docs(progress): record the R3 batch and its gate results
一些检测仍在等待运行
ci / Portable quality checks (push) Waiting to run
ci / Windows Tauri build (push) Waiting to run

Batch R3 (F-14 through F-19) is implemented and gate-verified. Record
what changed, the measured test counts, and the two things it did not
cover: auto_connect_configured_slots is still serial, and the
operations switch guard still has no production writer.

Also record that the new tests for F-14, F-15 and F-16 were each run
against a mutation restoring the old behaviour and each fails under it,
so they are known not to be tautological.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
这个提交存在于:
YokiiroBW 2026-07-27 00:31:18 +08:00
父节点 49dccc5eac
当前提交 178fba70ea
共有 5 个文件被更改,包括 146 次插入 和 3 次删除

查看文件

@ -40,6 +40,27 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
### Fixed
- One unreachable connection no longer freezes the whole application. The reachability probe now
runs outside the runtime lock, so a slot that is connecting cannot block status, sending,
cancelling or disconnecting on the other slot, and the attempt itself can be interrupted.
- An activation rejected during validation no longer tears down a working connection or leaves the
slot stuck in "connecting" with nothing to explain it.
- Switching a session while the slot is being reconnected no longer writes the old session id onto
the new connection, and no longer copies it into that profile's sticky session.
- One connection profile can no longer be activated into both slots at once, which previously let
the two slots overwrite each other's sticky session.
- Disconnecting or re-pointing the operations slot is now refused while an operations task is
active, matching the existing rule for slot assignment.
- A rejected chat send no longer leaves the question in local history without a failure marker, and
retrying no longer appends a second copy of it. A send that fails after AstrBot accepted it keeps
its question, since the remote conversation already has it.
- Streaming replies are no longer doubled in development. Event listeners that resolved after their
component had unmounted were never unregistered, so a leaked listener applied every delta twice.
- The composer draft, pending attachments, and the error banner are now per slot. An attachment
uploaded on the companion slot no longer follows the user to the operations slot, and an error
from one slot no longer appears on the other.
- An error from sending or uploading is no longer erased half a second later by the connection
status poll.
- Scoped AstrBot stream event ids per chat request so the second and later messages on a
connection are no longer discarded by the client-side event deduplicator.
- Accepted Rust `null` values in the runtime-event, remote-session, and model-option

查看文件

@ -553,3 +553,56 @@ pending in `docs/q9-local-acceptance.md`. Certificate pinning, signed updates, a
remain externally blocked.
Next: batch R3 (isolation and state correctness), F-14 through F-19.
## 2026-07-27: batch R3, isolation and state correctness
Current stage: P02 dual connections, branch `feat/p02-dual-connections`. Six commits, one per fix.
### Completed
- **F-14** `activate_slot` no longer holds the runtime lock across the reachability probe. It is now
three phases: validate and build the client with no lock, claim the slot in one critical section
that installs the cancellation token up front, probe outside the lock under a biased `select!`,
then write back only if the slot's generation still matches. One unreachable address used to
freeze every command on both slots for up to 120 seconds, and the activation could not be
interrupted. This also fixed an unrecorded defect: validation ran after the `Connecting` snapshot
was written, so a rejected request tore down a working connection and parked the slot in
`Connecting` with no client and no error.
- **F-17** `switch_slot_session` captures the generation with the client and checks it before
writing `active_session_id`, so a slot reset mid-switch cannot receive the old session id.
- **F-19** One profile can no longer occupy both runtime slots. Enforced inside `activate_slot`
(which also covers startup auto-connect), cross-checked against `connection_slots` in the
command, and the operations switch guard now covers probe and disconnect. Both rules are
negative — they ask only whether the profile is bound to the *other* slot — because auto-connect
activates from `auto_connect_slot` and never writes `connection_slots`.
- **F-18** Sends are validated before anything is written to history, via a new read-only
`ensure_slot_can_send`. A failure that the server never received marks the row rather than
leaving it clean; `cancelled` and `streamInterrupted` keep the question because AstrBot already
has it.
- **F-15** All four `listen()` call sites now unregister correctly when the promise resolves after
the effect was cleaned up. Under StrictMode the first listener leaked on every mount, which
doubled every character of a streamed reply.
- **F-16** `draft`, `pendingAttachments`, and the error banner are per slot, and every async action
pins its slot at the start. The error state was split in two: the 500ms status poll used to clear
the shared error on success, silently erasing action errors.
### Test status
Rust 107 passed / 2 ignored (`astrbot-desktop` unit tests 76 → 87). Frontend 24 passed / 5 files
(19 → 24, including a new `ChatPanel.test.tsx`, which the component previously had none of).
Playwright 5 passed. `npm run build` and `npm run tauri:build` pass.
The new tests for F-14, F-15, and F-16 were each checked against a mutation restoring the old
behaviour, and each fails under it, so they are not tautological.
### Not covered
`auto_connect_configured_slots` is still serial, so a dead profile delays the second profile by up
to 120 seconds — it no longer freezes the UI, but the delay remains. The operations switch guard
still has no production writer (F-74, audit P3-7); F-19 extends only its read side.
Real-hardware acceptance for R3 has not been performed; the steps are listed in
`docs/q9-local-acceptance.md`.
Next: batch R4 (data and credential boundaries) or R5 (protocol hardening). Neither depends on R3,
so they can be dispatched in parallel.

查看文件

@ -106,6 +106,31 @@ Not executed, not claimed. Each item names the fix it is the acceptance evidence
Items 1 and 2 need only the installer, not a container.
## 2026-07-27 verification update (R3 isolation and state correctness)
### Gate results
`cargo fmt`, `cargo clippy --workspace --all-targets -D warnings`, and `cargo check` pass.
`cargo test --workspace`: 107 passed, 0 failed, 2 ignored (`astrbot-desktop` unit tests 76 to 87).
`npm run typecheck`, `npm run lint`, `npm run build` pass. `npm run test -- --run`: 24 passed across
5 files, up from 19 across 4 — `ChatPanel.test.tsx` is new, the component had no tests before.
`npm run test:e2e`: 5 passed. `npm run tauri:build`: passed;
`target/release/bundle/nsis/AstrBot Desktop_0.1.0-alpha.2_x64-setup.exe`, 8,874,747 bytes, built
2026-07-27 00:30:22 +0800.
The new tests for F-14, F-15 and F-16 were each run against a mutation restoring the old behaviour
and each fails under it, so they exercise the invariant rather than restating the implementation.
### Pending, requires a live AstrBot container
1. With one slot pointed at an unreachable address, confirm the other slot's status, sending and
disconnect stay responsive, and that the connecting slot can be interrupted with 断开 — F-14.
2. Upload an attachment on the companion slot, switch to operations, send; the attachment must not
follow. Type a draft on one slot and confirm it does not appear on the other — F-16.
3. Send while disconnected and confirm no user row appears in history — F-18.
4. Probe the same profile into both slots in turn; the second must be refused — F-19.
5. Confirm a streamed reply renders each character once in a production build — F-15.
## Explicitly pending external acceptance
- HTTPS certificate pinning requires a real HTTPS endpoint and first-trust confirmation.

查看文件

@ -16,7 +16,8 @@
|---|---|---|
| R1 发布阻断批 | **已完成,门槛已验证** | 9 项全部实现。2026-07-26 首次跑完整门槛,发现 2 处编译/lint 失败并修复(详见下方「R1 门槛验证结果」),补写 F-05 缺失的单测。现全绿。 |
| R2 传输层批 | **已完成,门槛已验证** | F-10 – F-13 全部实现,新增 9 条 Rust 测试(`astrbot-desktop` 单元测试 67 → 76)。顺带并入 F-56(P3-4,取消不再置 Disconnected)。全绿。 |
| R3 – R8 | 未开始 | 下一批为 R3。依赖关系见文末批次依赖图。 |
| R3 隔离与状态正确性批 | **已完成,门槛已验证** | F-14 – F-19 全部实现,新增 11 条 Rust 测试(`astrbot-desktop` 单元测试 76 → 87)与 5 条前端测试(19 → 24,含新建的 `ChatPanel.test.tsx`)。全绿。详见下方「R3 实现记录」。 |
| R4 – R8 | 未开始 | 下一批为 R4 或 R5,两者与 R3 无依赖,可并行分派。 |
> 计数订正(2026-07-27 实测):本表此前记为「67 → 73(+6)」,与
> `r1-r2-status-2026-07-26.md` 的「67 → 76(+9)」冲突。实跑 `cargo test --workspace` 的结果是
@ -74,6 +75,45 @@
在把它二分到具体注入行之后才定位到真因。**F-69 不存在,已撤销。**
后续审查必须只读运行,或用独立 worktree 隔离。
### R3 实现记录(2026-07-27)
- **F-14** `activate_slot` 拆成三段:段 0 校验与建 client(无锁),段 1 单个临界区内 bump generation、
**当场安装** `connection_cancellation`、写 `Connecting` 快照,段 2 锁外 `tokio::select!`(`biased;`)
探测,段 3 重新取锁后先 `ensure_generation` 再回写。监听器收到的是段 1 那个 Arc——若在此新建第二个
token,`deactivate_slot` 将取消一个无人监听的 token,desktop-v1 的重连循环会比断开活得更久。
**顺带修掉一个队列未记的缺陷**:原实现把校验放在 `Connecting` 快照之后,`?` 早返回会拆掉一个正常
连接并把槽永久留在 `Connecting`(无 client、无 `last_error`)。
- **F-17** `switch_slot_session` 在第一临界区连同 client 捕获 `generation`,回写抽成
`commit_session_id()` 并先校验。
- **F-19** 三层:`activate_slot` 段 1 的运行时不变量(覆盖 `auto_connect_configured_slots`,它直接调
`activate_slot`)、`probe_connection_slot` 对 `connection_slots` 的交叉校验、运维守卫扩展到
probe / disconnect。**两条规则都写成否定式**(只问「是否绑在另一槽」):auto-connect 依据
`profile.auto_connect_slot` 激活且从不写 `connection_slots`,肯定式规则会打死重连。
运维守卫今日仍无生产写入方(F-74),因此附带一条直接翻转守卫的单测,避免复刻死分支。
- **F-18** 新增只读预检 `ensure_slot_can_send()`;条件抽成 `slot_send_precondition()` 供发送路径复用。
`message_id` 只铸一次,失败且 `!server_received(code)` 时以 `error = true` 幂等重写该行。
`server_received()` 只认 `cancelled` / `streamInterrupted`——这两种情况 AstrBot 已收到消息。
配套一条穷举 `astrbot_send_error` 全部变体的测试,防止 R2 错误码改名而静默破坏分类。
**R2 的重试循环与错误映射逐字未动**,五条 R2 哨兵测试复跑通过。
- **F-15** 四处 `listen()` 统一改「`active` 标志 + 迟到 resolve 立即注销 + `.catch()`」。
测试替身 `src/test/tauriEvent.ts` 采用**延迟 resolve**,否则竞态无法表达;`setup.ts` 每例后重置
注册表(`clearMocks` 不管普通模块状态)。新建 `ChatPanel.test.tsx`。
- **F-16** `draft` / `pendingAttachments` / 错误全部改 `Record<ChatSlot, …>`,每个异步动作开头
`const slot = selectedSlot;` 固定槽位。**错误必须拆成两个 state**:`refresh()` 每 500ms 在成功时
`setError(null)`,此前会在半秒内静默抹掉任何发送/上传错误横幅;轮询失败归 `runtimeError`
(描述 IPC 通道),用户动作归 `slotError`。
**测试有效性**:F-14、F-15、F-16 的新测试都做过变异检验——分别注入「探测期间持锁」「旧的无条件
赋值 unlisten」「按固定槽位读取」三种回退,对应测试全部失败,确认非同义反复。
**门槛结果(2026-07-27)**:`cargo fmt` / `clippy -D warnings` 通过;`cargo test --workspace`
107 通过 / 0 失败 / 2 ignored(`astrbot-desktop` 单元 76 → 87);`npm run typecheck` / `lint`
通过;`npm run test` 19 → 24 通过 / 5 文件;`npm run build`、`npm run test:e2e`(5/5)、
`npm run tauri:build`(NSIS)通过。
**R3 未覆盖、仍开放**:`auto_connect_configured_slots` 仍串行(`commands.rs`),一个死 profile 会把
第二个 profile 延迟至多 120s——但不再冻结 UI。属遗留限制,本批不修。
### R1 已改动的文件
新增:`src/lib/errorMessage.ts`、`src/features/pet/types.test.ts`

查看文件

@ -14,14 +14,18 @@
| --- | --- | --- | --- | --- |
| R1 发布阻断批 | 9 | 已完成 | 已通过 | **未做** |
| R2 传输层批 | 4 | 已完成 | 已通过 | **未做** |
| R3 隔离与状态正确性 | 6 | 未开始 | — | — |
| R3 隔离与状态正确性 | 6 | 已完成 | 已通过 | **未做** |
| R4 数据与凭据安全 | 6 | 未开始 | — | — |
| R5 协议健壮与契约治理 | 13 | 未开始 | — | — |
| R6 体验、无障碍与产品红线 | 10 | 未开始 | — | — |
| R7 性能与工程化 | 8(1 项已提前完成) | 未开始 | — | — |
| R8 剩余 P3 | 17(1 项已提前完成) | 未开始 | — | — |
**已完成 15 项,剩余 58 项。** 共 10 个提交,工作区干净。
**已完成 21 项,剩余 52 项。** 工作区干净。
> 2026-07-27 更新:R3(F-14 – F-19)已实现并通过门槛,实现记录见
> [fix-queue-2026-07-26.md](./fix-queue-2026-07-26.md) 的「R3 实现记录」。R3 的真机验收项
> 已并入 `docs/q9-local-acceptance.md` 的待真机清单。
> 2026-07-27 订正:R5 由 9 项增至 13 项、R8 由 16 项增至 17 项。初版队列遗漏了审计中的
> P2-6、P2-7、P2-8、P2-29、P3-7 五条发现,现补编为 F-70 – F-74(`F-69` 保持永久空缺)。