astrbot_plugin_private_comp.../dreaming.py
2026-07-30 01:13:23 +08:00

1118 行
51 KiB
Python

此文件含有模棱两可的 Unicode 字符

此文件含有可能会与其他字符混淆的 Unicode 字符。 如果您是想特意这样的,可以安全地忽略该警告。 使用 Escape 按钮显示他们。

# -*- coding: utf-8 -*-
from __future__ import annotations
import random
import re
from difflib import SequenceMatcher
from datetime import datetime
from typing import Any
from .helpers import _now_ts, _safe_float, _safe_int, _single_line, _today_key
_ABSTRACT_DREAM_FRAGMENT_MARKERS = (
"状态", "情绪", "心情", "感觉", "余韵", "碎片", "生活感", "日程", "计划", "总结",
"今天", "明天", "用户", "主动", "消息", "回复", "关系", "陪伴", "模型", "生成",
)
_DIARY_STATUS_BROADCAST_MARKERS = (
"今天偏", "当前天气", "状态确认", "今天状态", "能量", "适合推进",
"平稳推进", "没有什么特别重的话想说", "醒来后慢慢把自己拢回",
"梦里的雾还没散", "等晚一点遇到合适的小事再讲", "今日状态",
)
_DIARY_CONCRETE_ACTION_MARKERS = (
"放", "拿", "翻", "写", "擦", "收", "整理", "拉开", "关上", "停", "等",
"看", "听", "闻", "走", "坐", "喝", "热", "晾", "找", "碰", "摸", "回",
)
def _diary_reads_like_status_broadcast(payload: Any) -> bool:
if not isinstance(payload, dict):
return True
summary = _single_line(payload.get("summary"), 180)
body = _single_line(payload.get("body"), 600)
if not summary or not body:
return True
text = f"{summary} {body}"
marker_hits = sum(1 for marker in _DIARY_STATUS_BROADCAST_MARKERS if marker in text)
has_concrete_action = any(marker in body for marker in _DIARY_CONCRETE_ACTION_MARKERS)
return marker_hits >= 2 or (marker_hits >= 1 and not has_concrete_action)
def _clean_dream_fragment_text(text: Any, limit: int = 28) -> str:
raw = _single_line(text, 80)
if not raw:
return ""
raw = raw.replace(",", ",").replace("。", ",").replace(";", ",").replace("、", ",")
parts = [part.strip(" ,.!!??::()()[]【】\"'“”") for part in raw.split(",") if part.strip()]
if parts:
parts = sorted(parts, key=lambda item: (len(item) > limit, len(item)))
raw = parts[0]
raw = _single_line(raw, limit).strip(" ,.!!??::()()[]【】\"'“”")
if len(raw) <= 1:
return ""
return raw
def _dream_fragment_is_useful(text: str) -> bool:
cleaned = str(text or "").strip()
if not cleaned or cleaned in {"没有记住梦", "平稳", "暂无天气信息", "无明确碎片"}:
return False
if len(cleaned) > 32:
return False
abstract_hits = sum(1 for marker in _ABSTRACT_DREAM_FRAGMENT_MARKERS if marker in cleaned)
concrete_markers = (
"光", "雨", "风", "水", "纸", "书", "门", "窗", "杯", "碗", "路", "影", "声", "味",
"颜色", "蓝", "红", "白", "黑", "暖", "冷", "手", "衣", "鞋", "车", "灯", "雾", "床",
"被子", "手机", "屏幕", "钥匙", "包装", "饮料", "猫", "楼梯", "走廊",
)
has_concrete = any(marker in cleaned for marker in concrete_markers)
if abstract_hits >= 2 and not has_concrete:
return False
return True
def recent_diary_tags(plugin, /) -> set[str]:
diaries = plugin.data.get("bot_diaries", [])
tags: set[str] = set()
if not isinstance(diaries, list):
return tags
for diary in diaries[-3:]:
if not isinstance(diary, dict):
continue
raw_tags = diary.get("tags", [])
if isinstance(raw_tags, list):
tags.update(str(tag) for tag in raw_tags)
return tags
def recent_diary_context(plugin, count: int = 3) -> str:
diaries = plugin.data.get("bot_diaries", [])
if not isinstance(diaries, list) or not diaries:
return "(暂无)"
recent = [diary for diary in diaries[-max(count * 2, count):] if isinstance(diary, dict)]
repeated_food_tokens: set[str] = set()
food_seen: dict[str, int] = {}
for diary in recent:
text = " ".join(
_single_line(diary.get(key), 160)
for key in ("summary", "share_seed", "body")
if _single_line(diary.get(key), 160)
)
for token in _diary_food_motif_tokens(text):
food_seen[token] = food_seen.get(token, 0) + 1
repeated_food_tokens = {token for token, total in food_seen.items() if total >= 2}
lines = []
for diary in diaries[-count:]:
if not isinstance(diary, dict):
continue
tags = diary.get("tags", [])
tag_text = "、".join(str(tag) for tag in tags[:4]) if isinstance(tags, list) else ""
summary = _single_line(diary.get("summary"), 120)
if repeated_food_tokens:
summary = _soften_repeated_diary_food_motifs(summary, repeated_food_tokens)
if summary:
date_text = _single_line(diary.get("date"), 16)
age_text = _diary_age_label(plugin, date_text)
suffix = f"({age_text},只作余味和避重)" if age_text else "(只作余味和避重)"
continuity = diary.get("continuity_thread") if isinstance(diary.get("continuity_thread"), dict) else {}
motif = _single_line(continuity.get("motif"), 60)
status = _single_line(continuity.get("status"), 16)
thread_text = f";线索={motif}({status or '出现'})" if motif else ""
lines.append(f"- {date_text} {suffix}:{summary} {tag_text}{thread_text}".strip())
return "\n".join(lines) if lines else "(暂无)"
def _diary_age_label(plugin, date_text: str) -> str:
if not date_text:
return ""
try:
diary_date = datetime.strptime(date_text[:10], "%Y-%m-%d").date()
today = plugin._environment_now().date() if hasattr(plugin, "_environment_now") else datetime.now().date()
days = max(0, (today - diary_date).days)
except Exception:
return ""
if days <= 0:
return "今天"
if days == 1:
return "昨天"
return f"{days}天前"
def _diary_food_motif_tokens(text: Any) -> list[str]:
cleaned = _single_line(text, 500)
if not cleaned:
return []
food_tokens = (
"糖醋排骨", "排骨", "螺蛳粉", "锅包肉", "烤肠", "豆花", "冰粉", "奶茶",
"豆浆", "夜宵", "便当", "饭团", "甜口", "软糖",
)
return [token for token in food_tokens if token in cleaned]
def _soften_repeated_diary_food_motifs(text: str, repeated_tokens: set[str]) -> str:
softened = _single_line(text, 140)
if not softened:
return ""
changed = False
for token in sorted(repeated_tokens, key=len, reverse=True):
if token and token in softened:
softened = softened.replace(token, "近期重复食物意象")
changed = True
if changed:
softened += "(不要复刻具体菜名)"
return _single_line(softened, 140)
def _compact_diary_text(text: Any, limit: int = 220) -> str:
raw = _single_line(text, limit)
if not raw:
return ""
chars: list[str] = []
for char in raw:
if "\u4e00" <= char <= "\u9fff" or char.isascii() and char.isalnum():
chars.append(char.lower())
return "".join(chars)
_DIARY_DUPLICATE_KEYWORDS = (
"梦", "梦里", "梦见", "学校", "教室", "窗台", "窗边", "窗", "猫", "橘猫", "星图",
"发夹", "书包", "餐桌", "糖", "软糖", "花", "雨", "伞", "走廊", "床", "枕头",
)
def _diary_keyword_overlap(left: Any, right: Any) -> int:
a = _compact_diary_text(left, limit=520)
b = _compact_diary_text(right, limit=520)
if not a or not b:
return 0
return sum(1 for keyword in _DIARY_DUPLICATE_KEYWORDS if keyword in a and keyword in b)
def _diary_text_similarity(left: Any, right: Any) -> float:
a = _compact_diary_text(left)
b = _compact_diary_text(right)
if len(a) < 8 or len(b) < 8:
return 0.0
return SequenceMatcher(None, a, b).ratio()
def _recent_diary_avoid_context(plugin, count: int = 3) -> str:
diaries = plugin.data.get("bot_diaries", [])
if not isinstance(diaries, list) or not diaries:
return "(暂无)"
lines: list[str] = []
for diary in diaries[-count:]:
if not isinstance(diary, dict):
continue
date_text = _single_line(diary.get("date"), 16)
summary = _single_line(diary.get("summary"), 70)
share_seed = _single_line(diary.get("share_seed"), 90)
body = _single_line(diary.get("body"), 120)
fragments = []
for item in diary.get("dream_fragments", []) if isinstance(diary.get("dream_fragments"), list) else []:
if not isinstance(item, dict):
continue
text = _single_line(item.get("text"), 24)
if text:
fragments.append(text)
if len(fragments) >= 4:
break
parts = []
if summary:
parts.append(f"摘要={summary}")
if share_seed:
parts.append(f"分享句={share_seed}")
if body:
parts.append(f"正文片段={body}")
if fragments:
parts.append(f"梦境碎片={','.join(fragments)}")
if parts:
lines.append(f"- {date_text or '近期'}:" + ";".join(parts))
return "\n".join(lines) if lines else "(暂无)"
def _recent_diary_duplicate_hit(plugin, payload: dict[str, Any], count: int = 3) -> tuple[bool, str]:
diaries = plugin.data.get("bot_diaries", [])
if not isinstance(diaries, list) or not diaries:
return False, ""
current_share = _single_line(payload.get("share_seed"), 140)
current_summary = _single_line(payload.get("summary"), 180)
current_body = _single_line(payload.get("body"), 520)
current_all = " ".join(part for part in (current_share, current_summary, current_body) if part)
for diary in reversed(diaries[-count:]):
if not isinstance(diary, dict):
continue
prior_share = _single_line(diary.get("share_seed"), 140)
prior_summary = _single_line(diary.get("summary"), 180)
prior_body = _single_line(diary.get("body"), 520)
prior_all = " ".join(part for part in (prior_share, prior_summary, prior_body) if part)
share_ratio = _diary_text_similarity(current_share, prior_share)
all_ratio = _diary_text_similarity(current_all, prior_all)
cross_ratio = max(
_diary_text_similarity(current_share, prior_summary),
_diary_text_similarity(current_share, prior_body),
_diary_text_similarity(current_summary, prior_share),
)
keyword_overlap = max(
_diary_keyword_overlap(current_share, prior_share),
_diary_keyword_overlap(current_all, prior_all),
)
if share_ratio >= 0.58 or all_ratio >= 0.48 or cross_ratio >= 0.62 or keyword_overlap >= 4:
return True, _single_line(diary.get("date"), 16) or "近期日记"
return False, ""
def _repair_duplicate_daily_diary(plugin, payload: dict[str, Any], matched_date: str) -> dict[str, Any]:
state = plugin.data.get("daily_state", {})
mood = state.get("mood_bias", "平稳") if isinstance(state, dict) else "平稳"
weather = _single_line(plugin._weather_summary_text(plugin.data.get("daily_weather", {})), 48)
note = "今天脑子里还残留着前几天梦里的画面,像醒来后还留着的一点余温。"
if weather and weather != "暂无天气信息":
note += f"外面的{weather}让这种余韵更明显了一点。"
note += f"整个人偏{mood},但已经不太想继续在旧梦里打转,就把注意力慢慢放回今天新的小事。"
repaired = dict(payload)
repaired["summary"] = "今天有一点梦境余韵,但更想把注意力放回新的小事上。"
repaired["body"] = note
repaired["share_seed"] = "今天梦里的余韵还在,不过我想等遇到新的小事再讲给你听"
repaired["tags"] = payload.get("tags") if isinstance(payload.get("tags"), list) else ["平稳"]
return repaired
def normalize_dream_fragment_item(plugin, raw: Any) -> dict[str, Any] | None:
now_ts = _now_ts()
if isinstance(raw, str):
text = _clean_dream_fragment_text(raw)
if not text or not _dream_fragment_is_useful(text):
return None
return {
"text": text,
"weight": 1.0,
"created_ts": now_ts,
"source": "legacy",
}
if not isinstance(raw, dict):
return None
text = _clean_dream_fragment_text(raw.get("text") or raw.get("keyword") or raw.get("label"))
if not text or not _dream_fragment_is_useful(text):
return None
weight = float(_safe_float(raw.get("weight"), 1.0))
created_ts = _safe_float(raw.get("created_ts"), now_ts)
if created_ts <= 0:
created_ts = now_ts
return {
"text": text,
"weight": max(0.2, min(6.0, weight)),
"created_ts": created_ts,
"source": _single_line(raw.get("source"), 20) or "diary",
"date": _single_line(raw.get("date"), 16) or _today_key(),
}
def dream_fragment_effective_weight(plugin, fragment: dict[str, Any], now_ts: float | None = None) -> float:
now_ts = now_ts or _now_ts()
base_weight = max(0.2, min(6.0, _safe_float(fragment.get("weight"), 1.0)))
created_ts = _safe_float(fragment.get("created_ts"), now_ts)
age_hours = max(0.0, (now_ts - created_ts) / 3600.0)
decay = pow(0.72, age_hours / 24.0)
return base_weight * decay
def normalize_dream_fragment_pool(plugin, fragments: Any, *, now_ts: float | None = None) -> list[dict[str, Any]]:
now_ts = now_ts or _now_ts()
if not isinstance(fragments, list):
return []
deduped: dict[str, dict[str, Any]] = {}
fuzzy_seen: set[str] = set()
for raw in fragments:
item = plugin._normalize_dream_fragment_item(raw)
if not item:
continue
text = item["text"]
fuzzy_key = re.sub(r"[^\u4e00-\u9fffA-Za-z0-9]+", "", text).lower()[:36]
if not fuzzy_key or fuzzy_key in fuzzy_seen:
continue
item["effective_weight"] = plugin._dream_fragment_effective_weight(item, now_ts=now_ts)
if item["effective_weight"] < 0.12:
continue
existing = deduped.get(text)
if not existing or item["effective_weight"] > existing.get("effective_weight", 0):
deduped[text] = item
fuzzy_seen.add(fuzzy_key)
ranked = sorted(
deduped.values(),
key=lambda item: (float(item.get("effective_weight", 0)), float(item.get("created_ts", 0))),
reverse=True,
)
for item in ranked:
item.pop("effective_weight", None)
return ranked[:48]
def extract_weighted_dream_fragments(plugin, payload: Any) -> list[dict[str, Any]]:
raw_items = []
if isinstance(payload, dict):
raw_items = payload.get("dream_fragments") or []
if not isinstance(raw_items, list):
raw_items = []
items: list[dict[str, Any]] = []
for raw in raw_items[:12]:
if isinstance(raw, str):
normalized = plugin._normalize_dream_fragment_item({"text": raw, "weight": 1.0, "source": "diary"})
elif isinstance(raw, dict):
normalized = plugin._normalize_dream_fragment_item(
{
"text": raw.get("text") or raw.get("keyword") or raw.get("label"),
"weight": raw.get("weight", 1.0),
"source": raw.get("source") or "diary",
"date": _today_key(),
"created_ts": _now_ts(),
}
)
else:
normalized = None
if normalized:
items.append(normalized)
return items[:8]
def fallback_dream_fragments_for_diary(plugin, state: dict[str, Any]) -> list[dict[str, Any]]:
items: list[dict[str, Any]] = []
seed_candidates = [
_single_line(state.get("dream"), 36),
_single_line(state.get("mood_bias"), 20),
_single_line(plugin._weather_summary_text(plugin.data.get("daily_weather", {})), 36),
]
current_item = plugin._get_current_plan_item(plugin.data.get("daily_plan", {}))
if isinstance(current_item, dict):
seed_candidates.extend(
[
_single_line(current_item.get("activity"), 36),
_single_line(current_item.get("message_seed"), 30),
]
)
seen: set[str] = set()
for index, text in enumerate(seed_candidates):
text = _clean_dream_fragment_text(text)
if not text or text in seen or not _dream_fragment_is_useful(text):
continue
seen.add(text)
items.append(
{
"text": text,
"weight": max(1.0, 2.4 - index * 0.4),
"created_ts": _now_ts(),
"source": "fallback_diary",
"date": _today_key(),
}
)
return items[:5]
def merge_dream_fragment_pool(plugin, new_items: list[dict[str, Any]]) -> list[dict[str, Any]]:
existing = plugin._normalize_dream_fragment_pool(plugin.data.get("dream_fragments", []))
merged = existing + [item for item in new_items if isinstance(item, dict)]
return plugin._normalize_dream_fragment_pool(merged)
def weighted_unique_fragment_sample(plugin, fragments: list[dict[str, Any]], *, count: int) -> list[str]:
if not fragments or count <= 0:
return []
remaining = [dict(item) for item in fragments if isinstance(item, dict)]
picked: list[str] = []
while remaining and len(picked) < count:
weights = [max(0.01, plugin._dream_fragment_effective_weight(item)) for item in remaining]
total = sum(weights)
if total <= 0:
break
chosen = random.choices(remaining, weights=weights, k=1)[0]
text = _single_line(chosen.get("text"), 40)
if text and text not in picked:
picked.append(text)
remaining = [item for item in remaining if _single_line(item.get("text"), 40) != text]
return picked
def build_dream_memory_fragments(plugin, count: int = 8) -> list[str]:
fragment_pool = plugin._normalize_dream_fragment_pool(plugin.data.get("dream_fragments", []))
picked = plugin._weighted_unique_fragment_sample(fragment_pool, count=min(count, 6))
if len(picked) >= count:
return picked[:count]
fragments: list[str] = []
diaries = plugin.data.get("bot_diaries", [])
if isinstance(diaries, list):
for diary in diaries[-4:]:
if not isinstance(diary, dict):
continue
for candidate in (
_single_line(diary.get("share_seed"), 80),
_single_line(diary.get("summary"), 80),
):
if candidate:
cleaned = _clean_dream_fragment_text(candidate)
if cleaned and _dream_fragment_is_useful(cleaned):
fragments.append(cleaned)
current_plan = plugin.data.get("daily_plan", {})
if isinstance(current_plan, dict):
items = current_plan.get("items", [])
if isinstance(items, list):
for item in items[-3:]:
if not isinstance(item, dict):
continue
for candidate in (
_single_line(item.get("activity"), 80),
_single_line(item.get("message_seed"), 80),
):
cleaned = _clean_dream_fragment_text(candidate)
if cleaned and _dream_fragment_is_useful(cleaned):
fragments.append(cleaned)
can_do = plugin.data.get("can_do", [])
if isinstance(can_do, list):
for item in can_do[:4]:
candidate = _single_line(item, 60)
cleaned = _clean_dream_fragment_text(candidate)
if cleaned and _dream_fragment_is_useful(cleaned):
fragments.append(cleaned)
for entry in plugin._get_relevant_important_dates()[:3]:
if not isinstance(entry, dict):
continue
joined = _single_line(
f"{entry.get('title', '')} {entry.get('note', '')}",
80,
)
cleaned = _clean_dream_fragment_text(joined)
if cleaned and _dream_fragment_is_useful(cleaned):
fragments.append(cleaned)
yesterday = plugin.data.get("yesterday_conversation_summary", {})
if isinstance(yesterday, dict) and yesterday.get("date") == _today_key():
for candidate in (
_single_line(yesterday.get("dream_reference"), 100),
_single_line(yesterday.get("summary"), 100),
):
cleaned = _clean_dream_fragment_text(candidate)
if cleaned and "无明确" not in candidate and _dream_fragment_is_useful(cleaned):
fragments.append(cleaned)
residues = yesterday.get("residues", [])
if isinstance(residues, list):
for item in residues[:4]:
if not isinstance(item, dict):
continue
content = _single_line(item.get("content"), 80)
cleaned = _clean_dream_fragment_text(content)
if cleaned and _dream_fragment_is_useful(cleaned):
fragments.append(cleaned)
weather = _single_line(plugin._weather_summary_text(plugin.data.get("daily_weather", {})), 60)
weather_fragment = _clean_dream_fragment_text(weather)
if weather_fragment and _dream_fragment_is_useful(weather_fragment):
fragments.append(weather_fragment)
deduped: list[str] = []
seen: set[str] = set()
for fragment in fragments:
if not fragment or fragment in seen:
continue
seen.add(fragment)
deduped.append(fragment)
random.shuffle(deduped)
for fragment in deduped:
if fragment not in picked:
picked.append(fragment)
if len(picked) >= count:
break
return picked[:count]
def dream_theme_specs(plugin) -> list[tuple[str, str]]:
default_specs = {
"温柔日常": "梦像从白天的普通片段里慢慢渗出来,柔软、安静、带一点生活气。",
"奇幻": "现实里的东西轻轻偏离常理,带一点不合逻辑的发光感或变形感。",
"恐怖": "不是血腥惊吓,而是熟悉场景里多出一点说不清的不安和压迫。",
"追逐": "一直在赶什么、找什么、错过什么,节奏偏紧,醒来会残留一点慌。",
"悬疑": "细节像有答案却总差一点,梦里会反复回头、确认、怀疑。",
"荒诞": "东西会莫名其妙地接到一起,逻辑松掉,带一点好笑又奇怪的偏移。",
"怀旧": "梦会把旧场景、旧物件、旧关系轻轻翻出来,但不一定讲得明白。",
"暧昧春梦": "梦里会有一点亲密、靠近、心跳变快的错觉,但保持含蓄,不写露骨内容。",
}
raw = str(plugin.dream_theme_candidates or "").strip()
names = [name.strip() for name in raw.split(",") if name.strip()]
if not names:
names = list(default_specs.keys())
specs: list[tuple[str, str]] = []
for name in names:
if name == "暧昧春梦" and not plugin.enable_intimate_dream_theme:
continue
specs.append((name, default_specs.get(name, f"梦整体偏{name},但仍然要从具体生活碎片出发,保留一条能读懂的梦内情绪线。")))
if not specs:
specs.append(("温柔日常", default_specs["温柔日常"]))
return specs
async def generate_enhanced_dream_pick(plugin, weather: dict[str, Any] | None = None) -> tuple[str, str, int, int] | None:
fragments = plugin._build_dream_memory_fragments()
if not fragments:
persona_hint = _single_line(plugin._get_default_persona_prompt(), 80)
weather_hint = _single_line(plugin._weather_summary_text(weather or plugin.data.get("daily_weather", {})), 60)
can_do = plugin.data.get("can_do", [])
activity_hint = ""
if isinstance(can_do, list) and can_do:
activity_hint = _single_line(random.choice(can_do), 60)
fragments = [
item
for item in (persona_hint, weather_hint, activity_hint, "醒来后只剩一点断续的画面")
if item and item != "暂无天气信息"
]
dream_themes = plugin._dream_theme_specs()
primary_name, primary_hint = random.choice(dream_themes)
theme_name = primary_name
theme_hint = primary_hint
if plugin.enable_mixed_dream_themes and len(dream_themes) >= 2 and random.random() < 0.35:
alt_name, alt_hint = random.choice([item for item in dream_themes if item[0] != primary_name])
theme_name = f"{primary_name}+{alt_name}"
theme_hint = f"主调偏{primary_name},但中途会混进一点{alt_name}的质感。{primary_hint} 同时,{alt_hint}"
persona = plugin._get_default_persona_prompt()
worldview_adaptation = ""
formatter = getattr(plugin, "_format_worldview_adaptation_prompt", None)
if callable(formatter):
worldview_adaptation = formatter()
weather_text = plugin._weather_summary_text(weather or plugin.data.get("daily_weather", {}))
prompt = f"""
你现在是 Private Companion 的梦境生成器。请根据本次输入的记忆碎片,写一个拟人化 Bot 今早残留的完整梦境。
这个梦可以跳接、荒诞、前后不完全合逻辑,但读起来必须摸得到一条“梦里的情绪线”:她在找什么、躲什么、靠近什么、误认了什么,或为什么醒来后还残留那种感觉。不要只把碎片随机拼贴。
要求:
1. 梦境内容要像把记忆碎片在梦里重新变形,允许断裂和跳场,但必须有“发生了什么”和“为什么醒来后还记得”。
2. 尽量保留一点真实生活残影,不要纯奇幻大场面;如果出现奇幻,也要让它从生活物件、聊天残留或身体感受里长出来。
3. 不要写成日程、日记、设定说明或心理分析。梦里可以有人、物、地点变化,但不要解释得太清楚。
4. 如果主题偏温柔,energy_delta 可以略微为正;如果主题偏压迫/追赶/恐怖,可以略微为负。
5. 如果主题涉及暧昧或春梦,保持含蓄,只写心跳、靠近、错觉感,不要露骨。
6. 如果碎片很少,也要用已有的人格、天气、最近日记补出一个完整梦,不能输出“没有梦”“记不清”“什么都没有”。
7. 梦境不是现实复盘,但要有现实残留:物件、颜色、声音、气味、身体感受、半句话、聊天余味都可以以变形方式出现。
8. 不要让梦境像宏大奇幻设定简介。即使有不现实元素,也要从房间、桌面、手机、路口、雨声、光线、衣物、食物、课本、屏幕等具体生活物里长出来。
9. 梦境可以有突兀转场,但每个转场前后都要能被读者想象到画面。
10. 输出必须是 JSON,不要 Markdown,不要解释,不要在 JSON 外补充任何内容。
11. factors 必须是可感知的小碎片,例如物件、颜色、声音、气味、触感、半句话;不要输出“情绪很好”“今天很累”“日程残留”这类抽象标签。
12. content 至少包含三个连续梦内节点:起始画面、变形/转场、醒前一瞬。可以不讲现实逻辑,但要讲梦内因果。
13. 不要把“资料室、发光、迷路、追逐、水光、草稿纸”等词当作固定模板反复使用;只有输入碎片里真的有相近材料时才用。
只输出 JSON:
{{
"dream_type": "梦境类型,例如温柔日常/奇幻/追逐/悬疑/荒诞/怀旧/混合类型",
"factors": ["梦境因子或碎片,3到8个,可以是物件/颜色/声音/气味/半句话/动作"],
"content": "180到600字的梦境内容,写成完整一段梦;要有起始画面、变形/转场、醒前一瞬和清楚的梦内情绪线",
"afterglow": "醒来后的梦境余韵,20到120字,说明身体或情绪残留",
"label": "20到50字的短标签,概括这个梦留在身上的感觉",
"mood": "平稳/恍惚/柔和/低落/敏感/轻快 之一",
"energy_delta": -12到6之间的整数",
"duration_hours": 3到8之间的整数
}}
【本次输入】
【人格参考】
{persona}
{worldview_adaptation}
【梦境主题】
{theme_name}:{theme_hint}
【碎片记忆】
{chr(10).join(f"- {item}" for item in fragments)}
【天气】
{weather_text}
""".strip()
raw_text = await plugin._llm_call(
prompt,
max_tokens=1050,
task="dream",
provider_id=plugin._task_provider(
getattr(plugin, "dream_provider_id", ""),
getattr(plugin, "diary_provider_id", ""),
getattr(plugin, "mai_style_provider_id", ""),
),
)
payload = plugin._extract_json_payload(raw_text or "")
if not isinstance(payload, dict):
return None
content = _single_line(payload.get("content"), 900)
factors_raw = payload.get("factors")
factors = []
if isinstance(factors_raw, list):
factors = [_single_line(item, 30) for item in factors_raw[:8] if _single_line(item, 30)]
label = _single_line(payload.get("label"), 80)
if not label and content:
label = _single_line(content, 80)
if not label:
return None
mood = _single_line(payload.get("mood"), 12) or "恍惚"
energy_delta = _safe_int(payload.get("energy_delta"), -6, -12, 6)
duration_hours = _safe_int(payload.get("duration_hours"), 5, 3, 8)
if not content:
content = f"梦里只剩下一段很断续的画面:{label}"
plugin._last_generated_dream_payload = {
"dream_type": _single_line(payload.get("dream_type"), 40) or theme_name,
"factors": factors or fragments[:8],
"content": content,
"afterglow": _single_line(payload.get("afterglow"), 180) or label,
"label": label,
"mood": mood,
"energy_delta": energy_delta,
"duration_hours": duration_hours,
"raw": raw_text or "",
}
return label, mood, energy_delta, duration_hours
def _diary_entry_is_today(plugin, value: Any) -> bool:
today = _today_key()
if isinstance(value, (int, float)):
try:
stamp = float(value)
if stamp > 100_000_000_000:
stamp /= 1000.0
return datetime.fromtimestamp(stamp).strftime("%Y-%m-%d") == today
except Exception:
return False
return str(value or "").strip().startswith(today)
def _daily_diary_evidence_ledger(plugin) -> tuple[str, list[dict[str, str]]]:
data = plugin.data if isinstance(getattr(plugin, "data", None), dict) else {}
evidence: list[dict[str, str]] = []
def add(level: str, source: str, text: Any) -> None:
cleaned = _single_line(text, 180)
if not cleaned or any(item["text"] == cleaned for item in evidence):
return
evidence.append({"level": level, "source": source, "text": cleaned})
adjustments = data.get("schedule_adjustments") if isinstance(data.get("schedule_adjustments"), list) else []
for item in adjustments[-12:]:
if not isinstance(item, dict):
continue
stamp = item.get("created_at") or item.get("updated_at") or item.get("ts") or item.get("date")
if not stamp or not _diary_entry_is_today(plugin, stamp):
continue
text = item.get("summary") or item.get("reason") or item.get("adjustment") or item.get("text")
add("planned", "用户调整后的计划", text)
audits = data.get("proactive_audit_log") if isinstance(data.get("proactive_audit_log"), list) else []
for item in audits[-30:]:
if not isinstance(item, dict) or str(item.get("status") or "").lower() not in {"sent", "success", "completed"}:
continue
stamp = item.get("updated_at") or item.get("created_at") or item.get("ts") or item.get("sent_at")
if not stamp or not _diary_entry_is_today(plugin, stamp):
continue
text = item.get("final_text_preview") or item.get("text_preview") or item.get("topic") or item.get("note")
add("confirmed", "已执行主动", text)
for state_key, label in (("web_exploration", "主动搜索"), ("news_integration", "新闻阅读")):
source_state = data.get(state_key) if isinstance(data.get(state_key), dict) else {}
stamp = source_state.get("last_explore_at") or source_state.get("last_read_at") or source_state.get("updated_at")
digest = source_state.get("last_digest") if isinstance(source_state.get("last_digest"), dict) else {}
if stamp and _diary_entry_is_today(plugin, stamp):
add("confirmed", label, digest.get("topic") or digest.get("headline") or digest.get("note"))
for method_name, label in (
("_self_timeline_from_creative", "实际创作记录"),
("_self_timeline_from_private_reading", "实际阅读记录"),
("_self_timeline_from_photo_generation", "实际生图记录"),
("_self_timeline_from_qzone_publish", "实际空间发布"),
):
collector = getattr(plugin, method_name, None)
if not callable(collector):
continue
try:
entries = collector(data)
except Exception:
continue
for entry in entries[-6:] if isinstance(entries, list) else []:
if not isinstance(entry, dict) or not _diary_entry_is_today(plugin, entry.get("ts") or entry.get("date")):
continue
summary = _single_line(entry.get("summary"), 100)
detail = _single_line(entry.get("detail"), 140)
if "tid:" in detail:
detail = detail.split("tid:", 1)[0].rstrip(";; ")
add("confirmed", label, ";".join(part for part in (summary, detail) if part))
goals = data.get("personal_goals") if isinstance(data.get("personal_goals"), list) else []
for goal in goals[-8:]:
if not isinstance(goal, dict):
continue
title = _single_line(goal.get("title") or goal.get("name"), 60)
logs = goal.get("recent_logs") if isinstance(goal.get("recent_logs"), list) else []
for log in logs[-3:]:
if not isinstance(log, dict) or not _diary_entry_is_today(plugin, log.get("ts")):
continue
evidence_text = _single_line(log.get("evidence"), 100)
progress = _safe_int(log.get("progress"), -1, -1, 100)
suffix = f",进度 {progress}%" if progress >= 0 else ""
add("simulated", "个人目标运行记录", f"{title or '个人目标'}:{evidence_text}{suffix}")
enhanced = data.get("detail_enhanced_segments") if isinstance(data.get("detail_enhanced_segments"), dict) else {}
enhanced_day = str(data.get("detail_enhanced_day") or "")[:10]
for segment_key, snapshot in list(enhanced.items())[-8:]:
if not isinstance(snapshot, dict) or snapshot.get("status") != "done":
continue
if enhanced_day != _today_key() and not str(segment_key).startswith(_today_key()):
continue
add("simulated", "运行细化", snapshot.get("summary"))
for item in (snapshot.get("today_events") if isinstance(snapshot.get("today_events"), list) else [])[:2]:
if isinstance(item, dict):
add("simulated", "运行细化", item.get("event"))
plan = data.get("daily_plan") if isinstance(data.get("daily_plan"), dict) else {}
if str(plan.get("date") or "")[:10] != _today_key():
plan = {}
now = plugin._environment_now() if hasattr(plugin, "_environment_now") else datetime.now()
now_minutes = now.hour * 60 + now.minute
for item in plan.get("items", []) if isinstance(plan.get("items"), list) else []:
if not isinstance(item, dict):
continue
raw_end = str(item.get("end") or item.get("time") or "")
match = re.match(r"^(\d{1,2}):(\d{2})", raw_end)
if match and int(match.group(1)) * 60 + int(match.group(2)) <= now_minutes:
add("planned", "已过去的计划段", item.get("activity"))
if not evidence:
state = data.get("daily_state") if isinstance(data.get("daily_state"), dict) else {}
add("state", "当前状态", state.get("mood_bias") or state.get("summary") or "只留下了很少的具体记录")
level_labels = {
"confirmed": "已确认发生",
"simulated": "运行推演,不能直接声称真实发生",
"planned": "原计划,不能直接声称已经完成",
"state": "状态底色,不是事件",
}
lines = [f"- [{level_labels.get(item['level'], item['level'])}] {item['source']}:{item['text']}" for item in evidence[:16]]
return "\n".join(lines), evidence[:16]
def _daily_diary_form_instruction(plugin, evidence: list[dict[str, str]]) -> tuple[str, str]:
configured = str(getattr(plugin, "daily_diary_form", "auto") or "auto").strip().lower()
forms = {
"scene": "场景短记:围绕一个确有依据的场景写清当时的动作和注意力变化。",
"fragments": "碎片手记:允许两到四个短段或断句,不强求完整起承转合,但彼此要有同一天的气息。",
"inner_voice": "心绪自述:从一个真实触发点写内心反应,不写空泛情绪总结。",
"observation": "观察记录:抓住一个具体对象、声音、文字或细节,少解释,多保留当时的目光。",
}
if configured not in forms:
choices = ["scene", "fragments", "inner_voice", "observation"]
confirmed = sum(1 for item in evidence if item.get("level") == "confirmed")
seed = sum(ord(char) for char in _today_key()) + confirmed
configured = choices[seed % len(choices)]
return configured, forms[configured]
def _daily_diary_length_instruction(plugin) -> tuple[int, int]:
mode = str(getattr(plugin, "daily_diary_length", "standard") or "standard").strip().lower()
return {"short": (60, 130), "long": (180, 360)}.get(mode, (110, 240))
def _daily_diary_creativity_instruction(plugin) -> str:
mode = str(getattr(plugin, "daily_diary_creativity", "balanced") or "balanced").strip().lower()
if mode == "strict":
return "严格写实:只写已确认发生的事实;材料不足就写短,不补场景。"
if mode == "expressive":
return "表达可以更有个人色彩和节奏,但只能放大感受与观察,不能虚构人物、事件或完成结果。"
return "写实为主:允许对已确认事实做轻微感官化表达,不得把计划或运行推演写成真实经历。"
def _daily_diary_quality_issues(
plugin,
payload: Any,
evidence: list[dict[str, str]],
min_chars: int,
max_chars: int,
) -> list[str]:
if not isinstance(payload, dict):
return ["没有返回 JSON 对象"]
summary = _single_line(payload.get("summary"), 180)
body = _single_line(payload.get("body"), 800)
issues: list[str] = []
if not summary or not body:
issues.append("摘要或正文为空")
return issues
body_len = len(re.sub(r"\s+", "", body))
if body_len < max(28, min_chars // 2):
issues.append("正文过短且没有形成有效记录")
if body_len > max_chars + 120:
issues.append("正文明显超出所选篇幅")
internal_markers = ("系统", "模型", "提示词", "JSON", "运行推演", "状态数值", "主动消息", "日程字段", "evidence")
if any(marker in f"{summary} {body}" for marker in internal_markers):
issues.append("正文泄露后台术语")
if _diary_reads_like_status_broadcast(payload):
issues.append("正文像状态播报而不是私人日记")
unsupported = [item for item in evidence if item.get("level") in {"planned", "simulated"}]
completion_markers = ("完成了", "做完了", "已经做", "去了", "看完了", "写完了", "收拾好了", "结束了")
if unsupported and any(marker in body for marker in completion_markers):
compact_body = _compact_diary_text(body, 800)
for item in unsupported:
compact_evidence = _compact_diary_text(item.get("text"), 180)
trigrams = {compact_evidence[index : index + 3] for index in range(max(0, len(compact_evidence) - 2))}
if any(token in compact_body for token in trigrams):
issues.append("把未确认计划或推演写成了已完成经历")
break
compact_summary = _compact_diary_text(summary, 180)
compact_body = _compact_diary_text(body, 800)
summary_bigrams = {compact_summary[index : index + 2] for index in range(max(0, len(compact_summary) - 1))}
if len(compact_summary) >= 6 and not any(token in compact_body for token in summary_bigrams):
issues.append("摘要没有落在正文内容上")
duplicate_hit, _ = _recent_diary_duplicate_hit(plugin, payload)
if duplicate_hit:
issues.append("与近期日记过于相似")
return issues
async def _rewrite_daily_diary_once(
plugin,
payload: Any,
issues: list[str],
evidence_text: str,
form_instruction: str,
min_chars: int,
max_chars: int,
) -> dict[str, Any]:
current = payload if isinstance(payload, dict) else {}
prompt = f"""
请修订下面这篇私人日记,只修一次。保留有事实依据的部分,删除虚构和模板化补景。
问题:{';'.join(issues)}
写作方式:{form_instruction}
篇幅:{min_chars}-{max_chars} 个中文字符左右;材料不足可以更短。
今日经历账本:
{evidence_text}
原稿:
摘要:{_single_line(current.get('summary'), 180)}
正文:{_single_line(current.get('body'), 800)}
只有“已确认发生”可作为真实经历。运行推演和原计划不能改写成已经完成。
不要补桌面、窗光、茶水、便签等无来源场景。实在没有材料,就坦白今天具体记录很少。
只输出 JSON:{{"summary":"15-55字题眼","body":"日记正文","tags":["1-4个正文标签"]}}
""".strip()
try:
raw = await plugin._llm_call(
prompt,
max_tokens=520,
task="diary_rewrite",
provider_id=plugin._task_provider(
getattr(plugin, "diary_provider_id", ""),
getattr(plugin, "mai_style_provider_id", ""),
),
)
parsed = plugin._extract_json_payload(raw or "")
return parsed if isinstance(parsed, dict) else {}
except Exception:
return {}
async def _extract_daily_diary_derivatives(plugin, payload: dict[str, Any]) -> dict[str, Any]:
body = _single_line(payload.get("body"), 700)
if not body:
return {}
share_enabled = bool(getattr(plugin, "daily_diary_generate_share_seed", True))
prompt = f"""
请从这篇已经写好的私人日记中提取后台结构,不要改写日记正文,也不要新增事件。
日记:
{body}
只输出 JSON:
{{
"share_seed": "{('从正文真实出现的细节延伸出一句自然分享;不适合分享则留空' if share_enabled else '必须留空')}",
"dream_fragments": [{{"text": "正文里真实出现的物件/声音/动作/颜色/半句话", "weight": 0.6}}],
"continuity_thread": {{"motif": "值得跨日延续的具体线索,没有则留空", "status": "出现/变化/淡出", "next_hint": "以后只在自然有依据时承接"}},
"long_term_events": [{{"title": "正文里确实未完成且可能跨日的事项", "status": "当前状态", "next_hint": "下一步"}}]
}}
要求:dream_fragments 0–6 个,只提取正文确实存在的碎片,不足时留空;long_term_events 0–2 个。不要生成主动计划、今日事件或不存在的后续剧情。
""".strip()
try:
raw = await plugin._llm_call(
prompt,
max_tokens=320,
task="diary_derivatives",
provider_id=plugin._task_provider(getattr(plugin, "diary_provider_id", ""), getattr(plugin, "mai_style_provider_id", "")),
)
parsed = plugin._extract_json_payload(raw or "")
return parsed if isinstance(parsed, dict) else {}
except Exception:
return {}
async def generate_daily_diary(plugin) -> dict[str, Any]:
today = _today_key()
state = plugin.data.get("daily_state", {})
persona = plugin._get_default_persona_prompt()
schedule_persona = _single_line(getattr(plugin, "schedule_persona_prompt", ""), 1200)
schedule_worldview = _single_line(getattr(plugin, "schedule_worldview_prompt", ""), 1200)
calendar_context = plugin._format_calendar_context_for_prompt()
recent_diary_avoid_context = _recent_diary_avoid_context(plugin)
evidence_text, evidence = _daily_diary_evidence_ledger(plugin)
diary_form, form_instruction = _daily_diary_form_instruction(plugin, evidence)
min_chars, max_chars = _daily_diary_length_instruction(plugin)
creativity_instruction = _daily_diary_creativity_instruction(plugin)
custom_direction = _single_line(getattr(plugin, "daily_diary_custom_direction", ""), 500)
worldview_adaptation = ""
formatter = getattr(plugin, "_format_worldview_adaptation_prompt", None)
if callable(formatter):
worldview_adaptation = formatter()
prompt = f"""
请以当前人格的第一人称,写一篇今天的私人日记。只写日记,不安排主动消息、梦境素材或后续剧情。
写作方式:{form_instruction}
篇幅:{min_chars}–{max_chars} 个中文字符左右。
事实边界:{creativity_instruction}
{f'用户指定方向:{custom_direction}' if custom_direction else ''}
规则:
1. “已确认发生”可以写成经历;“运行推演、原计划、状态底色”只能影响语气或成为未确认的念头,绝不能写成已经发生。
2. 材料少就写短,允许今天没有戏剧性;禁止用桌面、窗光、凉茶、旧便签等通用小物件自行补场景。
3. 不固定“场景→发现→余韵”的三段式,不总结人生,不把普通小事拔高成道理。
4. 保持当前人格的词汇、观察角度和关系边界。不要写系统、模型、状态数值、日程字段或后台功能。
5. 最近日记只用于承接或避重;没有新变化时让旧线索淡出,不强行续篇。
只输出 JSON:
{{
"summary": "正文中最具体的一幕或题眼,15–55字",
"body": "私人日记正文",
"tags": ["正文确实体现的1–4个短标签"]
}}
【本次输入】
日期:{today}
【AstrBot 默认人格】
{persona}
【生活身份补充】
{schedule_persona or "(无)"}
【生活/世界观补充】
{schedule_worldview or "(无)"}
{worldview_adaptation}
日期语境:
{calendar_context}
【今日经历账本】
{evidence_text}
最近日记:
{plugin._recent_diary_context()}
近期需要避免复用的具体素材:
{recent_diary_avoid_context}
近期重要日期:
{plugin._format_important_dates_for_prompt()}
""".strip()
try:
raw_text = await plugin._llm_call(
prompt,
max_tokens=620,
task="diary",
provider_id=plugin._task_provider(
getattr(plugin, "diary_provider_id", ""),
getattr(plugin, "mai_style_provider_id", ""),
),
)
except Exception:
raw_text = ""
payload = plugin._extract_json_payload(raw_text or "")
used_fallback = False
quality_issues = _daily_diary_quality_issues(plugin, payload, evidence, min_chars, max_chars)
if quality_issues and isinstance(payload, dict):
payload = await _rewrite_daily_diary_once(
plugin, payload, quality_issues, evidence_text, form_instruction, min_chars, max_chars
)
quality_issues = _daily_diary_quality_issues(plugin, payload, evidence, min_chars, max_chars)
if quality_issues:
payload = plugin._fallback_diary_payload(evidence=evidence)
used_fallback = True
polisher = getattr(plugin, "_polish_diary_payload", None)
if callable(polisher):
payload = polisher(payload)
tags = payload.get("tags", [])
if not isinstance(tags, list):
tags = []
derivatives = {} if used_fallback else await _extract_daily_diary_derivatives(plugin, payload)
if not isinstance(derivatives, dict):
derivatives = {}
share_seed = _single_line(derivatives.get("share_seed"), 120) if getattr(plugin, "daily_diary_generate_share_seed", True) else ""
continuity_thread = derivatives.get("continuity_thread") if isinstance(derivatives.get("continuity_thread"), dict) else {}
derivative_payload = {
"dream_fragments": derivatives.get("dream_fragments", []),
"long_term_events": derivatives.get("long_term_events", []),
"today_events": [],
"proactive_events": [],
}
return {
"date": today,
"generated_at": datetime.now().strftime("%Y-%m-%d %H:%M"),
"summary": _single_line(payload.get("summary"), 160),
"body": _single_line(payload.get("body"), 500),
"share_seed": share_seed,
"tags": [_single_line(tag, 20) for tag in tags[:6] if _single_line(tag, 20)],
"diary_form": diary_form,
"evidence": evidence,
"continuity_thread": {
"motif": _single_line(continuity_thread.get("motif"), 80),
"status": _single_line(continuity_thread.get("status"), 20),
"next_hint": _single_line(continuity_thread.get("next_hint"), 100),
},
"dream_fragments": plugin._extract_weighted_dream_fragments(derivative_payload),
"story_plan": plugin._normalize_story_plan(derivative_payload),
"raw": raw_text or "",
}
def fallback_diary_payload(plugin, evidence: list[dict[str, str]] | None = None) -> dict[str, Any]:
state = plugin.data.get("daily_state", {})
energy = state.get("energy", 70) if isinstance(state, dict) else 70
tags = ["平稳"]
if _safe_int(energy, 70) < 45:
tags.append("低能量")
for key, tag in (("sleep", "失眠"), ("health", "生病"), ("dream", "好梦")):
value = str(state.get(key, "")) if isinstance(state, dict) else ""
if tag == "好梦" and "梦见" in value:
tags.append(tag)
elif tag != "好梦" and ("失眠" in value or "低烧" in value or "头重" in value):
tags.append(tag)
conditions = state.get("conditions", []) if isinstance(state, dict) else []
if isinstance(conditions, list):
phases = {str(cond.get("phase") or "") for cond in conditions if isinstance(cond, dict)}
kinds = {str(cond.get("kind") or "") for cond in conditions if isinstance(cond, dict)}
if "afterglow" in phases or {"recovery_afterglow", "sleep_afterglow", "soft_afterglow"} & kinds:
tags.append("回弹")
if "tail" in phases or {"health_tail", "sleep_tail"} & kinds:
tags.append("恢复期")
usable = [item for item in (evidence or []) if isinstance(item, dict) and _single_line(item.get("text"), 180)]
confirmed = next((item for item in usable if item.get("level") == "confirmed"), None)
uncertain = next((item for item in usable if item.get("level") in {"planned", "simulated"}), None)
if confirmed:
fact = _single_line(confirmed.get("text"), 150)
summary = fact
body = f"今天能确定留下来的记录是:{fact}。除此之外没有足够具体的细节,就先记到这里。"
elif uncertain:
fact = _single_line(uncertain.get("text"), 150)
summary = "今天只留下一条尚未确认的线索"
body = f"今天原本记着:{fact}。后来是否照计划发生,我这里没有足够记录,所以不把它写成已经做过的事。"
else:
summary = "今天留下的具体记录不多"
body = "今天没有留下足够具体、可以确认的经历。与其补出一个看似自然的小场景,不如先如实记到这里。"
return {
"summary": summary,
"body": body,
"share_seed": "",
"tags": tags,
"today_events": [],
"proactive_events": [],
"dream_fragments": [],
"long_term_events": [],
}