整状态,无需靠事件刷新。\n(c) 全仓 grep \"node_material_bindings_updated\" 只有 1 处 publish(service:210)+ 审计报告自身,订阅端为 0;全仓除 core/event_bus.py 定义外没有任何 subscribe( 调用。也就是说这是一条当前无人消费的死通知。\n(d) 唯一消费者是前端 5 秒轮询,App.vue:4582 用的是粗粒度前缀匹配 `eventType.startsWith(\"workflow.\")`,后续 4588-4601 仅在 activeNav 为 overview/performance/logs 时刷新 dashboard/logs,并不存在任何\"素材绑定专属的增量刷新\"。而保存素材绑定发生在工作流/生成视图,用户当时根本不在这些页面。App.vue:4603-4605 还有 catch 注释明确写着保留手动刷新与轮询作为静默兜底。\n\n结论:这是一处真实的装配不一致(wiring defect),会在未来有人真正订阅该事件时才爆发,属于潜伏的可维护性/可观测性问题,不是当前的功能性或数据性缺陷。原报告的 P0/P1 级定性不成立,修正为 P3-轻微(若把 4 处同类漏传作为系统性一致性问题合并看待,最高可辩护到 P2-中等)。修复方案很简<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "database/repositories/workflow_round_repository.py:17(同 workflow_cross_review_repository.py:17、profile_card_repository.py:11)",
|
|||
|
|
"reason": "【事实核对:全部属实,非幻觉】三处 import 逐字逐行对得上:\n\n- `database/repositories/workflow_round_repository.py:17` → `from backend.services.branch_lineage_service import branch_state_for, with_branch_state`\n- `database/repositories/workflow_cross_review_repository.py:17` → 同一行同一句\n- `database/repositories/profile_card_repository.py:11` → `from engines.character_flow import canonical_gradual_node_key`\n- 反方向 `backend/services/branch_lineage_service.py:17` → `from database.repositories.workflow_repository import WorkflowRepository`\n\n所以\"database 包与 backend.services 包互指\"这个静态事实成立,不是误报,不能整条推翻。但**声称的影响里有 3/4 条经代码验证不成立**,严重度被明显夸大:\n\n**1)\"只要再多 import 一个 repository 就立即真成环\"——风险被高估,且现结构有结构性缓冲。**\n读了 `database/repositories/workflow_repository.py` 的全部 import,只有两行:`from backend.dto.workflow_dto import ...` 和 `from database.db_helper import Database`,不回指任何 service。更关键的是我核对了包初始化文件:`database/repositories/__init__.py` 全文只有一行 docstring `\"\"\"Repository implementations for ImageStudio persistence.\"\"\"`(62 字节),`backend/services/__init__.py` 同样只有 `\"\"\"Backend service layer.\"\"\"`(29 字节),两者都**没有做任何 re-export**。这意味着即使将来 `branch_lineage_service` 真去 import `WorkflowRoundRepository`,也不会因为 `__init__` 的连带加载而放大成难以定位的部分初始化错误——包级双向指向在这个仓库里始终只是\"目录名义上的环\",实际加载图是 DAG。\n\n**2)\"branch_state_for/with_branch_state 是分支血缘业务规则\"——这条是错的,属于对被调函数的误判。**\n读了 `branch_lineage_service.py:262-306` 的实现,这两个函数体内**没有任何 SQL、没有任何 repository 调用、没有 I/O**,是纯 dict 函数:\n```python\ndef branch_state_for(metadata, branch_name):\n source = metadata if isinstance(metadata, dict) else {}\n branch_states = source.get(\"branch_state\")\n if isinstance(branch_states, dict) and isinstance(branch_states.get(branch_name), dict):\n return deepcopy(branch_states[branch_name])\n return {key: deepcopy(source[key]) for key in BRANCH_STATE_FIELDS if key in source}\n```\n它们做的是\"workflow metadata_json 里 branch_state 子字典的读取/合并 + 1.0.16 之前的根字段回退\",本质是**元数据 schema 序列化助手**。真正的血缘业务逻辑是同文件的 `BranchLineageResolver` 类(`__init__(self, workflow_repository)`,第 65-259 行,做 `branch_chain` / `resolve_dependencies` / `_unwrap_locked_copy` 的父分支回溯),而 `WorkflowRepository` 的反向 import 只被这个类用到——**两个 repository 从未引用 `BranchLineageResolver`**。所以\"血缘策略调整要改仓储\"不成立:调整回溯策略改的是 Resolver,与 `workflow_round_repository.py:1053/1097`、`workflow_cross_review_repository.py:657/734/877/898` 这些调用点无关。核对调用现场也印证了这点,1053-1104 行只是把 `character_execution` / `stage_inputs` / `stage_review_handoffs` 这几个 metadata 字段读出来改完再合并回去。\n\n**3)\"仓储层无法脱离服务层单独测试\"——被现有约定证伪。**\n`grep '^from (backend|engines)\\.' database/` 显示仓储层**本来就系统性依赖 backend**:`image_repository.py:8`、`material_repository.py:8`、`job_repository.py:8`、`profile_repository.py:8`、`wardrobe_repository.py:8`、`advanced_material_repository.py:9`、`workflow_repository.py:8`、`workflow_round_repository.py:11` 全部 import `backend.dto.*`。也就是说\"database 不许指向 backend\"这条纯洁性在本仓库从来就不是既定架构约束,DTO 反指是全局既有约定。多出来的这一步只是把依赖面从 `backend.dto` 扩到了 `backend.ser
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_location": "backend/services/model_provider_runtime.py:285(代码定位属实,但影响判定错误)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "一、代码事实核对(部分属实):`backend/services/model_provider_runtime.py:285` 确实是 `resolved = self.model_provider_service.resolve_route(route_name)`,其后 `provider = resolved.get(\"provider\") or {}` 直接进入 `resolve_secret_value(...)`,没有 provider_type 判断;对照的 `:131-136` 与 `:200` 也确如描述。行号与代码文本对得上,不是幻觉。但\"与同文件其他两处不一致\"被夸大了:`_structured_text_callable_for_single_route(:196)` 同样是 `resolve_route(route_name)` 不带 modality,且全库 model_routes 表只存在 `image_generation` / `vision_recognition` 两种 modality(`model_provider_bootstrap.py:75,245`),\"结构化 callable 不传 modality\"是这一族函数的既有写法,真正唯一缺失的只有 provider_type 一道门。\n\n二、\"安全审核静默绕过\"这一条是彻底错误的,方向恰好相反。`backend/services/prompt_governance_service.py:356-366`:\n```\nif self.reviewer is None or not bool(self.config.get(\"enabled\", False)):\n return {\"status\": \"not_configured\", \"action\": \"allow\", ...}\ntry:\n result = dict(self.reviewer({...}))\nexcept Exception:\n return {\"status\": \"review_failed\", \"action\": \"quarantine\", ...}\n```\n也就是说,`review_with_fallback(:271)` 抛出的 RuntimeError 会被上游捕获并**降级为 quarantine(fail-closed)**,`generation_service.py:966-971` 随后把该图标记 `quarantined=True` / `gallery_visibility=\"restricted\"` 并放进 `restricted_images`,绝不会当作放行。发现里\"上层若把审核不可用当放行即静默绕过\"的假设在本仓库不成立。\n\n更关键的是,如果按发现建议补上 modality 与 provider_type 前置校验,`_structured_vision_callable_for_single_route` 会返回 None → `structured_vision_callable` 返回 None → `node_routes.py:345-348` 得到 `image_moderation_callable = None` → `ImageModerationService.reviewer is None` → 走 `{\"status\": \"not_configured\", \"action\": \"allow\"}`,**全部图片直接放行**。即:现状的宽松解析是 fail-closed,而所谓的\"修复\"才会制造真正的审核静默绕过。安全论证被写反了。\n\n三、可达性也被高估。`config.yaml:120-122` 中 `prompt_governance.image_moderation.enabled: false`、`model_route: \"\"`,`node_routes.py:344-348` 对空路由直接不构造 callable,功能默认关闭。要触发 modality 串号,还必须管理员在 `image_generation` 和 `vision_recognition` 两个 modality 下配置**同名**路由(bootstrap 的 `_config_route_id` 以 modality 为前缀,同名跨模态是被设计允许的),此时 `model_provider_repository.py:266` 的 `ORDER BY route.priority ASC, route.route_id ASC` 在优先级相同时会让 `route_config_image_generation_*` 胜出。要触发 provider_type 问题,还需管理员在设置页把适配器类型选成\"自定义 HTTP\"(SettingsPage.vue 的 select 仅 openai_compatible / custom),而 custom 适配器在 `_image_provider_from_resolved(:79-80)` 里本来就返回 None、连生图都不可用。两个前提都是明确的管理员误配。\n\n四、真实残余影响只有一条:上述误配下每张图会被误判为 quarantine(可用性问题),且图片 metadata 里会留下 `image_moderation.status = \"review_failed\"`,可诊断、可回溯,既不泄密也不绕过安全。综上:定位属实、影响与严重度判定错误,按\"安全审核绕过\"定级不成立,最多算 P3 级的一致性/健壮性小瑕疵,且修复时必须同时处理\"reviewer 为 None 即放行\"的语义,否则会引入比原问题更严重的缺陷。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "engines/single_flow.py:61(缺陷根因为缺失 size→aspect_ratio 映射,写入侧位于 backend/services/chat_workflow_service.py:63 与 :929)",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"refuted": false,
|
|||
|
|
"reason": "尝试推翻失败,该发现成立,且实际影响比描述的还略重一点。\n\n一、代码真实存在,行号完全对得上(非幻觉)\nengines/single_flow.py:56-61:\n```\ndef generate(self, node: WorkflowNode, task: GenerationTask, request: dict[str, Any]) -> list[MockGeneratedAsset]:\n count = _bounded_count(request.get(\"count\", 1))\n ...\n width, height = _dimensions(str(request.get(\"aspect_ratio\") or \"1:1\"))\n```\nengines/single_flow.py:248-255:\n```\ndef _dimensions(aspect_ratio: str) -> tuple[int, int]:\n if aspect_ratio == \"3:4\": return 768, 1024\n if aspect_ratio == \"4:3\": return 1024, 768\n if aspect_ratio == \"16:9\": return 1280, 720\n return 1024, 1024\n```\nengines/single_flow.py:154-159 是全仓库生产代码中唯一构造 ImageGenerationRequest 的地方(grep `ImageGenerationRequest(` 仅命中 single_flow.py:155 与 tests/test_image_provider_timeout.py),`size=f\"{width}x{height}\"`。\n\n二、写入侧确认写的是 size,不是 aspect_ratio\nbackend/services/chat_workflow_service.py:59-64(/yp 单图):\n```\npayload={\n \"prompt\": command.prompt,\n \"negative_prompt\": command.negative_prompt,\n \"model_route\": command.model_route,\n \"size\": command.size,\n \"count\": command.count,\n```\nbackend/services/chat_workflow_service.py:925-930(高清修复):\n```\ngeneration = self._data(self.dispatcher(ctx, \"POST\", f\"nodes/{node['node_id']}/generate\", payload={\n \"prompt\": \"high resolution restoration, ...\",\n \"model_route\": assignments.get(\"hires_fix\"),\n \"size\": size,\n```\nbackend/dto/chat_command_dto.py:8:`SIZE_ALIASES = {\"1k\": \"1920x1080\", \"2k\": \"2560x1440\", \"4k\": \"3840x2160\"}`,adapters/astrbot_command_parser.py:49-52 确实做了 1K/2K/4K 校验并 raise「尺寸仅支持 1K、2K、4K」。\n\n三、不存在任何上游守卫/转换让问题不可达(逐环节 Read 验证)\n1) backend/routes/node_routes.py:132-146 的 generate_node 只做 require_write(ctx) 后直接 `resolved_generation_service.generate_node(node_id, request_payload, ctx)`,payload 是自由 dict,没有任何 schema 校验或键名归一化,未知键 size 不会被拒绝也不会被翻译。\n2) backend/services/generation_service.py:733 `runtime_generation_request = {**payload, ...}` 原样透传 size,随后写入 task.request(queue_service.create_task 的 \"request\": generation_request)。\n3) 队列执行路径 backend/services/generation_service.py:410-433:`runtime_request = copy.deepcopy(task.request)`,之后只覆写 resolved_model_route / count / input_images 并 pop(\"image\"),随即 `self.single_flow_engine.generate(node, task, runtime_request)`(第440行)。同步路径第882行同理。全程无 size→aspect_ratio 映射。\n4) 全仓库(排除 .tmp/、tests/、__pycache__)grep `aspect_ratio` 的顶层写入只有 advanced_material_service.py:912、advanced_style_set_service.py:284/293/301/1896/2049/2615/2629、node_injection_template_service.py(模板 schema 字面量)、prompt_expansion_service.py:143(且该处是嵌在 \"stage_prompt_guard\" 字典里的,根本不是顶层键),无一来自 size。\n5) grep 生产 Python 中 `get(\"size\")` / `[\"size\"]` / `payload.get(\"size\")` 的消费点:零命中。backend/services/generation_service.py 全文 grep \"size\" 除 size_bytes 外无任何命中——即 size 在服务层是彻底的孤儿键。\n\n四、框架层面无消解机制,反而有反证\nfrontend/standalone-web/src/App.vue:5858 对同一个 nodes/{id}/generate 端点发的是 `aspect_ratio: spec.aspectRatio`,证明 aspect_ratio 才是该端点的既定契约键,chat 适配器写的 size 是自造的、无人识别的键。\n\n五、影响属实,且比原描述更糟(用户被明确告知了错误的尺寸)\nmain.py:1136-1155:\n```\nnormalized_size = {\"1K\": \"1920x1080\", \"2K\": \"2560x1440\", \"4K\": \"3840x2160\"}.get(size.upper())\n...\nyield event.plain_result(f\"✅ 高清修复请求已进入后台
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/chat_media_import_service.py:33-36",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "尝试推翻失败,该发现成立(仅严重度需从\"致命\"下调到\"严重\")。\n\n一、代码真实存在,行号精确对得上,非幻觉。`backend/services/chat_media_import_service.py:21-47` 中:\n```\nelif text.startswith((\"http://\", \"https://\")):\n body = read_http_response_limited(text, timeout_seconds=30, max_bytes=MAX_CHAT_IMAGE_BYTES)\n extension = Path(urlsplit(text).path).suffix.lstrip(\".\").lower() or \"jpg\"\nelse: # 第33行\n source = Path(text).expanduser().resolve() # 第34行\n body = source.read_bytes() # 第35行\n extension = source.suffix.lstrip(\".\").lower() or \"jpg\"\n```\n证据片段与文件逐字一致,确无任何根目录约束、白名单或魔数校验。\n\n二、调用链真实且可达,不存在让它不可达的上游守卫:\n- `adapters/astrbot_media_adapter.py:15-19` 逐个取消息 Image 组件的 `(\"url\",\"file\",\"path\",\"image_url\")` 属性字符串,首个非空即返回,无任何格式校验;\n- `main.py:1161 references = image_references_from_event(event)` → `main.py:1167 await asyncio.to_thread(self.chat_workflow_service.import_material, reference, category.strip(), context)`;\n- `backend/services/chat_workflow_service.py:948 uri, filename, size_bytes = self.media_importer.import_image(image_reference)`,字符串原样透传,中间无净化。\n关键:`main.py:1160 imagestudio_import_material` 这个处理函数**完全没有** `command_allowed(config, action, context)` 权限门,而同文件 `main.py:508/537/1201/1255` 的其它命令都有;也没有调用 `ChatCommandGuardService.check_generation/record`。`main.py:1445-1446` 还有短命令别名直连同一函数。即任意聊天用户可无门槛触发。\n\n三、无框架层机制消解。落盘后走 `POST materials/upload`,而 `backend/services/material_service.py:285-324` 的强校验(`_decode_upload_data_url` 探测真实 content_type、比对声明 size、比对扩展名、`UPLOAD_ALLOWED_CONTENT_TYPES` 白名单)**只在 `image_data_url` 分支执行**;聊天导入走的是 `image_uri` 分支,全部绕过。随后 `materials/{id}/recognize` 与素材列表接口即可回读,外带路径成立。\n\n四、同仓库自身基线证明这是缺口而非有意取舍:`core/storage.py:100-115 resolve_uri` 对 storage:// 做了严格根目录逃逸校验(`if resolved_path != resolved_root and resolved_root not in resolved_path.parents: raise ApiError(ErrorCode.FORBIDDEN, \"storage uri escapes its root\")`);同函数 http 分支的 `core/remote_fetch.py:17-46` 有 `validate_http_url` 私网/凭据 URL 拦截;`engines/single_flow.py:443-455 _require_image_format` 对下游图片做魔数检测。三处防护齐备,唯独本地路径分支裸奔,且扩展名被 `chat_media_import_service.py:39-40` 强制改写为 jpg,非图片内容会被当图片存储。\n\n五、下调严重度的依据(唯一能替被审代码辩护之处):适配器按 url 优先取值,主流 OneBot 收图带 http url 会走受保护的 http 分支;且部分平台适配器下载到本地临时文件后把路径写进 `file`,说明该 else 分支本有正当用途。要真正读到宿主任意文件,需发送方能影响到非-URL 字段(依赖具体协议端实现是否透传 sender 提供的 file 值),并非所有部署都能直接利用,故定 P1 而非 P0。另外发现中\"超大 data URI 先完整解码进内存\"一条影响被夸大:`base64.b64decode` 相对原始文本仅约 4/3 放大,攻击者本就需先传入等量文本,不构成独立 DoS 面。异常会被 `main.py:1168-1170` 的 `except Exception as exc: yield event.plain_result(f\"❌ 素材导入或识别失败:{exc}\")` 捕获并把原始异常(含解析后的绝对路径)回显到聊天,这反而额外提供了路径探测 oracle。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/features/materials/AdvancedMaterialImportPage.vue:1304-1317(模板渲染点 :1656)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "经 Read 与 od -c 逐字节核验,缺陷属实,非幻觉。\\n\\n【1. 代码真实存在,行号对得上】AdvancedMaterialImportPage.vue:1304 起的函数体:\\n```\\nfunction styleValidationReviewTrust(validationCase: AdvancedStyleValidationCase) {\\n const coverage = Number(calibration.dimension_coverage)\\n const inconsistent = Boolean(calibration.review_inconsistent || calibration.dimension_consistency_capped)\\n const risks = Array.isArray(review.semantic_risks) && review.semantic_risks.length > 0\\n if (inconsistent || risks) return \\\"????????????????\\\"\\n if (Number.isFinite(coverage)) return `????? ? ???? ${Math.round(coverage * 100)}%`\\n return \\\"?????\\\"\\n}\\n```\\n`od -c` 确认为真正的 ASCII 0x3F,而非 Read 工具渲染假象。紧邻函数中文完好(styleValidationScenarioLabel 的 \\\"异角色头像\\\"、styleValidationStatusLabel 的 \\\"待审核\\\"、styleValidationReviewSummary 的 \\\"主要偏差:\\\"),证明是局部转码事故。grep 全仓库 src/ backend/ core/,`????` 字面量仅此三行。\\n\\n【2. 不存在守卫使其不可达】模板 :1656 `<small v-if=\\\"styleValidationFidelityPercent(validationCase) != null\\\" class=\\\"advanced-style-review-trust\\\">{{ styleValidationReviewTrust(validationCase) }}</small>` 无条件渲染返回值。后端确实产出全部依赖字段:backend/services/advanced_style_set_service.py:2345 `\\\"fidelity_score\\\": score`、:2347 `\\\"score_calibration\\\": score_meta`、:3346 `\\\"dimension_coverage\\\": round(available_weight, 4)`、:2353 `\\\"semantic_risks\\\": semantic_risks`。函数内的 typeof===\\\"object\\\" 仅防属性访问异常,不影响返回值;无 try-catch 或上游过滤能拦截。\\n\\n【3. 已进入构建产物】web/assets/index-BeGrmMHB.js 命中 `...y.semantic_risks.length>0;return Je||W?\\\"????????????????\\\":Number.isFinite(xe)?`????? ? ???? ${Ma...`,pages/imagestudio/assets/index-3MfZqRFA.js 同样命中,即已打包发布,不只是源码问题。\\n\\n【4. 严重度被夸大,下调为 P2】报告称「这条关键信息完全丢失,等于在缺少可信度提示的情况下做封存决策」不成立:同卡片 :1655 `自动还原度:{{ styleValidationFidelityPercent(validationCase) }}%` 与 :1657 `主要偏差:${mismatches.slice(0,2).join(\\\";\\\")}` 中文完好,审核人仍可见数值分数与偏差摘要,丢失的只是「维度覆盖率/评分一致性」辅助提示。且表现为肉眼可辨的乱码(显式降级),不会静默给出错误可信度误导决策。纯展示层,无数据损坏、无安全面、无崩溃,故为 P2-中等而非更高。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "database/repositories/profile_repository.py:46 / :73 / :99(触发点为同文件 get() :186-194、list() :218-232)",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "我按\"默认是误报\"的立场逐条查证,但这条推翻不了,结论是成立(仅在错误类型描述上有小瑕疵)。\n\n1) 代码真实存在,行号完全对得上(无幻觉)\n`D:\\CODEX\\projects\\AstrBot ImageStudio\\database\\repositories\\profile_repository.py` 第 46、73、99 行逐字重复三份:\n```\n AND COALESCE(legacy_card_workflow.metadata_json, '') LIKE '%\"final_archive_history_id\"%'\n```\n分别位于 PROFILE_GENERATION_KIND_COLUMN(32-57)、PROFILE_CATEGORY_COLUMN(59-84)、DETAILED_WORKFLOW_STATE_COLUMN(86-117)。全仓 grep `LIKE '%` 只有这三处(其余命中都在 .tmp/ 历史发布包里),说明这是孤例而非全局风格。\n\n2) 确实带参执行,%-展开路径可达\nget()(186-194 行)与 list()(218-232 行)都把这三个常量拼进 SQL 且传了参数元组:\n```python\nrow = conn.execute(\n f\"\"\"\n SELECT {PROFILE_COLUMNS}, {PROFILE_GENERATION_KIND_COLUMN},\n {PROFILE_CATEGORY_COLUMN}, {DETAILED_WORKFLOW_STATE_COLUMN}\n FROM character_profiles\n WHERE profile_id = {placeholder}\n \"\"\",\n (profile_id,),\n).fetchone()\n```\nlist() 的第二条查询同样 `tuple(params)`;即使 params 为空,`_Psycopg2ConnectionAdapter.execute`(db_helper.py:265-268)写的是 `cursor.execute(query, tuple(params or ()))`,永远传的是元组而不是 None,所以 psycopg2 的 %-展开无法被\"无参跳过\"绕开。\n\n3) placeholder 确为 %s,PG 分支真实可达\ndb_helper.py:288-290:\n```python\n @property\n def _placeholder(self) -> str:\n return \"%s\" if self.backend == \"postgresql\" else \"?\"\n```\nDatabaseSettings.selected_backend(第 108-117 行)在 `backend: postgresql/postgres/pg` 或 `postgresql.is_configured`(enabled=true 且有 dsn/host+db+user)时返回 \"postgresql\";config.yaml 第 151-167 行把 backend: auto/sqlite/postgresql 和 postgresql.enabled 作为对外配置暴露。connect() 第 158-171 行在非 sqlite 时走 `_connect_postgresql()`,优先 psycopg(3),ImportError 才回退 psycopg2。\n\n4) 没有任何上游守卫/框架机制消解它\n- 没有 try/except 包裹(get/list 里只有 `finally: conn.close()`,不吞异常)。\n- 没有 ORM:全是手写 f-string SQL + DBAPI 参数化,驱动层的 %-模板处理无法回避。\n- 同文件 _build_filters(305-310 行)恰恰用的是正确写法 `like_value = f\"%{keyword}%\"` 放进 params,反证作者本来知道该怎么写,这三个常量是漏网。\n- 调用面极广且都是核心读路径:ProfileRepository 在 branch/history/integration/node/node_material/profile/profile_card/review/wardrobe 等 9+ 个 routes 中实例化,profile_repository.get() 在 generation_service、workflow_service、profile_card_service、resource_access_service、wardrobe_service 等处被调用,一旦 PG 生效就是全线崩。\n- PG 不是\"名存实亡\"的分支:migrations 目录 18 个脚本 grep 不到 AUTOINCREMENT/json_extract/INSERT OR IGNORE 等 SQLite 专有语法,且 tests/test_workflow_rounds.py:142 有 `test_round_migration_stays_sqlite_and_postgresql_portable`,说明 PG 兼容性是被刻意维护的目标。\n\n5) 唯一需要修正的地方:错误类型与严重度\n- psycopg2 侧原报告正确:整条 query 被当作 %-模板,`%\"` 会抛 `ValueError: unsupported format character '\"' (0x22)`。\n- psycopg3 侧原报告说的 IndexError 不准确:psycopg3 的占位符解析在 `%` 后取一个字符再校验,非 s/b/t 会抛 `ProgrammingError: only '%s', '%b', '%t' are allowed as placeholders, got '%\"'`。失败结论一致,只是异常类型不同(本机 psycopg/psycopg2 均未安装,requirements.txt:6-7 里两者都是注释掉的可选依赖,无法实机复现,此处依据驱动实现推断)。\n- 严重度我下调到 P1:默认 `backend: auto` + `postgresql.enabled: false` 时走 SQLite,完全不受影响,必须用户主动开启 PG 才 100% 触发;不是开箱即崩,够不上 P0。但一旦开启即角色列表/详<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "database/db_helper.py:159-171",
|
|||
|
|
"reason": "代码事实属实但影响链条不成立,按描述的缺陷判为误报。\\n\\n1) 事实核对通过:database/db_helper.py:159-171 与证据片段逐字一致,`conn = sqlite3.connect(db_path, timeout=30.0)` 后只有 `PRAGMA foreign_keys = ON` 和 `PRAGMA busy_timeout = 30000`;全仓库 backend/core/database/tests 内 grep `journal_mode` 零命中,确实未启用 WAL。BEGIN IMMEDIATE 的两处行号(workflow_round_repository.py:313、:815)也对得上。\\n\\n2) SQLite 锁语义被说错。rollback journal 模式下 BEGIN IMMEDIATE 只拿 RESERVED 锁,RESERVED 不阻塞读者(读者继续持 SHARED 读原始 db 文件,写者脏页在 page cache/journal 中),只有 commit 瞬间升级 PENDING/EXCLUSIVE 才短暂挡读。\"写事务持有 RESERVED/EXCLUSIVE 锁会阻塞全库所有读\"不成立。\\n\\n3) 所谓\"长事务\"里没有任何慢 I/O。通读 confirm_candidate_atomic(database/repositories/workflow_round_repository.py:801-1131)全文,事务内只有本地 SELECT/UPDATE/INSERT(如 `UPDATE workflow_round_candidate_pool SET is_final = 0 ... WHERE round_id = ?`、`INSERT INTO workflow_history (...)`),配合 `datetime.now(timezone.utc)`、`uuid4()`、json 序列化,没有 provider 调用/网络/文件 I/O;get_consistent_round_snapshot(:300-354)是 3 条 SELECT 后立即 `conn.commit()`。事务寿命毫秒级,与 30s busy_timeout 相差约 4 个数量级,\"provider 调用稍慢就 30s 超时\"无代码依据。\\n\\n4) 声称的并发场景是虚构的。grep 全仓库,confirm_candidate_atomic 唯一调用方是 backend/services/workflow_round_service.py:694(用户确认候选的 API),get_consistent_round_snapshot 唯一调用方是 :473 create_next_round,二者均为交互式单用户动作;而 chat_workflow_service.py:216 的 4 线程提交的是 `executor.submit(self.generate_node, str(node[\"node_id\"]), chat)` → `nodes/{id}/generate` 链路,根本不会进入 confirm_candidate_atomic。\"任一线程进入 confirm_candidate_atomic 阻塞其余 3 线程\"不可能发生。\\n\\n5) except 引用不准且不相关。真实位置是 generation_service.py:1865-1866 与 1887-1888(:1890 是 `\"generation.node_failed\"` 事件名,不是 except),且两处都在 `_record_generation_failure` 内,仅在生成已失败后的失败记录路径执行,不覆盖正常写入路径。\\n\\n6) 另佐证:image_repository.create 等均为 connect→execute→commit→close 的短连接模式(database/repositories/image_repository.py:44-48),generation_service 自身不持有 connection 跨越 provider 调用;仓库内无任何 `database is locked`/OperationalError 处理或文档记录,无该问题实际发生的痕迹。\\n\\n结论:开启 WAL 仍是合理的低优先级调优(提升并发读写吞吐、减少 fsync),但不构成所描述的\"生成静默失败\"级缺陷,故降级为 P3 并判定该发现站不住脚。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_location": "backend/services/generation_service.py:1869-1886(残留的弱化问题:check-then-act 未用 update_node_if_status)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "代码片段本身不是幻觉,但这条发现的**因果链和影响是编造的**,且完全漏掉了服务层已有的合并守卫。\n\n**1. 代码确实存在(唯一成立的部分)**\n`database/repositories/workflow_repository.py:298-318` 的 `update_node` 和 `:154-178` 的 `update_workflow` 确实是整列覆盖 `metadata_json = ?`,无 merge/版本号。行号对得上。\n\n**2. 声称的影响路径在类型上不可达 —— 这是决定性的反驳**\n发现称\"`generation_service._prepare_quick_character` 用旧快照 `update_workflow`,把 `confirm_candidate_atomic` 写的 `stage_inputs` 抹掉\"。但这两条路径永远不会碰同一行工作流:\n\n`backend/services/generation_service.py:612` 只有 `is_quick_character_node(workflow, node)` 为真才进入 `_prepare_quick_character_context`,而 `backend/services/quick_character_consistency.py:49-53`:\n```python\ndef is_quick_character_node(workflow: Any, node: Any) -> bool:\n return (\n str(getattr(workflow, \"workflow_type\", \"\") or \"\").strip() == \"character_fast\"\n and str(getattr(node, \"node_key\", \"\") or \"\").strip() in QUICK_CHARACTER_TARGETS\n )\n```\n即该路径**只作用于 `character_fast`**。而回合/候选确认体系(`stage_inputs`/`branch_state`/`character_execution` 的写入者)在 `backend/services/workflow_round_service.py:1130` 与 `:147` 被硬性拒绝非渐进类型:\n```python\nif workflow.workflow_type != \"character_gradual\":\n raise ApiError(..., {\"workflow_id\": workflow.workflow_id, \"workflow_type\": workflow.workflow_type})\n```\n`character_fast` 工作流根本不存在 `stage_inputs`,\"确认了候选图但下一节点拿不到上游图\"的症状无法由此产生。\n\n**3. \"用更早读到的快照\"这句是事实错误**\n`generation_service.py:1145-1149` 恰恰是在锁内**重新读取**,紧接着 :1246 写回,中间没有跨请求的陈旧快照:\n```python\nlock_entry.lock.acquire()\ntry:\n current_workflow = self.workflow_service.get_workflow(workflow.workflow_id).workflow\n metadata = dict(current_workflow.metadata or {})\n```\n而且 :1246 的 `update_workflow` 只在 `else`(尚未 managed,即批次首次初始化)分支执行,二次进入走 `managed` 分支根本不写。\n\n**4. 服务层存在被完全忽略的合并守卫**\n`backend/services/workflow_service.py:234` 先 `before = self.get_workflow(workflow_id).workflow`,随后 `:268-275` 强制走 `_normalize_workflow_metadata`。该函数 `:1027-1031` 与 `:1058-1068`:\n```python\nif workflow.workflow_type == \"character_gradual\" and not allow_character_execution_update:\n if \"character_execution\" in current_metadata:\n normalized[\"character_execution\"] = current_metadata[\"character_execution\"]\n...\nif workflow.workflow_type == \"character_gradual\":\n if isinstance(current_metadata.get(\"branch_state\"), dict):\n # Clients edit only the active projection. Sibling branch\n # snapshots are server-owned and must survive stale PUTs.\n normalized[\"branch_state\"] = current_metadata[\"branch_state\"]\n```\n也就是说,发现点名的 `branch_state / character_execution / completed_node_ids` 恰恰是**从新鲜读取的 `before` 强制回填**的,调用方传入的整包 metadata 覆盖不掉它们。仓储层的\"整列覆盖\"在这几个键上已被服务层消解。\n\n**5. 写入方本身是原子的,且仓储已提供 CAS 变体**\n`workflow_round_repository.py:812-818` 的 `confirm_candidate_atomic` 用 `BEGIN IMMEDIATE`(sqlite)/ `FOR UPDATE`(postgresql)在同一事务内读 + 写 `:1112-1120`,本身无丢更新。仓储也已经有 CAS 版本 `update_workflow_if_status`(`workflow_repository.py:180-231`,`cursor.rowcount` 判定)和 `update_node_if_status`(`:324` 起),并非\"只靠进程内锁\"。\n\n**残留的、被大幅夸大后剩下的真实点**\n`generation_service.py:1869-1886` 的失败回写确实是 check-then-act 却用了普通 `update_node` 而非同文件仓储
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/features/quick-profile/profile-import.ts:149-172 (fallback 定义) + :336 (return { value: fallback, recognized: false }) + frontend/standalone-web/src/App.vue:9710-9739 (applyExternalProfileImport 未按 recognizedKeys 过滤)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "经逐行核对,该发现成立,非幻觉、非误报。\\n\\n【1. 代码真实存在,行号对得上】profile-import.ts:149-166 确为:\\nconst personality = resolveImportedProfileField(source, [\\\"性格\\\",\\\"人格\\\",\\\"个性\\\",\\\"personality\\\",\\\"trait\\\",\\\"traits\\\"], [...], \\\"性格关键词待从原文继续整理\\\")\\nconst appearance = resolveImportedProfileField(..., \\\"外观特征待从原文继续整理\\\")\\nconst outfit = resolveImportedProfileField(..., \\\"服装与配饰待从原文继续整理\\\")\\nconst scene = resolveImportedProfileField(..., \\\"常用场景待补充\\\")\\n:327-337 的 resolveImportedProfileField 三段式确认:label 命中→recognized:true;关键词行命中→recognized:true;否则 `return { value: fallback, recognized: false }`(:336)。即\"未识别\"时提示语被当成真实值返回,并在 :295-298 原样写进 draft 的 personality/appearance/outfit/scene。\\n\\n【2. 无上游守卫,路径可达】App.vue:9710 applyExternalProfileImport() 把 draft 的全部字段无条件赋给 quickProfileQuestionnaire.value(:9721 `personality: draft.personality`、:9722 `appearance: draft.appearance`、:9723 outfit、:9724 scene);recognizedKeys 仅在 :9740-9759 用于判断是否切换到 advanced 模式,从未用于过滤 value。全仓 grep \\\"待从原文继续整理\\\" 只命中源文件与 .tmp/deploy-* 构建快照,前后端均无占位符剔除逻辑。\\n\\n【3. 下游污染链路属实】\\n- App.vue:9568-9571 `personality: [questionnaire.personality, questionnaire.likes, questionnaire.dislikes].map(i=>i.trim()).filter(Boolean).join(\\\";\\\") || \\\"性格待细化\\\"` —— 占位串非空,因此 `|| \\\"性格待细化\\\"` 不会触发,占位串直接成为角色 personality。\\n- App.vue:9367 `base_prompt: [profile.name, profile.identity, profile.personality ? \\\\`personality: ${profile.personality}\\\\` : \\\"\\\"]` —— 确认以 `personality: 性格关键词待从原文继续整理` 形式进入 base_prompt。\\n- 外观/服装同样受害::9554 appearanceSource 由 questionnaire.appearance 拼成;:9573-9577 `hairstyle: extractQuestionnaireTrait(appearanceSource, [...], appearanceSource || \\\"发型待细化\\\")` 的兜底参数就是 appearanceSource 本身,占位串会被写进 hairstyle;:9579 `clothing: questionnaire.outfit.trim() || \\\"服装与配饰待细化\\\"` 同理;二者再经 :9370 拼成 appearance_prompt。\\n- 即使走后端草稿分支(:9364 `backendDraft?.profile_preview ?? {...}`),:9407-9434 buildQuestionnairePayload 仍把 personality/appearance/outfit/scene 原样序列化发给后端 LLM,污染文本一样进入模型输入。\\n\\n【4. 需要修正的两处,故降级为 P2 而非 P1】\\n(a) 原描述称\"personality/appearance/outfit/scene/avoid 五个字段\"均为提示语,不准确:avoid 的 fallback 是 \\\"低质量、结构错误、过度复杂、与角色设定冲突\\\",这是一条语义合法的负面提示词,不构成污染。真正有害的只有 personality/appearance/outfit/scene 四个。\\n(b) 存在部分缓解,削弱\"不可控\"强度:App.vue:947-956 生成预览行并计算 recognized;:13044 `:class=\\\"{ pending: externalProfileImportSource && !row.recognized }\\\"`、:13048 `{{ row.recognized ? \\\"已识别\\\" : \\\"待整理\\\" }}` 会把未识别字段明确标注为\"待整理\";且 :9762 applyExternalProfileImport 末尾调用 openPersonaQuestionnaire(),用户在保存前会进入可编辑问卷。因此不是静默直达生成,而是\"默认值看起来已填好、用户可能不修改\"的风险,原描述中\"用户看到已填好也不会去清理\"属未经验证的行为假设。\\n\\n结论:真实缺陷,性质为数据质量/提示词污染,非崩溃、非安全、非数据丢失,且有 UI 标注与人工确认关卡,定级 P2-中等。修复建议:recognized:false 时写入空串并把提示语改为输入框 placeholder,或在 applyExternalProfileImport <EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/features/settings/SettingsPage.vue:133(错误的 UI 承诺文案)+ :98(saveAndProbeAdapter 注入 api_key_plaintext)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "尝试推翻失败,代码链路逐环确认成立,但严重度被夸大,应降为 P2。\n\n【1. 代码真实存在,行号完全对得上】\n`SettingsPage.vue:98` 确为 `saveAndProbeAdapter`,逐字确认:\n```\nconst key=probeApiKeyValue(a.id).trim();\nif(Array.isArray(generation.adapters)){generation.adapters=generation.adapters.map(item=>{const adapter=obj(item);if(String(adapter.id||adapter.provider_id)!==a.id)return item;return{...adapter,...(key?{api_key_plaintext:key}:{})}})}\nconst provider=obj(generation.provider);if(String(provider.id||provider.provider_id)===a.id&&key)generation.provider={...provider,api_key_plaintext:key};\n...const saved=await postJson<SaveSettingsResult>(endpoints.settingsSave,{config});\n```\n`:133` 的文案也确为该输入框自身的 `<small>`:`placeholder=\"仅本次页面会话使用\" ... <small>点击测试时会直接读取输入框当前值,兼容浏览器自动填充;不写入配置或日志。</small>`,且 `:value=\"probeApiKeys[a.id]\"`,与 `probeApiKeyValue()` 读取的是同一状态。不是幻觉。\n\n【2. 不存在上游守卫使其不可达】\n后端 `SettingsService.save()`(settings_service.py:415-445)中:\n- `_merge_masked`(:702-713)只跳过字面量 `\"***MASKED***\"/\"***PATH***\"`,不识别也不剥离 `api_key_plaintext`;\n- `_restore_runtime_managed_fields`(:812-824)只覆盖 `get_runtime_config_overrides()` 里的路径,且 `_runtime_override_changes`(:798-801)在存在 `generation.adapters` 时反而把 `generation.provider.*` 排除在托管字段之外,不会抹掉明文;\n- 最终 `write_text_atomic(self.config_service.config_path, dump_yaml(merged))`(:445)。`_mask_sensitive` 只出现在 :453/:454/:456 的审计与返回值上,确实不作用于落盘内容。\n\n【3. 序列化层也不会消解】\n`core/config_writer.py:_format_scalar`(:79-83)对 list 中的 Mapping 走 `json.dumps(value, ensure_ascii=False)`,适配器对象连同 `api_key_plaintext` 以 JSON flow 形式原样写进 YAML。`config.yaml` 经 `git ls-files` 确认被跟踪。落盘属实。\n\n【4. 但严重度应下调,理由有三】\n(a) 明文持久化是**既定设计**而非意外泄漏:`backend/services/model_provider_runtime.py:350` 的 docstring 就是 \"Read the UI-saved adapter credential without copying it into the DB.\",:366 从落盘配置读回 `api_key_plaintext`;而 :98 的探测请求带 `use_saved_provider:true`,逻辑上必须先落盘。真正的缺陷是 `<small>` 文案陈旧/错误,而非行为越权。\n(b) 用户在动作点被正确告知:同一对话框 `:134` 的探测面板正文写着「会先保存当前适配器和 API Key,再由后端重新读取已保存配置执行测试。」——即用户点击的那个按钮旁边明确说明了会保存密钥。两处文案自相矛盾,用户不至于完全被蒙蔽,不构成 P1「凭据泄漏」。\n(c) `duplicateAdapter`(:89)复制的 `probeApiKeys[duplicateId]` 仅为前端内存态,只有用户再对副本点「保存并测试」才会落盘,「扩大扩散面」的说法偏弱,不足以抬高定级。\n\n结论:真实缺陷,但性质是「UI 承诺与实际持久化行为不一致(误导性文案)」,定 P2-中等。修复应二选一:改文案并加二次确认,或改用 `api_discovery_service.py:229` 已支持的 `api_key_plaintext` 直传探测端点、去掉先 save 的链路。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"corrected_location": "backend/services/asset_import_service.py:141-152(配合 :340、:361-366)",
|
|||
|
|
"reason": "试图推翻失败,代码与逻辑链条逐条核实成立,仅严重度和一处细节需修正。\n\n一、行号与代码属实,非幻觉。asset_import_service.py:141-145 原文即为:\n existing = self.material_service.repository.get_material(material_payload[\"material_id\"])\n if existing is not None:\n self.material_service.get_material(material_payload[\"material_id\"])\n self.material_service.validate_material_payload(material_payload, creating=existing is None)\n planned_materials.append((material_payload, existing is None))\n:152 `material = self.material_service.update_material(material_payload[\"material_id\"], material_payload, ctx)`;:340 `\"recognition_status\": _optional_text(item.get(\"recognition_status\")) or \"pending_recognition\"`;:361-366 `_material_id` 中 `explicit = _optional_text(item.get(\"material_id\") or item.get(\"id\")); if explicit: return explicit` 全部一致。\n\n二、找不到任何上游守卫让它不可达。\n1) 路由 backend/routes/asset_routes.py:27 只有 `require_permission(ctx, \"can_manage_materials\")`,之后直接把 package_uri 交给 plan_import_pack;route_registry.py:304 已注册 `POST assets/import-pack`,是活路由,不是死代码。\n2) 唯一的\"状态守卫\"在 MaterialService.confirm_material:603 `if material.recognition_status not in {\"pending_confirm\", \"confirmed\"}: raise ...`,而导入路径根本不经过 confirm_material,是直接把 recognition_status 塞进 create_material/update_material。validate_material_payload:273-274 只调用 _validate_status,而 _validate_status:663-669 确实只判断 `status not in MATERIAL_STATUSES`(:42 集合含 \"confirmed\")。所以清单写 \"recognition_status\": \"confirmed\" 可造出零 VLM 调用、prompt 完全由清单作者书写的\"已确认\"素材。\n3) 覆盖确为整体替换,且比原描述更彻底:MaterialService.update_material:239-250 的合并分支 `structured_keys & update_payload.keys()` 只匹配**顶层**的 raw_analysis/normalized_asset_profile/generation_capsules,而导入构造的 payload 把它们藏在 `metadata` 里(:341-348 直接重建 metadata dict),因此该分支不触发,直接落到 material_repository.update_material:262-264 `column = \"metadata_json\"; value = _json_or_none(value)` —— 整列覆盖,旧的 normalized_asset_profile / recognition / confirmed 标记被无声抹掉。\n4) 静默性也属实:dry_run 预览里 `applied` 是 plan_import_pack:49 写死的全零 dict,_manifest_summary 只回传 name/version/materials 计数/errors,既不列出 material_id,也不区分 create 还是 update,操作者在执行前看不到\"将覆盖你已确认的 N 条素材\"。\n5) 信任标记消费点属实:generation_reference_service.py:95 `if str(material.recognition_status or \"\").strip() != \"confirmed\": raise`,workflow_node_material_service.py:293 `if material.recognition_status != \"confirmed\": raise`,wardrobe_service.py:68/194/233 与 generation_service.py:1659 同样只认这一个状态。伪造出的 confirmed 素材会被生成链路当成可信参考直接使用。\n6) 无任何测试覆盖(tests/ 下无 asset import 用例,全仓 grep 只命中 .tmp 部署副本),也无 ORM/框架层机制消解——SQL 虽是参数化的,但参数化只防注入,防不住\"合法字段被合法覆盖\"。\n\n三、需要修正的地方(这是原描述唯一夸大之处):\n\"_asset_uri 还允许清单写 storage://<任意根>\" 不成立。VfsStorage.split_uri:117-129 会拒绝非 storage:// 前缀、拒绝绝对路径与 `..`,resolve_uri:100-109 通过 _root_paths() 查 _storage_roots(config),白名单仅 images/thumbnails/materials/advanced_materials/styles/archives/temp 七个根,未知根抛 MEDIA_NOT_FOUND,越界抛 FORBIDDEN;且 :138 `self.storage.ensure_file(...)` 要求文件真实存在。所以它只能引用 ImageStudio 自有存储根内已存在的文件,不是任意文件系统读取/注入原语,\"绕过包目录范围\"成立但危<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/features/quick-profile/profile-import.ts:32-59(数据源)+ frontend/standalone-web/src/App.vue:759(注入点)+ frontend/standalone-web/src/App.vue:34(未被调用的空表单工厂导入)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "我尝试推翻这条发现,但四个方向都没能推翻,核心事实成立;只有\"影响\"描述有夸大,需要下修严重度。\n\n1)代码真实存在,行号完全对得上,不是幻觉。\n`frontend/standalone-web/src/features/quick-profile/profile-import.ts:32` 起确实是:\n```ts\nexport function createDefaultQuickProfileQuestionnaire(): QuickProfileQuestionnaire {\n return {\n roleCore: \"近未来港区巡查员,负责夜间巡逻和异常记录。\",\n genderPresentation: \"female\",\n ageStage: \"adult\",\n species: \"human\",\n bodyType: \"athletic\",\n distinctiveFeatures: \"发光臂章,青绿色眼睛\",\n personality: \"冷静、可靠、行动快,外冷内热。\",\n appearance: \"黑色短发,青绿色眼睛,轻机能外套。\",\n outfit: \"短靴、发光臂章、便携通讯设备。\",\n scene: \"星港、雨夜街区、巡查终端、航灯背景。\",\n avoid: \"过度复杂、低清晰度、肢体结构错误。\",\n```\n`App.vue:759` 也确实是 `const quickProfileQuestionnaire = ref<QuickProfileQuestionnaire>(createDefaultQuickProfileQuestionnaire())`,且表单是 `v-model=\"quickProfileQuestionnaire.roleCore\"`(App.vue:13097 等)真实双向绑定,不是 placeholder 占位提示——placeholder 是另写的 `placeholder=\"角色名、身份、核心定位,可以只写一句。\"`。所以用户看到的是被填满的输入框。\n\n2)不存在让问题不可达的上游守卫/重置。我通读了所有打开路径:\n- `openPersonaQuestionnaire()`(App.vue:8031-8035)只做 `closeAllOverlays` + 清错误 + `showPersonaQuestionnaire.value = true`,**没有任何重置**;\n- `createProfileFromProfilesPage()`(App.vue:6399-6403)\"新建角色\"入口直接 `void nextTick(openPersonaQuestionnaire)`,同样不重置;\n- 全仓库只有 App.vue:9689/9694(文档导入)和 App.vue:9714(文本导入)会覆写这个 ref,都是导入路径,走不到就保持演示值。\n\n更有力的旁证:`createEmptyQuickProfileQuestionnaire`(profile-import.ts:61)在 App.vue 里**只在第 34 行 import、全文件零次调用**(grep 计数为 1,即只有 import 那一行)。说明\"空表单\"是写好了却从未接线,默认值仍是演示角色。\n\n3)没有框架层机制消解它。这是纯前端 ref 初值,Vue 不会替你清空;后端也管不到表单初值。\n\n4)防线缝隙属实。`frontend/tests/production-data-boundary.test.mjs:5-9` 的扫描面确实只有两个文件:\n```js\nconst appPath = new URL(\"../standalone-web/src/App.vue\", import.meta.url)\nconst materialDialogPath = new URL(\"../standalone-web/src/components/MaterialDetailDialog.vue\", import.meta.url)\n```\n我实跑了该测试:`✔ production UI never injects disconnected demo images, profiles, materials, or models`,1 passed 0 failed——即这批演示数据确实在测试全绿的情况下活着。与 `ProfilesPage.vue:156` 的空态文案「新建或导入角色后即可管理生成设定。」自相矛盾也属实。\n\n需要修正的部分(这是我唯一推翻掉的地方):声称的\"直接点『生成』就会消耗真实 API 额度生成这个虚构角色的四张图\"是夸大的。问卷 footer 的按钮文案是「生成档案」(App.vue:13323),点击走 `generateQuickProfileArchiveFromQuestionnaire`(App.vue:9251),它 POST 的是 `endpoints.quickProfileDraft` 与 `endpoints.quickProfileDraftSave`,落库的是文本档案 + quick_profile 工作流,footer 自己也写明「生成后会调用后端保存角色档案、创建 quick_profile 工作流,并自动选中角色卡。」,路径里没有四张出图;而且前面还有 `canGenerateQuickProfileArchive` 和 `backendWriteBlockMessage(\"生成档案\")` 两道闸。所以实际后果是\"一键把一条虚构演示角色写进真实档案库\",属于数据卫生/UX 问题,不是烧图片额度的致命问题——故定为 P2-中等而非更高。",
|
|||
|
|
"refuted": false
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:4576",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "推翻失败,该缺陷成立,但严重度需下调至 P2。\n\n【1. 代码真实存在,行号精确对得上】App.vue:4572-4577:\nconst previousIndex = hadPrevious\n ? stream.events.findIndex((event) => event.event_id === lastEventStreamId)\n : -1\nconst freshEvents = hadPrevious\n ? stream.events.slice(previousIndex >= 0 ? previousIndex + 1 : -1)\n : []\n同一三元的另一分支是 previousIndex + 1(语义为\"游标之后的全部\"),说明作者意图就是取整段新事件;游标缺失时正确写法应为 slice(0),而 slice(-1) 只取最后 1 条。这不可能是有意的保守回退——freshEvents 只用于触发 loadJobs/loadDashboard/loadGalleryImages/loadLogs,重复处理完全幂等,少取没有任何收益。\n\n【2. 前提条件逐条验证通过】\n- 窗口大小属实:前端固定传 { limit: 20 }(App.vue:4568)。\n- 顺序假设属实:backend/services/event_stream_service.py 的 recent_events 用 `selected = events[-normalized_limit:]` 返回升序尾部切片,故 stream.events.at(-1) 确为最新,findIndex 找不到即代表游标已滑出窗口。\n- 场景①属实:poll 首行 `if (backendConnection.value.state !== \"connected\" || document.hidden) return`(App.vue:4564),标签页隐藏时游标完全停止推进。\n- 兜底缺失属实:reconnectBackendWhenVisible(App.vue:4471-4479)在 state === \"connected\" 时只 `loadWorkflowRoundDetail(roundId, true)` 后 return,确实补不了 jobs/gallery/logs/dashboard;setActiveNav(App.vue:5102)切页不触发任何 reload;全仓前端仅 3 处 setInterval(App.vue:4002 头像轮播、4608 事件流、AdvancedMaterialImportPage.vue:316),无其他兜底轮询。\n\n【3. 无上游守卫/框架机制消解】唯一的 try/catch(App.vue:4603)是空吞异常的 `// Keep the existing manual refresh and polling paths as a silent fallback.`,反而加重了\"静默失败\"。\n\n【4. 严重度被夸大,应为 P2 而非更高】\n- 原描述称\"队列、图库、日志全部不刷新、工作流停更\",但 scheduleWorkflowRoundPolling(App.vue:6999-7008)对活跃轮次有独立的 4 秒 setTimeout 自续轮询,生成过程中最关键的工作流工作台视图并不会冻结,这一条不成立。\n- 自愈性:lastEventStreamId 已被更新为 newest,只要后续再到任意一条 generation.* 事件,下一轮 poll 游标必在 20 条窗口内,UI 自动恢复。真正长期卡死需同时满足\"游标滑出窗口 + 该批最后一条恰为非 generation.* + 此后不再有新事件\",属批量生图收尾时的特定组合,而非常态。\n- 影响仅为 UI 数据陈旧,无数据损坏、无安全影响,手动刷新或触发新生成即可恢复。\n\n修复建议:slice(previousIndex >= 0 ? previousIndex + 1 : 0)。",
|
|||
|
|
"refuted": false
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:11146(displayMediaUrl)+ :11203-11225(resolveSignedMediaUrl)+ :997-1011(galleryImages computed)+ :4041 (唯一 revoke 点)",
|
|||
|
|
"corrected_severity": "P1-严重",
|
|||
|
|
"reason": "尝试推翻失败,该发现成立,且行号逐条核对无误、无幻觉。\n\n一、代码真实存在,位置精确\n`App.vue:11146` 确为 `function displayMediaUrl(rawUrl?: string | null): string`,读 `resolvedMediaUrls.value[value]` 建立响应式依赖(:11150 / :11163),未命中即在函数体内直接发起 `resolveStorageMediaUrl(value)`(:11153)或 `resolveSignedMediaUrl(value, token)`(:11167)。`:11215-11221` 确实是 `URL.createObjectURL(...)` → `ownedMediaObjectUrls.add(objectUrl)` → `resolvedMediaUrls.value = { ...resolvedMediaUrls.value, [rawUrl]: objectUrl, [sourceKey]: objectUrl }` 整体替换 ref。调用点 `:997`(galleryImages computed,对全部图片逐张 `displayMediaUrl(image.image_url)` / `(image.thumbnail_url)`)、`:2341`、`:2352` 全部核实;模板侧 `:11668 / :11693 / :12826 / :12865 / :13961` 通过 `imageUrl()`(:11137,内部串联 4 次 displayMediaUrl)在 render 中调用,且 `v-if=\"imageUrl(image)\"` 与 `:src=\"imageUrl(image)\"` 同一节点调用两次。声明处 `:686-688` 确认 `resolvedMediaUrls` 是 ref、`resolvingMediaUrls`/`ownedMediaObjectUrls` 是非响应式 Map/Set。\n\n二、不存在让问题不可达的守卫,反而后端保证必然走副作用分支\n我原以为可以靠\"后端返回普通 http URL、signedMediaToken() 返回空字符串直接 return value\"来推翻。但 `backend/routes/image_routes.py:213-232` 的 `_serialize_image` 显式把每条记录改写为签名 URL:`payload[\"image_url\"] = _signed_url(media_service, image.image_uri)`,而 `core/media_proxy.py:49` 生成的正是 `f\"{MEDIA_FETCH_PATH}?token={token}\"`,MEDIA_FETCH_PATH = `/api/plug/astrbot_plugin_imagestudio/media/fetch`,恰好命中 `signedMediaToken()` 的 `url.pathname.endsWith(\"/media/fetch\")` 判定(:11196)。所以图库每一条 image_url/thumbnail_url 都必定进入发请求分支,不可达路径不存在。\n\n三、无框架层机制消解,反而放大\n- `frontend/shared/api/client.ts:30` 的 `getJson` 只是 `request()` 薄封装,无缓存/去重层。\n- `resolvingMediaUrls` Map 只去重\"同一 URL 的并发在途请求\",`.finally(() => resolvingMediaUrls.delete(value))` 且两个 resolve 函数的 catch 都只 `console.warn` 不写缓存(:11188-11190 / :11222-11224)——失败的 URL 每次重渲染都会重新补发,构成真实的重试风暴。\n- Vue computed 惰性求值不能救场:`watch(galleryImages, ..., { immediate: true })`(:4184-4196)主动读取,每次 `resolvedMediaUrls` 被整体替换都强制 galleryImages 重跑(uniqueImages + sort + 2N 次 displayMediaUrl),N 张图 → ~2N 次失效 → O(N²),与报告一致。\n- `loadGalleryImages`(:4511-4542)确实 `pageSize = 100` 循环 `do...while (loaded.length < total)` 拉全库所有页,无虚拟滚动,N 就是全库规模。\n\n四、严重度不但没被夸大,报告还低估了\n`core/media_proxy.py` 的 `fetch_payload` 返回 `\"body_base64\": _b64encode(path.read_bytes())` —— 即 media/payload 返回的是**整张原图的完整字节 base64**(体积 +33%,走 JSON)。因此进入图库不是\"泄漏若干 Blob\",而是把全库每张图的原图+缩略图完整二进制通过 JSON 下载进内存(同一张图还被 <img src> 用签名 URL 又下了一遍,因为签名分支 :11174 先返回原 URL 作为过渡),并且 `ownedMediaObjectUrls` 仅在 `onBeforeUnmount`(:4026、:4041-4042)统一 revoke,SPA 生命周期内不释放。1000 张图的库 = 约 2000 次全文件 base64 请求 + 2000 个常驻 Blob,足以打爆标签页内存。\n\n唯一可以对报告做的修正是措辞层面:报告第 3 点\"任何无关响应式变更重渲染都可能补发媒体请求\"在解析成功后不成立(成功后已进缓存命中直接返回),仅在解析失败时成立;但这属于细节偏差,不影响结论。综合判定为 P1-严重(资源耗尽/可用性级缺陷,非安全或数据正确性问题,故不到 P0)。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "database/repositories/advanced_material_repository.py:346(模式散布于 database/repositories/ 下 13 个仓储 + core/audit_service.py)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "我试图推翻这条发现,但事实层面推不倒——每一项可验证的断言都对得上真实代码;只有「影响」被明显夸大,故降级为 P3。\n\n**1. 行号与证据片段完全属实(非幻觉)。** `awk NR>=346` 确认 346 行正是 `def get_import(self, import_id: str) -> AdvancedMaterialImport | None:`,347 行 `conn = self.database.connect()`,348 行 `try:`,后续 `finally: conn.close()` 与引用片段逐字一致。紧邻的 `get_import_by_idempotency_key` 也确如所述是同一模板换个 WHERE 列。\n\n**2. 重复计数属实(甚至比原文略多)。** 按文件统计 `conn.close()`:advanced_material 37、workflow 16、workflow_round 14、job 12、model_provider 10、material 9、image 7、capability/profile_card/wardrobe/cross_review 各 5、profile 4、integration 2、core/audit_service.py 1 = **132**;`finally:` 合计 **132**(原文说 131,误差 1,不影响结论)。\n\n**3. `_rollback` 确被独立定义 3 次,且实现并不一致**(这点原文还说轻了):\n- `advanced_material_repository.py:2119` → `try: conn.rollback() except Exception: return`(吞异常)\n- `workflow_cross_review_repository.py:1464` 与 `workflow_round_repository.py:1378` → `rollback = getattr(conn, \"rollback\", None); if callable(rollback): rollback()`(**不吞异常**)\n同名函数三份拷贝、语义分叉,是真实的收敛点。\n\n**4. `__init__` / `placeholder` 逐字重复属实。** 例如 `profile_repository.py:121-126` 与 `capability_repository.py:12-17` 完全相同:`def __init__(self, database: Database): self.database = database` + `@property def placeholder(self): return self.database._placeholder`。且这些仓储彼此无公共基类(`wardrobe_repository.py:158` 的 `super().__init__` 只是同文件内父类,非统一 BaseRepository),说明确实缺一层抽象。\n\n**5. 「连接策略变更需改 132 处」属实。** `database/db_helper.py:159 def connect()` 里 sqlite 分支硬编码 `timeout=30.0`、`PRAGMA foreign_keys = ON`、`PRAGMA busy_timeout = 30000` 后 `return conn`——它只返回裸连接,**没有任何 contextmanager/transaction 封装**(全库 grep `contextmanager|def transaction` 在 database/*.py 零命中)。所以「引入连接池就得动每个调用点」是成立的。\n\n**——但影响被夸大,这是我唯一能推翻的部分:**\n\n原文称「忘记 rollback 只能靠人工 review 保证」,并以 `profile_repository.update`(242-282 行,确实只有 `finally: conn.close()`,无 `except: _rollback`)对比 `advanced_material_repository.update_import`(396-407 行,确有 `except Exception: _rollback(conn); raise`)。对比属实,**但不构成任何运行期缺陷**:这里是「每次调用新建连接、用完立即 close」的短连接模型,而 sqlite3 与 psycopg 在 `close()` 时都会隐式回滚未提交事务;`update` 中 `conn.execute(...)` 后紧跟 `conn.commit()`,异常路径下连接必经 `finally: conn.close()` 销毁,未提交数据无法泄漏到后续请求。因此缺失的 `_rollback` 分支是**冗余写法差异,而非数据一致性风险**,「靠人工 review 兜底」的危言不成立。\n\n综上:这是一条纯粹的可维护性/DRY 观察(约 500 行样板 + 3 份分叉的 `_rollback`),事实准确、值得重构,但零运行期后果、无正确性或安全影响,定 P3-轻微而非更高。",
|
|||
|
|
"refuted": false
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:4561-4609(轮询)+ 4511-4542(loadGalleryImages 整库翻页)+ 4600(gallery 触发点)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "核心缺陷成立,但影响描述有多处夸大,故降级为 P2。\\n\\n【1. 代码真实存在,行号完全对得上,非幻觉】\\nApp.vue:4561 `function startEventStreamPolling()`,4608 `eventStreamTimer = window.setInterval(poll, 5000)`,中间 4602 `await Promise.allSettled(refreshes)`,4600 `if (activeNav.value === \\\"gallery\\\" && generationChanged) refreshes.push(loadGalleryImages())`。loadGalleryImages 在 4511,4524-4534 的整库翻页循环与证据片段一字不差:`loaded.push(...result.items.filter(isAuthenticGenerationImage))` / `if (!result.items.length) break` / `} while (loaded.length < total)`。\\n\\n【2. 确无重入守卫】全仓 grep `loadingGalleryImages` 仅 4 处:App.vue:870 声明、4517 置 true、4540 置 false、13765 模板 `:loading=\\\"loadingDashboard || loadingGalleryImages\\\"`。它只是 UI loading 标志,从未被读作并发守卫,finding 这一点准确。frontend/shared/api/client.ts 的 getJson→request 也无请求去重/缓存/AbortController,纯透传 fetch 或 plugin bridge。\\n\\n【3. 唯一真实守卫在描述场景下不成立】App.vue:4570 `if (!newest || newest === lastEventStreamId) return` 使得只有出现新事件才刷新,因此标题\"每 5 秒触发\"是无条件的说法略过头。但 backend/services/queue_service.py:81/227/424/690 与 generation_service.py:787/860/1067 持续发布 `generation.task_created`/`generation.task_updated`/`generation.node_running`/`generation.node_completed`,批量生图时每 5s 轮询几乎必然命中新事件,finding 描述的\"后台批量生图 + 停留图库页\"场景下该守卫失效。\\n\\n【4. 框架层不但没消解反而放大】features/gallery/GalleryPage.vue:236 是 `v-for=\\\"image in filteredImages\\\"` 全量渲染,无虚拟滚动/无分页;App.vue:997 的 galleryImages computed 依赖 backendGalleryImages,而 4535 `backendGalleryImages.value = uniqueBy(loaded, ...)` 每轮整体替换引用,必然触发全量重算与整表重渲。\\n\\n【5. 需要修正的夸大部分】\\n(a) \"叠加 displayMediaUrl 的 Blob 解析\"不成立:App.vue:11146-11171 的 displayMediaUrl 有 `resolvedMediaUrls` 结果缓存和 `resolvingMediaUrls` Map 在途去重,同一 URL 不会因反复 reload 重复解析,\"几十秒内卡死\"无依据。\\n(b) \"每 5 秒发起 10 个请求\"是串行而非并发——do-while 内是 `await getJson(...)` 顺序翻页;叠加只在单轮耗时 >5s 时才发生,并非必然;plugin-bridge 路径还有 PLUGIN_BRIDGE_GET_TIMEOUT_MS=15000 兜底。\\n(c) \"循环必然多翻一整轮直到返回空页\"不准确:过滤造成 loaded.length < total 时,下一页返回空 items 即被 `if (!result.items.length) break` 截断,只多 1 个请求,不是一整轮,也不存在死循环。backend/routes/image_routes.py:49 的 `payload[\\\"total\\\"] = max(len(visible), int(payload.get(\\\"total\\\") or 0) - payload[\\\"restricted_hidden\\\"])` 还会逐页下调 total,进一步削弱该效应。\\n\\n综上:无 in-flight 守卫 + 图库页每事件 tick 整库分页重扫 + 全量无虚拟化重渲,是真实可复现的性能退化,值得修(加 in-flight 标志 / 用 loadingGalleryImages 早退 / 给 setInterval 换成自调度 setTimeout),但不是崩溃级或正确性级问题,定 P2-中等。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/routes/params.py:20-22",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "我尝试推翻但失败了:核心事实经 Read 逐条核实,全部属实,不是幻觉。但严重度被夸大,且细节有多处数字错误。\n\n【1. 代码真实存在,行号完全对得上】\nbackend/routes/params.py:11-22 原文:\n```\ndef optional_str(value: Any) -> str | None:\n \"\"\"Return a stripped string, or ``None`` for missing/blank values.\"\"\"\n\n if value is None:\n return None\n text = str(value).strip()\n return text or None\n\n\n# Both names are used for this coercion across the backend. Keep the alias so\n# route and service code can use the vocabulary that best fits the field.\noptional_text = optional_str\n```\n行号 20-22 与描述一致。\n\n【2. 别名确实是死的、注释确实是错的 —— 这条无法推翻】\n`grep -rn \"optional_text\" --include=\"*.py\" .` 过滤掉 `_optional_text` 后,全仓库仅剩 9 处命中,且全部是 `params.py:22` 这一行定义本身(其中 8 处是 .tmp/deploy-* 里的历史打包副本,同一文件的拷贝)。**零个消费者**。\n交叉验证 import:全仓库 16 处 `from backend.routes.params import ...`(data_backup/advanced_material/asset/cost/image/material/job/history/log/event/round/media/storage/settings/wardrobe/profile_routes),导入的是 `as_bool / int_or_default / optional_str / optional_int / optional_float`,**没有任何一处导入 optional_text**;也没有 `params.optional_text` 形式的属性访问。params.py 无 `__all__`,所以也不是\"导出给外部插件\"。\n结论:注释「Both names are used for this coercion across the backend」是**事实错误**,别名是死代码。这是可证伪的客观陈述,我无法推翻。\n\n【3. 30 处独立定义属实,模块分布也对】\n`^def _optional_text\\(` 精确匹配到 30 处,分布与描述完全一致:backend/services 19、database/repositories 4、engines 4、adapters 2、core 1。\n证据片段引用的三处逐字节核对无误:\n- backend/services/queue_service.py:713-717 `if value is None: return None / text = str(value).strip() / return text or None`\n- core/astrbot_provider_bridge.py:349-351 `text = str(value or \"\").strip()`\n\n【4. 但发现方的细节数字有错,需要修正】\n(a) 「与另外 26 份逐字节相同」**不实**。真正与 queue_service 版本逐字节相同的(`if value is None` 形式)只有 18 份,不是 26。\n(b) 「3 份变体用 str(value or \"\")」**严重少算**,实际是 9 份:adapters/wechat_adapter.py:77、adapters/onebot11_adapter.py:96、core/astrbot_provider_bridge.py:349、engines/vision_provider.py:450、engines/image_provider.py:404、database/repositories/model_provider_repository.py:383、backend/services/advanced_style_set_service.py:4283、backend/services/asset_import_service.py:373、backend/services/material_service.py:1046。(这点是低估,不构成推翻。)\n(c) 把 30 处全算成「同一个 5 行函数」**不准确**:至少 2 处根本不是重复实现——\n - backend/services/workflow_feedback_interpreter.py:479 签名不同且行为不同:`def _optional_text(value: Any, *, field: str)`,非 str 会 `raise WorkflowFeedbackInterpretationError(\"invalid_llm_result\", ...)`,是带校验的 LLM 输出守卫。\n - backend/services/prompt_governance_service.py:452 是 `text = _text(value); return text or None`,委托给 `_text()`,属于该文件自有的两层抽象。\n 真正可无损合并的同构副本约 28 处而非 30 处。\n\n【5. 严重度被夸大 —— 这是我下调评级的主要依据】\n声称的影响「跨模块搬运代码时会静默改变行为」是**推测性的、当前不可达**:\n- 我抽查了 `str(value or \"\")` 变体的全部调用点,传入的都是字符串型字段,不存在 0 / False / [] 实际流入的路径。例如 material_service.py:348/760/832-834/860-879 传的是 `recognition_model_route`、`content_type`、`provider_id`、`error_type`;model_provider_repository.py:45-46 传的是 `api_key_env_var`、`api_key_secret_ref`;
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:6126",
|
|||
|
|
"reason": "代码确实存在(App.vue:6126 的 `const instructions` 表、character_workflow_service.py:280 的 `fast = {`,行号完全对得上,非幻觉)。但该发现的**核心影响与一半前提都不成立**,判为误报,仅残留一个 P3 级 DRY 小问题。\n\n**1)「前端把 _FIELD_ALIASES 重抄一遍」——事实错误。**\n前端 `quickCharacterIdentityFields()`(App.vue:6100)产出的是**原始字段值**,键名恰好是后端别名表能识别的别名,并作为 `identity_fields` 随 payload/metadata 上传。这正是后端**明文定义的输入契约**,`build_identity_lock()`(quick_character_consistency.py:56-68)把它们列为显式数据源:\n```python\nsources = [\n _record(profile_metadata.get(\"questionnaire\")),\n _record(profile_metadata.get(\"identity_fields\")),\n profile_metadata,\n _record(request_metadata.get(\"identity_fields\")),\n _record(payload.get(\"identity_fields\")),\n _lock_fields(payload.get(\"identity_lock\")),\n]\n```\n别名归一、`_normalize_gender`/`_normalize_age_group`、缺省推断、canonical 字段集、锁文案与 `identity_lock_hash` **全部仍在后端**。前端没有复制任何一张别名表。\n\n**2)「身份锁在前端拼、四联图一致性不可复现」——直接被推翻。**\ngeneration_service.py:691 对**每个** quick-character 节点无条件套用服务端身份锁,代码注释明写此意图:\n```python\n# Direct mode intentionally bypasses LLM rewriting, but it must\n# never bypass the immutable server-owned character identity lock.\nprompt_expansion = _locked_quick_character_prompt_expansion(...)\n```\n其内部 `append_identity_lock(base_prompt, quick_context[\"identity_lock\"])`,而锁文案(`_identity_prompt`,:180-190)是:\n```\nMANDATORY CHARACTER IDENTITY LOCK; depict the same individual in every image in this quick-character batch; do not change gender presentation, apparent age, species, ancestry, skin tone, facial identity, hair, eyes, body build, distinctive features, or core outfit\n```\n前端 quick_avatar 少写的「keep identity consistent」这几个词,被这条强度高出一个数量级的强制锁**完全覆盖**。且 `identity_lock_hash()` 对 canonical JSON 做 sha256,一致性是**确定性可复现**的——恰恰与「不可复现」的断言相反。\n\n**3)「prompt 治理会被前端拼串绕过」——错误。**\n`_govern_prompt()`(:1456)显式读取前端串:`original = _optional_text(payload.get(\"user_prompt\") or payload.get(\"prompt\") or payload.get(\"prompt_patch\")) or ...`,随后审核(moderation)、路由裁决、`route.allowed` 为假时抛 `VALIDATION_FAILED` 阻断,**全部照常作用于前端串**。`prompt_processing_mode:\"direct\"` 只跳过 LLM 翻译/改写,而这是任何调用方都能传的显式开关,与本条「指令表重抄」无因果关系。\n\n**4)「前端触发 vs 后端自身触发产生不同提示词」——缺乏第二条路径。**\n`_step_instruction()` 只被 `_node_prompt()` 调用,而 `_node_prompt()` 仅在 `_execute_sequenced`/`_execute_unsequenced`(:123、:201)中用于**准备 node.prompt**,并不存在一条为 quick 四联图独立出图的后端路径。且在非 direct 模式下,`prompt_expansion_service.merge_prompt_sections` 同时并入 `{\"parts\": [node.prompt]}` 与 `_payload_parts(payload)`(含 prompt_patch),后端指令是**叠加而非被替换**。\n\n**结论**:真实残留问题仅为「4 条短指令字符串在前后端各写一份、措辞已漂移」的可维护性瑕疵,属 P3-轻微。所声称的身份一致性失效、治理被绕过、行为不可复现均不成立。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/services/material_binding_schema_service.py:47(快捷分支)+ :329(实际抛 KeyError 处)+ backend/services/workflow_service.py:729-732(先落库后同步)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "【技术事实:成立,非幻觉,行号完全对得上】\n\n1) `backend/services/material_binding_schema_service.py:47-58` 确实是原文所述的快捷分支,`return normalized` 前从未调用 `_normalize_rule`:\n```python\nif isinstance(payload, dict) and payload.get(\"schema_version\") == MATERIAL_BINDING_SCHEMA_VERSION:\n normalized = deepcopy(payload)\n if workflow_nodes:\n rules = normalized.get(\"active_rules\")\n ...\n normalized[\"by_node_id\"] = _index_by_node_id(rules, workflow_nodes)\n return normalized\n```\n2) `:326-331` 的 `_index_by_node_key` 确实用下标 `for node_key in rule[\"target_node_keys\"]`(非 `.get`),`_index_by_node_id`(:338)第一行就调它。\n3) `backend/services/workflow_service.py:729-732` 顺序确认无误:`after = self.workflow_repository.update_node(...)`(`database/repositories/workflow_repository.py:280-322` 内部 `conn.commit()` 已提交)之后才 `self._sync_material_bindings_from_node(after)`;`_sync_material_bindings_from_node`(:1071-1096)中 `nodes = list_nodes(...)` 必然非空(至少含 material_bind 节点自身),`normalize_material_bindings(raw_bindings, nodes)`(:1081)在 `update_workflow`(:1093)之前。\n4) 我实跑复现,确认 KeyError 真实:\n```\nempty-rule RAISED KeyError 'target_node_keys'\npoison-prompt OK # material.prompt='INJECTED' 原样进入 by_node_id\n```\n5) 上游无守卫:`backend/routes/node_routes.py:117-129` 只有 `require_write(ctx)`,payload 是裸 dict 直接透传,无 pydantic/schema 校验;`backend/routes/common.py:12-16` + `backend/dto/common.py:75-83` 把非 ApiError 异常统一转成 `UNKNOWN_ERROR`「服务器内部错误」,即 500。\n\n【但声称的影响严重失实,须降级】\n\nA) 「永久毒化、无法自愈的死节点」是错的。`workflow_repository.update_node` 是整列覆盖 `metadata_json = ?`(不是深合并),而且崩溃点在第 729 行落库**之后**——这恰恰意味着下一次 PUT 一定能写进去:新 metadata 若不含 `material_bindings`,`_sync_material_bindings_from_node` 在 :1075 `if raw_bindings is None: return` 直接返回;若含合法绑定则正常同步。前端 `App.vue:saveWorkflowMaterialBindings` 的保存动作本身就是 `{...旧metadata, material_bindings: 新payload}`,一键即可自愈。所以不存在「只能读不能改的死节点」。\n\nB) 读路径完全没被污染。脏数据永远进不了 workflow metadata:`_sync` 在 :1081 归一化时就抛了,:1093 的 `update_workflow` 根本没执行;而 `material_bindings_for_node`(:102)读的是 `workflow_metadata.get(\"material_bindings\")`,仍是上一版合法快照。依赖上下文/素材抽屉等读接口不受影响。\n\nC) 工作流级 PUT 无任何持久化风险:`workflow_service.py:268-271` 中 `_normalize_workflow_metadata` 在 `repository.update_workflow` 之前执行,:1053 抛异常时什么都没写。\n\nD) 「未经校验的任意文本注入通道」不构成安全边界穿越。触发者是持有 `can_write`(`core/security.py:27`)的用户,而 `prompt_expansion_service` 本来就把 `node.prompt` 原样并入最终提示词(`\"parts\": [node.prompt]`)——同一个用户直接改节点 prompt 即可,无需绕 schema。这不是提权,只是同权限用户的另一条等价写入路径。\n\nE) 触发不可能来自正常客户端。`App.vue:9117-9144` 的 `buildWorkflowMaterialBindingsPayload` 返回的对象里**没有** `schema_version`(它作为兄弟键 `material_binding_schema_version` 放在 metadata 上),所以真实前端永远走慢速路径 `_normalize_rule`。这条快捷分支的设计意图是对「已归一化对象的幂等回写」(对应 :1036 注释所说的 stale whole-metadata PUT),此时 rule 必然带 `target_node_keys`,不会崩。必须手工构造畸形 payload 才能触发。\n\n【结论】这是一个真实的健壮性缺陷:受信任写权限用户用畸形 payload 可造成 500,并留下「节点已写库<EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "backend/services/workflow_round_service.py:1544 (调用点 163);core/astrbot_provider_bridge.py:344 (调用点 278);backend/services/advanced_style_set_service.py:4288 (调用点 1537)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"refuted": true,
|
|||
|
|
"reason": "代码片段本身抄写无误、行号对得上,但\"18 次\"和三条声称的影响全部站不住脚。\n\n(1)计数错误:真实源码树 `grep -rn \"def _string_list\" --include=*.py backend core engines` 只有 17 处;多出来的是 `.tmp/deploy-1.0.35/`、`.tmp/deploy-clean-1.0.36/` 等构建产物快照里的重复文件,不是源码。\n\n(2)\"把逗号分隔的 tags 传给 workflow_round_service 版本会得到空列表\"——该调用路径不存在。workflow_round_service.py 全文只有一处调用:\n source_candidate_ids = _string_list(payload.get(\"source_candidate_ids\"))[:3] # 第 163 行\n处理的是内部 payload 里的候选 ID JSON 数组,用于 optimization_source_candidate_ids / source_candidate_id,tags 永不经过此处。同理 prompt_expansion_service.py:257 全文只用于第 109 行的 selected_draft_image_ids;generation_service.py:2417 只用于 1462 行 locked_terms 和 2250 行 reference_image_uris——均无逗号字符串来源。\n\n(3)\"prompt 片段顺序在 advanced_style_set_service 与 astrbot_provider_bridge 之间不可预测\"——两者不处理同一个值。bridge 唯一调用是 `tags=_string_list(payload.get(\"tags\"))`(第 278 行,视觉 provider 的 JSON);advanced_style_set 的调用全部是 representative_source_ids / comparison_source_ids / 反馈计划的 increase|decrease|remove|preserve / 自己那份 tags。prompt_parts 这个字段根本不经过任何一个 _string_list。\n\n(4)不存在\"跨 service 流转\":`grep \"import _string_list|from ... import ..._string_list\"` 零结果,17 个实现全部是下划线私有、无一被跨模块导入。每个调用点都是在本模块自己的信任边界上对刚解析的 dict 做一次本地归一化,而非接力传递已归一化列表,因此不同实现之间不会互相污染。\n\n(5)唯一真实的不一致:VisionRecognitionResult.tags 有三个构造点——engines/vision_provider.py:357(sorted)、core/astrbot_provider_bridge.py:278(sorted)、advanced_style_set_service.py:1537(dict.fromkeys 保序)。三者中两者一致,第三者顺序不同。但下游消费(material_service.py:472 generated_tags、advanced_style_set_service.py:1765 \"tags\": recognition.tags 等)只把 tags 当元数据集合存储/展示,无任何依赖顺序的逻辑,影响是装饰性的。\n\n综上:DRY 违规客观存在、值得统一收敛,但数据损坏 / 静默空列表 / 顺序不可预测三条影响均无可达路径,P1 定级被严重夸大,实际为 P3 可维护性异味。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "backend/routes/branch_routes.py:62-68",
|
|||
|
|
"reason": "代码属实,但影响被严重夸大,严重度应从高降至 P3。\n\n【成立部分】\n1) 行号精确对得上,非幻觉。branch_routes.py:62-68 确实缺 event_bus,且该文件 import 段(1-14 行)压根没引入 get_default_event_bus,是真实遗漏:\n def _create_workflow_service() -> WorkflowService:\n database = Database(load_database_settings())\n return WorkflowService(\n workflow_repository=WorkflowRepository(database),\n profile_repository=ProfileRepository(database),\n audit_service=AuditService(database),\n )\n对照另外三份均带 event_bus=get_default_event_bus():workflow_routes.py:143-150、node_routes.py:305-312、profile_routes.py:174-181。四选一不对称成立。\n\n2) 路径可达,无上游守卫消解。adapters/astrbot_adapter.py:255-260 的 _load_handler 用 importlib 按点号路径取函数,_build_args 只填有值的参数,service: WorkflowService | None = None 保持 None,因此 `service or _create_workflow_service()` 生产环境必然落到无 event_bus 那份。_publish 是空守卫(workflow_service.py:1012-1014 `if self.event_bus:`),workflow.branch_created 确实被静默吞掉。\n\n【推翻部分 —— 描述存在两处硬伤】\n3) switch_branch 根本不调用 _publish。通读 workflow_service.py:541-588,函数以 self.audit_service.record(...) 结尾后直接 `return after`,全程无 publish。原文引用的 \"workflow_service.py:640\" 实际落在 create_node 内部,发的是 \"workflow.node_created\",属张冠李戴。\n\n4) workflow.branch_switched 这个事件全仓库不存在。全库 grep(py/vue/ts/js,排除 .tmp 与构建产物)对 branch_switched 零命中;branch_created 仅一处命中,就是 workflow_service.py:515 的发布点本身——没有任何订阅方。\n(唯一真实但原文未指出的连带损失:switch_branch 第 570 行调 self.update_workflow,后者第 299 行发 \"workflow.updated\",这条也一起丢。)\n\n5) \"前端 SSE 永远看不到分支变更\"不成立。event_routes.py 不是推送通道,EventStreamService.recent_events(event_stream_service.py:22-32)只是读 event_bus.history() 的轮询快照。App.vue:4561-4608 每 5 秒轮询,只做粗粒度前缀匹配 eventType.startsWith(\"workflow.\"),且仅在 activeNav 为 overview/performance/logs 时触发 dashboard/logs 刷新。分支 UI 完全不依赖事件流:createBranchFromSelectedWorkflowNode(App.vue:8407+)直接用 POST 返回值写入 workflowBranchesByWorkflowId 再调 activateWorkflowBranch。审计链路不受影响——create_branch 与 switch_branch 的 audit_service.record 均无条件执行。\n\n6) tests/ 下对 event_bus / branch_created 零命中,确无回归保护。\n\n结论:一行一致性遗漏客观存在、值得修,但丢失的事件零消费者,用户可见影响仅为 3 个导航页上一次可能的自动刷新延迟,不构成功能故障,定级 P3。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/tests/production-data-boundary.test.mjs:33-35(整体问题覆盖 frontend/tests/ 下 8 个测试文件中的 7 个)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "我以\"默认误报\"立场逐条核验,但**每一项事实都被真实代码证实,无法推翻**。\n\n**1. 代码真实存在,行号对得上,非幻觉。**\n`frontend/tests/production-data-boundary.test.mjs` 第 33、35 行原文:\n```js\nassert.match(app, /const imageGenerationImages[\\s\\S]*localGenerationImages\\.value\\.filter\\(isAuthenticGenerationImage\\)/)\nassert.match(app, /const profilePageItems = computed<CharacterProfile\\[\\]>\\(\\(\\) => backendProfiles\\.value\\)/)\n```\n文件头部 `import { readFile } from \"node:fs/promises\"`,第 5 行 `const appPath = new URL(\"../standalone-web/src/App.vue\", import.meta.url)`——确实把 `.vue` 当纯文本 `readFile(appPath, \"utf8\")` 读入做正则断言,全程没有任何组件实例化。\n`workflow-cross-review-wiring.test.mjs` 第 19、27 行也逐字一致:`assert.equal((app.match(/<WorkflowGenerationReviewDialog\\b/g) ?? []).length, 1)` 与 `assert.match(dialog, /model_routes: \\[\\.\\.\\.reviewModelRoutes\\.value\\]/)`。\n\n**2. 7/8 比例属实。** `grep -l readFile *.mjs` 命中 7 个:advanced-material-upload、production-data-boundary、task-timeout-consistency、ui-redesign-smoke、ui-theme-contract、workflow-cross-review-wiring、workflow-profile-card-wiring。唯一真正 import 被测模块的是 `workflow-model-route.test.mjs`,它导入 `workflow-model-route.ts`(88 行) 和 `workflow-lock-state.ts`(27 行)——两者均无 `from 'vue'` 导入,确系纯函数模块,共 115 行。\n\n**3. 不存在任何上游守卫/框架机制消解该问题。** 这一条是我最想找到的反驳点,但没有:\n- `frontend/package.json` 第 13 行 `\"test:contracts\": \"node --test tests/*.test.mjs\"`,`\"devDependencies\": {}` 完全为空;\n- 在 `package-lock.json` 中 grep `vitest|@vue/test-utils|jsdom|happy-dom|@testing-library`,命中数为 **0**;\n- 全仓库 `find` 搜索 `vitest.config*` / `playwright.config*` / `cypress.config*` / `*.spec.ts`,**零结果**;`.github` 下也无 CI workflow。\n即不存在\"其实别处有组件测试兜底\"的可能。\n\n**4. 规模数据属实。** `find -name \"*.vue\"` = **36 个文件、35638 行**;`.ts` = **3616 行**。实跑 `node --test tests/*.test.mjs` 输出 `tests 26 / pass 26 / duration_ms 323.47`——323ms 跑完 26 个用例,正是\"只做字符串匹配\"的特征。\n\n**5. 假阴性/假阳性推理成立。** 第 35 行正则 `computed<CharacterProfile\\[\\]>\\(\\(\\) => backendProfiles\\.value\\)` 对空格和换行完全刚性,Prettier 换行或改写成等价的 `computed(() => backendProfiles.value)`(去掉泛型标注)都会让断言崩溃而功能无损;反之组件运行时抛异常、props 契约破裂、事件不触发,纯文本正则一律照过。\n\n**关于严重度**:我下调为 P2-中等而非更高。理由是这是**测试基建缺口而非线上运行缺陷**——它不会导致任何生产故障,描述的是风险敞口而不是已发生的损害;\"35638 行实际覆盖率为 0\"表述准确但属于风险陈述。不过考虑到涉及整个 UI 层且无任何 CI 兜底,不应低于 P2。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:5786-6076(状态推断 5945-5979,失败补偿 6046-6069)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "核心成立,但影响描述有一半被夸大,故降级为 P2。\\n\\n【一、代码属实,行号对得上】generateQuickCharacterImages 位于 App.vue:5786-6076,共 291 行。证据逐字吻合:5950-5955 前端自行推断 `const nextWorkflowStatus = generationCompleted ? \\\"completed\\\" : pendingTargetKeys.length ? \\\"running\\\" : \\\"failed\\\"`;5978-5979 `if (normalizeStatus(workflow.status) !== nextWorkflowStatus) workflowUpdate.status = nextWorkflowStatus; workflow = await putJson<ProfileWorkflow>(endpoints.workflow(workflow.workflow_id), workflowUpdate)`;6051 catch 块 `putJson(..., { status: \\\"failed\\\", metadata: {...} })`。非幻觉。\\n\\n【二、无守卫使其不可达,后端确实没有兜底】(1) endpoints.workflowStart 对应 backend/routes/workflow_routes.py:79-84 `def start_workflow(...): return _set_workflow_status(ctx, workflow_id, \\\"running\\\", service)`,仅改状态,不托管推进。(2) 后端唯一自动推终态的 _advance_character_gradual_progression(workflow_service.py:760-802,第 801 行 `if progression.all_completed and workflow.status in {\\\"draft\\\",\\\"active\\\",\\\"running\\\"}: update_payload[\\\"status\\\"] = \\\"completed\\\"`)在 766 行被 `if workflow is None or workflow.workflow_type != \\\"character_gradual\\\": return` 排除,而 App.vue:926 `const QUICK_CHARACTER_WORKFLOW_TYPE = \\\"character_fast\\\"`。(3) 全仓唯一的僵尸清理 queue_service.recover_stale_tasks(439 行)是 task 级、`source: str | None = \\\"workflow_round\\\"`,唯一调用方 workflow_round_worker_runtime.py:55,不覆盖 character_fast 工作流状态。(4) App.vue:5816 传 `forceNew: true`,ensureCharacterWorkflowForProfileOnce(7339-7364)在 forceNew 下必定 POST 新建 status:\\\"draft\\\" 工作流,因此每次被中断的运行都新增一条僵尸记录,累积是真实的。(5) 后端 update_workflow 的完成度校验只针对 gradual:workflow_service.py:236 `if before.workflow_type == \\\"character_gradual\\\" and payload.get(\\\"status\\\") == \\\"completed\\\"`,character_fast 可被前端任意标 completed 而不校验节点。\\n\\n【三、被推翻的部分:metadata 契约并非“完全由前端定义、后端被动接收”】workflow_service.py:64-72 定义 QUICK_CHARACTER_IMMUTABLE_METADATA_KEYS = {quick_character_identity_managed, quick_character_batch_id, identity_lock_hash, identity_lock, quick_character_route_snapshot, quick_character_target, quick_character_targets};1032-1042 对 character_fast 强制回填服务端值,代码注释即 “Browser clients may PUT a stale whole metadata object after the four requests finish. Preserve server-owned batch identity and route fields once initialized so clients cannot clear or mutate the lock”。generation_service.py:1236-1257 由后端自行生成 batch_id/identity_lock 并写 quick_character_identity_managed=\\\"generation_service_v1\\\"。anchor 语义也由后端强校验:generation_service.py:1337-1347 `raise ApiError(..., \\\"quick-character downstream generation requires the avatar image_id as an identity anchor\\\")`,1299-1329 还校验引用图归属工作流与 batch/lock 一致。故 anchor_image_id、quick_character_batch_id 并不是前端定义的隐式契约。真正纯前端定义的只有 quick_character_pending_target_keys、generated_target_keys、quick_character_schema、quick_character_operation_id 等展示字段;跨端被消费的仅 quick_character_failure(generation_job_retry_service.py:107 与 database/repositories/job_repository.py:540 的 metadata.pop)。\\n\\n【四、严重度】无数据丢失、无损坏、无鉴权/注入问题;卡在 running 不阻断后续生成(generation_service.py:595 反而要求 active/running),用户还可通过 App.vue:8312 的 start/cancel/archive 手动收尾。实际代价是工作流状态失真 + draft/running 记录堆积 + 前后端职责倒置的架构债,定 P2-中等而非 P0/P1。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "core/queue_manager.py:1-112(整文件)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "我以\"默认误报\"立场逐条核验,但四项核查均未能推翻该发现的事实主张,因此判定成立——只是严重度被显著夸大,需下调至 P3。\n\n【核查 1:代码是否真实存在、行号是否对得上】\n真实存在,非幻觉。`wc -l core/queue_manager.py` = 112 行,与标题一致。Read 全文确认结构与证据片段逐字吻合:\n- 第 1 行 docstring `\"\"\"In-memory generation queue primitives.\"\"\"`\n- 第 11 行 `from core.fsm import task_transition`\n- 第 18-19 行 `@dataclass` / `class QueueItem:`\n- 第 42-46 行 `class QueueManager:` 的三个字段 `_pending: deque[str]` / `_items: dict[str, QueueItem]` / `_idempotency_index: dict[str, str]`,与证据片段完全一致\n- 方法齐全:`enqueue`(48) / `dequeue`(62) / `complete`(72) / `fail`(79) / `retry`(87) / `get`(95) / `summary`(101) / `_set_status`(107)\n- 第 107-112 行确为 FSM 状态校验:\n ```\n def _set_status(self, item: QueueItem, status: str) -> None:\n transition = task_transition(item.status, status)\n if not transition.allowed:\n raise ValueError(transition.reason)\n ```\n\n【核查 2:零引用是否属实 / 是否有动态加载让它实际可达】\n属实。`grep -rn \"queue_manager|QueueManager|QueueItem\"` 全仓(不限 .py,排除 .tmp/.git/node_modules)的命中只有四类:\n(a) `core/queue_manager.py` 自身的 11 处;\n(b) `core/__pycache__/queue_manager.cpython-311.pyc`(陈旧字节码,非引用);\n(c) 前端 `frontend/standalone-web/src/App.vue` 的 `filteredQueueItems` / `overviewQueueItems`——纯属命名巧合的 Vue computed,与本模块无关;\n(d) 设计文档 `imagestudio_full_plugin_dev_doc_v3_ui_adjusted.md:164` 的目录树、`imagestudio_codex_hermes_dev_flow_v3.md:235` 的 \"S4-3 queue_manager 基础队列\",以及 `release/*.zip` 历史产物。\n我特意排查了动态加载这条\"实际可达\"的翻案路径:全仓 `importlib` 只有两处——`adapters/astrbot_adapter.py:257` 的 `_load_handler` 按 `handler_path.rsplit(\".\", 1)` 反射加载路由 handler,以及 `main.py:93` 的 `importlib.reload(_database_db_helper)`。前者依赖路由表字符串,而上述 grep 已覆盖 yaml/toml/json/txt 等所有清单文件、无一提及 `queue_manager`,故反射也够不到它。`core/__init__.py` 仅有一行 docstring、不做任何 re-export。结论:该模块从未被导入,代码从未被执行。\n\n【核查 3:是否有框架机制消解 / 真实队列是否另有其人】\n真实队列确实是持久化实现,与发现描述一致。`backend/services/queue_service.py`(29952 字节)+ `database/repositories/job_repository.py`(22687 字节)承担全部实际职责,且语义确实与死代码平行重复:\n- 幂等:真实实现走 `self.repository.get_by_idempotency_key(idempotency_key)`(queue_service.py:53、128-132),是 DB 查询;死代码走进程内 `self._idempotency_index` 字典(queue_manager.py:49-50、57-58)。\n- 状态机:真实实现是 `_validate_task_status_transition`(queue_service.py:673-686),先校验 `target not in TASK_STATUSES` 再调 `task_transition`,失败抛结构化 `ApiError(ErrorCode.VALIDATION_FAILED, ...)`;死代码则无白名单校验、失败抛裸 `ValueError`。两套语义确有分叉,\"影子实现\"的说法站得住。\n不存在任何守卫/框架机制\"消解\"它——但反过来说,正因为它完全不可达,也不存在可被触发的缺陷。\n\n【核查 4:严重度是否被夸大——此处发现方明显夸大,需下调】\n必须下调至 P3-轻微,理由有二:\n(1) 这是纯死代码,零运行时影响:没有错误路径、没有数据正确性问题、没有安全面、不影响打包(未进任何 manifest)。所谓\"新人极易误用\"是推测性的可维护性论点,不是可触发的失效场景。按\"缺陷\"标准衡量,它没有 failure scenario。\n(2) 发现方的一条子论断确属夸大,应予纠正:\"它单独依赖 core.fsm.task_tran
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "仓库根目录 D:\\CODEX\\projects\\AstrBot ImageStudio(缺失 conftest.py / pytest.ini / pyproject.toml),影响 tests/ 下全部 45 个测试文件",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "这条发现属实,不是幻觉,我已直接复现,但其\"声称的影响\"有明显夸大,故严重度下调为 P3。\n\n一、事实核验(全部成立)\n\n1. 配置确实一个都不存在。`git ls-files | grep -iE \"conftest|pytest.ini|pyproject|setup.cfg|tox.ini|setup.py\"` 返回 EXIT=1(零命中);Glob `**/conftest.py` 返回 \"No files found\"。同时 `tests/__init__.py` 也不存在(ls: cannot access 'tests/__init__.py'),因此 pytest 的 prepend 导入模式取的 basedir 是 `tests/` 目录本身,永远不会把仓库根插进 sys.path。\n\n2. 现场复现,且比报告更严重。报告只举了单文件,我跑了全量:\n 裸 pytest:`pytest tests/ -q --collect-only` →\n `!!!!!!!! Interrupted: 45 errors during collection !!!!!!!!`\n `no tests collected, 45 errors in 1.50s`\n python -m pytest:`python -m pytest tests/ -q --collect-only` →\n `263 tests collected in 0.48s`\n 单文件也与证据片段逐字吻合:`tests\\test_profile_delete_route.py:3: in <module> / from adapters.astrbot_adapter import resolve_route / E ModuleNotFoundError: No module named 'adapters'`,而 `python -m pytest` 同一文件 `1 passed in 0.16s`。\n\n3. 不存在任何上游守卫使其不可达。我逐文件循环检查 `^from (backend|adapters|core|engines|database|main)`,输出为空,即 45 个文件无一例外都用仓库根绝对导入。全仓 `grep -rn \"sys.path\" tests/` 只命中 `tests/test_plugin_bootstrap.py:22`、`:25`:\n `f\"sys.path.insert(0, {str(tmp_path)!r})\",`\n `f\"sys.path.insert(0, {str(plugin_root)!r})\",`\n 这两行是 f-string 拼接出来喂给**子进程脚本**的字符串,不是该测试模块自身的 import 引导,对收集阶段毫无作用。环境 PYTHONPATH 为空 `[]`,site-packages 下的 .pth 也只有 distutils-precedence / pywin32 / _virtualenv / hermes_agent,没有任何把本仓库映射进去的 editable 安装。\n\n4. 没有框架层机制消解它。pytest 8.4.1/9.1.1 均按 rootdir+basedir 规则工作,缺 conftest.py 就是拿不到仓库根。\n\n二、为何下调到 P3(\"声称的影响\"站不住的部分)\n\n1. \"任何 CI 模板、tox、pre-commit 都会全军覆没\"——本仓库根本不存在这些东西。实测 `ls -d .github .gitea .forgejo .gitlab-ci.yml .pre-commit-config.yaml .vscode .idea` 七项全部 \"No such file or directory\"。没有 CI、没有 tox、没有 pre-commit、也没有提交任何 IDE 配置,所以这部分影响是纯假设,不是现实损失。\n\n2. 项目已有文档化的可用命令。`INSTALL.md:221` 明确写着 `python -m pytest -q --basetemp=.pytest_tmp_release`,即官方发布流程本来就走 `-m` 形式,现有工作流并未被阻断。\n\n3. 零运行时影响。这只影响测试执行方式,不触及 backend/adapters/core 任何生产代码路径,无数据、无安全、无用户可见后果。\n\n4. 修复成本近乎为零:仓库根加一个空 conftest.py 即可(或加 pytest.ini 设 rootdir/pythonpath)。\n\n结论:问题客观存在且 100% 可复现(裸 pytest 收集阶段 45/45 全废),值得修,但它是开发体验/工具链层面的小瑕疵,不是 P0/P1 级缺陷。附带一提,报告顺带提出的\"缺 --strict-markers / 无超时保护 / 无覆盖率门槛\"属于工程增强诉求,不构成本条缺陷本身的证据。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": ".env.imagestudio_codex_handoff.yaml:25(api_key_plaintext 实际在第 25 行,原报告写 :26 差一行;:45/:57-61/:73/:76 均准确)",
|
|||
|
|
"corrected_severity": "P0-致命",
|
|||
|
|
"reason": "我以\"默认误报\"立场逐条验证,但每一条关键主张都被真实文件证实,无法推翻。\n\n**1. 代码/文件真实存在,不是幻觉。** `.env.imagestudio_codex_handoff.yaml` 确实存在于仓库根目录,3506 字节,Read 后内容与描述完全一致(仅行号差一行):\n- L25 `api_key_plaintext: \"sk-3af1f9dd...f3bcc6\"`(64 hex,weilai.chat)\n- L45 `api_key_plaintext: \"sk-B6KqCtlc...\"`(botcf.com)\n- L57-61 `host: \"<PG_HOST:内网IP:已脱敏>\"` / `port: 25008` / `database: \"imagestudio\"` / `user: \"<PG_USER:已脱敏>\"` / `password: \"<PG_PASSWORD:已脱敏>\"`(明文)\n- L73 `admin_token_plaintext: \"eyJhbGciOiJIUzI1NiI...\"`(HS256 JWT)\n- L75-76 `admin_username: \"<ADMIN_USERNAME:已脱敏>\"` / `admin_password_plaintext:` 明文\n- L80 `redact_secrets: true`——讽刺点属实。\n\n**2. 不存在任何上游守卫,问题完全可达。** \n- `.gitignore:39` 的 `.env*` 确实生效:`git check-ignore -v` 返回 `.gitignore:39:.env*`,`git ls-files | grep env` 只有 `frontend/env.d.ts`,`git log --all --diff-filter=A` 为空。**但这恰恰是误导性的\"安全感\"**——泄漏通道根本不走 git。\n- 打包用整目录复制而非 `git archive`,`.gitignore` 对它零约束。8 个暂存目录各含一份 3506 字节副本,全部实测确认:`deploy-1.0.35`、`deploy-clean-1.0.35/36/37`、`deploy-stage-1.0.38~41`。\n- `rules.redact_secrets: true` **不是任何代码强制的机制**,只是写给 AI 的自然语言备注:`grep -rn \"redact_secrets\" --include=*.py` 在全仓库返回空。没有 ORM、装饰器、try-catch 之类的框架层机制能消解文件被物理复制这件事。\n\n**3. 归档泄漏实测确认,且原报告是低估的。** 我对 `.tmp/` 下所有 `.tar.gz`/`.tar`/`.zip` 逐个跑 `tar -tf`/`unzip -l`,得到 **21 个**泄漏归档(原报告说 19,漏掉了 `astrbot_plugin_imagestudio-1.0.62.zip` 和 `1.0.63.zip`)。从 1.0.67 归档中原地解出的内容确认是**未脱敏的真实明文**,与工作区副本逐字节一致。对照组也成立:1.0.32、1.0.69、`astrbot_plugin_imagestudio-a03a25d.tar` 是干净的,说明这不是我的检测方法误报。其中 `nas-deploy-worktree-20260724.tar` 达 572 MB。\n\n**4. 确实已上传 NAS,且被二次扩散。** `.tmp/upload_via_ssh.py` 把泄漏包 scp 到 `agent@<PG_HOST:内网IP:已脱敏>:/tmp/`;`deploy_remote.sh` 随后 `tar -xzf \"$PKG\"` 解到 `/volume2/Dockers/qqbot/astrbot-ops-pro/data/plugins/astrbot_plugin_imagestudio`,凭据文件因此落盘到 NAS 插件目录;更糟的是脚本里 `cp -a \"$PLUGIN_DIR\" \"$BACKUP\"` 会把它连同凭据一起复制进 `plugin_backups/`,形成多份带时间戳的历史副本,清理时极易漏掉。\n\n**5. 严重度没有被夸大,P0 成立。** 唯一能为原报告打的折扣是:文件自身 L36/L52 的 notes 记录 2026-07-05 探测两个站点均返回 `401 Unauthorized`,所以\"直接盗刷模型额度\"这一条可能已因 key 失效而打折;JWT 的 `exp=1783872299`(约 2026-07-12)按今天 2026-07-27 看也已过期。但这不足以降级:PostgreSQL 明文密码与 `admin_password_plaintext` 没有任何失效证据,且如原报告正确指出的——明文管理员密码永不过期,拿到它即可随时重签 JWT 接管 AstrBot 实例(该实例托管本插件全部 API)。凭据泄漏的定级应以最强的那条为准,而非最弱的那条。\n\n**建议处置(超出本次复核范围,仅提示)**:这些凭据必须视为已泄漏并全部轮换(两个站点 API key、PG 密码、AstrBot 管理员密码及 JWT 签名密钥),而不只是删文件;同时清理 NAS 上 `data/plugins/` 与 `plugin_backups/` 中的副本,并把打包改为 `git archive` 或加显式 `--exclude='.env*'`。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "tests/(测试目录整体),非 backend/route_registry.py:1",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"refuted": true,
|
|||
|
|
"reason": "【结论:误报 / 不成立】原始统计数字能复现,但定位、证据与影响描述三处均不成立。\n\n一、定位是伪造的。`backend/route_registry.py:1` 是文件 docstring `\"\"\"Route registry manifest for AstrBot Web API mounting.\"\"\"`。整个文件(377 行)是一张纯声明式的 RouteSpec 清单加 `route_manifest()` / `assert_unique_routes()` / `assert_known_permissions()` 三个工具函数,代码本身没有任何缺陷。所谓\"缺陷\"是 `tests/` 目录的属性,不是这个文件的属性。把测试债挂到一个正确无误的清单文件第 1 行,不构成一条代码缺陷。\n\n二、最有杀伤力的那条证据是事实错误。发现称 storage_routes 是\"签名 URL / 文件下发,直接决定前端不能看到本地绝对路径这条安全约束\"。实读 `backend/routes/storage_routes.py`,全文只有两个 handler:`get_storage_status` 和 `migrate_storage`;registry 里也只挂了两条(route_registry.py:167-168):\n```\nRouteSpec(\"GET\", \"storage/status\", \"...storage_routes.get_storage_status\", \"read\"),\nRouteSpec(\"POST\", \"storage/migrate\", \"...storage_routes.migrate_storage\", \"config\"),\n```\n它既不下发文件也不签 URL。签名与取流在 media_routes(route_registry.py:169-173,`media/sign`、`media/fetch`),是另一个模块。旗舰级安全论据属于臆造。\n\n三、\"路由层的参数校验与权限装饰完全裸奔\"是假的。所有被点名的\"高危未测模块\"每个 handler 首行都显式调了 `core.security` 的守卫:\n- round_routes.py:33/55/71/88/108/128 → `require_authenticated(ctx)` / `require_write(ctx)`\n- data_backup_routes.py:11/15/23/27 → 四个 handler 全部 `require_config_manager(ctx)`\n- settings_routes.py:23/36 `require_authenticated`,50/64/78/91/104/118/133/147/161 共 9 处 `require_config_manager`\n- integration_routes.py:32/47 `require_authenticated`,63/77 `require_config_manager`\n- media_routes.py:26/68/86/107 `require_authenticated`\ncore/security.py:17-24 的 `require_permission` 会先 `require_authenticated` 再校验 `ctx.has_permission(permission)`,不通过直接抛 `ApiError(FORBIDDEN)`。守卫齐整存在,谈不上\"裸奔\"。storage_routes 自身同样有 `require_authenticated` / `require_config_manager`。\n\n四、度量方法系统性低估。用 endpoint 字面量全文 grep 会漏掉直接 import handler 的测试。我实测:endpoint 字面量命中 33,而 handler 函数名命中 50,模块名命中 8(非 7)。实证例子:tests/test_character_workflow_progression.py:6 `from backend.routes.node_routes import generate_node, retry_node`;tests/test_generation_node_completion_guard.py:6 导入 `mark_node_completed`;tests/test_profile_delete_route.py:12 走 adapter 的 `resolve_route` 断言 `handler_path == \"backend.routes.profile_routes.delete_profile_via_post\"`;tests/test_profile_card_finalize.py:290 与 test_workflow_generation_cross_review.py:9 直接对 ROUTES 做清单断言。19.9% 是下界而非真实覆盖率,要给结论必须跑 coverage.py。\n\n五、严重度严重夸大。这是一条测试债观察,不是缺陷:没有指出任何真实 bug、不可达守卫或权限漏洞,当前行为没有任何一处是错的。它描述的是\"未来重构可能没人拦\"的假设性风险。真要保留,合理形态是一条 P3 改进建议(例如补一个遍历 ROUTES、断言每个 handler 源码内出现 require_* 的元测试),而非现在这条带伪造定位和伪造安全论据的发现。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"corrected_location": ".gitignore(缺 release/ 规则)+ release/ 下 20 个 zip;注意 web/ 与 pages/imagestudio/ 不可加入 ignore,见 backend/services/web_hosting_service.py:15 与 main.py:230",
|
|||
|
|
"reason": "核心事实全部实测复现,成立,但影响被夸大两处,故降级为 P2。\n\n【一、.gitignore 确认非幻觉】文件共四个分区,Python / Node / runtime data / editor noise,确实只有 `.tmp/`、`.tmp_pytest*/`、`frontend/**/dist/`、`data/db/*` 等,全文没有任何 `release/` 规则。\n\n【二、数字逐条复现】git ls-files 455;git ls-files release=20、web=19、pages=19 → 58/455=12.7% ✓;du -sh .git=51M,git count-objects -vH 显示 size-pack 44.03 MiB、in-pack 12788 ✓;git rev-list --all --count=230,git log --oneline --all -- web pages | wc -l=86(37.4%)✓;git log --all --name-only -- web pages | sort -u | grep -c . = 439,而当前 git ls-files web pages = 38,确有约 401 个死 blob ✓;ls -l release 合计 37,948 KB,最大 1.0.0.zip = 4,229,098 B ✓。\n\n【三、运行期数据进发布包成立】解 release/astrbot_plugin_imagestudio-1.0.5.zip,按体积排序前列为 data/temp/provider_probe_B_image2_generation.json 1.76MB、同名 .png 1.32MB、s81_provider_runtime_b.png 0.76MB、data/db/imagestudio.db 0.29MB。该 JSON 解析后确是真实上游响应:data[0].url = \"https://image.cjyyswq.com/uploads/2026/07/05/28230b126fcc48778575e3e07e926940.png\",data[0].b64_json 长 1844600,usage.output_tokens=1650。这与 DEVELOPMENT.md:36 原文「不要向插件源码目录写数据库、图片、日志或缓存。测试可使用临时目录,且不能把真实运行数据加入发布包。」直接冲突。20 个 zip 均含 9~22 个 data/ 运行期文件。\n\n【四、两处夸大,必须修正】\n(1)「imagestudio.db 是真实 SQLite 库」不成立:把它从 zip 解出后用 sqlite3 逐表 count,只有 schema_migrations 有 13 行,其余业务表全空,是空壳 schema,不含任何用户数据。\n(2)「应把 web/assets/、pages/imagestudio/assets/ 加进 .gitignore」这条建议是错的,照做会破坏插件。backend/services/web_hosting_service.py:15 有 `STANDALONE_WEB_DIR = REPO_ROOT / \"web\"`,main.py:230 有 `\"web\": web_manifest`;AstrBot 插件为 clone 仓库安装、无构建步骤,web/ 与 pages/imagestudio/ 必须入库才能被运行时托管,它们不是「没人会用的产物」。content-hash 造成历史死 blob 堆积这一副作用真实存在,但正解是 gc/filter-repo 或固定输出文件名,而非 ignore。\n(3) 我对全部 20 个 zip 正则扫 `sk-[A-Za-z0-9]{16,}` 与 api_key 模式,0 命中,不存在凭据泄漏,「永久留在历史无法撤销」的安全后果远轻于描述。\n\n综上:属纯仓库卫生/可维护性问题(clone 成本、git log -p / bisect 噪声、违反自家发布规范),无运行时故障、无安全泄漏、DB 为空壳,且原建议有一半会破坏运行时,故 P2 而非更高。实际可执行动作仅限:为 release/ 加 ignore 并从索引移除、清理 zip 内 data/temp 与 data/db。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:10837-10859(判定函数整体,10843 是其中一行);对照 backend/services/workflow_progression_service.py:175-199",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "我逐条读了代码,四项事实全部核实通过,不是幻觉,也没有上游守卫或框架机制消解它。但影响被夸大,应降级为 P2。\n\n一、代码真实存在,行号基本对得上\n1) backend/services/workflow_progression_service.py 中 `is_character_gradual_node_completed` 定义在第 175 行,函数体正是 180-199 行:\n```python\nif node.status in PROGRESSION_COMPLETED_NODE_STATUSES: # line 180,第19行 frozenset({\"selected\",\"skipped\"})\n return True\nmetadata = node.metadata if isinstance(node.metadata, dict) else {}\nif node.status == \"completed\" and (\n node.node_type in {\"planning\", \"reference_binding\"}\n or canonical_gradual_node_key(node.node_key) == \"profile_card\"\n):\n return True\nif allow_legacy_planning_metadata:\n if node.node_key == \"prompt_plan\":\n return \"prompt_plan\" in metadata and metadata.get(\"prompt_plan\") is not None\n if node.node_key == \"material_bind\":\n return \"material_bindings\" in metadata and metadata.get(\"material_bindings\") is not None\nreturn False\n```\n且 235-240 行 `_allows_legacy_planning_metadata_completion` 确实以 `execution.get(\"progression_version\") != CHARACTER_GRADUAL_PROGRESSION_VERSION` 作为 legacy 开关。\n\n2) App.vue:10837-10859 是逐层对应的第二实现,10843 起:\n```ts\nconst status = normalizeStatus(backendNode.status)\nif (status === \"selected\" || status === \"skipped\") return true\nif (status === \"completed\" && (backendNode.node_type === \"planning\" || backendNode.node_type === \"reference_binding\" || canonicalCharacterWorkflowNodeKey(backendNode.node_key) === \"profile_card\")) return true\nconst execution = selectedCharacterWorkflow.value?.metadata?.character_execution\nconst legacyProgression = isRecord(execution) && execution.progression_version !== CHARACTER_WORKFLOW_PROGRESSION_VERSION\nif (!legacyProgression) return false\nif (backendNode.node_key === \"prompt_plan\") return backendNode.metadata?.prompt_plan != null\nif (backendNode.node_key === \"material_bind\") return backendNode.metadata?.material_bindings != null\n```\n3) App.vue:927 确为 `const CHARACTER_WORKFLOW_PROGRESSION_VERSION = \"character_gradual_sequential_v1\"`,与后端第 18 行常量字面量重复。\n4) node_routes.py:59-61 `list_workflow_nodes` 的 action 确实只返回 `[node.to_dict() for node in workflow_service.list_nodes(workflow_id)]`,无 progression 字段。\n\n二、\"后端不下发权威结果\"这一根因成立\ngrep 全仓库,`progression.to_dict()` 唯一调用点确为 character_workflow_service.py:339(`_execution_result`,execute-character 响应)。我进一步排查了\"前端能否改读工作流元数据里的权威快照\"这条反驳路径:`character_execution_metadata()`(workflow_progression_service.py:202-220)写入的 `completed_node_ids/current_node_id/all_completed` 只在 character_workflow_service.py:146-165 的 execute-character 路径落库;replay 分支(第 68-74 行)直接 early-return 不刷新;`allow_character_execution_update=True` 的另两处(workflow_service.py:578、807)是分支切换/创建,只改 `current_node_id` 不重算 completed 集合。因此该元数据在节点逐个完成后必然过期,前端确实拿不到可用的权威进度,只能自算。反驳不成立。\n\n三、没有守卫使其不可达,但后端是硬兜底\n`ensure_character_gradual_node_access`(同文件 104-172)在服务端强制抛 `character_gradual_progression_locked`,前端偏差无法被用来绕过规则;前端 6477/8705/8853/8957 等处也都会把锁定原因写进 `workflowNodeActionError`。所以不存在越权或数据损坏,属于 UI 与服务端可能不一致的可维护性缺陷。\n\n四、严重度应降级的理由\n我逐条比对了两份实现,当前语义完全等价(`\"prompt_plan\" in metadata and ... is not None` 与 `metadata?.prompt_plan != null` 在 JSON 语境下同义;legacy 开关条件、白名单集合、profile_card 别名归一均一致)。
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "INSTALL.md(发布包白名单/最小运行时清单)+ frontend/standalone-web/vite.config.ts:12",
|
|||
|
|
"reason": "代码引用属实但结论站不住脚,判定为误报(最多是 P3 流程 nit)。\n\n一、引用的代码确实存在,行号完全对得上:\n- `frontend/standalone-web/vite.config.ts:11-15`:`base: \"/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/\"` / `outDir: resolve(rootDir, \"../../web\")` / `emptyOutDir: true`\n- `frontend/plugin-page/vite.config.ts:11-15`:`base: \"./\"` / `outDir: resolve(rootDir, \"../../pages/imagestudio\")`\n- `backend/services/web_hosting_service.py:14-15`、`manifest()` 只判断 `(dir/\"index.html\").exists()`\n- `.gitignore:16` 只有 `frontend/**/dist/`\n- `git ls-files web/ pages/` = 38 个文件,`git log -- web/assets` = 80 次提交\n以上事实核对无误,不是幻觉。\n\n二、致命反驳:这两个目录不是\"误提交的编译产物\",而是**被文档明确定义的发布负载**,必须入库。\n`INSTALL.md` 的\"最小运行时清单\"把 `pages/imagestudio/` 和 `web/` 与 `main.py`、`metadata.yaml` 并列列为安装后必需内容;\"发布包白名单\"章节要求正式 ZIP 归档 `pages/` 和 `web/`。ImageStudio 是 AstrBot 插件(`metadata.yaml: astrbot_version: \">=4.25.0,<5\"`),装机路径是把仓库/ZIP 铺到 `data/plugins/astrbot_plugin_imagestudio/`,**安装期没有任何 npm/vite 步骤**。若按发现的建议把 `web/` `pages/` 加进 .gitignore,`WebHostingService._resolve` 会直接抛 `ErrorCode.RESOURCE_NOT_FOUND, \"web asset not built\"`,插件页面全线不可用。.gitignore 忽略 `frontend/**/dist/`(中间产物)而保留这两个目录,恰恰是正确的区分,不是遗漏。\n\n三、\"两份重复 dist\"这一核心指控不成立。两份构建的 `base` 不同且必须不同:`web/index.html` 里是绝对路径 `src=\"/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/assets/index-CvjPX4dT.js\"`,`pages/imagestudio/index.html` 里是相对路径 `src=\"./assets/index-Dp4kkpJ0.js\"`。两者挂载于不同 URL 根(`PLUGIN_PAGE_URL = \"/plugin/imagestudio\"` vs `STANDALONE_WEB_URL = \"/api/v1/...\"`),一份产物物理上无法同时服务两个挂载点。`plugin-page/src/App.vue` 复用 `standalone-web/src/App.vue` 正是为避免真正的重复。`DEVELOPMENT.md:57` 与 `INSTALL.md` 都说明了两者定位(`pages/imagestudio/` 是正式插件页,`web/` 是兼容产物)。\n\n四、\"必须构建两次、漏一次\"的风险已有机制消解:`frontend/package.json` 提供 `\"build\": \"npm run build:plugin && npm run build:web\"`,`INSTALL.md:223-224` 与 `DEVELOPMENT.md:72-73` 的验证清单都要求两条一起跑。\n\n五、影响被显著夸大,实测数据不支持\"2MB 噪音淹没 diff\":\n最近 5 次 dist 提交的实际 stat 为 `24 files changed, 59 insertions(+), 69 deletions(-)`(878f1f6)、`24 files changed, 37 insertions(+), 27 deletions(-)`(2dbf90b)。git 的重命名检测把内容哈希改名识别成 `web/assets/{index-CumvEEtc.js => index-CvjPX4dT.js}`,压缩产物是单行文件,因此每次只有几十行 diff,不是 2MB。\n\"git 累积大量永不复用的 blob\"归因也错了:`git count-objects -vH` 显示 size-pack 44.03 MiB,而 `git ls-files release/` 有 20 个 ZIP、仅当前树就占 37.0 MiB——仓库体积的主因是被跟踪的 release 包,不是 web/assets。\n\n六、残余的真实问题很小:确实没有 CI 校验产物新鲜度(仓库根本没有 `.github/workflows`),历史上有 24 次只更新 `web/assets`、3 次只更新 `pages/imagestudio/assets` 的提交。但 `manifest()` 从未声称做版本一致性校验,它只是\"是否已构建\"的状态位;`_resolve` 另有 `candidate != root and root not in candidate.parents` 的路径逃逸防护,功能上没有可达缺陷。把 `fix: recover material library after stale asset token`(3edec9a)当作双 dist 错位的症状也属牵强——该提交同时改了 `frontend/standalone-web/src/App.vue`、`AdvancedMaterialImportPage.vue` 与两个 assets 目录,是源码修复带构建,不是漏构建导<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:4575-4582;backend/services/event_stream_service.py:22-32",
|
|||
|
|
"reason": "代码事实部分属实(行号对得上),但**因果链在关键一步断裂,声称的影响不成立**,判定为误报。\n\n一、代码确实存在,无幻觉\n`backend/services/event_stream_service.py:22-32` 确实只有 `event_type`/`limit`,没有 `after_event_id`:\n```python\ndef recent_events(self, *, event_type: str | None = None, limit: int = 50) -> dict[str, Any]:\n events = self.event_bus.history(event_type)\n normalized_limit = _limit(limit)\n selected = events[-normalized_limit:]\n```\n`App.vue:4575-4577` 也确实是那段 `slice(previousIndex >= 0 ? previousIndex + 1 : -1)`。\n\n二、决定性反驳:全仓库事件类型只有 `generation.` / `workflow.` 两个前缀,OR 判定不可能因丢事件而翻假\n我把 `backend/services/` 下所有 `_publish(` / `event_bus.publish(` 的首参字面量全量抽取,共 17 个,**无一例外**:\n`generation.node_cancelled / node_completed / node_failed / node_queued / node_running / task_created / task_deleted / task_status_changed / task_updated`、`workflow.branch_created / created / node_created / node_material_bindings_updated / node_status_changed / status_changed / updated / node_updated`(含 `workflow_node_material_service.py:209`、`generation_job_retry_service.py:269,278`、`generation_service.py:1417,1426` 这些间接发布点)。\n\n而判定逻辑是 `.some()` 的 OR 归约:\n```js\nconst generationChanged = eventTypes.some((t) => t.startsWith(\"generation.\"))\nconst workflowChanged = eventTypes.some((t) => t.startsWith(\"workflow.\"))\n```\n既然总线上**每一条**事件都满足这两个谓词之一,那么退化到 `slice(-1)` 只留 1 条时,`generationChanged || workflowChanged` **恒为 true**。发现里\"eventTypes 只有一个元素 → 基于 1/20 样本判定 → 该刷的不刷\"是错的:把 20 条样本砍到 1 条,无法把一个所有候选都为真的 OR 归约翻成假。生成爆发期最后一条几乎必然是 `generation.task_updated`/`task_status_changed`,`generationChanged` 为真,`loadJobs()`/`loadDashboard()` 照常触发。\n\n三、刷新是全量重取而非增量应用,丢中间事件不产生状态失配\n`loadJobs()` / `loadDashboard()` / `loadWorkflowRoundDetail()` 都是重新 GET 全量状态,事件只当\"触发器\"用。因此丢掉 19 条冗余触发器不会让 UI 停在旧状态——只要有 1 条触发,取回的就是最新完整状态。\n\n四、workflow 轮次面板另有一条与事件流完全无关的 4 秒自续轮询,直接推翻\"生成完了界面不动的结构性来源\"\n`App.vue:6999-7008`:\n```js\nfunction scheduleWorkflowRoundPolling(detail = selectedWorkflowRound.value) {\n stopWorkflowRoundPolling()\n if (!detail || !workflowRoundIsActive(detail.round.status) || document.hidden) return\n workflowRoundPollTimer = window.setTimeout(() => { ... void loadWorkflowRoundDetail(roundId, true) }, 4000)\n}\n```\n它在轮次入队时就被武装(`App.vue:8538`,紧跟 `已加入生成队列` 分支),并在每次 `loadWorkflowRoundDetail` 成功后自续(`App.vue:7017`),轮次一旦不再 active 才停。另有 `reconnectBackendWhenVisible()`(`App.vue:4471-4477`)在页面重新可见时强制刷新活动轮次。即使事件流一条都不处理,轮次面板也是 4 秒一刷。\n\n五、爆发结束后必然自愈\n`lastEventStreamId` 每轮都推进到 `newest`;爆发一停,下一次 5 秒轮询窗口内新增 <20 条,`findIndex` 就能命中锚点,恢复精确增量。不存在\"直到用户手动刷新\"的持久卡死。\n\n六、\"窗口太小\"也不是接口设计缺陷\n`_limit` 上限是 200(`min(max(limit,1),200)`),总线 `EventBus(history_limit=500)`。前端自己传的 `{ limit: 20 }`,要扩窗口改一个数字即可,不需要引入游标。\n\n七、真正站得住的只有边角问题\n(1) 接口确实无游标;(2) `body`/`content_type` 预渲染的 SSE 文本被前端丢弃,属白费序列化——但每 5 秒 20 条小 payload,量级可忽略;(3) `sli
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/App.vue:3073-3077 / 3136-3140(仅为文案/tone 层面的一致性问题)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "代码本身不是幻觉,但\"声称的影响\"链条站不住脚,判为误报。\n\n一、引用的代码确实存在,行号对得上:\n- backend/route_registry.py:22-27 确为 `\"retry\": {... \"required_permission\": \"can_retry_tasks\", \"required_permissions\": [\"can_write\", \"can_retry_tasks\"] ...}`;\n- core/security.py:31-33 确为 `def require_task_retry(ctx): require_write(ctx); require_permission(ctx, \"can_retry_tasks\")`;\n- App.vue:2786 `const canRetry = Boolean(permissions?.can_retry_tasks)`,3076 与 3139 两处 retry 卡片确实只判 `canRetry && !selectedNodeWriteLocked && !workflowNodeActionExecuting`;3372-3373 队列面板确实是 `canManageQueue = canWrite && canRetry`。\n\n二、但前提条件在本仓库中不可产生(不可达)。全仓 RequestContext 的构造只有三处非测试路径,没有任何一处能产出\"有 can_retry_tasks 而无 can_write\"的组合:\n- main.py:104-106 与 2337:`BRIDGE_PERMISSIONS = frozenset({\"can_write\", \"can_manage_config\", \"can_manage_materials\", \"can_retry_tasks\"})`,`permissions = BRIDGE_PERMISSIONS if username or user_id else frozenset()` —— 要么四个全给,要么一个不给。全不给时 `canRetry` 为 false,卡片就显示\"无权限\"且 disabled,正是期望行为;\n- backend/services/chat_workflow_service.py:1160 `permissions={\"can_write\", \"can_retry_tasks\", \"can_manage_materials\"}` —— 同样含 can_write;\n- backend/services/chat_command_guard_service.py:42 `permissions=set()` —— 空集。\nbootstrap 的权限来源 backend/services/ui_service.py:24-29 只是逐项透传 `ctx.has_permission(...)`,不会额外派生。也就是说\"持有 can_retry_tasks 但无 can_write 的账号\"在当前代码里根本造不出来。\n\n三、即便假设该组合存在,点击路径也有上游守卫,\"一律被 require_write 拒为 FORBIDDEN\"的描述是错的。App.vue:8092-8094 的 workflowNodeRetry 走 `runSelectedWorkflowNodeGeneration({ retry: true })`,而该函数 8560 行第一件事就是:\n const blocked = backendWriteBlockMessage(options.retry ? \"节点重跑\" : \"节点生成\")\n if (blocked) { workflowNodeActionError.value = blocked; return false }\n而 backendWriteBlockMessage(4376-4387)里明确含 `if (!bootstrap.value?.permissions.can_write) return \\`${action}需要后端写入权限。\\``。所以无 can_write 时前端本地就拦下,根本不会发出 nodes/{id}/retry 或 nodes/{id}/generate 请求,也就没有所谓的 FORBIDDEN 回包。发现描述的核心危害链在此断裂。\n(唯一未被该守卫覆盖的是 8559 行 `if (isSelectedWorkflowRoundNode.value) return submitSelectedWorkflowRound(options)` 的提前分支,submitSelectedWorkflowRound(8462-)只判 workflowLockedNodeWriteMessage 而不判 can_write;但它打的是 rounds 接口,route_registry.py:92/95 标注为 \"write\" 而非 \"retry\",与本条发现主张的 retry 复合权限无关,且同样受前述\"权限不可分离\"约束。)\n\n四、严重度被夸大。真正剩下的只是 UI 文案层面的不一致:3073-3074 用 `canRetry ? \"可重试\" : \"无权限\"` 作 state/tone,理论上会在一个当前造不出的权限组合下显示乐观文案,按钮点下去也只会得到一句本地提示。另外 App.vue:3818-3821 也是同样的 canRetry-only 写法,说明这是节点级卡片的一贯写法(节点级操作的写权限统一由 backendWriteBlockMessage 兜底),而 3373 队列面板加 canWrite 是因为队列卡片直接调接口、没有同样的兜底函数——这是两类不同上下文,不能据此推断\"遗漏\"。无安全影响,无越权,无功能故障,最多是 P3 的文案精确性问题。",
|
|||
|
|
"refuted": true
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/features/workflow/workflow-lock-state.ts:12,21-27(对应后端 backend/services/workflow_round_service.py:1145-1155)",
|
|||
|
|
"reason": "结论:核心事实成立(不是幻觉),但\"声称的影响\"里最主要的那条触发路径是错的,严重度被明显夸大,应从原判降为 P3-轻微。\n\n一、代码真实存在,行号对得上\nfrontend/standalone-web/src/features/workflow/workflow-lock-state.ts 第 9-13 行确实是:\n return (\n status === \"archived\" ||\n status === \"cancelled\" ||\n Boolean(metadata?.final_archive_history_id || metadata?.sealed_at)\n )\n第 21-27 行 isWorkflowNodeWriteLocked 确实只判 `normalizedStatus(nodeStatus) === \"selected\" || isWorkflowSealed(...)`。\nbackend/services/workflow_round_service.py:1145-1155 也确实是 `node.status == \"selected\" or node.selected_image_id or (progression is not None and node.node_id in progression.completed_node_ids)`,且该守卫被 create_round(:81) 和 next_round(:497) 两个入口都调用,前端 App.vue:7697 走的 endpoints.roundNext 正好命中 :497。\n\n二、sealed_at 确认是死判据(但无害)\n`grep -rn \"sealed_at\" --include=*.py`(排除 .tmp/ 部署快照)零命中,全仓仅 workflow-lock-state.ts:12 一处。但这是 OR 链里的一个恒 undefined 项,只可能\"多加锁\"不可能\"漏加锁\",行为上完全无副作用,属于代码卫生问题而非缺陷。\n\n三、发现里最主要的触发场景是错的\n原文称\"adopt-image 后 status 仍为 completed\"。实测 backend/services/history_service.py:55 一次性写入两个字段:\n {\"status\": \"selected\", \"selected_image_id\": image.image_id}\n即采用路径两边一致,不产生分歧。此外 generation_service.py:473/955 生成候选时写死 `\"is_selected\": False`,正常轮次流程也不会凭空产生 selected_image_id。\n\n四、分歧确实存在,但路径窄\n真正能造成 selected_image_id 已设而 status 仍为 completed 的,是 backend/services/image_service.py:150-159 的 select_image:它只 `update_node(after.node_id, {\"selected_image_id\": after.image_id})` 而不改 status;且 :249-255 的守卫明确放行 completed:\n if node and node.status in LOCKED_NODE_STATUSES and node.status != \"completed\":\n raise locked_node_edit_error(node)\n(注释自称这是\"tentative selection\"设计。)前端 GalleryPage.vue:121 的批量\"标记选中\"按钮调用 endpoints.imageSelect 即可走到这里。此时 App.vue:2467 canCreateSelectedWorkflowRound 仍返回 true(写锁 false,且 workflowProgressionBlockerFor(App.vue:10830-10835) 只检查该节点之前的节点,不会锁住节点自身),按钮可点,后端 :1145 报错。所以分歧客观存在、可达。\n至于 completed_node_ids 那一条:对 node_type==\"generation\" 的节点,is_character_gradual_node_completed(:180-199) 实际只剩 status in {\"selected\",\"skipped\"} 一种可能(planning/reference_binding/profile_card 分支和 legacy prompt_plan/material_bind 分支都不适用于 generation 节点),因此额外差异仅是 \"skipped\",而 UI 并未提供把生成节点置为 skipped 的入口,属理论洞。\n\n五、影响被夸大\n\"用户看到无法自恢复的报错\"不成立。App.vue:7788-7795 actionableWorkflowApiError 会读取 err.details.required_action,:7783 已有映射 `create_branch: \"当前节点已锁定或工作流已封存,请先创建可编辑分支\"`,用户拿到的是带明确恢复指引的中文提示,且 WorkflowBranchPanel 提供建分支入口。方向上前端偏宽松、后端为权威判定,属 fail-safe,不涉及数据损坏、越权或状态不一致落库。\n\n综上:判据不一致 + 死字段属实,但可达路径窄、后果仅为\"按钮该禁用却没禁用 + 一次带指引的报错\",定为 P3-轻微。修复建议:把 isWorkflowNodeWriteLocked 的签名扩成接收整个 backend node,加上 `Boolean(node.selected_image_id) || status === \"skipped\"`,同时删掉 metadata.sealed_at 这个不存在的字段。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"corrected_location": "frontend/standalone-web/src/features/workflow/workflow-model-route.ts:36-88(代码属实,但所述影响不可达)",
|
|||
|
|
"reason": "代码位置属实(前端 36-48 的候选链、59-88 的 assignment 映射,后端 generation_service.py:169-176 / 617-624 / 2022-2045 全部对得上,不是幻觉),但三条\"不一致\"的结论逐条站不住脚,核心影响不可达。\n\n【1. \"前端多了 round 层\" —— 错误,backend 确实有这一层】\n指控者只看了 generation_service.py,漏了续轮真正的服务。backend/services/workflow_round_service.py:596-597:\n \"model_route\": _optional_text(payload.get(\"model_route\"))\n or _optional_text(parent.prompt_snapshot.get(\"model_route\")),\n即\"本次覆盖 → 父轮次冻结路由\",与前端 [\"override\", …] → [\"round\", …] 的顺序完全一致。前端 selectedWorkflowRoundFrozenRoute(App.vue:2202-2210)读的正是 detail.round.prompt_snapshot 的 model_route,就是后端这个 parent.prompt_snapshot。round 层是如实镜像,不是前端发明的。\n\n【2. \"少了 image_low 兜底\" —— 是刻意设计,且 fail-safe,不会说谎】\nworkflow-model-route.ts:28-31 的注释明写 \"Mirrors the server-owned route chain without inventing a frontend default.\",第 48 行返回 { route: \"\", source: \"backend\" }。这个空值不是被静默吞掉,而是被上游当作硬闸门:\n- App.vue:8492-8495:if (!persistentParent && !selectedWorkflowModelResolution.value.route) { workflowRoundError.value = \"当前节点和后端默认均未提供生图路由…\"; return false }\n- 模板 12929 / 12935-12936::model-status-message 与 :submit-disabled 都绑定 !selectedWorkflowModelResolution.route\n所以缺 image_low 的后果是\"标签显示后端自动 + 禁止提交\",而不是显示错误路由。硬编码 image_low 反而会让 UI 声称一个后端可能拒绝的路由(require_resolved_model_route 会抛 MODEL_CAPABILITY_UNCONFIRMED)。\n\n【3. \"快速角色标签说谎 / quick_avatar 查不到分配\" —— 不可达,这是本条发现的致命硬伤】\n该标签链的输入全部来自 selectedWorkflowBackendNode ← activeWorkflowBackendNodes ← selectedQuickProfileLibraryItem.value?.workflowId(App.vue:1768-1775)。而 workflowId 字段只可能是 character_gradual,四处赋值全查过:\n- App.vue:4943-4954:const detailedWorkflow = availableWorkflows.find(w => w.workflow_type === CHARACTER_WORKFLOW_TYPE),然后 workflowId: detailedWorkflow?.workflow_id;快速工作流被单独放进 quickCharacterWorkflowId 字段。\n- App.vue:7410-7419:只有 if (workflowType === CHARACTER_WORKFLOW_TYPE) 才调 updateQuickProfileWorkflowLink(7524 写 workflowId);else 分支只写 quickCharacterWorkflowId。\n- App.vue:9455:workflowId: saved.workflow && isCharacterWorkflow(saved.workflow) ? … : undefined —— 该 saved.workflow 由 quick_profile_service.py:140 创建,workflow_type 为 \"quick_profile\",isCharacterWorkflow(7533-7535 只认 character_gradual / character_fast)返回 false,故为 undefined。\n- App.vue:9489:唯一调用点 4873 传入 characterWorkflow = undefined。\nCHARACTER_WORKFLOW_TYPE = \"character_gradual\"(App.vue:925)。结论:quick_avatar/quick_bust/quick_stand/quick_sheet 永远不会成为 selectedWorkflowBackendNode,这个 Composer 标签根本不为快速角色流渲染,\"UI 显示本次覆盖而后端按 assigned_route 出图\"的场景无法构造。\n\n而且快速角色流有自己独立且正确的解析器 App.vue:6086-6098:\n return recordString(assignments, \"quick_generation\") || recordString(generation, \"default_model\") || selectedImageModelRoute()\n与后端 _assignment_key 把 quick_* 映射到 quick_generation 完全对齐;且 5813-5815 在其为空时直接 throw \"未配置 quick_generation 统一生图路由…\",连提交都进不去。\n\n【4. 实际可达路径下前后端结论完全一致】\ngradual 工作流的生图节点键(engines/character_flow.py:51-105)只有 headshot / upper_body / lower_body / full_body / turnaround / design_sheet / style_lock,这七个在前端 76-84 与后端 2026-20
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"corrected_location": "backend/services/integration_token_service.py:1(若保留,应收窄为“9 个模块零执行”,且必须删除 branch_lineage / workflow_progression 两项)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"refuted": true,
|
|||
|
|
"reason": "## 判定:误报(方法论无效,核心量化结论与最耸动的子论断均被实测推翻)\n\n### 1. 该发现的取证方法是「在 tests/ 里 grep 模块文件名字符串」,这不是覆盖率\n我先复现了它的脚本,结果确实一致(68 个 service,23 个模块名在 tests/ 全文中零出现),类名同样为 0:\n`IntegrationTokenService 0 / ChatCommandGuardService 0 / DeletionSafetyService 0 / BranchLineageService 0 / WorkflowProgressionService 0 ...`\n\n但这只证明「测试没有直接 import 这个文件名」,不能证明「零测试」。我随即安装 coverage 7.15.2,对全量套件做了真实行覆盖测量:\n`python -m coverage run --source=backend/services -m pytest tests/ -q` → **263 passed in 177.79s**。\n\n### 2. 实测结果:23 个模块中有 8 个被实质性执行,其中 5 个覆盖率 74%~94%\n```\nbranch_lineage_service stmts=172 covered=160 pct=93%\nworkflow_progression_service stmts= 90 covered= 85 pct=94%\nprompt_plan_schema_service stmts=147 covered=123 pct=84%\nmaterial_binding_schema_service stmts=140 covered=104 pct=74%\nnode_injection_template_service stmts= 27 covered= 25 pct=93%\nmodel_provider_bootstrap stmts=126 covered= 41 pct=33%\nreference_binding_mapper stmts= 40 covered= 22 pct=55%\nchat_session_service stmts= 66 covered= 34 pct=52%\n```\n我进一步用 AST 剔除 import/def/class 头部行,只统计**函数体内真实执行的语句**,排除「仅 import 噪音」的可能:\n```\nbranch_lineage_service body_executed=133\nprompt_plan_schema_service body_executed= 97\nmaterial_binding_schema_service body_executed= 84\nworkflow_progression_service body_executed= 67\n```\n这是货真价实的行为覆盖,不是导入副作用。\n\n### 3. 最耸动的那句「同名测试不测同名 service,制造假象」——事实完全相反\n发现称 `test_workflow_branch_lineage.py`、`test_character_workflow_progression.py` 是「假象」。实测这两个模块恰恰是覆盖率最高的两个(93% / 94%)。原因是它们是**被再导出的 helper 模块**,测试通过前门 service 调用:\n```python\n# backend/services/workflow_service.py:10\nfrom backend.services.branch_lineage_service import (\n# backend/services/workflow_node_material_service.py:17\nfrom backend.services.branch_lineage_service import workflow_metadata_for_branch\n# backend/services/character_workflow_service.py:11\nfrom backend.services.workflow_progression_service import (\n# backend/services/image_service.py:11\nfrom backend.services.workflow_progression_service import ensure_character_gradual_node_access\n```\n而 `test_workflow_branch_lineage.py` 第 6-8 行正是 `from backend.services.character_workflow_service import CharacterWorkflowService` / `workflow_service import WorkflowService`。同名测试**正在**测同名 service,只是走的是调用方入口。把这称为「比明摆着没测试更危险的假象」是对代码的误读,且是本条发现中最不成立的部分。\n\n### 4. 行数被夸大约一倍\n被声称的 4579 行中,实质有覆盖的 8 个模块合计约 **2279 行**(356+240+379+408+366+375+69+86),即约一半。真正零执行的只有 9 个模块(api_discovery 351、asset_import 406、contract 30、dashboard 117、deletion_safety 75、history 314、integration 171、integration_token 202、log 300 ≈ **1966 行**),不是 23 个 / 4579 行。\n\n### 5. 残余的真实内核(但不足以支撑原发现,且非代码缺陷)\n公允地说,实测确认了三点中的两点:\n- `integration_token_service` executed=0(129 statements 全未执行)——属实;\n- `deletion_safety_service` executed=0——属实;\n- `chat_command_guard_service` 实际是 30%,但 body_executed 仅 4 行,`check_generation` 主体(21-39 行)与 `_daily_count`/`_daily_group_count` 全在 missing 中,行为层面确实未测——这一条方向上成立,但发现写的「零测试」与实测的 30% 不<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/routes/common.py:12-16(配合 backend/dto/common.py:75-83)",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "发现成立,非幻觉,但严重度应下调为 P2-中等。\n\n【1. 代码真实性:完全对得上】backend/routes/common.py 全文仅 17 行,第 12-16 行逐字符与证据一致:\n```\ndef handle_request(ctx: RequestContext, callback: Callable[[], Any]) -> dict[str, Any]:\n try:\n return success_response(callback(), request_id=ctx.request_id)\n except Exception as exc:\n return response_from_exception(exc, request_id=ctx.request_id)\n```\nbackend/dto/common.py:75-83 亦一致,第 80 行在源码中是转义形式 `message=\"服务器内部错误\"`(即\"服务器内部错误\"),details 只有 `{\"exception\": exc.__class__.__name__}`,无栈、无消息、无上下文。\n\n【2. 数量断言精确】`python -c \"from backend.route_registry import ROUTES; print(len(ROUTES))\"` 输出 166;backend/routes/ 下除 common.py/params.py/__init__.py 外正好 32 个 route 文件,handle_request 在其中出现 189 次。\n\n【3. 无上游守卫,异常确实无法冒泡】adapters/astrbot_adapter.py:171-187 的 dispatch_extension_request 末行是 `return match.handler(*_handler_args(match, ctx, query or {}, payload or {}))`,无 try/except、无日志;main.py 的 dispatch_bridge_request 只做 `_to_media_http_response` / `_to_bridge_response` 响应转换。异常在 handle_request 内即被吞,AstrBot 宿主日志同样看不到。\n\n【4. 全仓确无 API 层日志】grep \"logging\\.|getLogger|logger\\.|traceback|format_exc\" 覆盖 backend/ core/ adapters/ database/ main.py,仅命中 backend/services/worker_service.py(worker 路径有 logger.exception)与 core/config_writer.py 的一行注释;无 print、无 sys.stderr。\n\n【5. log_service 断言属实】backend/services/log_service.py:170-211 的 error_logs 只 SELECT `FROM generation_tasks WHERE status ...` 与 `FROM integration_logs WHERE ...` 两张表,确实不含 API 层错误。\n\n【6. 影响面比原描述还略宽】backend/services/ 下有 92 处 raw builtin raise,例如 advanced_style_set_service.py:1258 `raise RuntimeError(\"style source recognition no longer owns the source\")`,走 HTTP 路径时其消息文本同样被丢弃,只剩 \"RuntimeError\" 类名。\n\n【需要修正的定性与严重度】这是可观测性/可维护性缺陷,不是功能或安全缺陷:响应结构仍正确(ok:false + error_code + request_id),468 处 `raise ApiError` 的显式错误码与消息全部完整保留(response_from_exception 第 76-77 行的 isinstance 分支),向客户端隐藏内部异常消息本身是正确做法——真正的缺陷只是服务端一侧零留痕。不造成数据损坏、不绕过鉴权、不产生错误结果,判 P0/P1 属夸大,P2-中等 合适。修复成本也极低:在 except 分支对非 ApiError 加一行 logger.exception 即可。"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": false,
|
|||
|
|
"corrected_location": "backend/routes/common.py:15(真正根因);backend/services/worker_service.py:145 为次要根因",
|
|||
|
|
"corrected_severity": "P2-中等",
|
|||
|
|
"reason": "【核心事实成立,但定位错、影响描述有多处夸大,严重度需下调】\n\n一、统计数字逐条复核,完全属实\n`grep --include=\"*.py\" \"import logging|getLogger|logger\\.|logging\\.\"` 全仓命中 **8 处,全部集中在 backend/services/worker_service.py**(Grep 输出:`Found 8 total occurrences across 1 file`)。行号对得上:第 5 行 `import logging`、第 17 行 `logger = logging.getLogger(__name__)`。源码树文件数 backend 118 + core 24 + database 18 + engines 8 + adapters 10 = **178**,与报告一致。另外补充:全仓连 `print(` 都是 0。`except Exception` 实测 180 处,其中直接 `pass/return None/continue` 实测 30 处(报告说 33,误差可忽略)。**不是幻觉。**\n\n二、我找到了三层反证,必须先摆出来\n1. `backend/services/log_service.py` 存在一整套 DB 落库的可观测层:`generation_tasks`(error_message/status/progress)、`workflow_history`、`integration_logs`(含 `request_json`/`response_json`/`error_message`)、`audit_logs`。\n2. `core/audit_service.py:38-63` 的 `AuditService.record()` 写入 `request_id / actor_type / actor_id / action / before_snapshot / after_snapshot / metadata_json`,且被 **17/18 个 route 文件 + 十余个 service** 引用。\n3. `backend/dto/common.py:44-72`,**每一个成功和失败响应都带 `request_id`**,由 `core/identity.py:16 request_id: str = field(default_factory=new_request_id)` 生成;`route_registry.py:329-333` 还暴露了 `/logs/errors`、`/logs/generation`、`/security/audit-logs` 查询端点。\n\n所以报告中「无 request_id」「无入参」「生产环境完全不可观测」**是错的**,标题属于夸大。\n\n三、但缺陷本身依然成立,且根因不在报告指的位置\n真正的根因是 `backend/routes/common.py:12-16`:\n```python\ndef handle_request(ctx, callback):\n try:\n return success_response(callback(), request_id=ctx.request_id)\n except Exception as exc:\n return response_from_exception(exc, request_id=ctx.request_id)\n```\n配合 `backend/dto/common.py:75-83`,非 ApiError 的异常被压成 `{\"ok\": false, \"error_code\": \"UNKNOWN_ERROR\", \"message\": \"服务器内部错误\", \"details\": {\"exception\": exc.__class__.__name__}}` —— **只剩一个类名,栈被彻底丢弃,且此处没有任何 logging 调用**。我确认了 `backend/routes/` 下 **158 处** `handle_request(ctx, ...)` 调用全部走这条路径。\n\n关键在于这个 `except` 让框架兜底也失效了:`adapters/astrbot_adapter.py:171-187` 的 `dispatch_extension_request` 和 `main.py:315-344` 的 `dispatch_bridge_request` 都**没有** try/except,本来未捕获异常会冒泡到 Quart 由框架打出 traceback;但 `handle_request` 在更内层就吃掉了,Quart 永远看不到。所以第 3 问「是否有框架机制消解」的答案是否定的——反而是这段代码主动屏蔽了框架机制。\n\n四、报告漏掉了一个更严重的点(对其有利)\n唯一有 logger 的文件,其实也没记业务失败。`worker_service.py:145-146`:\n```python\nexcept Exception as exc:\n return self._handle_failure(task, exc, ctx)\n```\n而 `_handle_failure`(223-294 行)只对**二次失败**调 `logger.exception`(\"could not reload task\"、\"could not mark task as failed\"),对真正的生成异常 `exc` 只取 `_safe_error_message(exc)` 文本入库,从不打栈。第 198 行那条 `logger.exception` 仅在 `_claim_next`/`_handle_failure` 自身抛错时才触发。也就是说**连队列 worker 的生成失败也没有栈**。\n\n五、严重度:从标题暗示的 P0/P1 下调为 P2-中等\n理由:不涉及数据丢失、越权、崩溃或功能不可用;`audit_logs` + `integration_logs` + `generation_tasks.error_message` + `metadata.worker.error_type` 已提供了带 request_id 的结构化追踪面,运维并非「一无所有」。真实损害是 158 个端点上非预期异常(KeyError/AttributeError/TypeError)的栈信息不可恢复,<EFBFBD><EFBFBD>
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"refuted": true,
|
|||
|
|
"corrected_location": "backend/services/chat_workflow_service.py:903(唯一绕过 dispatcher 的直连调用,对应 main.py:1134 `/yp 审核`)",
|
|||
|
|
"corrected_severity": "P3-轻微",
|
|||
|
|
"reason": "判定:误报(核心影响链不成立)。\n\n一、机制本身属实,但只是\"潜在\"而非\"已触发\"\nbackend/dto/common.py:29-33 的代码确实存在,行号完全对得上。我本机实测(Python 3.11.15)确认位置参数构造时 `BaseException.__new__` 会填充 args:\n POS str = \"(<ErrorCode.VALIDATION_FAILED: 'VALIDATION_FAILED'>, '内容审核未通过', {'body_preview': 'secret'})\"\n所以\"str(exc) 返回元组 repr\"这一点我不否认。但这只在 ApiError **作为异常对象本身**流到 f-string 时才有害。\n\n二、决定性反证:main.py 里根本没有 ApiError\n`git ls-files \"*.py\" | xargs grep -ln ApiError` 的结果里**没有 main.py**;`grep -c ApiError main.py` = 0。main.py 那 40 处 `{exc}`(我逐条核对,515/669/1591 等行号确实存在)捕获到的不是 ApiError。\n\n三、框架层已经消解了它——两道闸门\n闸门1,backend/routes/common.py:12-16:\n def handle_request(ctx, callback):\n try: return success_response(callback(), request_id=ctx.request_id)\n except Exception as exc: return response_from_exception(exc, request_id=ctx.request_id)\n`response_from_exception` 走 `exc.to_response()`,用的是 **self.message 字段**而非 str(exc),details 被装进 dict 的 \"details\" 键。\n我用脚本遍历了 `backend.route_registry.ROUTES` 全部 166 条路由 / 163 个 distinct handler:**0 个未被包裹**。初筛出的 8 个(start_workflow、archive_profile、get_error_logs 等)经查全部委托给 `_set_workflow_status` / `delete_profile` / `_read_logs`,这些内部都以 `return handle_request(ctx, action)` 收尾。\n\n闸门2,backend/services/chat_workflow_service.py:1166-1170:\n @staticmethod\n def _data(response):\n if not response.get(\"ok\"):\n raise ValueError(str(response.get(\"message\") or response.get(\"error_code\") or \"ImageStudio 请求失败\"))\n return response.get(\"data\")\n即 details 在这里被**整体丢弃**,只取 message,且 ValueError 是单个位置参数 → str() 干净。\n\n端到端实测(模拟含 provider_id/model_id/body_preview 的 ApiError):\n route response dict: {'ok': False, ..., 'details': {'provider_id':..., 'body_preview': '{\"error\":{\"message\":\"sk-leak...\"}}'}}\n type reaching main.py: ValueError\n main.py renders: 错误:生成失败:上游返回错误\n聊天用户拿到的是干净中文 message,无元组语法、无 details。所谓\"中文提示被 Python 元组语法淹没\"与\"details 推送到群聊\"两条影响均不成立。\n\n四、证据片段有事实错误\n发现称 engines/image_provider.py:253,289,324 会把 body_preview 塞进 ApiError。实际读该文件:`grep -c ApiError engines/image_provider.py` = **0**。那三处抛的是 ImageProviderError,且 engines/image_provider.py:96-110 定义为:\n class ImageProviderError(RuntimeError):\n def __init__(self, message, *, provider_id, status_code=None, error_type=\"provider_error\", details=None):\n super().__init__(message) # 单参数 → str(exc) 就是 message\n self.details = details or {}\ndetails 是普通属性,永远不进 str()。body_preview 泄漏链完全不存在。\n\n五、异步推送路径也已加固\nworker_service.py:230 已经是 `message = _safe_error_message(exc)`,514-534 的实现只保留 message + code + reason/finish_reason/provider/model 白名单提示并截断 500 字。发现说\"作者知道坑但聊天路径没复用\"——实际是聊天路径用了另一套等效防护(_data 直接丢 details),不需要复用。\n\n六、唯一残留(远低于所称严重度)\nchat_workflow_service.py:903 是全仓唯一绕过 dispatcher 的进程内直调:\n model_reviews = {node_id: self.cross_reviewer_factory().review_node(node_id) for node_id in node_ids}\nCrossReviewService.review_node 内部经 workflow_progression_service.py:168-172 可抛 ApiError,此时会以元组 repr 落到 main.py:1134 `❌ 审核失败:{exc}`<EFBFBD><EFBFBD>
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
}
|