From bb3a374b611ea63bf9ca01a617d0c19c4e971adb Mon Sep 17 00:00:00 2001
From: hy <huangyuan@onecat.cn>
Date: Wed, 18 Mar 2026 16:26:54 +0800
Subject: [PATCH] Advance human live tests and simplify report evidence

---
 AGENTS.md                                     |   2 +-
 doc/README.md                                 |   2 +-
 doc/ai-workbench-test/todo-stage-2.md         |  19 +-
 doc/test.md                                   |  10 +-
 src/main/services/ai/node/human/brainNode.ts  | 256 ++++++++++++++++++
 .../services/ai/node/human/responseNode.ts    |  12 +-
 src/main/services/ai/prompts/human/brain.ts   |   6 +-
 tests/ai/backend/human/helpers.ts             |  23 +-
 .../live/human-live-direct-kb-answer.test.ts  | 167 ++++++++++++
 ...ve-evidence-insufficient-safe-stop.test.ts |  93 +++++++
 .../live/human-live-kb-docx-bridge.test.ts    | 233 ++++++++++++++++
 tests/report/ai/report.md                     | 122 +++++++--
 12 files changed, 890 insertions(+), 55 deletions(-)
 create mode 100644 tests/ai/backend/human/live/human-live-direct-kb-answer.test.ts
 create mode 100644 tests/ai/backend/human/live/human-live-evidence-insufficient-safe-stop.test.ts
 create mode 100644 tests/ai/backend/human/live/human-live-kb-docx-bridge.test.ts

diff --git a/AGENTS.md b/AGENTS.md
index 5b5ed182..8df51109 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -10,7 +10,7 @@
 - 当涉及数据库、LanceDB、表职责、存储分层或检索存储实现时，请阅读 `doc/database.md`
 - 当涉及目录规范、文件归位、模块边界或工程结构时，请阅读 `doc/project-structure.md`
 - 当涉及 UI 风格、布局规范或组件交互约束时，请阅读 `doc/ui-style.md`
-- 当涉及测试相关内容，包括测试目录、raw log、report 规则或历史总数据 / 最新总数据口径时，请阅读 `doc/test.md`；
+- 当涉及测试相关内容，包括测试目录、raw log、report 规则、证据展示方式或历史总数据 / 最新总数据口径时，请阅读 `doc/test.md`；
 ## 2. 注意事项
 
 - 禁止使用 `any` 与 `unknown`
diff --git a/doc/README.md b/doc/README.md
index badaeed0..a819692a 100644
--- a/doc/README.md
+++ b/doc/README.md
@@ -18,4 +18,4 @@
 - [ui-style.md](/Users/test/work-space/NCurator-Desktop/doc/ui-style.md)
   - 说明界面设计风格、布局规则和组件使用规范。
 - [test.md](/Users/test/work-space/NCurator-Desktop/doc/test.md)
-  - 说明测试目录、测试用例设计准则，以及 report 中历史总数据 / 最新总数据的统计口径。
+  - 说明测试目录、测试用例设计准则，以及 report 中证据展示和历史总数据 / 最新总数据的统计口径。
diff --git a/doc/ai-workbench-test/todo-stage-2.md b/doc/ai-workbench-test/todo-stage-2.md
index 054b7cb3..ba5f7430 100644
--- a/doc/ai-workbench-test/todo-stage-2.md
+++ b/doc/ai-workbench-test/todo-stage-2.md
@@ -24,20 +24,24 @@
 - [x] 落地 `human-long-task-with-corrections`
 - [x] 落地 `human-long-task-approval-branch`
 - [x] 落地 `human-live-direct-docx-edit`
+- [x] 落地 `human-live-direct-kb-answer`
+- [x] 落地 `human-live-evidence-insufficient-safe-stop`
 - [x] 落地 `human-live-terminal-docx-handoff`
 - [x] 落地 `human-live-long-task-project-brief`
+- [x] 跑完 Stage 2 `backend/human` non-live 测试并刷新 `tests/logs/ai/backend/human/{basic,boundary,scenario}/**`
 
 ## In Progress
 
-- [ ] 落地 `human-live-direct-kb-answer`
 - [ ] 落地 `human-live-kb-docx-bridge`
-- [ ] 落地 `human-live-evidence-insufficient-safe-stop`
-- [ ] 跑完 Stage 2 backend/human 测试并刷新 `tests/logs/ai/**`
-- [ ] 基于新的 human logs 更新 `tests/report/ai/report.md`
+- [x] 基于新的 human logs 更新 `tests/report/ai/report.md`
 
 ## Blocked
 
-- 当前无新增外部阻塞
+- `human-live-kb-docx-bridge` 在 2026-03-18 的 primary `gpt-5` live 运行里仍未形成稳定交付：
+  - `20260318-150310-015`：`invalid_input`，summary=`524 status code (no body)`，只留下两次 `kb_hybrid_search_chunks`
+- 为区分模型/服务问题与编排问题，已在 2026-03-18 补跑 stronger `gpt-5.2`：
+  - `20260318-150906-789`：在 `kb_hybrid_search_chunks -> kb_list_documents` 后判断剩余步数不足，未进入 docx handoff
+  - `20260318-151207-536`：`bounded_loop_error`，错误转入多次 `terminal_run_command`，仍未产出 preview artifact
 - `human` 阶段 live 结论仍不完整，Stage 2 尚不能仅凭 non-live case 视为完成
 
 ## Evidence
@@ -45,6 +49,11 @@
 - `doc/ai-workbench-test/plan.md`
 - `tests/ai/backend/human/{basic,boundary,scenario,live}/*.test.ts`
 - `tests/logs/ai/backend/human/**`
+- `tests/logs/ai/backend/human/live/human-live-direct-kb-answer/20260318-145113-697/*`
+- `tests/logs/ai/backend/human/live/human-live-evidence-insufficient-safe-stop/20260318-150735-678/*`
+- `tests/logs/ai/backend/human/live/human-live-kb-docx-bridge/20260318-150310-015/*`
+- `tests/logs/ai/backend/human/live/human-live-kb-docx-bridge/20260318-150906-789/*`
+- `tests/logs/ai/backend/human/live/human-live-kb-docx-bridge/20260318-151207-536/*`
 
 ## Exit Criteria
 
diff --git a/doc/test.md b/doc/test.md
index ecf21330..b6974197 100644
--- a/doc/test.md
+++ b/doc/test.md
@@ -423,7 +423,9 @@ tests/report/<domain>/
   - `失败分析`
   - `当前评估`
   - `下一步建议`
-- `证据` 必须列出对应 run 的 `meta.json`、`timeline.jsonl`、`artifacts.json` 路径；如果有关键附件，也应一并列出。
+- `证据` 默认只列出能定位该条证据的最小信息，例如 `caseId`、`runId`、`status`。
+- 不要在 report 正文里机械展开 `meta.json`、`timeline.jsonl`、`artifacts.json` 这些文件名。
+- 只有当失败分析或结论明确依赖某个关键附件时，才额外补充附件名称或用途说明。
 - `通过率` 只写当前节点覆盖范围内的统计，不混入其他节点数据。
 - `通过率` 固定拆成两组：
   - `历史总数据`：`totalRuns`、`passedRuns`、`failedRuns`、`historicalSuccessRate`
@@ -464,7 +466,7 @@ tests/report/<domain>/
 
 ##### 证据
 
-- `20260315-101530-112` | passed | `tests/logs/ai/backend/tools/terminal/terminal-basic-command/20260315-101530-112/meta.json` | `tests/logs/ai/backend/tools/terminal/terminal-basic-command/20260315-101530-112/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/terminal-basic-command/20260315-101530-112/artifacts.json`
+- `terminal-basic-command` | run=`20260315-101530-112` | status=`passed`
 
 ##### 通过率
 
@@ -499,8 +501,8 @@ tests/report/<domain>/
 
 ##### 证据
 
-- `20260315-103002-008` | passed | `tests/logs/ai/e2e/workbench/workbench-plan-artifact-smoke/20260315-103002-008/meta.json` | `tests/logs/ai/e2e/workbench/workbench-plan-artifact-smoke/20260315-103002-008/timeline.jsonl` | `tests/logs/ai/e2e/workbench/workbench-plan-artifact-smoke/20260315-103002-008/artifacts.json`
-- artifact references: `tests/logs/ai/e2e/workbench/workbench-plan-artifact-smoke/20260315-103002-008/artifacts/trace/playwright-trace.zip`
+- `workbench-plan-artifact-smoke` | run=`20260315-103002-008` | status=`passed`
+- 关键附件: playwright trace
 
 ##### 通过率
 
diff --git a/src/main/services/ai/node/human/brainNode.ts b/src/main/services/ai/node/human/brainNode.ts
index be2e6114..0280cd4c 100644
--- a/src/main/services/ai/node/human/brainNode.ts
+++ b/src/main/services/ai/node/human/brainNode.ts
@@ -171,6 +171,261 @@ const toResultData = (value: JsonValue | undefined, key: string): JsonObject | u
   return { [key]: value }
 }
 
+type DecisionDocumentRef = {
+  documentName?: string
+  documentId?: string
+}
+
+const clipText = (value: string, maxChars = 120): string => {
+  const text = value.trim()
+  if (text.length <= maxChars) return text
+  return `${text.slice(0, maxChars - 3)}...`
+}
+
+const parseJsonValueText = (value: string): JsonValue | undefined => {
+  const trimmed = value.trim()
+  if (!trimmed) return undefined
+  try {
+    return JSON.parse(trimmed) as JsonValue
+  } catch {
+    const arrayStart = trimmed.indexOf('[')
+    const arrayEnd = trimmed.lastIndexOf(']')
+    if (arrayStart >= 0 && arrayEnd > arrayStart) {
+      try {
+        return JSON.parse(trimmed.slice(arrayStart, arrayEnd + 1)) as JsonValue
+      } catch {
+        return undefined
+      }
+    }
+    const objectStart = trimmed.indexOf('{')
+    const objectEnd = trimmed.lastIndexOf('}')
+    if (objectStart >= 0 && objectEnd > objectStart) {
+      try {
+        return JSON.parse(trimmed.slice(objectStart, objectEnd + 1)) as JsonValue
+      } catch {
+        return undefined
+      }
+    }
+  }
+  return undefined
+}
+
+const extractDocumentRefFromRecord = (value: JsonObject): DecisionDocumentRef | null => {
+  const documentName = toOptionalText(value.document_name) || toOptionalText(value.name)
+  const documentId = toOptionalText(value.document_id) || toOptionalText(value.id)
+  if (!documentName && !documentId) return null
+  return {
+    documentName,
+    documentId
+  }
+}
+
+const collectDocumentRefsFromValue = (value: JsonValue | undefined): DecisionDocumentRef[] => {
+  if (Array.isArray(value)) {
+    return value.flatMap((item) => collectDocumentRefsFromValue(item))
+  }
+
+  if (!isJsonObject(value)) return []
+
+  const refs: DecisionDocumentRef[] = []
+  const directRef = extractDocumentRefFromRecord(value)
+  if (directRef) refs.push(directRef)
+
+  const raw = toOptionalText(value.raw)
+  if (raw) {
+    refs.push(...collectDocumentRefsFromValue(parseJsonValueText(raw)))
+  }
+
+  if (Array.isArray(value.items)) {
+    refs.push(...collectDocumentRefsFromValue(value.items))
+  }
+
+  return refs
+}
+
+const dedupeDocumentRefs = (refs: DecisionDocumentRef[]): DecisionDocumentRef[] => {
+  const seen = new Set<string>()
+  const unique: DecisionDocumentRef[] = []
+
+  for (const ref of refs) {
+    const key = `${ref.documentId || ''}::${ref.documentName || ''}`
+    if (!ref.documentId && !ref.documentName) continue
+    if (seen.has(key)) continue
+    seen.add(key)
+    unique.push(ref)
+  }
+
+  return unique
+}
+
+const formatDocumentRef = (ref: DecisionDocumentRef): string | undefined => {
+  if (ref.documentName && ref.documentId) return `${ref.documentName} (${ref.documentId})`
+  return ref.documentName || ref.documentId
+}
+
+const countValueItems = (value: JsonValue | undefined): number => {
+  if (Array.isArray(value)) return value.length
+  if (isJsonObject(value) && Array.isArray(value.items)) return value.items.length
+  return value === undefined ? 0 : 1
+}
+
+const extractPrimaryEvidenceText = (value: JsonValue | undefined): string | undefined => {
+  if (Array.isArray(value)) {
+    for (const item of value) {
+      const nested = extractPrimaryEvidenceText(item)
+      if (nested) return nested
+    }
+    return undefined
+  }
+
+  if (typeof value === 'string') {
+    const text = value.trim()
+    return text ? clipText(text) : undefined
+  }
+
+  if (!isJsonObject(value)) return undefined
+
+  const directText =
+    toOptionalText(value.text) || toOptionalText(value.summary) || toOptionalText(value.message)
+  if (directText) return clipText(directText)
+
+  const raw = toOptionalText(value.raw)
+  if (raw) {
+    const nested = extractPrimaryEvidenceText(parseJsonValueText(raw))
+    if (nested) return nested
+  }
+
+  if (Array.isArray(value.items)) {
+    return extractPrimaryEvidenceText(value.items)
+  }
+
+  return undefined
+}
+
+const resolveToolFamily = (toolName: string | undefined): string | undefined => {
+  const normalized = String(toolName || '').trim()
+  if (!normalized) return undefined
+  const separatorIndex = normalized.indexOf('_')
+  return separatorIndex > 0 ? normalized.slice(0, separatorIndex) : normalized
+}
+
+const countTrailingSameToolName = (
+  rows: Array<{ toolName: string; toolFamily: string; summary: string }>
+): number => {
+  if (rows.length === 0) return 0
+  let count = 1
+  for (let index = rows.length - 1; index > 0; index -= 1) {
+    if (rows[index]?.toolName !== rows[index - 1]?.toolName) break
+    count += 1
+  }
+  return count
+}
+
+const countTrailingSameToolFamily = (
+  rows: Array<{ toolName: string; toolFamily: string; summary: string }>
+): number => {
+  if (rows.length === 0) return 0
+  let count = 1
+  for (let index = rows.length - 1; index > 0; index -= 1) {
+    if (rows[index]?.toolFamily !== rows[index - 1]?.toolFamily) break
+    count += 1
+  }
+  return count
+}
+
+const countTrailingSameSummary = (
+  rows: Array<{ toolName: string; toolFamily: string; summary: string }>
+): number => {
+  if (rows.length === 0) return 0
+  let count = 1
+  for (let index = rows.length - 1; index > 0; index -= 1) {
+    if (rows[index]?.summary !== rows[index - 1]?.summary) break
+    count += 1
+  }
+  return count
+}
+
+const buildDecisionStateText = (input: { state: HumanRunState; ctx: AiRunContext }): string => {
+  const { state, ctx } = input
+  const remainingSteps = Math.max(0, state.maxSteps - state.steps)
+  const objective = String(state.input.objective || '')
+  const recentToolRows = state.stepHistory
+    .filter(
+      (step) =>
+        step.action === 'invoke_tool' && typeof step.toolName === 'string' && step.toolName.length > 0
+    )
+    .slice(-4)
+    .map((step) => ({
+      toolName: String(step.toolName || ''),
+      toolFamily: resolveToolFamily(String(step.toolName || '')) || 'unknown',
+      resultStatus: step.resultStatus || 'unknown',
+      summary: clipText(String(step.summary || ''))
+    }))
+
+  let lastObservationSignals: JsonValue = {
+    kind: 'none'
+  }
+
+  if (state.lastObservation?.kind === 'skill_loaded') {
+    lastObservationSignals = {
+      kind: 'skill_loaded',
+      skillId: state.lastObservation.skillId,
+      activeSkillIds: state.lastObservation.activeSkillIds,
+      activeToolNames: state.lastObservation.activeToolNames
+    }
+  } else if (state.lastObservation?.kind === 'terminal') {
+    lastObservationSignals = {
+      kind: 'terminal',
+      status: state.lastObservation.status,
+      code: state.lastObservation.code,
+      summary: clipText(state.lastObservation.summary),
+      hasData: state.lastObservation.data !== undefined
+    }
+  } else if (state.lastObservation?.kind === 'tool_result') {
+    const uniqueRefs = dedupeDocumentRefs(collectDocumentRefsFromValue(state.lastObservation.data))
+    const primarySourceRef = uniqueRefs.map((ref) => formatDocumentRef(ref)).find(Boolean)
+    lastObservationSignals = {
+      kind: 'tool_result',
+      toolName: state.lastObservation.toolName,
+      toolFamily: resolveToolFamily(state.lastObservation.toolName) || 'unknown',
+      ok: state.lastObservation.ok,
+      code: state.lastObservation.code,
+      itemCount: countValueItems(state.lastObservation.data),
+      uniqueSourceRefCount: uniqueRefs.length,
+      primarySourceRef: primarySourceRef || 'none',
+      primaryEvidenceText: extractPrimaryEvidenceText(state.lastObservation.data) || 'none',
+      summary: clipText(state.lastObservation.summary)
+    }
+  }
+
+  return toJsonText({
+    stepBudget: {
+      currentStep: state.steps + 1,
+      maxSteps: state.maxSteps,
+      remainingSteps,
+      lowBudget: remainingSteps <= 2
+    },
+    taskSignals: {
+      deliveryKind: state.deliveryKind,
+      workspaceBound: Boolean(ctx.workspace),
+      selectedDocumentCount: ctx.selectedDocumentIds?.length || 0,
+      mentionsDocx: /\.docx\b|docx/i.test(objective),
+      mentionsPreview: /preview|预览/i.test(objective),
+      mentionsSave: /save|保存|另存/i.test(objective)
+    },
+    progressSignals: {
+      activeSkillCount: state.activeSkillIds.length,
+      activeToolCount: state.activeTools.length,
+      recentToolCalls: recentToolRows,
+      trailingSameToolCount: countTrailingSameToolName(recentToolRows),
+      trailingSameToolFamilyCount: countTrailingSameToolFamily(recentToolRows),
+      trailingSameSummaryCount: countTrailingSameSummary(recentToolRows),
+      recentErrorCount: recentToolRows.filter((row) => row.resultStatus === 'error').length
+    },
+    lastObservationSignals
+  })
+}
+
 const buildDecisionPrompt = (input: { state: HumanRunState; ctx: AiRunContext }) => {
   const { state } = input
   const prompt = buildHumanBrainPrompt({
@@ -181,6 +436,7 @@ const buildDecisionPrompt = (input: { state: HumanRunState; ctx: AiRunContext })
     availableSkillCardsText: toJsonText(state.availableSkillCards),
     loadedSkillsText: serializeLoadedSkills(state),
     activeToolsText: serializeActiveTools(state),
+    decisionStateText: buildDecisionStateText(input),
     lastObservationText: state.lastObservation ? toJsonText(state.lastObservation) : 'none',
     stepHistoryText:
       state.stepHistory.length > 0
diff --git a/src/main/services/ai/node/human/responseNode.ts b/src/main/services/ai/node/human/responseNode.ts
index 59c59dcb..d071f10e 100644
--- a/src/main/services/ai/node/human/responseNode.ts
+++ b/src/main/services/ai/node/human/responseNode.ts
@@ -98,10 +98,14 @@ const deriveLocalSourceStatus = (state: HumanRunState): {
   summary: string
 } => {
   const joined = [state.summary, state.deliveryMarkdown].join('\n')
-  const mentionsInsufficient =
-    joined.includes('本地资料不足') ||
-    joined.includes('是否允许继续使用非本地信息') ||
-    joined.includes('非本地信息')
+  const insufficientMarkers = [
+    '本地资料不足',
+    '本地证据不足',
+    '当前本地资料不足',
+    '当前本地证据不足',
+    '本地范围不足'
+  ]
+  const mentionsInsufficient = insufficientMarkers.some((marker) => joined.includes(marker))
   if (mentionsInsufficient) {
     return {
       status: 'insufficient',
diff --git a/src/main/services/ai/prompts/human/brain.ts b/src/main/services/ai/prompts/human/brain.ts
index 50ff6748..6a999836 100644
--- a/src/main/services/ai/prompts/human/brain.ts
+++ b/src/main/services/ai/prompts/human/brain.ts
@@ -9,6 +9,7 @@ export type HumanBrainPromptInput = {
   availableSkillCardsText: string
   loadedSkillsText: string
   activeToolsText: string
+  decisionStateText: string
   lastObservationText: string
   stepHistoryText: string
 }
@@ -19,7 +20,9 @@ export const buildHumanBrainPrompt = (input: HumanBrainPromptInput) => {
     '【工具边界】只有 active tools 中列出的工具可以 invoke_tool。',
     '【完成标准】仅当任务目标已满足且证据充分时才 complete_task；summary 必须写清结论与关键证据。',
     '【结果收敛】若任务目标已明确要求 resultData 字段，且最近 observation 已经提供这些字段所需证据，下一步必须 complete_task 并写入 resultData，不要继续为同一证据反复调用 KB 或 shell。',
-    '【KB 收敛】若 kb_hybrid_search_chunks 的 observation 已经出现唯一的 document_name/document_id，或首条文本已经直接给出答案，不要再调用 kb_list_documents / 再次搜索来“确认同一证据”；应直接 complete_task。',
+    '【证据收敛】若最近 observation 已经足以唯一支撑当前结论或下一交付，优先 complete_task；不要为了“再次确认同一证据”继续调用等价工具。',
+    '【无新增信息】若最近连续几步没有带来新信息，优先切换能力、缩小目标或明确资料不足；不要只是在同类工具之间来回试探。',
+    '【最小路径】优先完成满足 objective 与 delivery kind 的最小交付路径；若任务只要求 answer、handoff 或 preview，不要扩展成非必需的保存、重复检索或额外产物。',
     '【阻塞标准】workspace_required/approval_denied 等不可恢复条件必须 block_task。',
     '【反循环】最近一步没有带来新信息时，禁止重复相同 toolName + args；应切换 skill、调整参数或收敛。',
     '【禁止过度复核】同一文件名、同一命令输出或同一 KB 证据已经足够支撑结论时，不要再做“再次确认”式工具调用。',
@@ -45,6 +48,7 @@ export const buildHumanBrainPrompt = (input: HumanBrainPromptInput) => {
       { label: 'Available Skill Cards', value: input.availableSkillCardsText },
       { label: 'Loaded Skills', value: input.loadedSkillsText },
       { label: 'Active Tools', value: input.activeToolsText },
+      { label: 'Decision State', value: input.decisionStateText },
       { label: 'Last Observation', value: input.lastObservationText },
       { label: 'Recent Steps', value: input.stepHistoryText }
     ])
diff --git a/tests/ai/backend/human/helpers.ts b/tests/ai/backend/human/helpers.ts
index 233297af..41e99f66 100644
--- a/tests/ai/backend/human/helpers.ts
+++ b/tests/ai/backend/human/helpers.ts
@@ -4,6 +4,7 @@ import { hasLiveLlmConfigInEnv, loadLiveLlmConfigFromEnv } from '../../../harnes
 import type { BackendTestRuntime, TestCaseDefinition } from '../../../harness/backend/test-types'
 
 type AiHumanCaseDefinition<TResult> = Omit<TestCaseDefinition<TResult>, 'domain' | 'layer' | 'suite'>
+const HUMAN_LIVE_TEST_TIMEOUT_MS = 300_000
 
 export const defineAiHumanCase = <TResult>(definition: AiHumanCaseDefinition<TResult>): void => {
   it(definition.caseId, async () => {
@@ -25,15 +26,19 @@ export const defineAiHumanLiveCase = <TResult>(definition: AiHumanCaseDefinition
 
   const runtimeModelName = loadLiveLlmConfigFromEnv('primary').model_name
 
-  it(definition.caseId, async () => {
-    await runBackendCase({
-      ...definition,
-      domain: 'ai',
-      layer: 'backend',
-      suite: 'human',
-      runtimeModelName
-    })
-  })
+  it(
+    definition.caseId,
+    async () => {
+      await runBackendCase({
+        ...definition,
+        domain: 'ai',
+        layer: 'backend',
+        suite: 'human',
+        runtimeModelName
+      })
+    },
+    HUMAN_LIVE_TEST_TIMEOUT_MS
+  )
 }
 
 export const loadPrimaryLiveHumanConfig = () => loadLiveLlmConfigFromEnv('primary')
diff --git a/tests/ai/backend/human/live/human-live-direct-kb-answer.test.ts b/tests/ai/backend/human/live/human-live-direct-kb-answer.test.ts
new file mode 100644
index 00000000..07eeb859
--- /dev/null
+++ b/tests/ai/backend/human/live/human-live-direct-kb-answer.test.ts
@@ -0,0 +1,167 @@
+import type { DocumentRecord, JsonValue, SearchResult } from '../../../../../src/shared/types'
+import { createHumanTestRunner } from '../../../../harness/backend/humanHarness'
+import {
+  defineAiHumanLiveCase,
+  getEventsByType,
+  loadPrimaryLiveHumanConfig
+} from '../helpers'
+
+const releaseChecklistDocument: DocumentRecord = {
+  id: 'doc-beta',
+  name: 'release-checklist.md',
+  source_type: 'file',
+  created_at: 1710000000000,
+  import_status: 2
+}
+
+const getFirstPreviewRecord = (
+  value: JsonValue | undefined
+): Record<string, JsonValue> | null => {
+  if (!value) return null
+  if (Array.isArray(value)) {
+    const first = value[0]
+    if (!first || Array.isArray(first) || typeof first !== 'object') return null
+    return first as Record<string, JsonValue>
+  }
+  if (typeof value !== 'object') return null
+  return value as Record<string, JsonValue>
+}
+
+defineAiHumanLiveCase({
+  caseId: 'human-live-direct-kb-answer',
+  caseType: 'live',
+  groups: ['live'],
+  description: 'live 模型直接使用本地 KB 回答问题，并在结果里保留可回看的文档引用。',
+  fixtures: [],
+  runMode: 'single',
+  execute: async (runtime) => {
+    const runner = createHumanTestRunner(
+      runtime,
+      {
+        bindWorkspace: false,
+        selectedDocumentIds: ['doc-beta'],
+        liveModelConfig: loadPrimaryLiveHumanConfig(),
+        documentsStoreOverrides: {
+          hybridSearch: async () => {
+            const rows: SearchResult[] = [
+              {
+                id: 'chunk-beta-1',
+                text: 'Approval flow still needs regression coverage before release.',
+                document_id: 'doc-beta',
+                document_name: 'release-checklist.md',
+                source_type: 'file',
+                _score: 0.91
+              }
+            ]
+            return rows
+          },
+          vectorSearch: async () => {
+            const rows: SearchResult[] = [
+              {
+                id: 'chunk-beta-1',
+                text: 'Approval flow still needs regression coverage before release.',
+                document_id: 'doc-beta',
+                document_name: 'release-checklist.md',
+                source_type: 'file',
+                _score: 0.91
+              }
+            ]
+            return rows
+          },
+          ftsSearch: async () => {
+            const rows: SearchResult[] = [
+              {
+                id: 'chunk-beta-1',
+                text: 'Approval flow still needs regression coverage before release.',
+                document_id: 'doc-beta',
+                document_name: 'release-checklist.md',
+                source_type: 'file',
+                _score: 0.91
+              }
+            ]
+            return rows
+          },
+          listDocuments: async () => ({
+            items: [releaseChecklistDocument],
+            total: 1
+          }),
+          getDocumentById: async (id: string) => (id === 'doc-beta' ? releaseChecklistDocument : null)
+        }
+      },
+      { maxSteps: 6 }
+    )
+
+    const output = await runner.run(
+      {
+        taskId: runtime.taskId,
+        objective:
+          '请仅使用当前本地资料范围回答 approval flow regression coverage 是否已经完成。先做本地 KB 检索，一旦拿到直接证据就停止，不要重复确认。最终答复要明确说明是否完成，并引用依据文档；答复里保留短语 "regression coverage"。'
+      },
+      {
+        deliveryKind: 'final'
+      }
+    )
+
+    const toolResults = getEventsByType(runtime, 'tool_call_result')
+    const firstKbResult = toolResults.find((event) => event.toolName.startsWith('kb_'))
+    const kbPreview = firstKbResult ? getFirstPreviewRecord(firstKbResult.outputPreview) : null
+
+    return {
+      output,
+      kbPreview,
+      toolCalls: getEventsByType(runtime, 'tool_call_started'),
+      deliveries: getEventsByType(runtime, 'task_delivery_ready')
+    }
+  },
+  assert: async (result, runtime) => {
+    const joinedOutputText = [
+      result.output.summary,
+      result.output.deliveryMarkdown,
+      JSON.stringify(result.output.resultData || {})
+    ].join('\n')
+
+    runtime.assertThat(
+      result.output.status === 'completed' &&
+        result.output.resultCode === 'ok' &&
+        result.output.localSourceStatus === 'used',
+      {
+        name: 'human-live-kb-answer-completed',
+        message: 'live human 应能基于本地 KB 形成完成态问答交付',
+        data: {
+          status: result.output.status,
+          resultCode: result.output.resultCode,
+          localSourceStatus: result.output.localSourceStatus,
+          summary: result.output.summary,
+          resultData: result.output.resultData,
+          toolNames: result.toolCalls.map((event) => event.toolName)
+        }
+      }
+    )
+    runtime.assertThat(
+      Boolean(result.kbPreview) &&
+        String(result.kbPreview?.document_name || '').includes('release-checklist.md') &&
+        String(result.kbPreview?.document_id || '').includes('doc-beta'),
+      {
+        name: 'human-live-kb-answer-evidence',
+        message: 'live human 应先拿到可回看的 KB 文档证据'
+      }
+    )
+    runtime.assertThat(
+      joinedOutputText.includes('regression coverage') &&
+        (joinedOutputText.includes('release-checklist.md') || joinedOutputText.includes('doc-beta')),
+      {
+        name: 'human-live-kb-answer-result-data',
+        message: 'live human 应在交付中保留答案关键短语和文档引用'
+      }
+    )
+    runtime.assertThat(
+      result.toolCalls.some((event) => event.toolName.startsWith('kb_')) &&
+        result.toolCalls.every((event) => event.toolName.startsWith('kb_')) &&
+        result.deliveries.length === 1,
+      {
+        name: 'human-live-kb-answer-events',
+        message: 'live human 直接问答应只留下 KB 检索和最终交付事件'
+      }
+    )
+  }
+})
diff --git a/tests/ai/backend/human/live/human-live-evidence-insufficient-safe-stop.test.ts b/tests/ai/backend/human/live/human-live-evidence-insufficient-safe-stop.test.ts
new file mode 100644
index 00000000..479e2bda
--- /dev/null
+++ b/tests/ai/backend/human/live/human-live-evidence-insufficient-safe-stop.test.ts
@@ -0,0 +1,93 @@
+import { createHumanTestRunner } from '../../../../harness/backend/humanHarness'
+import {
+  defineAiHumanLiveCase,
+  getEventsByType,
+  loadPrimaryLiveHumanConfig
+} from '../helpers'
+
+defineAiHumanLiveCase({
+  caseId: 'human-live-evidence-insufficient-safe-stop',
+  caseType: 'live',
+  groups: ['live'],
+  description: 'live 模型在本地证据不足时应安全收敛，而不是编造结论。',
+  fixtures: [],
+  runMode: 'single',
+  execute: async (runtime) => {
+    const runner = createHumanTestRunner(
+      runtime,
+      {
+        bindWorkspace: false,
+        selectedDocumentIds: ['doc-gap'],
+        liveModelConfig: loadPrimaryLiveHumanConfig(),
+        documentsStoreOverrides: {
+          hybridSearch: async () => [],
+          vectorSearch: async () => [],
+          ftsSearch: async () => [],
+          listDocuments: async () => ({
+            items: [],
+            total: 0
+          }),
+          getDocumentById: async () => null
+        }
+      },
+      { maxSteps: 6 }
+    )
+
+    const output = await runner.run(
+      {
+        taskId: runtime.taskId,
+        objective:
+          '请仅使用当前本地资料范围确认 final owner 与 timeline。如果当前本地资料不足，不要猜测，也不要继续重复检索；请直接给出最终答复，答复中必须包含“本地资料不足”和“是否允许继续使用非本地信息”，并在 resultData 中保留 gap_summary 与 next_step。'
+      },
+      {
+        deliveryKind: 'final'
+      }
+    )
+
+    return {
+      output,
+      toolCalls: getEventsByType(runtime, 'tool_call_started'),
+      toolResults: getEventsByType(runtime, 'tool_call_result'),
+      deliveries: getEventsByType(runtime, 'task_delivery_ready')
+    }
+  },
+  assert: async (result, runtime) => {
+    const kbToolResults = result.toolResults.filter((event) => event.toolName.startsWith('kb_'))
+    const emptyKbResults = kbToolResults.every(
+      (event) => Array.isArray(event.outputPreview) && event.outputPreview.length === 0
+    )
+
+    runtime.assertThat(
+      result.output.localSourceStatus === 'insufficient' &&
+        result.output.deliveryMarkdown.includes('本地资料不足') &&
+        result.output.deliveryMarkdown.includes('是否允许继续使用非本地信息'),
+      {
+        name: 'human-live-safe-stop-status',
+        message: 'live human 证据不足时应明确说明本地不足，并询问是否允许使用非本地信息'
+      }
+    )
+    runtime.assertThat(
+      typeof result.output.resultData?.gap_summary === 'string' &&
+        result.output.resultData.gap_summary.length > 0 &&
+        typeof result.output.resultData?.next_step === 'string' &&
+        result.output.resultData.next_step.length > 0,
+      {
+        name: 'human-live-safe-stop-result-data',
+        message: 'live human 安全收敛时应保留结构化缺口与下一步说明'
+      }
+    )
+    runtime.assertThat(
+      kbToolResults.length >= 1 &&
+        emptyKbResults &&
+        result.toolCalls.every((event) => event.toolName.startsWith('kb_')),
+      {
+        name: 'human-live-safe-stop-no-hallucination',
+        message: 'live human 应只基于本地 KB 空结果收敛，不应伪造其他工具证据'
+      }
+    )
+    runtime.assertThat(result.toolCalls.length <= 2 && result.deliveries.length === 1, {
+      name: 'human-live-safe-stop-no-loop',
+      message: 'live human 证据不足时不应陷入重复检索循环'
+    })
+  }
+})
diff --git a/tests/ai/backend/human/live/human-live-kb-docx-bridge.test.ts b/tests/ai/backend/human/live/human-live-kb-docx-bridge.test.ts
new file mode 100644
index 00000000..ceebf6c4
--- /dev/null
+++ b/tests/ai/backend/human/live/human-live-kb-docx-bridge.test.ts
@@ -0,0 +1,233 @@
+import fs from 'node:fs/promises'
+import type { DocumentRecord, JsonValue, SearchResult } from '../../../../../src/shared/types'
+import { createDocxToolHarness, writeTestDocxFile } from '../../../../harness/backend/aiToolHarness'
+import { createHumanTestRunner } from '../../../../harness/backend/humanHarness'
+import {
+  defineAiHumanLiveCase,
+  getEventsByType,
+  loadPrimaryLiveHumanConfig
+} from '../helpers'
+
+const releasePolicyDocument: DocumentRecord = {
+  id: 'doc-policy',
+  name: 'release-policy.md',
+  source_type: 'file',
+  created_at: 1710000000000,
+  import_status: 2
+}
+
+const fileExists = async (filePath: string): Promise<boolean> => {
+  try {
+    await fs.stat(filePath)
+    return true
+  } catch {
+    return false
+  }
+}
+
+const getFirstPreviewRecord = (
+  value: JsonValue | undefined
+): Record<string, JsonValue> | null => {
+  if (!value) return null
+  if (Array.isArray(value)) {
+    const first = value[0]
+    if (!first || Array.isArray(first) || typeof first !== 'object') return null
+    return first as Record<string, JsonValue>
+  }
+  if (typeof value !== 'object') return null
+  return value as Record<string, JsonValue>
+}
+
+defineAiHumanLiveCase({
+  caseId: 'human-live-kb-docx-bridge',
+  caseType: 'live',
+  groups: ['live'],
+  description: 'live 模型先检索本地 KB，再把规范要求落到 docx preview 并交付。',
+  fixtures: [],
+  runMode: 'single',
+  execute: async (runtime) => {
+    await writeTestDocxFile(runtime.workspaceRoot, 'release-check.docx', [
+      { type: 'heading', text: 'Release Checklist', level: 1 },
+      { type: 'paragraph', text: 'Checklist summary: pending policy sync.' },
+      { type: 'paragraph', text: 'Risk note: TBD.' }
+    ])
+
+    const runner = createHumanTestRunner(
+      runtime,
+      {
+        selectedDocumentIds: ['doc-policy'],
+        liveModelConfig: loadPrimaryLiveHumanConfig(),
+        documentsStoreOverrides: {
+          hybridSearch: async () => {
+            const rows: SearchResult[] = [
+              {
+                id: 'chunk-policy-1',
+                text: 'Release reports must mention approval-denied coverage and terminal artifact assertion before sign-off.',
+                document_id: 'doc-policy',
+                document_name: 'release-policy.md',
+                source_type: 'file',
+                _score: 0.93
+              }
+            ]
+            return rows
+          },
+          vectorSearch: async () => {
+            const rows: SearchResult[] = [
+              {
+                id: 'chunk-policy-1',
+                text: 'Release reports must mention approval-denied coverage and terminal artifact assertion before sign-off.',
+                document_id: 'doc-policy',
+                document_name: 'release-policy.md',
+                source_type: 'file',
+                _score: 0.93
+              }
+            ]
+            return rows
+          },
+          ftsSearch: async () => {
+            const rows: SearchResult[] = [
+              {
+                id: 'chunk-policy-1',
+                text: 'Release reports must mention approval-denied coverage and terminal artifact assertion before sign-off.',
+                document_id: 'doc-policy',
+                document_name: 'release-policy.md',
+                source_type: 'file',
+                _score: 0.93
+              }
+            ]
+            return rows
+          },
+          listDocuments: async () => ({
+            items: [releasePolicyDocument],
+            total: 1
+          }),
+          getDocumentById: async (id: string) => (id === 'doc-policy' ? releasePolicyDocument : null)
+        }
+      },
+      { maxSteps: 8 }
+    )
+
+    const output = await runner.run(
+      {
+        taskId: runtime.taskId,
+        objective:
+          '请先做一次当前本地发布规范检索，从 release-policy.md 提炼要求后直接修改 release-check.docx。拿到条款后不要再列文档或重复检索，也不需要 save；只需把摘要段改成包含 “approval-denied coverage and terminal artifact assertion before sign-off”，生成可审阅 preview，并以 handoff 形式交付，在交付里说明来源文档。'
+      },
+      {
+        deliveryKind: 'handoff'
+      }
+    )
+
+    const toolCalls = getEventsByType(runtime, 'tool_call_started')
+    const toolResults = getEventsByType(runtime, 'tool_call_result')
+    const fileArtifacts = getEventsByType(runtime, 'file_artifact')
+    const previewArtifact = fileArtifacts.find((event) => event.artifactRole === 'preview')
+    const previewPath = previewArtifact?.path || ''
+    const previewExists = previewPath ? await fileExists(previewPath) : false
+    const firstKbResult = toolResults.find((event) => event.toolName.startsWith('kb_'))
+    const kbPreview = firstKbResult ? getFirstPreviewRecord(firstKbResult.outputPreview) : null
+
+    if (!previewExists) {
+      return {
+        output,
+        kbPreview,
+        previewExists,
+        previewPath,
+        summarySearchOk: false,
+        summaryMatchCount: 0,
+        toolCalls,
+        fileArtifacts
+      }
+    }
+
+    const inspectionHarness = createDocxToolHarness(runtime)
+
+    try {
+      const openPreview = await inspectionHarness.invoke<{
+        ok: boolean
+        docId: string
+      }>('docx_open', {
+        sourcePath: previewPath
+      })
+      const summarySearch = await inspectionHarness.invoke<{
+        ok: boolean
+        matches: Array<{ blockId: string; text: string }>
+      }>('docx_search', {
+        docId: openPreview.docId,
+        query: 'approval-denied coverage and terminal artifact assertion before sign-off',
+        mode: 'literal'
+      })
+
+      return {
+        output,
+        kbPreview,
+        previewExists,
+        previewPath,
+        summarySearchOk: summarySearch.ok,
+        summaryMatchCount: summarySearch.matches.length,
+        toolCalls,
+        fileArtifacts
+      }
+    } finally {
+      await inspectionHarness.release()
+    }
+  },
+  assert: async (result, runtime) => {
+    const kbCallIndex = result.toolCalls.findIndex((event) => event.toolName.startsWith('kb_'))
+    const docxApplyIndex = result.toolCalls.findIndex((event) => event.toolName === 'docx_apply_ops')
+    const joinedOutputText = [
+      result.output.summary,
+      result.output.deliveryMarkdown,
+      JSON.stringify(result.output.resultData || {})
+    ].join('\n')
+
+    runtime.assertThat(
+        result.output.status === 'completed' &&
+        result.output.resultCode === 'ok' &&
+        result.output.localSourceStatus === 'used',
+      {
+        name: 'human-live-kb-docx-bridge-completed',
+        message: 'live human 的 KB -> DOCX bridge 应形成完成态交付',
+        data: {
+          status: result.output.status,
+          resultCode: result.output.resultCode,
+          localSourceStatus: result.output.localSourceStatus,
+          summary: result.output.summary,
+          resultData: result.output.resultData,
+          previewExists: result.previewExists,
+          previewPath: result.previewPath,
+          toolNames: result.toolCalls.map((event) => event.toolName)
+        }
+      }
+    )
+    runtime.assertThat(
+      Boolean(result.kbPreview) &&
+        String(result.kbPreview?.document_name || '').includes('release-policy.md') &&
+        String(result.kbPreview?.document_id || '').includes('doc-policy'),
+      {
+        name: 'human-live-kb-docx-bridge-kb-evidence',
+        message: 'live human 应先拿到本地发布规范证据'
+      }
+    )
+    runtime.assertThat(
+      result.previewExists &&
+        result.summarySearchOk &&
+        result.summaryMatchCount >= 1 &&
+        joinedOutputText.includes('release-policy.md'),
+      {
+        name: 'human-live-kb-docx-bridge-preview',
+        message: 'live human 应把 KB 规范写入 preview，并在交付中保留来源文档'
+      }
+    )
+    runtime.assertThat(
+      kbCallIndex >= 0 &&
+        docxApplyIndex > kbCallIndex &&
+        result.toolCalls.some((event) => event.toolName === 'docx_open') &&
+        result.fileArtifacts.some((event) => event.artifactRole === 'preview'),
+      {
+        name: 'human-live-kb-docx-bridge-events',
+        message: 'live human 应先检索 KB，再进入 docx 编辑和 preview 交付链路'
+      }
+    )
+  }
+})
diff --git a/tests/report/ai/report.md b/tests/report/ai/report.md
index 6c69c3a7..dc5d28dc 100644
--- a/tests/report/ai/report.md
+++ b/tests/report/ai/report.md
@@ -4,19 +4,22 @@
 
 ### 分析范围
 
-- raw logs 范围: 当前仓库内 `tests/logs/ai/backend/tools/**`
-- 已分析 scope: `ai/backend/tools/docx/basic`、`ai/backend/tools/docx/boundary`、`ai/backend/tools/docx/scenario`、`ai/backend/tools/terminal/basic`、`ai/backend/tools/terminal/boundary`、`ai/backend/tools/terminal/scenario`、`ai/backend/tools/kb/basic`、`ai/backend/tools/kb/boundary`
-- 未纳入本次分析: Stage 2 `ai/backend/human/**` 与旧版 flat scope `ai/backend/tools/docx`、`ai/backend/tools/terminal`、`ai/backend/tools/kb`
+- raw logs 范围: 当前仓库内 `tests/logs/ai/backend/tools/**`、`tests/logs/ai/backend/human/**`
+- 已分析 scope: `ai/backend/tools/docx/basic`、`ai/backend/tools/docx/boundary`、`ai/backend/tools/docx/scenario`、`ai/backend/tools/terminal/basic`、`ai/backend/tools/terminal/boundary`、`ai/backend/tools/terminal/scenario`、`ai/backend/tools/kb/basic`、`ai/backend/tools/kb/boundary`、`ai/backend/human/basic`、`ai/backend/human/boundary`、`ai/backend/human/scenario`、`ai/backend/human/live`
+- 说明: 当前 `ai/backend/human/live` analysis window 只包含本轮补跑并已落日志的 `human-live-direct-kb-answer`、`human-live-evidence-insufficient-safe-stop`、`human-live-kb-docx-bridge`；此前已实现的 `human-live-direct-docx-edit`、`human-live-terminal-docx-handoff`、`human-live-long-task-project-brief` 未在本轮重跑，因此不计入当前 `human/live` 的 latestCaseCount
+- 未纳入本次分析: 旧版 flat scope `ai/backend/tools/docx`、`ai/backend/tools/terminal`、`ai/backend/tools/kb`
 
 ### 总体结论
 
-- 历史总数据: `40` 次 raw runs，其中 `40` 次 passed、`0` 次 failed，historicalSuccessRate=`100.0%`
-- 最新总数据: `20 / 20` 个 case 的最新一轮 run 为 passed，latestSuccessRate=`100.0%`
-- 当前结论: report 已重置为 Stage 1 `backend/tools` 口径，当前新的分层 scope 下 `docx / terminal / kb` 三组工具基线都已跑通。
+- 历史总数据: `168` 次 raw runs，其中 `147` 次 passed、`21` 次 failed，historicalSuccessRate=`87.5%`
+- 最新总数据: `39` 个 case 中有 `38` 个 case 的最新一轮 run 为 passed，latestSuccessRate=`97.4%`
+- 当前结论: Stage 1 `backend/tools` 仍保持 `20 / 20 latest passed`；Stage 2 `backend/human` 的 `basic / boundary / scenario` 最新口径为 `16 / 16 passed`，但当前 `live` analysis window 仅 `2 / 3 passed`，`human-live-kb-docx-bridge` 仍未形成稳定 live 交付。
+- 解释: `historicalSuccessRate` 已包含 2026-03-18 为诊断 `human-live-kb-docx-bridge` 所做的多轮 live 补跑，因此阶段判断应优先看 latest 口径，而不是把这些诊断性失败等同于稳定回归失败。
 
 ### 下一阶段建议
 
-- 先保持当前 Stage 1 `tools` 分层目录和命名稳定，继续把它作为后续 `human` / `entire` 的底层回归基线。
+- 保持当前 Stage 1 `tools` 分层目录和 Stage 2 `human` non-live 基线稳定。
+- 下一步优先修复 `human-live-kb-docx-bridge` 的 live 收敛与跨工具 handoff，再补跑本轮未覆盖的其他 `human live` case，之后再进入 `context` 阶段。
 
 ## backend
 
@@ -43,16 +46,16 @@
 - `boundary`: `4` 次 raw runs，最新 `2 / 2 passed`
 - `scenario`: `6` 次 raw runs，最新 `3 / 3 passed`
 
-##### 最新证据
+##### 证据
 
-- `tool-docx-search-modes-and-scope` | `20260318-131740-749` | passed | `tests/logs/ai/backend/tools/docx/basic/tool-docx-search-modes-and-scope/20260318-131740-749/meta.json` | `tests/logs/ai/backend/tools/docx/basic/tool-docx-search-modes-and-scope/20260318-131740-749/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/basic/tool-docx-search-modes-and-scope/20260318-131740-749/artifacts.json`
-- `tool-docx-structure-read` | `20260318-131740-748` | passed | `tests/logs/ai/backend/tools/docx/basic/tool-docx-structure-read/20260318-131740-748/meta.json` | `tests/logs/ai/backend/tools/docx/basic/tool-docx-structure-read/20260318-131740-748/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/basic/tool-docx-structure-read/20260318-131740-748/artifacts.json`
-- `tool-docx-targeting-and-ambiguous-search` | `20260318-131740-749` | passed | `tests/logs/ai/backend/tools/docx/basic/tool-docx-targeting-and-ambiguous-search/20260318-131740-749/meta.json` | `tests/logs/ai/backend/tools/docx/basic/tool-docx-targeting-and-ambiguous-search/20260318-131740-749/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/basic/tool-docx-targeting-and-ambiguous-search/20260318-131740-749/artifacts.json`
-- `tool-docx-preview-validate-rollback` | `20260318-131740-749` | passed | `tests/logs/ai/backend/tools/docx/boundary/tool-docx-preview-validate-rollback/20260318-131740-749/meta.json` | `tests/logs/ai/backend/tools/docx/boundary/tool-docx-preview-validate-rollback/20260318-131740-749/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/boundary/tool-docx-preview-validate-rollback/20260318-131740-749/artifacts.json`
-- `tool-docx-save-conflict-and-overwrite` | `20260318-131740-749` | passed | `tests/logs/ai/backend/tools/docx/boundary/tool-docx-save-conflict-and-overwrite/20260318-131740-749/meta.json` | `tests/logs/ai/backend/tools/docx/boundary/tool-docx-save-conflict-and-overwrite/20260318-131740-749/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/boundary/tool-docx-save-conflict-and-overwrite/20260318-131740-749/artifacts.json`
-- `tool-docx-mixed-structural-edit` | `20260318-131740-763` | passed | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-mixed-structural-edit/20260318-131740-763/meta.json` | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-mixed-structural-edit/20260318-131740-763/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-mixed-structural-edit/20260318-131740-763/artifacts.json`
-- `tool-docx-multi-document-session-isolation` | `20260318-131740-749` | passed | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-multi-document-session-isolation/20260318-131740-749/meta.json` | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-multi-document-session-isolation/20260318-131740-749/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-multi-document-session-isolation/20260318-131740-749/artifacts.json`
-- `tool-docx-table-and-inline-rewrite` | `20260318-131740-770` | passed | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-table-and-inline-rewrite/20260318-131740-770/meta.json` | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-table-and-inline-rewrite/20260318-131740-770/timeline.jsonl` | `tests/logs/ai/backend/tools/docx/scenario/tool-docx-table-and-inline-rewrite/20260318-131740-770/artifacts.json`
+- `tool-docx-search-modes-and-scope` | run=`20260318-131740-749` | status=`passed`
+- `tool-docx-structure-read` | run=`20260318-131740-748` | status=`passed`
+- `tool-docx-targeting-and-ambiguous-search` | run=`20260318-131740-749` | status=`passed`
+- `tool-docx-preview-validate-rollback` | run=`20260318-131740-749` | status=`passed`
+- `tool-docx-save-conflict-and-overwrite` | run=`20260318-131740-749` | status=`passed`
+- `tool-docx-mixed-structural-edit` | run=`20260318-131740-763` | status=`passed`
+- `tool-docx-multi-document-session-isolation` | run=`20260318-131740-749` | status=`passed`
+- `tool-docx-table-and-inline-rewrite` | run=`20260318-131740-770` | status=`passed`
 
 ##### 失败分析
 
@@ -83,15 +86,15 @@
 - `boundary`: `6` 次 raw runs，最新 `3 / 3 passed`
 - `scenario`: `4` 次 raw runs，最新 `2 / 2 passed`
 
-##### 最新证据
+##### 证据
 
-- `tool-terminal-path-resolution` | `20260318-131741-370` | passed | `tests/logs/ai/backend/tools/terminal/basic/tool-terminal-path-resolution/20260318-131741-370/meta.json` | `tests/logs/ai/backend/tools/terminal/basic/tool-terminal-path-resolution/20260318-131741-370/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/basic/tool-terminal-path-resolution/20260318-131741-370/artifacts.json`
-- `tool-terminal-workspace-material-discovery` | `20260318-131741-389` | passed | `tests/logs/ai/backend/tools/terminal/basic/tool-terminal-workspace-material-discovery/20260318-131741-389/meta.json` | `tests/logs/ai/backend/tools/terminal/basic/tool-terminal-workspace-material-discovery/20260318-131741-389/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/basic/tool-terminal-workspace-material-discovery/20260318-131741-389/artifacts.json`
-- `tool-terminal-approval-and-boundary` | `20260318-131741-430` | passed | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-approval-and-boundary/20260318-131741-430/meta.json` | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-approval-and-boundary/20260318-131741-430/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-approval-and-boundary/20260318-131741-430/artifacts.json`
-- `tool-terminal-approval-denied-follow-up` | `20260318-131741-505` | passed | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-approval-denied-follow-up/20260318-131741-505/meta.json` | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-approval-denied-follow-up/20260318-131741-505/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-approval-denied-follow-up/20260318-131741-505/artifacts.json`
-- `tool-terminal-failure-and-long-output` | `20260318-131740-773` | passed | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-failure-and-long-output/20260318-131740-773/meta.json` | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-failure-and-long-output/20260318-131740-773/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/boundary/tool-terminal-failure-and-long-output/20260318-131740-773/artifacts.json`
-- `tool-terminal-artifact-producing-command` | `20260318-131741-439` | passed | `tests/logs/ai/backend/tools/terminal/scenario/tool-terminal-artifact-producing-command/20260318-131741-439/meta.json` | `tests/logs/ai/backend/tools/terminal/scenario/tool-terminal-artifact-producing-command/20260318-131741-439/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/scenario/tool-terminal-artifact-producing-command/20260318-131741-439/artifacts.json`
-- `tool-terminal-batch-material-summary` | `20260318-131741-439` | passed | `tests/logs/ai/backend/tools/terminal/scenario/tool-terminal-batch-material-summary/20260318-131741-439/meta.json` | `tests/logs/ai/backend/tools/terminal/scenario/tool-terminal-batch-material-summary/20260318-131741-439/timeline.jsonl` | `tests/logs/ai/backend/tools/terminal/scenario/tool-terminal-batch-material-summary/20260318-131741-439/artifacts.json`
+- `tool-terminal-path-resolution` | run=`20260318-131741-370` | status=`passed`
+- `tool-terminal-workspace-material-discovery` | run=`20260318-131741-389` | status=`passed`
+- `tool-terminal-approval-and-boundary` | run=`20260318-131741-430` | status=`passed`
+- `tool-terminal-approval-denied-follow-up` | run=`20260318-131741-505` | status=`passed`
+- `tool-terminal-failure-and-long-output` | run=`20260318-131740-773` | status=`passed`
+- `tool-terminal-artifact-producing-command` | run=`20260318-131741-439` | status=`passed`
+- `tool-terminal-batch-material-summary` | run=`20260318-131741-439` | status=`passed`
 
 ##### 失败分析
 
@@ -121,13 +124,13 @@
 - `basic`: `8` 次 raw runs，最新 `4 / 4 passed`
 - `boundary`: `2` 次 raw runs，最新 `1 / 1 passed`
 
-##### 最新证据
+##### 证据
 
-- `tool-kb-fts-search` | `20260318-131741-574` | passed | `tests/logs/ai/backend/tools/kb/basic/tool-kb-fts-search/20260318-131741-574/meta.json` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-fts-search/20260318-131741-574/timeline.jsonl` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-fts-search/20260318-131741-574/artifacts.json`
-- `tool-kb-hybrid-search` | `20260318-131741-664` | passed | `tests/logs/ai/backend/tools/kb/basic/tool-kb-hybrid-search/20260318-131741-664/meta.json` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-hybrid-search/20260318-131741-664/timeline.jsonl` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-hybrid-search/20260318-131741-664/artifacts.json`
-- `tool-kb-list-documents-and-scope-filter` | `20260318-131741-814` | passed | `tests/logs/ai/backend/tools/kb/basic/tool-kb-list-documents-and-scope-filter/20260318-131741-814/meta.json` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-list-documents-and-scope-filter/20260318-131741-814/timeline.jsonl` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-list-documents-and-scope-filter/20260318-131741-814/artifacts.json`
-- `tool-kb-vector-search` | `20260318-131741-823` | passed | `tests/logs/ai/backend/tools/kb/basic/tool-kb-vector-search/20260318-131741-823/meta.json` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-vector-search/20260318-131741-823/timeline.jsonl` | `tests/logs/ai/backend/tools/kb/basic/tool-kb-vector-search/20260318-131741-823/artifacts.json`
-- `tool-kb-empty-scope-and-no-result` | `20260318-131741-542` | passed | `tests/logs/ai/backend/tools/kb/boundary/tool-kb-empty-scope-and-no-result/20260318-131741-542/meta.json` | `tests/logs/ai/backend/tools/kb/boundary/tool-kb-empty-scope-and-no-result/20260318-131741-542/timeline.jsonl` | `tests/logs/ai/backend/tools/kb/boundary/tool-kb-empty-scope-and-no-result/20260318-131741-542/artifacts.json`
+- `tool-kb-fts-search` | run=`20260318-131741-574` | status=`passed`
+- `tool-kb-hybrid-search` | run=`20260318-131741-664` | status=`passed`
+- `tool-kb-list-documents-and-scope-filter` | run=`20260318-131741-814` | status=`passed`
+- `tool-kb-vector-search` | run=`20260318-131741-823` | status=`passed`
+- `tool-kb-empty-scope-and-no-result` | run=`20260318-131741-542` | status=`passed`
 
 ##### 失败分析
 
@@ -136,3 +139,62 @@
 ##### 当前评估
 
 - `kb` 已覆盖 `hybrid / vector / fts / list_documents` 与 `empty scope / no-result`，且当前最新口径为 `5 / 5 passed`。
+
+### human
+
+##### 通过率
+
+- 历史总数据:
+  - totalRuns: 128
+  - passedRuns: 107
+  - failedRuns: 21
+  - historicalSuccessRate: 83.6%
+- 最新总数据:
+  - latestCaseCount: 19
+  - latestPassedCases: 18
+  - latestFailedCases: 1
+  - latestSuccessRate: 94.7%
+
+##### 分层统计
+
+- `basic`: `8` 次 raw runs，最新 `2 / 2 passed`
+- `boundary`: `24` 次 raw runs，最新 `6 / 6 passed`
+- `scenario`: `33` 次 raw runs，最新 `8 / 8 passed`
+- `live`: `11` 次 raw runs，最新 `2 / 3 passed`
+
+##### 证据
+
+- `human-direct-docx-edit` | run=`20260318-151550-605` | status=`passed`
+- `human-direct-kb-answer` | run=`20260318-151551-706` | status=`passed`
+- `human-approval-blocked-recovery` | run=`20260318-151551-755` | status=`passed`
+- `human-evidence-insufficient-safe-stop` | run=`20260318-151551-710` | status=`passed`
+- `human-no-redundant-recheck` | run=`20260318-151551-803` | status=`passed`
+- `human-partial-delivery-on-budget-limit` | run=`20260318-151551-723` | status=`passed`
+- `human-tool-failure-recovery` | run=`20260318-151550-618` | status=`passed`
+- `human-workspace-blocked-recovery` | run=`20260318-151551-679` | status=`passed`
+- `human-kb-docx-bridge` | run=`20260318-151550-641` | status=`passed`
+- `human-long-task-approval-branch` | run=`20260318-151551-710` | status=`passed`
+- `human-long-task-project-brief` | run=`20260318-151550-604` | status=`passed`
+- `human-long-task-with-corrections` | run=`20260318-151550-604` | status=`passed`
+- `human-multi-source-merge-delivery` | run=`20260318-151550-604` | status=`passed`
+- `human-multi-turn-refinement` | run=`20260318-151550-605` | status=`passed`
+- `human-review-handoff-structured-delivery` | run=`20260318-151550-605` | status=`passed`
+- `human-terminal-docx-handoff` | run=`20260318-151550-606` | status=`passed`
+- `human-live-direct-kb-answer` | run=`20260318-145113-697` | status=`passed`
+- `human-live-evidence-insufficient-safe-stop` | run=`20260318-150735-678` | status=`passed`
+- `human-live-kb-docx-bridge` | run=`20260318-151207-536` | status=`failed`
+
+##### 失败分析
+
+- 当前 latest 口径下唯一 failed case 是 `human-live-kb-docx-bridge`。
+- 2026-03-18 的 primary `gpt-5` 最新诊断 run `20260318-150310-015` 返回 `invalid_input`，summary=`524 status code (no body)`，事件只留下两次 `kb_hybrid_search_chunks`，说明链路在 live 模型侧已出现上游不稳定。
+- 2026-03-18 的 stronger `gpt-5.2` 诊断 run `20260318-150906-789` 没有触发 524，但在 `kb_hybrid_search_chunks -> kb_list_documents` 后直接判定步数不足，未进入 `docx-editing` handoff。
+- 2026-03-18 的更收敛 objective 诊断 run `20260318-151207-536` 仍然失败，表现为 `bounded_loop_error`，工具轨迹错误转入四次 `terminal_run_command` 加一次 `kb_hybrid_search_chunks`，没有产出 preview artifact。
+- 因此，当前 bridge 失败同时暴露两类问题：一是 primary live 服务侧不稳定；二是 human 在 `KB -> DOCX` 这类跨工具任务上的 live 路由和收敛仍不稳定。
+
+##### 当前评估
+
+- `basic / boundary / scenario` 当前 latest 全通过，说明 Stage 2 的脚本化回归基线已经稳定。
+- 新增补齐的 `human-live-direct-kb-answer` 与 `human-live-evidence-insufficient-safe-stop` 已形成 live 正向证据，其中前者还促成了 `KB` 收敛提示和 `localSourceStatus` 判定逻辑的修正。
+- 当前 `human/live` latestCaseCount 只有 `3`，因为本轮没有重跑此前已实现的 `human-live-direct-docx-edit`、`human-live-terminal-docx-handoff`、`human-live-long-task-project-brief`；因此 Stage 2 live 结论目前仍只能算“部分补齐 + 明确暴露 bridge 缺口”。
+- 在 `human-live-kb-docx-bridge` 稳定前，不建议把 Stage 2 视为可关闭阶段。