0afe787c95
* fix: prevent authority-invalid HY retries Co-authored-by: Cursor <cursoragent@cursor.com> * fix: align coding authority with conversations Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: alvinfei <alvinfei@tencent.com> Co-authored-by: Cursor <cursoragent@cursor.com>
906 lines
40 KiB
Python
906 lines
40 KiB
Python
# -*- coding: utf-8 -*-
|
||
"""会话级事实 reconciler 测试：INSERT / 同任务推进 SUPERSEDE / 经验链过继"""
|
||
|
||
import json
|
||
import re
|
||
from datetime import datetime
|
||
|
||
import pytest
|
||
|
||
from hy_memory.coding.fact_reconciler import (
|
||
CHAIN_SUMMARY_PROMPT,
|
||
FACT_RECONCILE_PROMPT,
|
||
WorkingFactReconciler,
|
||
)
|
||
from hy_memory.coding.types import CodingMemory, WorkingFact, WorkingFactDraft
|
||
|
||
from tests.coding._helpers import FailingLLMProvider, MockLLMProvider
|
||
|
||
|
||
def _ops_resp(ops):
|
||
return json.dumps({"ops": ops})
|
||
|
||
|
||
# 发生 SUPERSEDE 时 reconciler 会再要一次 LLM 写演变总结，所以 mock 得备第二条
|
||
# 回复。不备的话总结那次会拿到 ops JSON 原文，测试现象会很迷惑。
|
||
_SUMMARY = "为修复 CI 构建，先定位 Node 18 错误，再升级 workflow；当前构建已通过。"
|
||
|
||
|
||
def _llm(ops, summary=_SUMMARY):
|
||
return MockLLMProvider([_ops_resp(ops), summary])
|
||
|
||
|
||
def _draft(statement, files=None):
|
||
return WorkingFactDraft(statement=statement, files=files or [])
|
||
|
||
|
||
class TestPromptEvidenceContract:
|
||
"""只验证 reconciler prompt 契约，不把它冒充为 LLM 行为证明。"""
|
||
|
||
def test_candidate_judgement_does_not_upgrade_conjecture(self):
|
||
assert "evidence-bounded text" in FACT_RECONCILE_PROMPT
|
||
assert "shared conjecture/error" in FACT_RECONCILE_PROMPT
|
||
assert "only about task continuation" in FACT_RECONCILE_PROMPT
|
||
|
||
def test_chain_summary_preserves_uncertainty_and_blockers(self):
|
||
assert "This is compression only" in CHAIN_SUMMARY_PROMPT
|
||
assert "inferred is not established" in CHAIN_SUMMARY_PROMPT
|
||
assert "unresolved stays unresolved" in CHAIN_SUMMARY_PROMPT
|
||
assert "search/permission/coverage limits" in CHAIN_SUMMARY_PROMPT
|
||
assert "Never upgrade it" in CHAIN_SUMMARY_PROMPT
|
||
assert "directly says that cause was verified" in CHAIN_SUMMARY_PROMPT
|
||
|
||
def test_chain_summary_preserves_stages_and_current_status(self):
|
||
assert "every important stage in chronological order" in CHAIN_SUMMARY_PROMPT
|
||
assert "actual action or change" in CHAIN_SUMMARY_PROMPT
|
||
assert "direct result" in CHAIN_SUMMARY_PROMPT
|
||
assert "unresolved issue or limiting condition" in CHAIN_SUMMARY_PROMPT
|
||
assert "explicit current status" in CHAIN_SUMMARY_PROMPT
|
||
assert "chains of 2-3 generations" in CHAIN_SUMMARY_PROMPT
|
||
assert "2-4 compact paragraphs" in CHAIN_SUMMARY_PROMPT
|
||
assert "Do not mechanically copy the full history" in CHAIN_SUMMARY_PROMPT
|
||
assert "Do not invent stage headings or dates" in CHAIN_SUMMARY_PROMPT
|
||
assert "earlier history was omitted" in CHAIN_SUMMARY_PROMPT
|
||
assert "important numeric values" in CHAIN_SUMMARY_PROMPT
|
||
|
||
|
||
@pytest.mark.asyncio
|
||
class TestInsert:
|
||
async def test_insert_all_drafts(self, coding_store, mock_embed):
|
||
llm = MockLLMProvider(_ops_resp([]))
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
facts, ops = await rec.reconcile(
|
||
[_draft("事实 A", ["a.py"]), _draft("事实 B")],
|
||
user_id="u1", agent_id="ag1", session_id="s1", workspace_id="ws/a",
|
||
)
|
||
|
||
assert [f.statement for f in facts] == ["事实 A", "事实 B"]
|
||
assert [o.action for o in ops] == ["INSERT", "INSERT"]
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert stored is not None
|
||
assert stored.status == "active"
|
||
assert stored.files == ["a.py"]
|
||
assert stored.agent_id == "ag1"
|
||
assert stored.session_id == "s1"
|
||
assert stored.workspace_id == "ws/a"
|
||
assert stored.fact_link is None
|
||
|
||
async def test_no_llm_call_without_candidates(self, coding_store, mock_embed):
|
||
"""库里一条候选都没有就没得判，不该白花一次 LLM 调用。"""
|
||
llm = MockLLMProvider(_ops_resp([]))
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
await rec.reconcile([_draft("第一条事实")], user_id="u1")
|
||
assert len(llm.calls) == 0
|
||
|
||
async def test_empty_drafts(self, coding_store, mock_embed):
|
||
llm = MockLLMProvider(_ops_resp([]))
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
facts, ops = await rec.reconcile([], user_id="u1")
|
||
assert facts == []
|
||
assert ops == []
|
||
|
||
|
||
@pytest.mark.asyncio
|
||
class TestSupersede:
|
||
async def _seed_old(self, store, *, derived=None):
|
||
old = WorkingFact(
|
||
fact_id="f_old", user_id="u1",
|
||
session_id="test-session",
|
||
statement=(
|
||
"用户要修复 CI 的 npm run build；检查 .github/workflows/ci.yml 后"
|
||
"定位到 Node 18 报 ERR_OSSL_EVP_UNSUPPORTED，尚未修改。"
|
||
),
|
||
derived_experience_ids=derived or [],
|
||
)
|
||
await store.insert_fact(old)
|
||
return old
|
||
|
||
async def test_supersede_marks_old_and_links(self, coding_store, mock_embed):
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([{
|
||
"draft_idx": 0,
|
||
"supersedes": ["f_old"],
|
||
"reason": "继续推进同一个 CI 构建修复",
|
||
}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
facts, ops = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续修复同一 CI 构建；把 .github/workflows/ci.yml 升到 Node 20，"
|
||
"npm run build 已通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
new_id = facts[0].fact_id
|
||
old = await coding_store.get_fact("f_old", user_id="u1")
|
||
assert old.status == "superseded"
|
||
assert old.fact_link == new_id
|
||
assert [o.action for o in ops] == ["SUPERSEDE", "INSERT"]
|
||
|
||
async def test_experience_chain_inherited(self, coding_store, mock_embed):
|
||
"""旧事实攒下的经验链要过继给新事实，两边都补上。"""
|
||
await coding_store.insert(CodingMemory(
|
||
experience_id="e1", user_id="u1", task="改 config", solution="S",
|
||
source_fact_ids=["f_old"],
|
||
))
|
||
await self._seed_old(coding_store, derived=["e1"])
|
||
llm = _llm([{"draft_idx": 0, "supersedes": ["f_old"]}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
facts, _ = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续修复同一 CI 构建；把 .github/workflows/ci.yml 升到 Node 20，"
|
||
"npm run build 已通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
new_id = facts[0].fact_id
|
||
|
||
# 正向：新事实拿到旧事实的经验链
|
||
assert facts[0].derived_experience_ids == ["e1"]
|
||
stored_new = await coding_store.get_fact(new_id, user_id="u1")
|
||
assert stored_new.derived_experience_ids == ["e1"]
|
||
# 反向：经验补上新事实，旧的也留着（链路可回溯）
|
||
exp = await coding_store.get_by_id("e1", user_id="u1")
|
||
assert exp.source_fact_ids == ["f_old", new_id]
|
||
|
||
async def test_link_chain_stays_walkable(self, coding_store, mock_embed):
|
||
"""历史全靠 fact_link 反查还原，所以连续覆盖必须一跳不断。"""
|
||
await self._seed_old(coding_store) # 第 1 代
|
||
gens = [
|
||
(
|
||
"用户继续修复同一 CI 构建；将 .github/workflows/ci.yml 升到 Node 20，"
|
||
"但 npm run build 仍受旧缓存影响。",
|
||
"f_old",
|
||
),
|
||
(
|
||
"用户继续验证同一 CI 修复；清理 npm 缓存后 npm run build 通过。",
|
||
None,
|
||
),
|
||
]
|
||
prev_id = None
|
||
for stmt, target in gens:
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": [target or prev_id]}]),
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[_draft(stmt)], user_id="u1", session_id="test-session"
|
||
)
|
||
prev_id = facts[0].fact_id
|
||
|
||
lineage = await coding_store.get_fact_ancestors(prev_id, user_id="u1")
|
||
assert [f.statement for f in lineage] == [
|
||
"用户要修复 CI 的 npm run build；检查 .github/workflows/ci.yml 后"
|
||
"定位到 Node 18 报 ERR_OSSL_EVP_UNSUPPORTED，尚未修改。",
|
||
"用户继续修复同一 CI 构建；将 .github/workflows/ci.yml 升到 Node 20，"
|
||
"但 npm run build 仍受旧缓存影响。",
|
||
]
|
||
|
||
async def test_reconcile_prompt_requires_same_task_not_same_file(
|
||
self, coding_store, mock_embed
|
||
):
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
await rec.reconcile(
|
||
[_draft(
|
||
"用户要给 .github/workflows/ci.yml 增加发布 job；添加 twine upload，"
|
||
"配置完成但尚未发布。"
|
||
)],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
|
||
prompt = llm.calls[0]["prompt"]
|
||
assert "继续推进某条已有 session fact 所记录的同一个用户" in prompt
|
||
assert "只是提到了相同 file" in prompt
|
||
assert "追求不同的用户目标" in prompt
|
||
assert "INSERT 是安全默认值" in prompt
|
||
assert "ERR_OSSL_EVP_UNSUPPORTED" in prompt
|
||
assert "twine upload" in prompt
|
||
|
||
async def test_summary_empty_when_nothing_superseded(self, coding_store, mock_embed):
|
||
"""没覆盖任何东西就没有演变可讲。"""
|
||
llm = MockLLMProvider(_ops_resp([]))
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
facts, _ = await rec.reconcile([_draft("全新的结论")], user_id="u1")
|
||
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert stored.fact_chain_summary == ""
|
||
|
||
async def test_old_statement_not_indexed_for_retrieval(self, coding_store, mock_embed):
|
||
"""被顶掉的说法留在库里但撤出向量库，不该重新具备召回力。"""
|
||
await self._seed_old(coding_store)
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed, _llm([{"draft_idx": 0, "supersedes": ["f_old"]}]),
|
||
)
|
||
await rec.reconcile(
|
||
[_draft("config 位于 /etc/app/config.yaml")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
|
||
# 拿旧说法原文去搜，只能落到新事实（靠两者本身相似），
|
||
# 且召回的 statement 一定是当前值，不是血统里那句
|
||
emb = await coding_store.embed_service.embed("config 位于 /opt/app/config.yaml")
|
||
hits = await coding_store.search_facts(emb, user_id="u1", top=5)
|
||
assert all(f.statement != "config 位于 /opt/app/config.yaml" for f, _ in hits)
|
||
|
||
async def test_ignores_id_outside_candidates(self, coding_store, mock_embed):
|
||
"""LLM 幻觉出一个不在候选里的 fact_id，不能真去动它。"""
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([{"draft_idx": 0, "supersedes": ["f_hallucinated"]}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
_, ops = await rec.reconcile(
|
||
[_draft("config 位于 /etc/app/config.yaml")], user_id="u1",
|
||
)
|
||
assert [o.action for o in ops] == ["INSERT"]
|
||
old = await coding_store.get_fact("f_old", user_id="u1")
|
||
assert old.status == "active"
|
||
|
||
async def test_near_duplicate_fact_cannot_cross_authority_scope(
|
||
self, coding_store, mock_embed
|
||
):
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old",
|
||
user_id="u1",
|
||
agent_id="scope:2026-10-18",
|
||
session_id="conversation-a",
|
||
workspace_id="repo-a",
|
||
statement="CI build now passes after upgrading Node",
|
||
))
|
||
llm = _llm([{"draft_idx": 0, "supersedes": ["f_old"]}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
_, ops = await rec.reconcile(
|
||
[_draft("CI build now passes after upgrading Node")],
|
||
user_id="u1",
|
||
agent_id="scope:2026-10-19",
|
||
session_id="conversation-b",
|
||
workspace_id="repo-a",
|
||
)
|
||
|
||
old = await coding_store.get_fact("f_old", user_id="u1")
|
||
assert old is not None
|
||
assert old.status == "active"
|
||
assert [op.action for op in ops] == ["INSERT"]
|
||
assert len(llm.calls) == 0
|
||
|
||
async def test_old_fact_superseded_only_once(self, coding_store, mock_embed):
|
||
"""两条新事实都声称覆盖同一条旧事实时，只让第一条生效。
|
||
|
||
否则 fact_link 会被后写的那条覆盖，前一条的过继链就断了。
|
||
"""
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([
|
||
{"draft_idx": 0, "supersedes": ["f_old"]},
|
||
{"draft_idx": 1, "supersedes": ["f_old"]},
|
||
])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
facts, ops = await rec.reconcile(
|
||
[
|
||
_draft("config 位于 /etc/app/config.yaml"),
|
||
_draft("config 位于 /srv/app/config.yaml"),
|
||
],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
assert len(facts) == 2
|
||
assert sum(1 for o in ops if o.action == "SUPERSEDE") == 1
|
||
old = await coding_store.get_fact("f_old", user_id="u1")
|
||
assert old.fact_link == facts[0].fact_id
|
||
|
||
|
||
@pytest.mark.asyncio
|
||
class TestChainSummary:
|
||
"""同一任务进展总结由 LLM 写，历史现取自 fact_link 链，不另存。"""
|
||
|
||
async def _seed_old(self, store):
|
||
await store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1",
|
||
session_id="test-session",
|
||
statement=(
|
||
"用户要迁移 app config；确认当前读取 /opt/app/config.yaml，"
|
||
"尚未修改配置加载逻辑。"
|
||
),
|
||
created_at=datetime(2026, 1, 1),
|
||
))
|
||
|
||
async def test_llm_summary_stored_and_used(self, coding_store, mock_embed):
|
||
await self._seed_old(coding_store)
|
||
written = (
|
||
"为迁移 app config，先确认 /opt/app/config.yaml，再切到 "
|
||
"/etc/app/config.yaml；当前加载测试通过。"
|
||
)
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_old"]}], summary=written),
|
||
)
|
||
|
||
facts, _ = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert stored.fact_chain_summary == written
|
||
|
||
async def test_prompt_carries_history_and_current(self, coding_store, mock_embed):
|
||
"""总结调用要同时看到旧 session 和当前 session，才能写任务进展。"""
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([{"draft_idx": 0, "supersedes": ["f_old"]}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
assert len(llm.calls) == 2, "SUPERSEDE 应触发第二次调用"
|
||
prompt = llm.calls[1]["prompt"]
|
||
assert "/opt/app/config.yaml" in prompt
|
||
assert "/etc/app/config.yaml" in prompt
|
||
assert "goal" in prompt
|
||
assert "逐阶段 progression" in prompt
|
||
assert "current status" in prompt
|
||
assert llm.calls[1]["max_tokens"] == 600
|
||
|
||
async def test_no_extra_call_without_supersede(self, coding_store, mock_embed):
|
||
"""没覆盖就没有演变可总结，不该白花一次调用。"""
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
await rec.reconcile(
|
||
[_draft("另一个话题的结论")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
assert len(llm.calls) == 1
|
||
|
||
async def test_prompt_carries_full_lineage(self, coding_store, mock_embed):
|
||
"""第三代的总结要看到头两代，不能只看到刚被顶掉的那一条。"""
|
||
await self._seed_old(coding_store)
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed, _llm([{"draft_idx": 0, "supersedes": ["f_old"]}]),
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
llm2 = _llm([{"draft_idx": 0, "supersedes": [facts[0].fact_id]}])
|
||
rec2 = WorkingFactReconciler(coding_store, mock_embed, llm2)
|
||
await rec2.reconcile(
|
||
[_draft(
|
||
"用户继续完善同一 config 迁移；改由环境变量 APP_CONFIG 指定路径，"
|
||
"回归测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
prompt = llm2.calls[1]["prompt"]
|
||
assert "/opt/app/config.yaml" in prompt, "第一代不该丢"
|
||
assert "/etc/app/config.yaml" in prompt
|
||
assert "APP_CONFIG" in prompt
|
||
assert prompt.index("/opt/") < prompt.index("/etc/"), "时间正序"
|
||
|
||
async def test_summary_failure_does_not_block_write(self, coding_store, mock_embed):
|
||
"""总结挂了，覆盖和入库照常；历史还在库里，下次覆盖会重新总结出来。"""
|
||
await self._seed_old(coding_store)
|
||
|
||
class _SummaryFails(MockLLMProvider):
|
||
async def complete(self, prompt, **kwargs):
|
||
if "ENTRIES" in prompt:
|
||
raise RuntimeError("summary boom")
|
||
return await super().complete(prompt, **kwargs)
|
||
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_SummaryFails(_ops_resp([{"draft_idx": 0, "supersedes": ["f_old"]}])),
|
||
)
|
||
facts, ops = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
assert sum(1 for o in ops if o.action == "SUPERSEDE") == 1
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert stored.fact_chain_summary == ""
|
||
# 关键：这一代没写成总结，但血统没断，还能查回去
|
||
lineage = await coding_store.get_fact_ancestors(stored.fact_id, user_id="u1")
|
||
assert [f.statement for f in lineage] == [
|
||
"用户要迁移 app config；确认当前读取 /opt/app/config.yaml，"
|
||
"尚未修改配置加载逻辑。"
|
||
]
|
||
|
||
async def test_summary_recovers_after_a_failed_generation(
|
||
self, coding_store, mock_embed
|
||
):
|
||
"""某一代没写成总结，下一代重走链子能把漏掉的那代补回来。"""
|
||
await self._seed_old(coding_store)
|
||
|
||
class _SummaryFails(MockLLMProvider):
|
||
async def complete(self, prompt, **kwargs):
|
||
if "ENTRIES" in prompt:
|
||
raise RuntimeError("summary boom")
|
||
return await super().complete(prompt, **kwargs)
|
||
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_SummaryFails(_ops_resp([{"draft_idx": 0, "supersedes": ["f_old"]}])),
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
assert facts[0].fact_chain_summary == ""
|
||
|
||
llm2 = _llm([{"draft_idx": 0, "supersedes": [facts[0].fact_id]}])
|
||
rec2 = WorkingFactReconciler(coding_store, mock_embed, llm2)
|
||
await rec2.reconcile(
|
||
[_draft(
|
||
"用户继续完善同一 config 迁移；改由环境变量 APP_CONFIG 指定路径，"
|
||
"回归测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
prompt = llm2.calls[1]["prompt"]
|
||
assert "/opt/app/config.yaml" in prompt, "上一代漏写的历史要能补回来"
|
||
|
||
async def test_deep_lineage_truncated_and_flagged(self, coding_store, mock_embed):
|
||
"""长命事实不该把整条链塞进提示词，但得说明省略了几代。"""
|
||
from hy_memory.coding.types import MAX_FACT_CHAIN_ENTRIES
|
||
|
||
prev = None
|
||
total = MAX_FACT_CHAIN_ENTRIES + 3
|
||
for i in range(total):
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id=f"g{i}", user_id="u1", session_id="test-session",
|
||
statement=f"第 {i} 代结论",
|
||
created_at=datetime(2026, 1, i + 1),
|
||
))
|
||
if prev:
|
||
await coding_store.supersede_fact(prev, f"g{i}")
|
||
prev = f"g{i}"
|
||
|
||
llm = _llm([{"draft_idx": 0, "supersedes": [prev]}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
await rec.reconcile(
|
||
[_draft("最新结论")], user_id="u1", session_id="test-session"
|
||
)
|
||
|
||
prompt = llm.calls[1]["prompt"]
|
||
assert "第 0 代结论" not in prompt, "最老的几代该被截掉"
|
||
assert f"第 {total - 1} 代结论" in prompt, "最近的要留着"
|
||
assert "omitted" in prompt, "截断了就得标出来"
|
||
# 别报代数。反查本身就只走 MAX+1 步，数不出真实代数，写个数字等于骗
|
||
# 下游——第 30 代覆盖时也只会算出"1 代"，而实际略掉了二十多代。
|
||
assert not re.search(r"\d+\s+earlier", prompt), "截断标注不该带代数"
|
||
|
||
async def test_no_truncation_flag_on_short_lineage(self, coding_store, mock_embed):
|
||
"""没截断就不该出现省略标注，否则 LLM 会以为还有没给它的历史。"""
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([{"draft_idx": 0, "supersedes": ["f_old"]}])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
entries = llm.calls[1]["prompt"].split(
|
||
"ENTRIES\n═══════════════════════════════════════════════════════════════════\n",
|
||
1,
|
||
)[1]
|
||
assert "older revisions omitted" not in entries
|
||
|
||
async def test_summary_rewritten_each_generation(self, coding_store, mock_embed):
|
||
"""第二代覆盖时总结要重写，而不是留着第一代那句。"""
|
||
await self._seed_old(coding_store)
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_old"]}], summary="第一版总结"),
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
rec2 = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": [facts[0].fact_id]}],
|
||
summary="第二版总结"),
|
||
)
|
||
facts2, _ = await rec2.reconcile(
|
||
[_draft(
|
||
"用户继续完善同一 config 迁移；改由环境变量 APP_CONFIG 指定路径，"
|
||
"回归测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
stored = await coding_store.get_fact(facts2[0].fact_id, user_id="u1")
|
||
assert stored.fact_chain_summary == "第二版总结"
|
||
|
||
async def test_disabled_skips_the_call(self, coding_store, mock_embed):
|
||
"""关掉之后不该再多花一次调用，覆盖本身照常。"""
|
||
await self._seed_old(coding_store)
|
||
llm = _llm([{"draft_idx": 0, "supersedes": ["f_old"]}])
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed, llm, enable_chain_summary=False,
|
||
)
|
||
facts, ops = await rec.reconcile(
|
||
[_draft(
|
||
"用户继续迁移 app config；将加载路径改为 /etc/app/config.yaml，"
|
||
"配置加载测试通过。"
|
||
)], user_id="u1", session_id="test-session",
|
||
)
|
||
|
||
assert len(llm.calls) == 1
|
||
assert sum(1 for o in ops if o.action == "SUPERSEDE") == 1
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert stored.fact_chain_summary == ""
|
||
|
||
|
||
class TestCleanSummary:
|
||
"""模型爱给总结加装饰，进库前剥掉。"""
|
||
|
||
@pytest.mark.parametrize("raw,expected", [
|
||
("```\n一句总结。\n```", "一句总结。"),
|
||
("```markdown\n一句总结。\n```", "一句总结。"),
|
||
("总结：一句总结。", "一句总结。"),
|
||
("Summary: 一句总结。", "一句总结。"),
|
||
("“一句总结。”", "一句总结。"),
|
||
('"一句总结。"', "一句总结。"),
|
||
("一句\n总结。", "一句 总结。"),
|
||
(" 一句总结。 ", "一句总结。"),
|
||
("", ""),
|
||
("```", ""),
|
||
])
|
||
def test_strips_decoration(self, raw, expected):
|
||
assert WorkingFactReconciler._clean_summary(raw) == expected
|
||
|
||
def test_keeps_inner_quotes(self):
|
||
"""只剥最外层的成对引号，句子里引的路径名不该被削掉。"""
|
||
s = WorkingFactReconciler._clean_summary('现在读 "APP_CONFIG" 这个变量。')
|
||
assert s == '现在读 "APP_CONFIG" 这个变量。'
|
||
|
||
|
||
@pytest.mark.asyncio
|
||
class TestFailSafe:
|
||
async def test_llm_failure_still_inserts(self, coding_store, mock_embed):
|
||
"""LLM 挂了宁可多留一条过期事实，也不能因为一次调用失败丢结论。"""
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", session_id="test-session",
|
||
statement="config 位于 /opt/app/config.yaml",
|
||
))
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, FailingLLMProvider())
|
||
|
||
facts, ops = await rec.reconcile(
|
||
[_draft("config 位于 /etc/app/config.yaml")], user_id="u1",
|
||
)
|
||
assert len(facts) == 1
|
||
assert [o.action for o in ops] == ["INSERT"]
|
||
old = await coding_store.get_fact("f_old", user_id="u1")
|
||
assert old.status == "active"
|
||
|
||
async def test_unparsable_llm_output(self, coding_store, mock_embed):
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", session_id="test-session",
|
||
statement="config 位于 /opt/app/config.yaml",
|
||
))
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed, MockLLMProvider("不是 JSON 的一段话")
|
||
)
|
||
|
||
facts, ops = await rec.reconcile(
|
||
[_draft("config 位于 /etc/app/config.yaml")], user_id="u1",
|
||
)
|
||
assert len(facts) == 1
|
||
assert [o.action for o in ops] == ["INSERT"]
|
||
|
||
async def test_markdown_fence_parsed(self, coding_store, mock_embed):
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", session_id="test-session",
|
||
statement="config 位于 /opt/app/config.yaml",
|
||
))
|
||
llm = MockLLMProvider([
|
||
"```json\n" + _ops_resp([{"draft_idx": 0, "supersedes": ["f_old"]}]) + "\n```",
|
||
_SUMMARY,
|
||
])
|
||
rec = WorkingFactReconciler(coding_store, mock_embed, llm)
|
||
|
||
_, ops = await rec.reconcile(
|
||
[_draft("config 位于 /etc/app/config.yaml")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
assert sum(1 for o in ops if o.action == "SUPERSEDE") == 1
|
||
|
||
|
||
@pytest.mark.asyncio
|
||
class TestSubjectGateSwitch:
|
||
"""subject 机械闸门的开关。
|
||
|
||
这个闸门是精确字符串相等，它的正确性完全取决于 subject 是不是稳定实体名。
|
||
Travel 是（人名，实测 933 组被拦对里明确误判 0 组）；Progressive Search 不是
|
||
（59.2% 是 'Subquery N' 这种位置标签，两个位置标签必然不相等所以必然互拦）。
|
||
所以它必须能按数据集关掉，而不是二选一定死。
|
||
"""
|
||
|
||
async def _seed(self, store, fact_id, subject, statement):
|
||
await store.insert_fact(WorkingFact(
|
||
fact_id=fact_id, user_id="u1", session_id="test-session",
|
||
subject=subject, statement=statement,
|
||
))
|
||
|
||
async def test_gate_on_blocks_cross_subject_supersede(self, coding_store, mock_embed):
|
||
"""开启时：主体不同的候选进不了 LLM 名单，所以覆盖不可能发生。"""
|
||
await self._seed(coding_store, "f_emma", "Emma", "Emma 的三日行程已排定，预算待定。")
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_emma"]}]),
|
||
subject_gate=True,
|
||
)
|
||
_, ops = await rec.reconcile(
|
||
[WorkingFactDraft(subject="Bart", statement="Bart 的三日行程已排定，预算待定。")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
assert [o.action for o in ops] == ["INSERT"]
|
||
assert (await coding_store.get_fact("f_emma", user_id="u1")).status == "active"
|
||
|
||
async def test_gate_off_allows_cross_subject_supersede(self, coding_store, mock_embed):
|
||
"""关闭时：候选原样交给 LLM，它判覆盖就真覆盖。"""
|
||
await self._seed(coding_store, "f_emma", "Emma", "Emma 的三日行程已排定，预算待定。")
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_emma"]}]),
|
||
subject_gate=False,
|
||
)
|
||
facts, ops = await rec.reconcile(
|
||
[WorkingFactDraft(subject="Bart", statement="Bart 的三日行程已排定，预算待定。")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
assert [o.action for o in ops] == ["SUPERSEDE", "INSERT"]
|
||
old = await coding_store.get_fact("f_emma", user_id="u1")
|
||
assert old.status == "superseded"
|
||
assert old.fact_link == facts[0].fact_id
|
||
|
||
async def test_gate_off_skips_pre_execute_recheck(self, coding_store, mock_embed):
|
||
"""关闭时执行前那道复查也要一起跳过，否则闸门没关干净。
|
||
|
||
直接调 _execute 绕过召回，模拟"候选闸门放行了但复查还在拦"这个漏关场景。
|
||
"""
|
||
await self._seed(coding_store, "f_emma", "Emma", "Emma 的行程。")
|
||
old = await coding_store.get_fact("f_emma", user_id="u1")
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed, _llm([]), subject_gate=False,
|
||
)
|
||
_, ops = await rec._execute(
|
||
[WorkingFactDraft(subject="Bart", statement="Bart 的行程。")],
|
||
[[old]],
|
||
{0: ["f_emma"]},
|
||
user_id="u1", agent_id="default_agent",
|
||
session_id="test-session", workspace_id=None,
|
||
)
|
||
assert "SUPERSEDE" in [o.action for o in ops]
|
||
|
||
async def test_empty_subject_never_blocked_either_way(self, coding_store, mock_embed):
|
||
"""有一侧 subject 为空时，开关无论怎么设都不该拦。
|
||
|
||
空不等于"主体不同"，等于"这次没能定出主体"（兜底存的原文、改动前的老行）。
|
||
拿一个未知去否决一个判断是错的，该交回 LLM 按语义判。
|
||
"""
|
||
for gate in (True, False):
|
||
await self._seed(coding_store, f"f_noneed_{gate}", "", "旧的部署脚本结论。")
|
||
old = await coding_store.get_fact(f"f_noneed_{gate}", user_id="u1")
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed, _llm([]), subject_gate=gate,
|
||
)
|
||
_, ops = await rec._execute(
|
||
[WorkingFactDraft(subject="deploy.sh", statement="新的部署脚本结论。")],
|
||
[[old]],
|
||
{0: [f"f_noneed_{gate}"]},
|
||
user_id="u1", agent_id="default_agent",
|
||
session_id="test-session", workspace_id=None,
|
||
)
|
||
assert "SUPERSEDE" in [o.action for o in ops], f"gate={gate} 时空 subject 被误拦"
|
||
|
||
async def test_env_var_controls_default(self, coding_store, mock_embed, monkeypatch):
|
||
"""环境变量能切默认值，显式构造参数优先于它。"""
|
||
monkeypatch.setenv("HY_FACT_SUBJECT_GATE", "0")
|
||
assert WorkingFactReconciler(coding_store, mock_embed, _llm([])).subject_gate is False
|
||
assert WorkingFactReconciler(
|
||
coding_store, mock_embed, _llm([]), subject_gate=True
|
||
).subject_gate is True
|
||
monkeypatch.setenv("HY_FACT_SUBJECT_GATE", "1")
|
||
assert WorkingFactReconciler(coding_store, mock_embed, _llm([])).subject_gate is True
|
||
|
||
|
||
@pytest.mark.asyncio
|
||
class TestDetailsInheritance:
|
||
"""被覆盖事实的 details 过继。
|
||
|
||
fact_chain_summary 是 LLM 压缩产物，实测保真度按链距衰减（一条 3 代链里 gen2
|
||
槽位保留 5/5，gen1 保留 0/2，丢的正是具体数值）。过继走复制不走生成，所以
|
||
预算内 100% 保真；且 details 不进向量，追加不改变召回打分。
|
||
"""
|
||
|
||
async def test_ancestor_details_copied_verbatim(self, coding_store, mock_embed):
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", subject="Emma",
|
||
session_id="test-session",
|
||
statement="Emma 的行程草案已出，预算待核。",
|
||
details="日均餐费 $32.4-$57.6\n酒店评分 3.2-3.8",
|
||
))
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_old"]}]),
|
||
subject_gate=False, inherit_details=True,
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[WorkingFactDraft(subject="Emma", statement="Emma 的行程已定稿。",
|
||
details="航班 CA1234")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
# 本代的 details 在，祖先的具体取值也一字不差地在
|
||
assert "航班 CA1234" in stored.details
|
||
assert "$32.4-$57.6" in stored.details
|
||
assert "3.2-3.8" in stored.details
|
||
assert "gen1" in stored.details and "Emma" in stored.details
|
||
|
||
async def test_no_exponential_growth_down_the_chain(self, coding_store, mock_embed):
|
||
"""连续覆盖三代，details 不该按代翻倍。
|
||
|
||
祖先的 details 里已经含着它自己那代过继来的块，直接追加会指数膨胀。
|
||
血统由 _collect_history 一次走全，所以每代只贡献自己那一份。
|
||
"""
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_g1", user_id="u1", subject="T",
|
||
session_id="test-session",
|
||
statement="第一代结论。", details="V1=" + "a" * 300,
|
||
))
|
||
prev = "f_g1"
|
||
sizes = []
|
||
for gen in range(2, 5):
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": [prev]}]),
|
||
subject_gate=False, inherit_details=True,
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[WorkingFactDraft(subject="T", statement=f"第{gen}代结论。",
|
||
details=f"V{gen}=" + "a" * 300)],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
prev = facts[0].fact_id
|
||
stored = await coding_store.get_fact(prev, user_id="u1")
|
||
sizes.append(len(stored.details))
|
||
# 每代的取值标记只出现一次，没有被重复过继
|
||
for g in range(1, gen + 1):
|
||
assert stored.details.count(f"V{g}=") == 1, (
|
||
f"第{gen}代里 V{g}= 出现 {stored.details.count(f'V{g}=')} 次，重复过继了"
|
||
)
|
||
# 线性增长：4 代最多约 4×(300+标记)，指数增长会是 300×2^3 以上
|
||
assert sizes[-1] < 2200, f"details 增长过快: {sizes}"
|
||
|
||
async def test_total_budget_enforced_and_loss_reported(self, coding_store, mock_embed):
|
||
"""预算是硬的，但被丢掉的代数必须写出来，不静默截断。"""
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", subject="T",
|
||
session_id="test-session",
|
||
statement="旧结论。", details="X=" + "b" * 8000,
|
||
))
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_old"]}]),
|
||
subject_gate=False, inherit_details=True,
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[WorkingFactDraft(subject="T", statement="新结论。", details="Y=1")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert "…(truncated)" in stored.details
|
||
assert len(stored.details) < 6000
|
||
|
||
async def test_switch_off_keeps_details_untouched(self, coding_store, mock_embed):
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", subject="T",
|
||
session_id="test-session",
|
||
statement="旧结论。", details="祖先取值 Z=9",
|
||
))
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_old"]}]),
|
||
subject_gate=False, inherit_details=False,
|
||
)
|
||
facts, ops = await rec.reconcile(
|
||
[WorkingFactDraft(subject="T", statement="新结论。", details="Y=1")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
# 覆盖照样发生，只是不过继
|
||
assert "SUPERSEDE" in [o.action for o in ops]
|
||
stored = await coding_store.get_fact(facts[0].fact_id, user_id="u1")
|
||
assert stored.details == "Y=1"
|
||
assert "Z=9" not in stored.details
|
||
|
||
async def test_inheritance_does_not_change_recall_vector(self, coding_store, mock_embed):
|
||
"""details 不进向量，所以过继多少都不该改变召回打分。
|
||
|
||
这是"追加不伤召回质量"这个论断的直接检查，不是推断。
|
||
"""
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_old", user_id="u1", subject="T",
|
||
session_id="test-session",
|
||
statement="旧结论。", details="无关噪声 " + "q" * 2000,
|
||
))
|
||
rec = WorkingFactReconciler(
|
||
coding_store, mock_embed,
|
||
_llm([{"draft_idx": 0, "supersedes": ["f_old"]}]),
|
||
subject_gate=False, inherit_details=True,
|
||
)
|
||
facts, _ = await rec.reconcile(
|
||
[WorkingFactDraft(subject="T", statement="新结论。", details="Y=1")],
|
||
user_id="u1",
|
||
session_id="test-session",
|
||
)
|
||
new_id = facts[0].fact_id
|
||
emb = await coding_store.embed_service.embed("T 新结论。")
|
||
hits = await coding_store.search_facts(emb, user_id="u1", top=5)
|
||
got = {f.fact_id: s for f, s in hits}
|
||
assert new_id in got
|
||
# 同样的 statement+subject 但不带过继，打分应当一致
|
||
await coding_store.insert_fact(WorkingFact(
|
||
fact_id="f_clean", user_id="u1", subject="T",
|
||
statement="新结论。", details="Y=1",
|
||
))
|
||
hits2 = await coding_store.search_facts(emb, user_id="u1", top=5)
|
||
got2 = {f.fact_id: s for f, s in hits2}
|
||
assert abs(got2["f_clean"] - got2[new_id]) < 1e-6, (
|
||
f"过继改变了召回打分: clean={got2['f_clean']} inherited={got2[new_id]}"
|
||
)
|