foxclaw 0.1.3 → 0.1.4

package/python/frontend/v3/index.html

@@ -6,8 +6,8 @@
   <title>Foxclaw</title>
   <!-- 内联 SVG 图标:不新增静态文件,也避免浏览器探 /favicon.ico 报 404 -->
   <link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 32 32'><rect width='32' height='32' rx='9' fill='%23179d5f'/><circle cx='16' cy='16' r='6.5' fill='%23f4fbf0'/></svg>" />
-  <script type="module" crossorigin src="/v3/assets/index-CQSBdxk4.js"></script>
-  <link rel="stylesheet" crossorigin href="/v3/assets/index-D5nIsjms.css">
+  <script type="module" crossorigin src="/v3/assets/index-tN90dpUH.js"></script>
+  <link rel="stylesheet" crossorigin href="/v3/assets/index-JKfHlb-a.css">
 </head>
 <body>
   <div id="app"></div>

package/python/frontend_src/package.json

@@ -1,7 +1,7 @@
 {
   "name": "foxclaw-deck-next",
   "private": true,
-  "version": "0.1.3",
+  "version": "0.1.4",
   "type": "module",
   "description": "Foxclaw Deck v3 - Vue 3 rendering layer (builds into frontend/v3)",
   "scripts": {

package/python/frontend_src/src/components/panels/MemoryPanel.vue

@@ -1,7 +1,7 @@
 <script setup>
 import { computed, onMounted, ref } from 'vue'
 import PanelBox from './PanelBox.vue'
-import { panels, loadPanel, refreshPanel, dropPreference } from '../../stores/panels.js'
+import { panels, loadPanel, refreshPanel, dropPreference, approveSensitiveWrite, rejectSensitiveWrite } from '../../stores/panels.js'
 import { openDoc } from '../../stores/app.js'
 import { fmtTime, fmtNum, clip } from '../../lib/fmt.js'
 
@@ -31,6 +31,16 @@ async function drop(pref) {
   try { await dropPreference(pref.id) } catch (e) { err.value = String(e.message || e) }
   busy.value = ''
 }
+
+async function decide(item, keep) {
+  busy.value = item.id
+  err.value = ''
+  try {
+    if (keep) await approveSensitiveWrite(item.id)
+    else await rejectSensitiveWrite(item.id)
+  } catch (e) { err.value = String(e.message || e) }
+  busy.value = ''
+}
 </script>
 
 <template>
@@ -47,10 +57,32 @@ async function drop(pref) {
       <div><b>{{ fmtNum(p.stats.shared ?? 0) }}</b><span>共享</span></div>
       <div><b>{{ fmtNum(p.stats.private ?? 0) }}</b><span>私有</span></div>
       <div><b>{{ fmtNum(p.prefs.length) }}</b><span>用户偏好</span></div>
+      <div v-if="p.pending.length"><b>{{ fmtNum(p.pending.length) }}</b><span>待确认</span></div>
     </div>
 
     <div v-if="err" class="pn-st err">{{ err }}</div>
 
+    <div v-if="p.pending.length" class="pn-sec pending">
+      <div class="pn-h">待确认写入 <em>{{ p.pending.length }}</em></div>
+      <div class="pn-empty">
+        这几条含凭据、证件、健康或账户信息,按规矩没有直接写进长期记忆。确认无误再点「记住」。
+      </div>
+      <div v-for="it in p.pending" :key="it.id" class="pn-row">
+        <span class="m">
+          <b>{{ clip(it.content, 60) }}</b>
+          <span class="s">
+            {{ (it.categories || []).join(' · ') }} · {{ it.agent_id || '—' }} · {{ it.id }}
+          </span>
+        </span>
+        <button class="act ok" :disabled="busy === it.id" title="确认写入长期记忆" @click.stop="decide(it, true)">
+          记住
+        </button>
+        <button class="act" :disabled="busy === it.id" title="丢弃,不写入" @click.stop="decide(it, false)">
+          丢弃
+        </button>
+      </div>
+    </div>
+
     <div class="pn-cols">
       <div class="pn-sec">
         <div class="pn-h">最近记忆 <em>{{ p.recent.length }}</em></div>
@@ -100,13 +132,21 @@ async function drop(pref) {
     <div class="pn-note">
       后端 <code>{{ p.stats.backend || '—' }}</code> v{{ p.stats.version || '—' }} ·
       向量 <code>{{ p.stats.embedding || '—' }}</code>。
-      数据来自 <code>/api/memory/stats</code>、<code>/api/memory/recent</code>、<code>/api/preferences</code>。
+      数据来自 <code>/api/memory/stats</code>、<code>/api/memory/recent</code>、<code>/api/preferences</code>、<code>/api/memory/pending</code>。
     </div>
   </PanelBox>
 </template>
 
 <style scoped>
 .sec-gap { margin-top: 14px; }
+.pending .pn-h em { color: var(--amber, #d99a2b); }
+.act {
+  flex-shrink: 0; padding: 4px 10px; border-radius: var(--r-pill);
+  font-size: 11px; font-weight: 600; color: var(--dim);
+  background: var(--surface-2); border: 1px solid var(--line-soft);
+}
+.act.ok { color: var(--accent, #3b82f6); border-color: currentColor; }
+.act:disabled { opacity: 0.5; }
 .del {
   flex-shrink: 0; padding: 4px 10px; border-radius: var(--r-pill);
   font-size: 11px; font-weight: 600; color: var(--dim);

package/python/frontend_src/src/stores/panels.js

@@ -15,7 +15,7 @@ export const panels = reactive({
     recent: [], recentTotal: 0,
     ingest: 'idle', ingestMsg: '', ingestErr: '',
   },
-  memory: { state: 'idle', stats: {}, recent: [], prefs: [] },
+  memory: { state: 'idle', stats: {}, recent: [], prefs: [], pending: [] },
   review: {
     state: 'idle',
     stats: { cards: 0, learned: 0, fresh: 0, due: 0, avg_mastery: 0, load7: {}, risk: [] },
@@ -121,15 +121,17 @@ const LOADERS = {
   },
 
   async memory(p) {
-    const [st, rc, pf] = await Promise.all([
+    const [st, rc, pf, pd] = await Promise.all([
       safe(getJSON('/memory/stats')),
       safe(getJSON('/memory/recent?limit=12')),
       safe(getJSON('/preferences')),
+      safe(getJSON('/memory/pending')),
     ])
     p.stats = st || {}
     p.recent = rc?.memories || []
     p.prefs = pf?.preferences || []
-    return !!(st || rc || pf)
+    p.pending = pd?.pending || []
+    return !!(st || rc || pf || pd)
   },
 }
 
@@ -215,6 +217,17 @@ export async function dropPreference(id) {
   await refreshPanel('memory')
 }
 
+// 高敏感内容被闸门拦下后停在这里,用户点头才落库(见 foxclaw/core/sensitive_write.py)
+export async function approveSensitiveWrite(id) {
+  await postJSON(`/memory/pending/${encodeURIComponent(id)}/approve`, {})
+  await refreshPanel('memory')
+}
+
+export async function rejectSensitiveWrite(id) {
+  await postJSON(`/memory/pending/${encodeURIComponent(id)}/reject`, {})
+  await refreshPanel('memory')
+}
+
 export async function openMdSkill(id) {
   const d = await getJSON(`/ahs/md-skills/${encodeURIComponent(id)}`)
   openDoc({

package/python/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "foxclaw"
-version = "0.1.3"
+version = "0.1.4"
 description = "Foxclaw - a personal agent with memory, built on the HMR memory runtime"
 requires-python = ">=3.11"
 
@@ -70,6 +70,13 @@ foxclaw = "foxclaw.cli:main"
 # 这里把它们真正打进 wheel,并让 package-data 一并带上 LICENSE/JSON 等非 .py 资源。
 include-package-data = true
 
+[tool.setuptools.package-data]
+# frontend/ 在仓库根、不属于任何包,packages.find 收不到;构建前由
+# tools/bundle_web.py 复制成 foxclaw/web/(已 gitignore),这里再把它带进 wheel/sdist。
+# 不复制的话 pip install 之后 foxclaw serve 首页 404(只有 API/CLI)。
+"foxclaw.web" = ["**/*"]
+
+
 [tool.setuptools.packages.find]
 # foxclaw 在仓库根;hmr 在 vendor/ 下;hkc_* 在 vendor/hkc/ 下 —— 三个搜索根都要给。
 where = [".", "vendor", "vendor/hkc"]
@@ -146,6 +153,10 @@ testpaths = ["tests"]
 # `python -m pytest <file> -q` 隔离)。本白名单确保任意单文件可被 pytest 收集/运行。
 python_files = [
     "test_agent_lifecycle.py",
+    "test_sensitive_write.py",
+    "test_content_guard.py",
+    "test_behavior_spec.py",
+    "test_agent_rules.py",
     "test_ahs_p2.py",
     "test_anthropic_tools.py",
     "test_api.py",

package/python/tests/run_all.py

@@ -93,6 +93,7 @@ TESTS = [
     ("test_negotiation.py",  "智能体层·跨域协商闭环(被拒→让步→再裁决·有界·拒绝不降级)"),
     ("test_tool_use.py",        "P2·function-calling 工具化(多轮工具循环/注册表/降级)"),
     ("test_injection_guard.py", "提示词注入纵深防御(检出/wrap_untrusted/内核前缀/隔离)"),
+    ("test_agent_rules.py",     "Agent 规则文件启动注入(按 AGENT_ID 读盘/顺序/回退开关/包内路径)"),
     ("test_sessions.py",        "会话存储(增删改查/标题提炼/路径穿越/并发追加)"),
     ("test_stats_and_events.py","记忆统计 + 活动事件(hmr.stats 计数修复/chat.completed)"),
     ("test_llm_planner.py",            "P0-1·LLM 驱动规划器(llm_plan/replan·opt-in+fail-safe+sv_replan)"),

package/python/tests/test_agent_rules.py

@@ -0,0 +1,81 @@
+"""
+Agent 规则文件加载测试 —— 全离线。
+
+覆盖:
+  1. agents/rules/<AGENT_ID>/*.md 在组装系统提示词时被读入(启动即生效,不靠代码常驻)。
+  2. 叠加顺序:内核前缀 → 角色职责 → 本 Agent 规则;规则位于最后。
+  3. 回退开关 FOXCLAW_AGENT_RULES=0 可整体关闭;无 AGENT_ID / 无目录时不注入。
+  4. 规则目录必须落在 foxclaw 包内(随 wheel 分发,不能像仓库根 skills/ 那样漏包)。
+"""
+import os, sys, types
+from pathlib import Path
+os.environ.setdefault("HF_HUB_OFFLINE", "1")
+os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
+os.environ["FOXCLAW_EMBEDDING_MODEL"] = "foxclaw-nonexistent-fast"
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from foxclaw.agents.base import KERNEL_PROMPT
+from foxclaw.agents.mixins.prompt import PromptMixin, RULES_DIR, load_agent_rules
+
+
+def _fake(role="我是财务角色ROLEXYZ。", agent_id="finance"):
+    return types.SimpleNamespace(_role_prompt=role, AGENT_ID=agent_id)
+
+
+def test_finance_rules_file_is_loaded():
+    text = load_agent_rules("finance")
+    assert "Personal Finance Agent Rules v1.0" in text
+    assert "不得编造" in text
+    assert "储蓄率" in text
+
+
+def test_unknown_agent_has_no_rules():
+    assert load_agent_rules("no_such_agent") == ""
+    assert load_agent_rules("") == ""
+
+
+def test_system_prompt_order_kernel_role_rules():
+    out = PromptMixin.system_prompt.fget(_fake())
+    assert KERNEL_PROMPT in out
+    assert out.index("认知操作系统") < out.index("ROLEXYZ") < out.index("Personal Finance Agent Rules")
+
+
+def test_rules_follow_kernel_when_role_is_short():
+    out = PromptMixin.system_prompt.fget(_fake(role="ROLEXYZ"))
+    assert "本 Agent 专业规则(必须遵守)" in out
+    assert "月度结余" in out
+
+
+def test_switch_disables_rules(monkeypatch):
+    monkeypatch.setenv("FOXCLAW_AGENT_RULES", "0")
+    out = PromptMixin.system_prompt.fget(_fake())
+    assert "ROLEXYZ" in out and "Personal Finance Agent Rules" not in out
+
+
+def test_switch_disables_rules_even_without_kernel(monkeypatch):
+    monkeypatch.setenv("FOXCLAW_KERNEL_PROMPT", "0")
+    monkeypatch.setenv("FOXCLAW_AGENT_RULES", "0")
+    monkeypatch.setenv("FOXCLAW_BEHAVIOR", "0")
+    assert PromptMixin.system_prompt.fget(_fake(role="ROLEXYZ")) == "ROLEXYZ"
+
+
+def test_rules_survive_kernel_off(monkeypatch):
+    monkeypatch.setenv("FOXCLAW_KERNEL_PROMPT", "0")
+    out = PromptMixin.system_prompt.fget(_fake(role="ROLEXYZ"))
+    assert KERNEL_PROMPT not in out
+    assert "Personal Finance Agent Rules" in out
+    # 行为规范是硬边界,不随内核前缀一起关掉
+    assert out.index("[行为规范") < out.index("ROLEXYZ")
+
+
+def test_agent_without_id_gets_no_rules():
+    fake = types.SimpleNamespace(_role_prompt="ROLEXYZ")
+    out = PromptMixin.system_prompt.fget(fake)
+    assert "ROLEXYZ" in out and "Personal Finance Agent Rules" not in out
+
+
+def test_rules_dir_lives_inside_package():
+    """规则文件必须在 foxclaw 包内:仓库根的 skills/ 就没被打进 wheel,别重演。"""
+    pkg = Path(RULES_DIR).resolve()
+    assert pkg.name == "rules" and pkg.parent.name == "agents"
+    assert (pkg / "finance" / "RULES.md").is_file()

package/python/tests/test_behavior_spec.py

@@ -0,0 +1,198 @@
+"""
+行为规范(Behavior Spec)测试 —— 全离线。
+
+覆盖:
+  1. 内置 behavior.yaml 可加载,id 唯一,每个 code_check 都解析到真实可调用对象。
+  2. 检查点不是装饰:policy_gate.check_action / confirms.consume / verify.verify 在业务代码里真有调用。
+  3. prompt 注入:行为规范在最前段(Markdown 内核之前、persona 之前),带 spec_version,可整体关闭。
+  4. 本地覆盖:可新增 / 可收紧;放宽或省略删除严格层规则 → BehaviorLoadError。
+  5. 本地覆盖的语法错、版本过期 → 降级为 warning 并忽略,不把应用锁死(内置文件的问题仍然硬失败)。
+"""
+import os
+import sys
+from pathlib import Path
+
+import pytest
+
+os.environ.setdefault("HF_HUB_OFFLINE", "1")
+os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
+os.environ["FOXCLAW_EMBEDDING_MODEL"] = "foxclaw-nonexistent-fast"
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from foxclaw.agents.mixins.prompt import PromptMixin
+from foxclaw.behavior import BehaviorLoadError, load, local_path, prompt_block
+
+ROOT = Path(__file__).parent.parent
+
+BASE_YAML = """
+version: 1
+spec_id: test-behavior
+prohibitions:
+  - id: no_x
+    rule: 不做 X
+    enforce: both
+    code_check: "foxclaw.core.verify:verify"
+confirmations:
+  - id: confirm_y
+    rule: 做 Y 前先确认
+    enforce: prompt
+uncertainty:
+  - id: zzz
+    rule: 不确定就问
+    enforce: prompt
+"""
+
+
+def _write(tmp_path, name, text):
+    p = tmp_path / name
+    p.write_text(text, encoding="utf-8")
+    return p
+
+
+def _loaded(tmp_path, local_text):
+    return load(_write(tmp_path, "behavior.yaml", BASE_YAML),
+                _write(tmp_path, "behavior.local.yaml", local_text))
+
+
+def test_builtin_spec_loads_with_unique_ids():
+    spec = load()
+    ids = [r.id for r in spec.rules]
+    assert len(ids) == len(set(ids))
+    assert spec.version == 1 and spec.spec_version
+    assert {"confirm_side_effect", "no_silent_failure"} <= set(ids)
+
+
+def test_every_code_check_resolves_to_callable():
+    spec = load()
+    checks = spec.checkpoints()
+    assert checks, "内置规范至少要登记一个真实检查点"
+    assert all(callable(fn) for fn in checks.values())
+
+
+def test_checkpoints_are_really_wired_in_business_code():
+    """反漂移:登记了检查点,业务代码里就必须真有调用,否则清单会退化成装饰。"""
+    blob = "\n".join(p.read_text(encoding="utf-8")
+                     for p in (ROOT / "foxclaw").rglob("*.py")
+                     if p.parts[len(ROOT.parts)] != "behavior")
+    assert "policy_gate.check_action(" in blob
+    assert "confirms.consume(" in blob
+    assert "verify.verify(" in blob
+
+
+def test_prompt_block_leads_with_spec_and_version():
+    spec = load()
+    block = prompt_block(spec)
+    assert block.startswith("[行为规范")
+    assert spec.spec_version in block
+    for rule in spec.injectable():
+        assert f"[{rule.id}]" in block
+    assert "不可协商" in block
+
+
+def test_behavior_precedes_kernel_and_role():
+    import types
+    fake = types.SimpleNamespace(_role_prompt="我是角色ROLEXYZ。", AGENT_ID="finance")
+    out = PromptMixin.system_prompt.fget(fake)
+    assert out.index("[行为规范") < out.index("认知操作系统") < out.index("ROLEXYZ")
+
+
+def test_switch_disables_behavior_block(monkeypatch):
+    monkeypatch.setenv("FOXCLAW_BEHAVIOR", "0")
+    assert prompt_block() == ""
+
+
+def test_local_can_add_and_tighten(tmp_path):
+    spec = _loaded(tmp_path, """
+version: 1
+confirmations:
+  - id: confirm_y
+    rule: 做 Y 前先确认(收紧)
+    enforce: both
+    code_check: "foxclaw.core.verify:verify"
+uncertainty:
+  - id: new_rule
+    rule: 新加的更严规则
+    enforce: prompt
+""")
+    assert spec.by_id("confirm_y").enforce == "both"
+    assert spec.by_id("new_rule") is not None
+    assert spec.by_id("no_x").enforce == "both"  # 未提及的保持不变
+
+
+def test_local_cannot_weaken_enforce(tmp_path):
+    with pytest.raises(BehaviorLoadError, match="不能放宽"):
+        _loaded(tmp_path, """
+prohibitions:
+  - id: no_x
+    rule: 不做 X
+    enforce: prompt
+""")
+
+
+def test_local_cannot_drop_strict_layer_rule(tmp_path):
+    with pytest.raises(BehaviorLoadError, match="不能通过省略删除"):
+        _loaded(tmp_path, "confirmations: []\n")
+
+
+def test_local_syntax_error_is_warning_not_fatal(tmp_path):
+    spec = _loaded(tmp_path, "prohibitions: [oops\n")
+    assert spec.warnings and "已忽略" in spec.warnings[0]
+    assert spec.by_id("no_x") is not None
+
+
+def test_local_stale_version_is_ignored(tmp_path):
+    spec = _loaded(tmp_path, "version: 99\nuncertainty: []\n")
+    assert any("已过期" in w for w in spec.warnings)
+    assert spec.by_id("zzz") is not None
+
+
+def test_builtin_version_mismatch_is_fatal(tmp_path):
+    p = _write(tmp_path, "behavior.yaml", "version: 99\nprohibitions: []\n")
+    with pytest.raises(BehaviorLoadError, match="version"):
+        load(p, tmp_path / "none.yaml")
+
+
+def test_enforce_without_code_check_fails(tmp_path):
+    p = _write(tmp_path, "behavior.yaml", """
+version: 1
+prohibitions:
+  - id: no_x
+    rule: 不做 X
+    enforce: both
+""")
+    with pytest.raises(BehaviorLoadError, match="没有 code_check"):
+        load(p, tmp_path / "none.yaml")
+
+
+def test_unknown_checkpoint_fails(tmp_path):
+    p = _write(tmp_path, "behavior.yaml", """
+version: 1
+prohibitions:
+  - id: no_x
+    rule: 不做 X
+    enforce: code
+    code_check: "foxclaw.core.verify:no_such_function"
+""")
+    with pytest.raises(BehaviorLoadError, match="不存在"):
+        load(p, tmp_path / "none.yaml")
+
+
+def test_duplicate_id_fails(tmp_path):
+    p = _write(tmp_path, "behavior.yaml", """
+version: 1
+prohibitions:
+  - id: no_x
+    rule: A
+    enforce: prompt
+uncertainty:
+  - id: no_x
+    rule: B
+    enforce: prompt
+""")
+    with pytest.raises(BehaviorLoadError, match="重复"):
+        load(p, tmp_path / "none.yaml")
+
+
+def test_local_path_follows_data_dir(monkeypatch, tmp_path):
+    monkeypatch.setenv("FOXCLAW_DATA_DIR", str(tmp_path))
+    assert local_path() == tmp_path / "behavior.local.yaml"

package/python/tests/test_conflict_dismiss.py

@@ -0,0 +1,143 @@
+"""误判冲突关闭(dismiss)与否定判定收紧的回归测试。
+
+背景:技术文档里「不变/不可观测/不同」这类描述性「不」被当成立场否定,只要两句
+同领域(cos>0.35)就判「矛盾」,实测攒下 1063 张伪冲突卡。dismiss 用于把这些卡关掉
+且**不废弃任何一方观点**;这里锁住三件事:
+  1. dismiss 后双方 claim_status 不变(尤其不能被标 superseded);
+  2. 因冲突被置 disputed 的一方,在没有其它 open 冲突时恢复 active(否则检索永久丢);
+  3. 仍牵涉其它 open 冲突的 claim 保持 disputed。
+"""
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+sys.path.insert(0, str(Path(__file__).parent.parent / "vendor" / "hkc"))
+
+from hkc_core.graph.sqlite_store import SQLiteGraphStore
+from hkc_core.models.enums import ClaimStatus, ConflictStatus
+from hkc_core.models.ku import ClaimKU, ConflictCard
+from hkc_core.utils.id_gen import IDGenerator
+from hkc_kee.kee import KnowledgeEvolutionEngine, _SimpleEmbedder
+from hkc_kep.event_bus import EventBus
+
+
+def _engine(tmp_path):
+    store = SQLiteGraphStore(str(tmp_path / "hkc.db"))
+    bus = EventBus(str(tmp_path / "events.db"))
+    gen = IDGenerator(str(tmp_path / "ids.db"))
+    return store, KnowledgeEvolutionEngine(store, bus, gen)
+
+
+def _claim(store, ku_id, status=ClaimStatus.ACTIVE, conf=0.8):
+    store.put(ClaimKU(
+        ku_id=ku_id, name=ku_id, statement=f"{ku_id} 的观点",
+        confidence=conf, domain="测试", claim_status=status,
+    ))
+
+
+def test_dismiss_keeps_both_claims_and_restores_disputed(tmp_path):
+    store, kee = _engine(tmp_path)
+    _claim(store, "CLM_A")
+    _claim(store, "CLM_B", status=ClaimStatus.DISPUTED)
+    store.save_conflict(ConflictCard(
+        conflict_id="CFT_1", claim_a_id="CLM_A", claim_b_id="CLM_B", domain="测试"))
+
+    assert kee.dismiss_conflict("CFT_1", note="误判") is True
+
+    card = store.get_conflict("CFT_1")
+    assert card.status == ConflictStatus.RESOLVED
+    assert card.resolved_by == "human"
+    assert card.resolution_note == "误判"
+    assert card.resolved_at
+    assert store.get("CLM_A").claim_status == ClaimStatus.ACTIVE
+    assert store.get("CLM_B").claim_status == ClaimStatus.ACTIVE, \
+        "disputed 且无其它 open 冲突 → 必须恢复 active,否则检索永久剔除"
+
+
+def test_dismiss_leaves_claim_still_disputed_if_other_conflict_open(tmp_path):
+    store, kee = _engine(tmp_path)
+    _claim(store, "CLM_A")
+    _claim(store, "CLM_B", status=ClaimStatus.DISPUTED)
+    _claim(store, "CLM_C")
+    store.save_conflict(ConflictCard(
+        conflict_id="CFT_1", claim_a_id="CLM_A", claim_b_id="CLM_B", domain="测试"))
+    store.save_conflict(ConflictCard(
+        conflict_id="CFT_2", claim_a_id="CLM_B", claim_b_id="CLM_C", domain="测试"))
+
+    assert kee.dismiss_conflict("CFT_1") is True
+    assert store.get("CLM_B").claim_status == ClaimStatus.DISPUTED, \
+        "还牵涉 CFT_2,不能提前恢复"
+    assert store.get_conflict("CFT_2").status == ConflictStatus.OPEN
+
+
+def test_dismiss_batch_and_idempotent(tmp_path):
+    store, kee = _engine(tmp_path)
+    _claim(store, "CLM_A")
+    _claim(store, "CLM_B")
+    store.save_conflict(ConflictCard(
+        conflict_id="CFT_1", claim_a_id="CLM_A", claim_b_id="CLM_B"))
+    store.save_conflict(ConflictCard(
+        conflict_id="CFT_2", claim_a_id="CLM_A", claim_b_id="CLM_B"))
+
+    assert kee.dismiss_conflicts(["CFT_1", "CFT_2", "CFT_2"]) == 2
+    assert kee.dismiss_conflicts(["CFT_1"]) == 0, "重复关闭应为 0 而非异常"
+    assert kee.dismiss_conflict("CFT_NOPE") is False
+
+
+def test_has_neg_ignores_descriptive_bu():
+    assert _SimpleEmbedder._has_neg("现有测试行为不变") is False
+    assert _SimpleEmbedder._has_neg("Director 不可观测") is False
+    assert _SimpleEmbedder._has_neg("这是「不管调度」原则的延续") is False
+    assert _SimpleEmbedder._has_neg("不同模块用不同 mock") is False
+    assert _SimpleEmbedder._has_neg("无论怎么改都不影响") is True, "无论剔除后残留的「不」仍算否定"
+    assert _SimpleEmbedder._has_neg("这个方案不好") is True
+    assert _SimpleEmbedder._has_neg("not good") is True
+
+
+def test_neg_flip_defaults_are_configurable(tmp_path):
+    store, _ = _engine(tmp_path)
+    emb = _SimpleEmbedder()
+    assert emb.NEG_FLIP_MIN_COS == 0.35
+    assert emb.NEG_FLIP_MIN_OVERLAP == 0.40
+    tight = _SimpleEmbedder(neg_flip_min_cos=0.9, neg_flip_min_overlap=0.8)
+    assert (tight.NEG_FLIP_MIN_COS, tight.NEG_FLIP_MIN_OVERLAP) == (0.9, 0.8)
+    engine = KnowledgeEvolutionEngine(
+        store, EventBus(":memory:"), IDGenerator(str(tmp_path / "ids2.db")),
+        neg_flip_min_cos=0.6, neg_flip_min_overlap=0.7)
+    assert (engine.neg_flip_min_cos, engine.neg_flip_min_overlap) == (0.6, 0.7)
+
+
+def test_same_proposition_requires_lexical_overlap():
+    emb = _SimpleEmbedder()
+    assert emb._same_proposition("这个方案是可行的，风险可控", "这个方案不可行，风险不可控") is True
+    assert emb._same_proposition(
+        "现有编译测试的输入都是高置信度正例，预期 open_questions 为空，行为不变。",
+        "依赖注入优于 spy-per-module 式模块级 mock") is False
+
+
+def test_score_ignores_same_domain_different_propositions():
+    from hkc_search.embedding_backends import get_st_model
+    if get_st_model(_SimpleEmbedder._MODEL_NAME) is None:
+        import pytest
+        pytest.skip("语义模型未缓存,跳过 score 级验证")
+
+    emb = _SimpleEmbedder()
+    near = ("现有编译测试的输入都是高置信度正例，预期 open_questions 为空，行为不变。",
+            "依赖注入优于 spy-per-module 式模块级 mock")
+    assert emb.score(*near) > -0.30, "同领域不同命题曾被判成矛盾(1063 张伪冲突的来源)"
+
+    contra = ("这个方案是可行的，风险可控", "这个方案不可行，风险不可控")
+    assert emb.score(*contra) < -0.30, "同一命题的正反两面仍须判成矛盾"
+
+
+if __name__ == "__main__":
+    import tempfile
+    for name, fn in list(globals().items()):
+        if name.startswith("test_") and callable(fn):
+            with tempfile.TemporaryDirectory() as d:
+                try:
+                    fn(Path(d))
+                except TypeError:
+                    fn()
+            print(f"✓ {name}")
+    print("\n误判冲突关闭 + 否定判定 回归通过")