feat: recover evolving threads from freeform notes
This commit is contained in:
wuyang
@@ -13,6 +13,8 @@ Agent 使用 DeepSeek 的思考模式，结合最近的连续原文与精简想
|
||||
- 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。
|
||||
- 后台的“想法”是广义的演化脉络，也包括项目、工作、学习、问题和现实承诺。
|
||||
- 后续输入让一个方向成形时，系统可以把先前暂存的原文补回完整证据链。
|
||||
- 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。
|
||||
|
||||
+49
-10
@@ -29,7 +29,7 @@ from .db import Database
|
||||
from .schemas import CuratorDecision
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
PROMPT_VERSION = "maturity-2026-07-28.v3"
|
||||
PROMPT_VERSION = "maturity-2026-07-29.v4"
|
||||
RECENT_CONTEXT_LIMIT = 12
|
||||
IDEA_CATALOG_LIMIT = 40
|
||||
|
||||
@@ -38,7 +38,9 @@ CURATOR_INSTRUCTIONS = """
|
||||
你是一个私人思想笔记本内部的“策展者”。用户界面必须安静，所以你的工作发生在幕后。
|
||||
|
||||
你的任务不是给句子贴标签，也不是强迫想法进入固定工作流，而是辨认：新片段是否属于一个
|
||||
持续演化的想法；如果属于，它此刻大致在哪里、正在怎样运动、内在张力是什么。
|
||||
持续演化的脉络；如果属于，它此刻大致在哪里、正在怎样运动、内在张力是什么。这里的“想法”
|
||||
是面向用户的安静称呼，后台必须把它理解得更广：观点、问题、项目、工作、学习、创作、决策和
|
||||
现实承诺，只要会随时间变化、值得未来回来，都可能是一个脉络。
|
||||
|
||||
判断“成熟度”时使用融合性的人类经验，而不是计数规则。综合考察：
|
||||
1. 想法是否已经有自己的身份、边界与可复述的核心；
|
||||
@@ -54,9 +56,23 @@ motion 是你对当下运动的自然语言压缩，例如“向现实探去”
|
||||
预设流水线。position 说明此刻所处的位置。trajectory 只描述最近发生的变化。possible_moves
|
||||
是 1—4 个根据当前张力即时生成的可能动作，不是任务清单，不承诺固定的下一关。
|
||||
|
||||
在决定 standalone 或新建前，必须把两个问题分开：
|
||||
1. 它是否延续某个已有脉络；
|
||||
2. 即使不延续，它是否带着值得未来回来的未完成方向。
|
||||
“不属于已有想法”绝不等于“应当 standalone”。如果片段已经显露出可命名的对象，并带有
|
||||
未解问题、现实承诺、明确意图或下一步动作之一，即使只有第一条，也通常应建立一个低位势的
|
||||
新脉络。尤其是用户明确说“新工作”“需要做”“我想研究”“我要先……”时，不要把它当作
|
||||
不值得跟踪的杂讯。只有纯粹随手事实、情绪或转瞬观察，既无可辨认对象也无未来张力，或信息
|
||||
少到无法诚实命名时，才 standalone。
|
||||
|
||||
谨慎合并。仅因主题词相似不能归到同一想法；关注它们是否共享同一个问题、张力或生成方向。
|
||||
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。
|
||||
一个片段最多关联两个想法。新建必须克制，但克制意味着不凭关键词重复建，不意味着压制真实
|
||||
的新方向。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。
|
||||
|
||||
如果 recent_context 中较早的原文与当前判断共同构成同一脉络，但它当时可能只被暂存，使用
|
||||
supporting_fragment_ids 把它补回证据链。只可填写 recent_context 中出现的准确 id，最多六条；
|
||||
当前 new_fragment 会被系统自动关联，不要填写它。不要为主题词相似而回收，也不要填写
|
||||
later_context 中的记录。
|
||||
|
||||
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
|
||||
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
|
||||
@@ -190,12 +206,10 @@ class CuratorRunHooks(RunHooks[CuratorContext]):
|
||||
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 = [
|
||||
def _prior_context(
|
||||
fragment: dict[str, Any], recent_fragments: list[dict[str, Any]]
|
||||
) -> list[dict[str, Any]]:
|
||||
return [
|
||||
{
|
||||
"id": item["id"],
|
||||
"created_at": item["created_at"],
|
||||
@@ -205,6 +219,14 @@ def build_curator_prompt(
|
||||
if item["id"] != fragment["id"]
|
||||
and item["created_at"] <= fragment["created_at"]
|
||||
][-RECENT_CONTEXT_LIMIT:]
|
||||
|
||||
|
||||
def build_curator_prompt(
|
||||
fragment: dict[str, Any],
|
||||
ideas: list[dict[str, Any]],
|
||||
recent_fragments: list[dict[str, Any]],
|
||||
) -> str:
|
||||
prior_context = _prior_context(fragment, recent_fragments)
|
||||
later_context = [
|
||||
{
|
||||
"created_at": item["created_at"],
|
||||
@@ -256,6 +278,9 @@ def build_curator_prompt(
|
||||
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"
|
||||
"先分别判断“是否延续已有脉络”与“是否形成新的未完成方向”。若较早的 recent_context "
|
||||
"与当前片段共同构成同一脉络，在对应 assessment 的 supporting_fragment_ids 中补回；"
|
||||
"没有确切证据时保持为空。\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)}"
|
||||
)
|
||||
@@ -635,6 +660,20 @@ class Curator:
|
||||
raise ValueError(
|
||||
"curator referenced an unknown existing idea"
|
||||
)
|
||||
eligible_fragment_ids = {
|
||||
item["id"]
|
||||
for item in _prior_context(fragment, recent_fragments)
|
||||
}
|
||||
unknown_fragment_ids = [
|
||||
supporting_id
|
||||
for item in decision.assessments
|
||||
for supporting_id in item.supporting_fragment_ids
|
||||
if supporting_id not in eligible_fragment_ids
|
||||
]
|
||||
if unknown_fragment_ids:
|
||||
raise ValueError(
|
||||
"curator referenced a fragment outside recent_context"
|
||||
)
|
||||
break
|
||||
except (ValidationError, ValueError, TypeError) as exc:
|
||||
decision = None
|
||||
|
||||
@@ -604,6 +604,25 @@ class Database:
|
||||
continue
|
||||
seen.add(idea_id)
|
||||
affected.append(idea_id)
|
||||
evidence_fragment_ids: list[str] = []
|
||||
for evidence_fragment_id in [
|
||||
fragment_id,
|
||||
*assessment.get("supporting_fragment_ids", []),
|
||||
]:
|
||||
if evidence_fragment_id in evidence_fragment_ids:
|
||||
continue
|
||||
evidence_exists = connection.execute(
|
||||
"""
|
||||
SELECT 1 FROM fragments
|
||||
WHERE id = ? AND user_id = ?
|
||||
""",
|
||||
(evidence_fragment_id, user_id),
|
||||
).fetchone()
|
||||
if not evidence_exists:
|
||||
raise ValueError(
|
||||
"supporting fragment not found for user"
|
||||
)
|
||||
evidence_fragment_ids.append(evidence_fragment_id)
|
||||
connection.execute(
|
||||
"""
|
||||
INSERT OR IGNORE INTO idea_fragments (
|
||||
@@ -612,7 +631,7 @@ class Database:
|
||||
""",
|
||||
(
|
||||
idea_id,
|
||||
fragment_id,
|
||||
evidence_fragment_id,
|
||||
assessment.get("relevance", 1),
|
||||
now,
|
||||
),
|
||||
|
||||
@@ -44,12 +44,33 @@ class IdeaAssessment(BaseModel):
|
||||
trajectory: str = Field(min_length=1, max_length=280)
|
||||
possible_moves: list[str] = Field(min_length=1, max_length=4)
|
||||
relevance: float = Field(default=1, ge=0, le=1)
|
||||
supporting_fragment_ids: list[str] = Field(
|
||||
default_factory=list,
|
||||
max_length=6,
|
||||
description=(
|
||||
"Ids of earlier fragments from recent_context that are substantive "
|
||||
"evidence for this same evolving thread. The new fragment is linked "
|
||||
"automatically and must not be included."
|
||||
),
|
||||
)
|
||||
|
||||
@field_validator("possible_moves")
|
||||
@classmethod
|
||||
def moves_are_brief(cls, moves: list[str]) -> list[str]:
|
||||
return [move.strip()[:100] for move in moves if move.strip()][:4]
|
||||
|
||||
@field_validator("supporting_fragment_ids")
|
||||
@classmethod
|
||||
def supporting_fragments_are_unique(
|
||||
cls, fragment_ids: list[str]
|
||||
) -> list[str]:
|
||||
unique: list[str] = []
|
||||
for fragment_id in fragment_ids:
|
||||
value = fragment_id.strip()
|
||||
if value and value not in unique:
|
||||
unique.append(value)
|
||||
return unique[:6]
|
||||
|
||||
|
||||
class CuratorDecision(BaseModel):
|
||||
standalone: bool = Field(
|
||||
|
||||
@@ -577,6 +577,12 @@ function EventPayload({ event }: { event: AgentEvent }) {
|
||||
{String(assessment.motion ?? "")}
|
||||
</span>
|
||||
<p>{String(assessment.position ?? "")}</p>
|
||||
{Array.isArray(assessment.supporting_fragment_ids) &&
|
||||
assessment.supporting_fragment_ids.length > 0 && (
|
||||
<small>
|
||||
同时补回 {assessment.supporting_fragment_ids.length} 条早期记录
|
||||
</small>
|
||||
)}
|
||||
</div>
|
||||
))}
|
||||
</div>
|
||||
|
||||
@@ -64,3 +64,71 @@ def test_tool_loop_uses_no_tool_convergence_repair(
|
||||
assert "convergence_repair_started" in [
|
||||
event["event_type"] for event in events
|
||||
]
|
||||
|
||||
|
||||
def test_curator_recovers_visible_prior_fragment(
|
||||
tmp_path: Path, monkeypatch
|
||||
):
|
||||
admin = UserSeed(
|
||||
id="admin",
|
||||
label="管理员",
|
||||
role="admin",
|
||||
access_key_hash="unused",
|
||||
)
|
||||
database = Database(tmp_path / "recovery.sqlite3")
|
||||
database.initialize([admin])
|
||||
earlier = database.create_fragment(
|
||||
"admin", "有个新工作可能要做小爱架构。"
|
||||
)
|
||||
database.apply_assessments("admin", earlier["id"], [])
|
||||
current = database.create_fragment(
|
||||
"admin", "确定要做，先熟悉 harness 5.0。"
|
||||
)
|
||||
settings = Settings(
|
||||
data_dir=tmp_path,
|
||||
database_path=tmp_path / "recovery.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,
|
||||
)
|
||||
|
||||
async def fake_run(*args, **kwargs):
|
||||
return SimpleNamespace(
|
||||
final_output=(
|
||||
"{"
|
||||
'"standalone":false,'
|
||||
'"reasoning_note":"明确的新工作，应形成可跟踪脉络",'
|
||||
'"assessments":[{'
|
||||
'"idea_id":null,'
|
||||
'"title":"小爱新架构开发",'
|
||||
'"summary":"熟悉现有方案并推进新架构开发。",'
|
||||
'"maturity":10,'
|
||||
'"confidence":0.9,'
|
||||
'"motion":"启动中",'
|
||||
'"position":"先熟悉 harness 5.0 核心方案。",'
|
||||
'"tension":"任务已明确，但范围与现有方案尚未理解。",'
|
||||
'"trajectory":"从可能插入的工作变成明确承诺。",'
|
||||
'"possible_moves":["阅读核心方案"],'
|
||||
'"relevance":1,'
|
||||
f'"supporting_fragment_ids":["{earlier["id"]}"]'
|
||||
"}]}"
|
||||
)
|
||||
)
|
||||
|
||||
monkeypatch.setattr(Runner, "run", fake_run)
|
||||
curator = Curator(database, settings)
|
||||
|
||||
asyncio.run(curator.analyze(current["id"]))
|
||||
|
||||
idea = database.get_idea(
|
||||
"admin", database.list_ideas("admin")[0]["id"]
|
||||
)
|
||||
assert idea["title"] == "小爱新架构开发"
|
||||
assert [item["id"] for item in idea["fragments"]] == [
|
||||
earlier["id"],
|
||||
current["id"],
|
||||
]
|
||||
|
||||
+60
-1
@@ -25,7 +25,11 @@ def database(tmp_path: Path) -> Database:
|
||||
return db
|
||||
|
||||
|
||||
def assessment(idea_id: str | None = None, maturity: float = 38):
|
||||
def assessment(
|
||||
idea_id: str | None = None,
|
||||
maturity: float = 38,
|
||||
supporting_fragment_ids: list[str] | None = None,
|
||||
):
|
||||
return {
|
||||
"idea_id": idea_id,
|
||||
"title": "无分类的笔记入口",
|
||||
@@ -38,6 +42,7 @@ def assessment(idea_id: str | None = None, maturity: float = 38):
|
||||
"trajectory": "从输入摩擦的感受，收束成了产品边界。",
|
||||
"possible_moves": ["做一个只保留原文的输入原型"],
|
||||
"relevance": 1,
|
||||
"supporting_fragment_ids": supporting_fragment_ids or [],
|
||||
}
|
||||
|
||||
|
||||
@@ -119,6 +124,60 @@ def test_curator_prompt_contains_recent_continuity_and_idea_catalog(
|
||||
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
|
||||
assert "supporting_fragment_ids" in prompt
|
||||
assert "是否形成新的未完成方向" in prompt
|
||||
|
||||
|
||||
def test_later_fragment_can_recover_earlier_standalone_evidence(
|
||||
tmp_path: Path,
|
||||
):
|
||||
db = database(tmp_path)
|
||||
earlier = db.create_fragment(
|
||||
ADMIN.id, "也许该重新设计小爱的架构，但现在还说不清。"
|
||||
)
|
||||
db.apply_assessments(ADMIN.id, earlier["id"], [])
|
||||
current = db.create_fragment(
|
||||
ADMIN.id, "这个新工作确定要做，我先熟悉 harness 5.0。"
|
||||
)
|
||||
|
||||
db.apply_assessments(
|
||||
ADMIN.id,
|
||||
current["id"],
|
||||
[assessment(supporting_fragment_ids=[earlier["id"]])],
|
||||
)
|
||||
|
||||
idea = db.get_idea(ADMIN.id, db.list_ideas(ADMIN.id)[0]["id"])
|
||||
assert [item["content"] for item in idea["fragments"]] == [
|
||||
earlier["content"],
|
||||
current["content"],
|
||||
]
|
||||
assert idea["snapshots"][-1]["fragment_ids"] == [
|
||||
earlier["id"],
|
||||
current["id"],
|
||||
]
|
||||
|
||||
|
||||
def test_recovered_evidence_cannot_cross_user_boundary(tmp_path: Path):
|
||||
db = database(tmp_path)
|
||||
member_fragment = db.create_fragment(MEMBER.id, "另一个人的记录")
|
||||
current = db.create_fragment(ADMIN.id, "管理员的新方向")
|
||||
|
||||
try:
|
||||
db.apply_assessments(
|
||||
ADMIN.id,
|
||||
current["id"],
|
||||
[
|
||||
assessment(
|
||||
supporting_fragment_ids=[member_fragment["id"]]
|
||||
)
|
||||
],
|
||||
)
|
||||
except ValueError as exc:
|
||||
assert str(exc) == "supporting fragment not found for user"
|
||||
else:
|
||||
raise AssertionError("cross-user supporting fragment was accepted")
|
||||
|
||||
assert db.list_ideas(ADMIN.id) == []
|
||||
|
||||
|
||||
def test_delayed_reanalysis_keeps_later_input_out_of_prior_context(
|
||||
|
||||
Reference in New Issue
Block a user