feat: give the curator continuous context and durable idea history

README.md

@@ -3,21 +3,23 @@
 一个“先保存，后理解”的私人 AI 笔记本。
 
 记录页只呈现用户写下的原文，输入历史像对话一样始终可回看。后台的单一策展
-Agent 使用 DeepSeek 的思考模式，按需搜索已有想法与轨迹，再给出结构化的“想法位势”
+Agent 使用 DeepSeek 的思考模式，结合最近的连续原文与精简想法目录，必要时再查看
-提案。Agent 不能直接写数据库；应用只在 schema 和引用完整性校验通过后，用事务提交
+已有想法的完整轨迹，最后给出结构化的“想法位势”提案。Agent 不能直接写数据库；
-更新。
+应用只在 schema 和引用完整性校验通过后，用事务提交更新。
 
 ## 设计边界
 
 - 原始片段先持久化，AI 失败不影响记录。
 - 捕获流不显示保存状态、AI 标签、关联或建议。
 - 不要求用户先建页面、取标题、选分类或填写日期。
+- 不要求用户建立会话；系统在后台保留滚动的连续语境。
 - 成熟度是可回退的连续位势，不是阶段、成绩或任务完成百分比。
 - 运动、张力和可能动作由模型结合上下文动态生成，不使用固定关卡。
 - 用户手动校准的位势优先展示，AI 估计仍独立保留。
 - 每个用户的记录、想法、Session 与 Agent 上下文都以 `user_id` 在 SQL 层隔离。
 - 管理员能看运行元数据；其他用户的原文与 AI 产物默认脱敏，只有用户主动开启
   “调试共享”后才可见。
+- 每次想法更新和人工校准都会保存完整版本；原始片段始终是不可替代的证据层。
 - 管理后台保存模型轮次、工具调用、耗时、token、错误和结构化决策产物，但不保存或
   暴露模型隐藏思维链。
 
@@ -33,11 +35,14 @@ React/Vite ── cookie session ── FastAPI ── SQLite
 核心数据流是：
 
 1. `POST /api/fragments` 先提交原文并立即返回；
-2. 后台 Agent 读取当前用户自己的想法上下文；
+2. 后台 Agent 读取当前用户最近的连续输入和精简想法目录；
-3. Agent 通过 `search_ideas`、`inspect_idea` 工具选择相关材料；
+3. Agent 通过 `search_ideas`、`inspect_idea` 工具补充相关材料；
 4. 应用校验结构化结果后，以事务写入想法、轨迹和片段关联；
 5. 每次运行同时形成 `agent_runs`、`agent_events`，供管理后台聚合分析。
 
+工具探索超过预算或输出结构无效时，系统会进入一次没有工具的收敛回合；失败运行也可以由
+管理员重新排队。模型用量通过逐回合钩子采集，因此异常退出不会被误记为零。
+
 ## 本地运行
 
 要求 Python 3.12+、Node.js 22+。
@@ -112,6 +117,7 @@ Session 会归属给配置清单中的第一位管理员；不会把旧数据复
 - `GET /api/admin/users`：用户空间、数据量与调试共享状态；
 - `GET /api/admin/runs?limit=80`：Agent 运行列表；
 - `GET /api/admin/runs/{run_id}`：一次运行的事件、工具、token 与结构化产物；
+- `POST /api/admin/runs/{run_id}/retry`：重新排队一次仍处于错误状态的分析；
 - `GET /api/admin/audit?limit=120`：登录、启动、记录、校准等应用审计事件。
 
 普通用户可调用 `PATCH /api/account/debug-sharing` 控制自己的调试内容是否向管理员

app/curator.py

@@ -3,19 +3,23 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
+import re
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from typing import Any
 
 from agents import (
     Agent,
     ModelSettings,
     OpenAIChatCompletionsModel,
+    RunHooks,
     RunContextWrapper,
     Runner,
     function_tool,
     set_tracing_disabled,
 )
+from agents.exceptions import MaxTurnsExceeded
+from agents.items import ModelResponse
 from openai import AsyncOpenAI
 from openai.types.shared import Reasoning
 from pydantic import ValidationError
@@ -25,7 +29,9 @@ from .db import Database
 from .schemas import CuratorDecision
 
 logger = logging.getLogger(__name__)
-PROMPT_VERSION = "maturity-2026-07-28.v2"
+PROMPT_VERSION = "maturity-2026-07-28.v3"
+RECENT_CONTEXT_LIMIT = 12
+IDEA_CATALOG_LIMIT = 40
 
 
 CURATOR_INSTRUCTIONS = """
@@ -52,8 +58,13 @@ motion 是你对当下运动的自然语言压缩，例如“向现实探去”
 一个片段最多关联两个想法。若现在还不值得形成或归入想法，让它 standalone，原文仍会保留。
 新建想法时 idea_id 必须为 null；更新时必须使用工具返回的准确 id。
 
-先查看候选想法；遇到可能相关或可能冲突的候选时，使用 inspect_idea。最终必须只返回 JSON，
+输入会直接提供最近的连续原文和精简想法目录。把“这个、上面、第二点、继续、刚才”等指代
-不能用 Markdown 包裹，也不能附加解释。JSON 必须符合输入中给出的 schema。
+放回连续语境中理解；不要因为新片段换了关键词就丢掉思想上的延续。目录中的 id 只供系统内部
+引用。遇到可能相关或冲突的候选时，使用 inspect_idea；search_ideas 只用于目录过大或需要
+补充候选时，最多调用两次；同一个想法最多 inspect 一次。搜索无结果时不要反复改写查询，
+也绝不能用 none、unknown 等占位符调用 inspect_idea。工具探索后必须及时收敛并返回最终判断。
+
+最终必须只返回 JSON，不能用 Markdown 包裹，也不能附加解释。JSON 必须符合输入中给出的 schema。
 不要在输出中评价用户、诊断心理、说教或伪造证据。
 """.strip()
 
@@ -67,6 +78,187 @@ class CuratorContext:
     recent_fragments: list[dict[str, Any]]
     search_count: int = 0
     inspection_count: int = 0
+    model_rounds: int = 0
+    input_tokens: int = 0
+    output_tokens: int = 0
+    reasoning_tokens: int = 0
+    cached_tokens: int = 0
+    llm_started_at: float | None = None
+    inspected_idea_ids: set[str] = field(default_factory=set)
+
+
+def _search_units(value: str) -> list[str]:
+    units: list[str] = []
+    for token in re.findall(r"[a-zA-Z0-9_-]+|[\u3400-\u9fff]+", value.lower()):
+        if token not in units:
+            units.append(token)
+        if re.fullmatch(r"[\u3400-\u9fff]+", token) and len(token) > 2:
+            for index in range(len(token) - 1):
+                pair = token[index : index + 2]
+                if pair not in units:
+                    units.append(pair)
+    return units[:24]
+
+
+def _rank_ideas(
+    ideas: list[dict[str, Any]], query: str
+) -> tuple[list[dict[str, Any]], bool]:
+    terms = _search_units(query)
+    ranked: list[tuple[int, str, dict[str, Any]]] = []
+    for idea in ideas:
+        evidence = " ".join(
+            str(fragment.get("content", ""))
+            for fragment in idea.get("recent_fragments", [])
+        )
+        haystack = " ".join(
+            str(idea.get(key, ""))
+            for key in (
+                "title",
+                "summary",
+                "position",
+                "tension",
+                "trajectory",
+            )
+        )
+        haystack = f"{haystack} {evidence}".lower()
+        score = sum(
+            (4 if term in str(idea.get("title", "")).lower() else 1)
+            * haystack.count(term)
+            for term in terms
+        )
+        if score:
+            ranked.append((score, str(idea.get("updated_at", "")), idea))
+    fallback = not ranked
+    if fallback:
+        ranked = [
+            (0, str(idea.get("updated_at", "")), idea) for idea in ideas
+        ]
+    ranked.sort(key=lambda item: (item[0], item[1]), reverse=True)
+    return [idea for _, _, idea in ranked[:8]], fallback
+
+
+class CuratorRunHooks(RunHooks[CuratorContext]):
+    async def on_llm_start(
+        self,
+        context: RunContextWrapper[CuratorContext],
+        agent: Agent[CuratorContext],
+        system_prompt: str | None,
+        input_items: list[Any],
+    ) -> None:
+        context.context.llm_started_at = time.perf_counter()
+
+    async def on_llm_end(
+        self,
+        context: RunContextWrapper[CuratorContext],
+        agent: Agent[CuratorContext],
+        response: ModelResponse,
+    ) -> None:
+        state = context.context
+        state.model_rounds += 1
+        usage = response.usage
+        input_tokens = int(usage.input_tokens or 0)
+        output_tokens = int(usage.output_tokens or 0)
+        reasoning_tokens = int(
+            usage.output_tokens_details.reasoning_tokens or 0
+        )
+        cached_tokens = int(
+            usage.input_tokens_details.cached_tokens or 0
+        )
+        state.input_tokens += input_tokens
+        state.output_tokens += output_tokens
+        state.reasoning_tokens += reasoning_tokens
+        state.cached_tokens += cached_tokens
+        duration_ms = (
+            round((time.perf_counter() - state.llm_started_at) * 1000)
+            if state.llm_started_at is not None
+            else None
+        )
+        state.db.add_agent_event(
+            state.run_id,
+            state.user_id,
+            "model_round_completed",
+            {
+                "round": state.model_rounds,
+                "agent": agent.name,
+                "input_tokens": input_tokens,
+                "output_tokens": output_tokens,
+                "reasoning_tokens": reasoning_tokens,
+                "cached_tokens": cached_tokens,
+            },
+            duration_ms=duration_ms,
+        )
+        state.llm_started_at = None
+
+
+def build_curator_prompt(
+    fragment: dict[str, Any],
+    ideas: list[dict[str, Any]],
+    recent_fragments: list[dict[str, Any]],
+) -> str:
+    prior_context = [
+        {
+            "id": item["id"],
+            "created_at": item["created_at"],
+            "content": str(item["content"])[:1_200],
+        }
+        for item in recent_fragments
+        if item["id"] != fragment["id"]
+        and item["created_at"] <= fragment["created_at"]
+    ][-RECENT_CONTEXT_LIMIT:]
+    later_context = [
+        {
+            "created_at": item["created_at"],
+            "content": str(item["content"])[:1_200],
+        }
+        for item in recent_fragments
+        if item["id"] != fragment["id"]
+        and item["created_at"] > fragment["created_at"]
+    ][-6:]
+    catalog = [
+        {
+            "id": idea["id"],
+            "title": idea["title"],
+            "summary": idea["summary"],
+            "motion": idea["motion"],
+            "position": idea["position"],
+            "tension": idea["tension"],
+            "updated_at": idea["updated_at"],
+        }
+        for idea in ideas[:IDEA_CATALOG_LIMIT]
+    ]
+    new_fragment = {
+        "id": fragment["id"],
+        "created_at": fragment["created_at"],
+        "content": fragment["content"],
+    }
+    output_schema = CuratorDecision.model_json_schema()
+    catalog_note = (
+        "目录已经覆盖全部已有想法。优先从目录直接判断候选，相关时用准确 id 调用 "
+        "inspect_idea；一般不需要 search_ideas。"
+        if len(ideas) <= IDEA_CATALOG_LIMIT
+        else "目录只包含最近更新的想法；找不到可能候选时可调用 search_ideas。"
+    )
+    delayed_note = (
+        "这是一次延迟重整。later_context 是这条记录之后才发生的输入，只用于避免把现有想法"
+        "的状态和轨迹倒退；不要把它们误当成当时的上文。应把当前片段作为迟到的证据，结合"
+        "想法目录中的最新状态作出增量判断。\n\n"
+        if later_context
+        else ""
+    )
+    return (
+        "请策展这个刚刚保存的新片段。最近原文按时间从早到晚排列，它们是理解“这个、"
+        "第二点、继续、刚才”等连续表达的第一依据；不要要求用户显式建立会话。\n\n"
+        f"<recent_context>{json.dumps(prior_context, ensure_ascii=False)}</recent_context>\n\n"
+        f"<new_fragment>{json.dumps(new_fragment, ensure_ascii=False)}</new_fragment>\n\n"
+        f"<later_context>{json.dumps(later_context, ensure_ascii=False)}</later_context>\n\n"
+        f"{delayed_note}"
+        f"<idea_catalog>{json.dumps(catalog, ensure_ascii=False)}</idea_catalog>\n\n"
+        f"当前共有 {len(ideas)} 个已有想法。{catalog_note}\n"
+        "如果新片段既验证了一个既有想法、又形成了一个新的独立问题，可以给出两个 assessment；"
+        "不要因为表面主题变化而遗漏它对上一段思想过程的反馈。\n\n"
+        "最终只输出符合以下 JSON Schema 的 JSON 对象：\n"
+        f"{json.dumps(output_schema, ensure_ascii=False)}"
+    )
 
 
 @function_tool(strict_mode=False)
@@ -76,17 +268,31 @@ async def search_ideas(
     """Search existing evolving ideas by a short concept, question, or tension."""
     ctx.context.search_count += 1
     started = time.perf_counter()
-    terms = [term.lower() for term in query.split() if term.strip()]
+    if ctx.context.search_count > 2:
-    ranked: list[tuple[int, dict[str, Any]]] = []
+        candidates = [
-    for idea in ctx.context.ideas:
+            {"id": item["id"], "title": item["title"]}
-        haystack = " ".join(
+            for item in ctx.context.ideas[:8]
-            str(idea.get(key, ""))
+        ]
-            for key in ("title", "summary", "position", "tension", "trajectory")
+        ctx.context.db.add_agent_event(
-        ).lower()
+            ctx.context.run_id,
-        score = sum(haystack.count(term) for term in terms)
+            ctx.context.user_id,
-        if score or not terms:
+            "tool_search_ideas",
-            ranked.append((score, idea))
+            {
-    ranked.sort(key=lambda item: (item[0], item[1].get("updated_at", "")), reverse=True)
+                "query": query,
+                "result_count": len(candidates),
+                "budget_exhausted": True,
+                "candidates": candidates,
+            },
+            duration_ms=round((time.perf_counter() - started) * 1000),
+        )
+        return json.dumps(
+            {
+                "notice": "搜索额度已用完。使用已有目录并立即形成最终判断。",
+                "candidates": candidates,
+            },
+            ensure_ascii=False,
+        )
+    ranked, fallback = _rank_ideas(ctx.context.ideas, query)
     compact = [
         {
             "id": idea["id"],
@@ -97,7 +303,7 @@ async def search_ideas(
             "position": idea["position"],
             "tension": idea["tension"],
         }
-        for _, idea in ranked[:8]
+        for idea in ranked
     ]
     ctx.context.db.add_agent_event(
         ctx.context.run_id,
@@ -106,6 +312,7 @@ async def search_ideas(
         {
             "query": query,
             "result_count": len(compact),
+            "fallback_to_recent": fallback,
             "candidates": [
                 {"id": item["id"], "title": item["title"]} for item in compact
             ],
@@ -134,7 +341,39 @@ async def inspect_idea(
             {"idea_id": idea_id, "found": False},
             duration_ms=round((time.perf_counter() - started) * 1000),
         )
-        return json.dumps({"error": "idea not found"})
+        return json.dumps(
+            {
+                "error": "idea not found",
+                "valid_candidates": [
+                    {"id": item["id"], "title": item["title"]}
+                    for item in ctx.context.ideas[:8]
+                ],
+            },
+            ensure_ascii=False,
+        )
+    if idea_id in ctx.context.inspected_idea_ids:
+        ctx.context.db.add_agent_event(
+            ctx.context.run_id,
+            ctx.context.user_id,
+            "tool_inspect_idea",
+            {
+                "idea_id": idea_id,
+                "found": True,
+                "reused": True,
+                "title": idea["title"],
+            },
+            duration_ms=round((time.perf_counter() - started) * 1000),
+        )
+        return json.dumps(
+            {
+                "notice": (
+                    "这个想法已经检查过。使用上一份结果并立即形成最终判断，"
+                    "不要再次调用工具。"
+                )
+            },
+            ensure_ascii=False,
+        )
+    ctx.context.inspected_idea_ids.add(idea_id)
     payload = {
         "id": idea["id"],
         "title": idea["title"],
@@ -244,22 +483,19 @@ class Curator:
                 "fragment_characters": len(fragment["content"]),
                 "idea_count": len(ideas),
                 "recent_fragment_count": len(recent_fragments),
+                "prior_fragment_count": sum(
+                    item["id"] != fragment_id
+                    and item["created_at"] <= fragment["created_at"]
+                    for item in recent_fragments
+                ),
+                "later_fragment_count": sum(
+                    item["id"] != fragment_id
+                    and item["created_at"] > fragment["created_at"]
+                    for item in recent_fragments
+                ),
             },
         )
-        output_schema = CuratorDecision.model_json_schema()
+        prompt = build_curator_prompt(fragment, ideas, recent_fragments)
-        lookup_instruction = (
-            "必须先调用 search_ideas 搜索共享的问题或张力；如果候选可能相关，"
-            "再调用 inspect_idea 查看它的真实轨迹后判断。\n\n"
-            if ideas
-            else "目前没有已有想法，不要调用搜索工具。\n\n"
-        )
-        prompt = (
-            "请策展这个刚刚保存的新片段：\n"
-            f"<fragment id=\"{fragment_id}\">{fragment['content']}</fragment>\n\n"
-            f"当前共有 {len(ideas)} 个已有想法。{lookup_instruction}"
-            "最终只输出符合以下 JSON Schema 的 JSON 对象：\n"
-            f"{json.dumps(output_schema, ensure_ascii=False)}"
-        )
 
         set_tracing_disabled(True)
         client = AsyncOpenAI(
@@ -270,63 +506,115 @@ class Curator:
             model=self.settings.deepseek_model,
             openai_client=client,
         )
-        agent: Agent[CuratorContext] = Agent(
-            name="私人思想策展者",
-            instructions=CURATOR_INSTRUCTIONS,
-            model=model,
-            tools=[search_ideas, inspect_idea],
         model_settings = ModelSettings(
             max_tokens=2_400,
             reasoning=Reasoning(effort="high"),
             extra_body={"thinking": {"type": "enabled"}},
             extra_args={"response_format": {"type": "json_object"}},
+        )
+        agent: Agent[CuratorContext] = Agent(
+            name="私人思想策展者",
+            instructions=CURATOR_INSTRUCTIONS,
+            model=model,
+            tools=[search_ideas, inspect_idea],
+            model_settings=model_settings,
+        )
+        repair_agent: Agent[CuratorContext] = Agent(
+            name="判断收敛器",
+            instructions=(
+                CURATOR_INSTRUCTIONS
+                + "\n\n这是收敛回合。不能调用任何工具；只使用输入中已经提供的连续原文、"
+                "想法目录和 schema，立即给出最终合法 JSON。"
             ),
+            model=model,
+            tools=[],
+            model_settings=model_settings,
         )
         decision: CuratorDecision | None = None
         last_error: Exception | None = None
         attempt_count = 0
-        model_rounds = 0
+        hooks = CuratorRunHooks()
-        usage = {
-            "input_tokens": 0,
-            "output_tokens": 0,
-            "reasoning_tokens": 0,
-            "cached_tokens": 0,
-        }
         try:
             for attempt in range(2):
                 attempt_count = attempt + 1
-                repair = (
+                is_repair = attempt > 0
-                    ""
-                    if attempt == 0
-                    else "\n\n上一次输出未通过结构校验。重新完整判断，并只返回合法 JSON。"
-                )
                 self.db.add_agent_event(
                     run_id,
                     user_id,
                     "attempt_started",
-                    {"attempt": attempt_count, "is_repair": attempt > 0},
+                    {"attempt": attempt_count, "is_repair": is_repair},
                 )
                 attempt_started = time.perf_counter()
+                rounds_before = context.model_rounds
+                usage_before = {
+                    "input_tokens": context.input_tokens,
+                    "output_tokens": context.output_tokens,
+                    "reasoning_tokens": context.reasoning_tokens,
+                    "cached_tokens": context.cached_tokens,
+                }
+                try:
                     result = await Runner.run(
-                    agent,
+                        repair_agent if is_repair else agent,
-                    input=prompt + repair,
+                        input=(
-                    context=context,
+                            prompt
-                    max_turns=4,
+                            + (
+                                "\n\n上一次工具探索或输出没有收敛。不要再探索，"
+                                "重新完整判断并只返回合法 JSON。"
+                                if is_repair
+                                else ""
                             )
+                        ),
+                        context=context,
+                        max_turns=2 if is_repair else 6,
+                        hooks=hooks,
+                    )
+                except MaxTurnsExceeded as exc:
                     attempt_duration = round(
                         (time.perf_counter() - attempt_started) * 1000
                     )
-                current_usage = self._result_usage(result)
+                    current_usage = {
-                model_rounds += current_usage.pop("model_rounds")
+                        key: getattr(context, key) - value
-                for key in usage:
+                        for key, value in usage_before.items()
-                    usage[key] += current_usage[key]
+                    }
                     self.db.add_agent_event(
                         run_id,
                         user_id,
                         "model_attempt_completed",
                         {
                             "attempt": attempt_count,
-                        "model_rounds": len(result.raw_responses),
+                            "model_rounds": (
+                                context.model_rounds - rounds_before
+                            ),
+                            "completed": False,
+                            "error_type": type(exc).__name__,
+                            **current_usage,
+                        },
+                        duration_ms=attempt_duration,
+                    )
+                    last_error = exc
+                    if not is_repair:
+                        self.db.add_agent_event(
+                            run_id,
+                            user_id,
+                            "convergence_repair_started",
+                            {"reason": "tool_loop_exceeded"},
+                        )
+                    continue
+                attempt_duration = round(
+                    (time.perf_counter() - attempt_started) * 1000
+                )
+                current_usage = {
+                    key: getattr(context, key) - value
+                    for key, value in usage_before.items()
+                }
+                self.db.add_agent_event(
+                    run_id,
+                    user_id,
+                    "model_attempt_completed",
+                    {
+                        "attempt": attempt_count,
+                        "model_rounds": context.model_rounds - rounds_before,
+                        "completed": True,
                         **current_usage,
                     },
                     duration_ms=attempt_duration,
@@ -336,9 +624,17 @@ class Curator:
                     if not isinstance(raw, str):
                         raise TypeError("curator output was not text")
                     decision = CuratorDecision.model_validate_json(raw)
-                    if ideas and context.search_count == 0:
+                    valid_idea_ids = {idea["id"] for idea in ideas}
-                        decision = None
+                    unknown_ids = [
-                        raise ValueError("curator skipped required idea search")
+                        item.idea_id
+                        for item in decision.assessments
+                        if item.idea_id
+                        and item.idea_id not in valid_idea_ids
+                    ]
+                    if unknown_ids:
+                        raise ValueError(
+                            "curator referenced an unknown existing idea"
+                        )
                     break
                 except (ValidationError, ValueError, TypeError) as exc:
                     decision = None
@@ -358,9 +654,16 @@ class Curator:
                         fragment_id,
                         attempt_count,
                     )
+                    if not is_repair:
+                        self.db.add_agent_event(
+                            run_id,
+                            user_id,
+                            "convergence_repair_started",
+                            {"reason": "output_validation_failed"},
+                        )
             if decision is None:
                 raise RuntimeError(
-                    "curator returned invalid JSON twice"
+                    "curator could not produce a valid decision"
                 ) from last_error
             assessments = (
                 []
@@ -386,11 +689,14 @@ class Curator:
                 status="success",
                 duration_ms=round((time.perf_counter() - run_started) * 1000),
                 attempt_count=attempt_count,
-                model_rounds=model_rounds,
+                model_rounds=context.model_rounds,
                 tool_calls=context.search_count + context.inspection_count,
                 search_calls=context.search_count,
                 inspection_calls=context.inspection_count,
-                **usage,
+                input_tokens=context.input_tokens,
+                output_tokens=context.output_tokens,
+                reasoning_tokens=context.reasoning_tokens,
+                cached_tokens=context.cached_tokens,
             )
             logger.info(
                 "Curator completed fragment %s with %s search and %s inspection calls",
@@ -413,32 +719,14 @@ class Curator:
                 status="error",
                 duration_ms=round((time.perf_counter() - run_started) * 1000),
                 attempt_count=attempt_count,
-                model_rounds=model_rounds,
+                model_rounds=context.model_rounds,
                 tool_calls=context.search_count + context.inspection_count,
                 search_calls=context.search_count,
                 inspection_calls=context.inspection_count,
                 error=exc,
-                **usage,
+                input_tokens=context.input_tokens,
+                output_tokens=context.output_tokens,
+                reasoning_tokens=context.reasoning_tokens,
+                cached_tokens=context.cached_tokens,
             )
             raise
-
-    @staticmethod
-    def _result_usage(result: Any) -> dict[str, int]:
-        totals = {
-            "model_rounds": len(result.raw_responses),
-            "input_tokens": 0,
-            "output_tokens": 0,
-            "reasoning_tokens": 0,
-            "cached_tokens": 0,
-        }
-        for response in result.raw_responses:
-            response_usage = response.usage
-            totals["input_tokens"] += response_usage.input_tokens
-            totals["output_tokens"] += response_usage.output_tokens
-            totals["reasoning_tokens"] += (
-                response_usage.output_tokens_details.reasoning_tokens or 0
-            )
-            totals["cached_tokens"] += (
-                response_usage.input_tokens_details.cached_tokens or 0
-            )
-        return totals

app/db.py

@@ -103,12 +103,18 @@ class Database:
                     id TEXT PRIMARY KEY,
                     idea_id TEXT NOT NULL REFERENCES ideas(id) ON DELETE CASCADE,
                     source_fragment_id TEXT REFERENCES fragments(id) ON DELETE SET NULL,
+                    title TEXT,
+                    summary TEXT,
                     maturity_ai REAL NOT NULL,
+                    maturity_override REAL,
+                    confidence REAL,
                     motion TEXT NOT NULL,
                     position TEXT NOT NULL,
                     tension TEXT NOT NULL,
                     trajectory TEXT NOT NULL,
                     possible_moves TEXT NOT NULL,
+                    change_kind TEXT NOT NULL DEFAULT 'legacy_partial',
+                    fragment_ids TEXT NOT NULL DEFAULT '[]',
                     created_at TEXT NOT NULL
                 );
 
@@ -185,6 +191,26 @@ class Database:
             self._ensure_column(connection, "fragments", "user_id", "TEXT")
             self._ensure_column(connection, "ideas", "user_id", "TEXT")
             self._ensure_column(connection, "sessions", "user_id", "TEXT")
+            self._ensure_column(connection, "idea_snapshots", "title", "TEXT")
+            self._ensure_column(connection, "idea_snapshots", "summary", "TEXT")
+            self._ensure_column(
+                connection, "idea_snapshots", "maturity_override", "REAL"
+            )
+            self._ensure_column(
+                connection, "idea_snapshots", "confidence", "REAL"
+            )
+            self._ensure_column(
+                connection,
+                "idea_snapshots",
+                "change_kind",
+                "TEXT NOT NULL DEFAULT 'legacy_partial'",
+            )
+            self._ensure_column(
+                connection,
+                "idea_snapshots",
+                "fragment_ids",
+                "TEXT NOT NULL DEFAULT '[]'",
+            )
 
             now = utc_now()
             for seed in seeds:
@@ -457,8 +483,10 @@ class Database:
             ).fetchall()
             snapshots = connection.execute(
                 """
-                SELECT s.maturity_ai, s.motion, s.position, s.tension,
+                SELECT s.title, s.summary, s.maturity_ai,
-                       s.trajectory, s.possible_moves, s.created_at
+                       s.maturity_override, s.confidence, s.motion,
+                       s.position, s.tension, s.trajectory, s.possible_moves,
+                       s.change_kind, s.fragment_ids, s.created_at
                 FROM idea_snapshots s
                 JOIN ideas i ON i.id = s.idea_id
                 WHERE s.idea_id = ? AND i.user_id = ?
@@ -471,6 +499,7 @@ class Database:
             {
                 **dict(row),
                 "possible_moves": _loads(row["possible_moves"], []),
+                "fragment_ids": _loads(row["fragment_ids"], []),
             }
             for row in snapshots
         ]
@@ -588,23 +617,55 @@ class Database:
                         now,
                     ),
                 )
+                version_state = connection.execute(
+                    """
+                    SELECT maturity_override FROM ideas
+                    WHERE id = ? AND user_id = ?
+                    """,
+                    (idea_id, user_id),
+                ).fetchone()
+                fragment_ids = [
+                    row["fragment_id"]
+                    for row in connection.execute(
+                        """
+                        SELECT l.fragment_id
+                        FROM idea_fragments l
+                        JOIN fragments f ON f.id = l.fragment_id
+                        WHERE l.idea_id = ? AND f.user_id = ?
+                        ORDER BY f.created_at ASC, f.rowid ASC
+                        """,
+                        (idea_id, user_id),
+                    ).fetchall()
+                ]
                 connection.execute(
                     """
                     INSERT INTO idea_snapshots (
-                        id, idea_id, source_fragment_id, maturity_ai, motion,
+                        id, idea_id, source_fragment_id, title, summary,
-                        position, tension, trajectory, possible_moves, created_at
+                        maturity_ai, maturity_override, confidence, motion,
-                    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                        position, tension, trajectory, possible_moves,
+                        change_kind, fragment_ids, created_at
+                    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                     """,
                     (
                         str(uuid.uuid4()),
                         idea_id,
                         fragment_id,
+                        assessment["title"],
+                        assessment["summary"],
                         assessment["maturity"],
+                        (
+                            version_state["maturity_override"]
+                            if version_state
+                            else None
+                        ),
+                        assessment["confidence"],
                         assessment["motion"],
                         assessment["position"],
                         assessment["tension"],
                         assessment["trajectory"],
                         _json(assessment["possible_moves"]),
+                        "analysis",
+                        _json(fragment_ids),
                         now,
                     ),
                 )
@@ -622,16 +683,66 @@ class Database:
     def update_idea_override(
         self, user_id: str, idea_id: str, maturity_override: float | None
     ) -> dict[str, Any] | None:
+        now = utc_now()
         with self.connect() as connection:
             result = connection.execute(
                 """
                 UPDATE ideas SET maturity_override = ?, updated_at = ?
                 WHERE id = ? AND user_id = ?
                 """,
-                (maturity_override, utc_now(), idea_id, user_id),
+                (maturity_override, now, idea_id, user_id),
             )
             if result.rowcount == 0:
                 return None
+            idea = connection.execute(
+                """
+                SELECT title, summary, maturity_ai, maturity_override,
+                       confidence, motion, position, tension, trajectory,
+                       possible_moves
+                FROM ideas WHERE id = ? AND user_id = ?
+                """,
+                (idea_id, user_id),
+            ).fetchone()
+            fragment_ids = [
+                row["fragment_id"]
+                for row in connection.execute(
+                    """
+                    SELECT l.fragment_id
+                    FROM idea_fragments l
+                    JOIN fragments f ON f.id = l.fragment_id
+                    WHERE l.idea_id = ? AND f.user_id = ?
+                    ORDER BY f.created_at ASC, f.rowid ASC
+                    """,
+                    (idea_id, user_id),
+                ).fetchall()
+            ]
+            connection.execute(
+                """
+                INSERT INTO idea_snapshots (
+                    id, idea_id, source_fragment_id, title, summary,
+                    maturity_ai, maturity_override, confidence, motion,
+                    position, tension, trajectory, possible_moves,
+                    change_kind, fragment_ids, created_at
+                ) VALUES (?, ?, NULL, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    str(uuid.uuid4()),
+                    idea_id,
+                    idea["title"],
+                    idea["summary"],
+                    idea["maturity_ai"],
+                    idea["maturity_override"],
+                    idea["confidence"],
+                    idea["motion"],
+                    idea["position"],
+                    idea["tension"],
+                    idea["trajectory"],
+                    idea["possible_moves"],
+                    "manual_calibration",
+                    _json(fragment_ids),
+                    now,
+                ),
+            )
         return self.get_idea(user_id, idea_id)
 
     @staticmethod
@@ -928,6 +1039,48 @@ class Database:
         ).hexdigest()[:16]
         return run
 
+    def prepare_agent_retry(self, run_id: str) -> dict[str, str] | None:
+        with self.connect() as connection:
+            row = connection.execute(
+                """
+                SELECT r.user_id, r.fragment_id, r.status AS run_status,
+                       f.analysis_status
+                FROM agent_runs r
+                JOIN fragments f ON f.id = r.fragment_id
+                WHERE r.id = ? AND f.user_id = r.user_id
+                """,
+                (run_id,),
+            ).fetchone()
+            if (
+                not row
+                or row["run_status"] != "error"
+                or row["analysis_status"] != "error"
+            ):
+                return None
+            newer = connection.execute(
+                """
+                SELECT 1 FROM agent_runs
+                WHERE fragment_id = ? AND user_id = ?
+                  AND id != ? AND status IN ('running', 'success')
+                LIMIT 1
+                """,
+                (row["fragment_id"], row["user_id"], run_id),
+            ).fetchone()
+            if newer:
+                return None
+            connection.execute(
+                """
+                UPDATE fragments
+                SET analysis_status = 'pending', analysis_error = NULL
+                WHERE id = ? AND user_id = ?
+                """,
+                (row["fragment_id"], row["user_id"]),
+            )
+        return {
+            "user_id": row["user_id"],
+            "fragment_id": row["fragment_id"],
+        }
+
     def list_agent_events(self, run_id: str) -> list[dict[str, Any]]:
         with self.connect() as connection:
             rows = connection.execute(

app/main.py

@@ -276,7 +276,12 @@ def _redacted_event(event: dict[str, Any]) -> dict[str, Any]:
         safe = payload
     elif event_type == "context_loaded":
         safe = payload
-    elif event_type in {"attempt_started", "model_attempt_completed"}:
+    elif event_type in {
+        "attempt_started",
+        "model_round_completed",
+        "model_attempt_completed",
+        "convergence_repair_started",
+    }:
         safe = payload
     elif event_type == "tool_search_ideas":
         safe = {"result_count": payload.get("result_count", 0)}
@@ -341,6 +346,28 @@ async def admin_run_detail(
     }
 
 
+@app.post(
+    "/api/admin/runs/{run_id}/retry",
+    status_code=status.HTTP_202_ACCEPTED,
+)
+async def retry_admin_run(
+    run_id: str, admin: AuthUser = Depends(administrator)
+):
+    retry = db.prepare_agent_retry(run_id)
+    if not retry:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="这次分析现在不能重新运行",
+        )
+    curator.enqueue(retry["fragment_id"])
+    db.add_audit_event(
+        "agent_run_requeued",
+        admin.id,
+        {"run_id": run_id, "owner_user_id": retry["user_id"]},
+    )
+    return {"queued": True}
+
+
 static_dir = Path(__file__).parent / "static"
 assets_dir = static_dir / "assets"
 if assets_dir.is_dir():

frontend/src/App.tsx

@@ -363,12 +363,41 @@ function IdeaDetail({
             </ul>
           </section>
 
+          {idea.snapshots && idea.snapshots.length > 1 && (
+            <details className="idea-history">
+              <summary>这个想法走过的路</summary>
+              <div>
+                {[...idea.snapshots].reverse().map((snapshot, index) => (
+                  <article key={`${snapshot.created_at}-${index}`}>
+                    <time>
+                      {new Intl.DateTimeFormat("zh-CN", {
+                        month: "long",
+                        day: "numeric",
+                        hour: "2-digit",
+                        minute: "2-digit",
+                      }).format(new Date(snapshot.created_at))}
+                    </time>
+                    <strong>
+                      {snapshot.change_kind === "manual_calibration"
+                        ? "你重新放置了它"
+                        : snapshot.motion}
+                    </strong>
+                    <p>{snapshot.trajectory}</p>
+                  </article>
+                ))}
+              </div>
+            </details>
+          )}
+
           {idea.fragments && idea.fragments.length > 0 && (
             <details className="evidence">
-              <summary>构成这个想法的片段</summary>
+              <summary>原始脉络</summary>
               <div>
                 {idea.fragments.map((fragment) => (
-                  <blockquote key={fragment.id}>{fragment.content}</blockquote>
+                  <article key={fragment.id}>
+                    <time>{dayLabel(fragment.created_at)}</time>
+                    <blockquote>{fragment.content}</blockquote>
+                  </article>
                 ))}
               </div>
             </details>
@@ -454,6 +483,8 @@ const eventNames: Record<string, string> = {
   tool_search_ideas: "搜索已有想法",
   tool_inspect_idea: "查看想法轨迹",
   model_attempt_completed: "模型返回",
+  model_round_completed: "完成一轮判断",
+  convergence_repair_started: "转入收敛判断",
   validation_failed: "结构校验未通过",
   decision_committed: "提交分析结果",
   run_failed: "运行失败",
@@ -466,6 +497,7 @@ const auditNames: Record<string, string> = {
   logout: "退出登录",
   fragment_created: "保存新片段",
   idea_position_calibrated: "人工校准位势",
+  agent_run_requeued: "重新分析",
   debug_sharing_changed: "调整调试共享",
 };
 
@@ -510,6 +542,18 @@ function EventPayload({ event }: { event: AgentEvent }) {
       </p>
     );
   }
+  if (event.event_type === "model_round_completed") {
+    return (
+      <p>
+        输入 {compactNumber(Number(payload.input_tokens ?? 0))} · 输出{" "}
+        {compactNumber(Number(payload.output_tokens ?? 0))} · 推理{" "}
+        {compactNumber(Number(payload.reasoning_tokens ?? 0))}
+      </p>
+    );
+  }
+  if (event.event_type === "convergence_repair_started") {
+    return <p>工具探索没有及时收束，系统改用已有上下文直接完成判断。</p>;
+  }
   if (event.event_type === "validation_failed") {
     return (
       <p>
@@ -561,11 +605,24 @@ function RunDetail({
   onClose: () => void;
 }) {
   const [detail, setDetail] = useState<AgentRunDetail | null>(null);
+  const [retrying, setRetrying] = useState(false);
+  const [requeued, setRequeued] = useState(false);
 
   useEffect(() => {
     api.adminRun(runId).then(setDetail);
   }, [runId]);
 
+  async function retry() {
+    if (retrying || requeued) return;
+    setRetrying(true);
+    try {
+      await api.retryAdminRun(runId);
+      setRequeued(true);
+    } finally {
+      setRetrying(false);
+    }
+  }
+
   return (
     <div className="sheet-backdrop" role="presentation" onMouseDown={onClose}>
       <article
@@ -614,8 +671,7 @@ function RunDetail({
                 <blockquote>{detail.fragment.content}</blockquote>
               ) : (
                 <div className="redacted-content">
-                  内容已隔离 · {detail.fragment.content_length} 字 · 指纹{" "}
+                  内容已隔离 · {detail.fragment.content_length} 字
-                  {detail.fragment.content_sha256}
                 </div>
               )}
               <small>{detail.privacy.reason}</small>
@@ -636,6 +692,19 @@ function RunDetail({
                 </article>
               ))}
             </section>
+            {detail.run.status === "error" && (
+              <button
+                className="retry-run"
+                onClick={() => void retry()}
+                disabled={retrying || requeued}
+              >
+                {requeued
+                  ? "已经重新交给策展者"
+                  : retrying
+                    ? "正在重新交付…"
+                    : "重新整理这条记录"}
+              </button>
+            )}
             <footer className="reasoning-boundary">
               不保存模型隐藏思维链；这里展示的是输入、工具行为、用量、校验与最终结构化产物。
             </footer>

frontend/src/api.ts

@@ -12,12 +12,17 @@ export type User = {
 };
 
 export type Snapshot = {
+  title: string | null;
+  summary: string | null;
   maturity_ai: number;
+  maturity_override: number | null;
+  confidence: number | null;
   motion: string;
   position: string;
   tension: string;
   trajectory: string;
   possible_moves: string[];
+  change_kind: "analysis" | "manual_calibration" | "legacy_partial";
   created_at: string;
 };
 
@@ -226,6 +231,10 @@ export const api = {
     request<{ items: AgentRun[] }>(`/api/admin/runs?limit=${limit}`),
   adminRun: (id: string) =>
     request<AgentRunDetail>(`/api/admin/runs/${id}`),
+  retryAdminRun: (id: string) =>
+    request<{ queued: boolean }>(`/api/admin/runs/${id}/retry`, {
+      method: "POST",
+    }),
   adminAudit: (limit = 80) =>
     request<{ items: AuditEvent[] }>(`/api/admin/audit?limit=${limit}`),
 };

frontend/src/styles.css

@@ -885,12 +885,18 @@ summary:focus-visible {
   color: var(--warm);
 }
 
+.idea-history,
 .evidence {
   margin-top: 50px;
   border-top: 1px solid var(--line);
   color: var(--muted);
 }
 
+.idea-history + .evidence {
+  margin-top: 0;
+}
+
+.idea-history summary,
 .evidence summary {
   padding: 18px 0;
   cursor: pointer;
@@ -898,12 +904,44 @@ summary:focus-visible {
   letter-spacing: 0.05em;
 }
 
+.idea-history article {
+  display: grid;
+  grid-template-columns: 92px 1fr;
+  gap: 4px 14px;
+  padding: 14px 0;
+  border-bottom: 1px solid rgba(23, 37, 31, 0.07);
+}
+
+.idea-history time,
+.evidence time {
+  color: #92958f;
+  font-size: 9px;
+}
+
+.idea-history strong {
+  color: var(--ink);
+  font-size: 12px;
+  font-weight: 500;
+}
+
+.idea-history p {
+  grid-column: 2;
+  margin: 2px 0 0;
+  color: #626b66;
+  font-size: 11px;
+  line-height: 1.7;
+}
+
+.evidence article {
+  padding: 13px 0;
+  border-bottom: 1px solid rgba(23, 37, 31, 0.07);
+}
+
 .evidence blockquote {
-  margin: 0;
+  margin: 5px 0 0;
-  padding: 15px 0;
+  padding: 0;
   color: #4e5752;
   font: 400 14px/1.85 "Songti SC", "STSong", serif;
-  border-bottom: 1px solid rgba(23, 37, 31, 0.07);
 }
 
 .admin-view {
@@ -1411,6 +1449,22 @@ summary:focus-visible {
   margin: 0 !important;
 }
 
+.retry-run {
+  margin-top: 28px;
+  padding: 9px 13px;
+  border: 1px solid rgba(153, 104, 72, 0.28);
+  border-radius: 4px;
+  background: transparent;
+  color: var(--warm);
+  cursor: pointer;
+  font-size: 10px;
+}
+
+.retry-run:disabled {
+  cursor: default;
+  opacity: 0.6;
+}
+
 .reasoning-boundary {
   margin-top: 45px;
   padding-top: 16px;

tests/test_api.py

@@ -82,9 +82,10 @@ def test_login_isolation_roles_and_debug_privacy(
             headers=WRITE_HEADERS,
         )
         assert member_fragment.status_code == 201
-        assert [item["content"] for item in client.get("/api/fragments").json()["items"]] == [
+        assert [
-            "用户二的记录"
+            item["content"]
-        ]
+            for item in client.get("/api/fragments").json()["items"]
+        ] == ["用户二的记录"]
         assert client.get("/api/admin/overview").status_code == 403
 
         run_id = database.start_agent_run(
@@ -105,9 +106,10 @@ def test_login_isolation_roles_and_debug_privacy(
         )
 
         client.cookies.set(COOKIE_NAME, admin_cookie)
-        assert [item["content"] for item in client.get("/api/fragments").json()["items"]] == [
+        assert [
-            "管理员的记录"
+            item["content"]
-        ]
+            for item in client.get("/api/fragments").json()["items"]
+        ] == ["管理员的记录"]
         redacted = client.get(f"/api/admin/runs/{run_id}").json()
         assert redacted["fragment"]["content_visible"] is False
         assert redacted["events"][-1]["payload"] == {
@@ -127,3 +129,44 @@ def test_login_isolation_roles_and_debug_privacy(
         visible = client.get(f"/api/admin/runs/{run_id}").json()
         assert visible["fragment"]["content_visible"] is True
         assert visible["fragment"]["content"] == "用户二的记录"
+
+        retry_error = RuntimeError("tool loop exceeded")
+        database.finish_agent_run(
+            run_id,
+            status="error",
+            duration_ms=100,
+            attempt_count=1,
+            model_rounds=4,
+            tool_calls=4,
+            search_calls=3,
+            inspection_calls=1,
+            input_tokens=400,
+            output_tokens=80,
+            reasoning_tokens=60,
+            cached_tokens=0,
+            error=retry_error,
+        )
+        database.set_fragment_status(
+            member_fragment.json()["id"], "error", str(retry_error)
+        )
+        queued: list[str] = []
+        monkeypatch.setattr(main.curator, "enqueue", queued.append)
+
+        client.cookies.set(COOKIE_NAME, member_cookie)
+        assert client.post(
+            f"/api/admin/runs/{run_id}/retry",
+            headers=WRITE_HEADERS,
+        ).status_code == 403
+
+        client.cookies.set(COOKIE_NAME, admin_cookie)
+        retried = client.post(
+            f"/api/admin/runs/{run_id}/retry",
+            headers=WRITE_HEADERS,
+        )
+        assert retried.status_code == 202
+        assert retried.json() == {"queued": True}
+        assert queued == [member_fragment.json()["id"]]
+        assert client.post(
+            f"/api/admin/runs/{run_id}/retry",
+            headers=WRITE_HEADERS,
+        ).status_code == 409

tests/test_curator.py

@@ -0,0 +1,66 @@
+import asyncio
+from pathlib import Path
+from types import SimpleNamespace
+
+from agents.exceptions import MaxTurnsExceeded
+
+from app.config import Settings, UserSeed
+from app.curator import Curator, Runner
+from app.db import Database
+
+
+def test_tool_loop_uses_no_tool_convergence_repair(
+    tmp_path: Path, monkeypatch
+):
+    admin = UserSeed(
+        id="admin",
+        label="管理员",
+        role="admin",
+        access_key_hash="unused",
+    )
+    database = Database(tmp_path / "repair.sqlite3")
+    database.initialize([admin])
+    fragment = database.create_fragment("admin", "一个需要收敛的念头")
+    settings = Settings(
+        data_dir=tmp_path,
+        database_path=tmp_path / "repair.sqlite3",
+        users=(admin,),
+        session_secret="unused",
+        deepseek_api_key="test-key",
+        deepseek_base_url="https://api.deepseek.com",
+        deepseek_model="deepseek-v4-pro",
+        cookie_secure=False,
+        auth_disabled=True,
+    )
+    calls = 0
+
+    async def fake_run(*args, **kwargs):
+        nonlocal calls
+        calls += 1
+        if calls == 1:
+            raise MaxTurnsExceeded("Max turns (6) exceeded")
+        assert kwargs["max_turns"] == 2
+        assert args[0].tools == []
+        return SimpleNamespace(
+            final_output=(
+                '{"standalone":true,"reasoning_note":"暂时独立保留",'
+                '"assessments":[]}'
+            )
+        )
+
+    monkeypatch.setattr(Runner, "run", fake_run)
+    curator = Curator(database, settings)
+
+    asyncio.run(curator.analyze(fragment["id"]))
+
+    assert calls == 2
+    assert database.get_fragment(fragment["id"], "admin")[
+        "analysis_status"
+    ] == "done"
+    run = database.list_agent_runs(1)[0]
+    events = database.list_agent_events(run["id"])
+    assert run["status"] == "success"
+    assert run["attempt_count"] == 2
+    assert "convergence_repair_started" in [
+        event["event_type"] for event in events
+    ]

tests/test_db.py

@@ -1,6 +1,7 @@
 from pathlib import Path
 
 from app.config import UserSeed
+from app.curator import _rank_ideas, build_curator_prompt
 from app.db import Database
 
 
@@ -97,6 +98,105 @@ def test_apply_assessment_builds_trajectory(tmp_path: Path):
     assert idea["maturity"] == 46
 
 
+def test_curator_prompt_contains_recent_continuity_and_idea_catalog(
+    tmp_path: Path,
+):
+    db = database(tmp_path)
+    first = db.create_fragment(
+        ADMIN.id,
+        "我先让 AI 完成综述，接下来会抽取几个感兴趣的地方。",
+    )
+    db.apply_assessments(ADMIN.id, first["id"], [assessment()])
+    current = db.create_fragment(
+        ADMIN.id,
+        "我识别出了两个新的认知，第一点让我耳目一新。",
+    )
+    ideas, fragments = db.curator_context(ADMIN.id)
+
+    prompt = build_curator_prompt(current, ideas, fragments)
+
+    assert "我先让 AI 完成综述" in prompt
+    assert "我识别出了两个新的认知" in prompt
+    assert "无分类的笔记入口" in prompt
+    assert "不要要求用户显式建立会话" in prompt
+
+
+def test_delayed_reanalysis_keeps_later_input_out_of_prior_context(
+    tmp_path: Path,
+):
+    db = database(tmp_path)
+    previous = db.create_fragment(ADMIN.id, "当时真正的上文")
+    target = db.create_fragment(ADMIN.id, "后来需要重新整理的这一条")
+    future = db.create_fragment(ADMIN.id, "这句话发生在目标片段之后")
+    timeline = (
+        (previous, "2026-07-28T10:00:00.000+00:00"),
+        (target, "2026-07-28T10:10:00.000+00:00"),
+        (future, "2026-07-28T10:20:00.000+00:00"),
+    )
+    with db.connect() as connection:
+        for fragment, created_at in timeline:
+            fragment["created_at"] = created_at
+            connection.execute(
+                "UPDATE fragments SET created_at = ? WHERE id = ?",
+                (created_at, fragment["id"]),
+            )
+    ideas, fragments = db.curator_context(ADMIN.id)
+
+    prompt = build_curator_prompt(target, ideas, fragments)
+    recent_section = prompt.split("<recent_context>", 1)[1].split(
+        "</recent_context>", 1
+    )[0]
+    later_section = prompt.split("<later_context>", 1)[1].split(
+        "</later_context>", 1
+    )[0]
+
+    assert previous["content"] in recent_section
+    assert future["content"] not in recent_section
+    assert future["content"] in later_section
+    assert "延迟重整" in prompt
+
+
+def test_idea_search_falls_back_to_recent_catalog():
+    idea = {
+        "id": "idea-one",
+        "title": "先综述再深入",
+        "summary": "一种理解策略",
+        "position": "正在实践",
+        "tension": "可靠性",
+        "trajectory": "从原则走向实践",
+        "updated_at": "2026-07-28",
+        "recent_fragments": [],
+    }
+
+    candidates, fallback = _rank_ideas(
+        [idea], "agent architecture long task reliability"
+    )
+
+    assert fallback is True
+    assert candidates == [idea]
+
+
+def test_idea_versions_preserve_complete_state(tmp_path: Path):
+    db = database(tmp_path)
+    first = db.create_fragment(ADMIN.id, "第一条原始证据")
+    db.apply_assessments(ADMIN.id, first["id"], [assessment(maturity=30)])
+    idea = db.list_ideas(ADMIN.id)[0]
+    first_version = db.get_idea(ADMIN.id, idea["id"])["snapshots"][-1]
+
+    assert first_version["title"] == "无分类的笔记入口"
+    assert first_version["summary"] == "先保留念头，再让结构在后台形成。"
+    assert first_version["confidence"] == 0.76
+    assert first_version["change_kind"] == "analysis"
+    assert first_version["fragment_ids"] == [first["id"]]
+
+    db.update_idea_override(ADMIN.id, idea["id"], 61)
+    versions = db.get_idea(ADMIN.id, idea["id"])["snapshots"]
+
+    assert versions[-1]["change_kind"] == "manual_calibration"
+    assert versions[-1]["maturity_override"] == 61
+    assert versions[-1]["fragment_ids"] == [first["id"]]
+
+
 def test_manual_position_is_authoritative(tmp_path: Path):
     db = database(tmp_path)
     fragment = db.create_fragment(ADMIN.id, "一个想法")
@@ -188,3 +288,41 @@ def test_agent_run_records_metrics_and_events(tmp_path: Path):
         "tool_search_ideas",
     ]
     assert overview["last_24h"]["success_rate"] == 100
+
+
+def test_failed_agent_run_can_be_safely_requeued(tmp_path: Path):
+    db = database(tmp_path)
+    fragment = db.create_fragment(ADMIN.id, "需要重新整理")
+    run_id = db.start_agent_run(
+        ADMIN.id,
+        fragment["id"],
+        "deepseek-v4-pro",
+        "prompt.v2",
+        "high",
+    )
+    error = RuntimeError("temporary failure")
+    db.finish_agent_run(
+        run_id,
+        status="error",
+        duration_ms=100,
+        attempt_count=1,
+        model_rounds=1,
+        tool_calls=0,
+        search_calls=0,
+        inspection_calls=0,
+        input_tokens=10,
+        output_tokens=2,
+        reasoning_tokens=1,
+        cached_tokens=0,
+        error=error,
+    )
+    db.set_fragment_status(fragment["id"], "error", str(error))
+
+    retry = db.prepare_agent_retry(run_id)
+
+    assert retry == {
+        "user_id": ADMIN.id,
+        "fragment_id": fragment["id"],
+    }
+    assert db.get_fragment(fragment["id"], ADMIN.id)["analysis_status"] == "pending"
+    assert db.prepare_agent_retry(run_id) is None