@thincoder/core 0.10.3 → 0.10.5

package/CHANGELOG.md

@@ -3,6 +3,68 @@
 All notable changes to the core package are documented here.
 Format: Keep a Changelog · 中文 · 号在发布时定（CalVer——见 `docs/RELEASE.md` §4）。
 
+## [0.10.5] — 2026-10-13
+
+> 0.10.4 → 0.10.5（月内 +1——发布时定号）
+
+### Added
+
+- **浏览器工具输入域（#1045）**：`press` ∥ `hover` ∥ `wheel` ∥ `mouse` ∥ `drag` ∥ `touch` ∥ `insert` ∥ `clipboard` 八动作面（回执带来源标记）∥ 异步任务 + kill 缝 ∥ 动作预算与会话自愈。
+- **`manifest` 工具（#1098）**：读 ∥ 初始化 ∥ 白名单点写（`docRoot`/`checkConfig`/`index` 子键；整对象写与表外键拒）；设计门辖区 = manifest 树（#1104）。
+- **索引自维护（#1096）**：幽灵行 GC ∥ 压实 ∥ 备份轮换 + `memory maintain` action；记忆索引行预算面向面收正（#1107）。
+- **disk-root 启动门（#1080/#1081）**：磁盘根启动拦截 ∥ 索引排除名。
+- **团队核件（B1）**：`teamVerify` 登录校验缝（三端共用）；**团队台账核工具 `team_ledger`**（五 action——写三仅主会话 ∥ 子代理只读变体 ∥ 仓派生原文上送 ∥ 登录态门零 fetch）。
+- **台账工具面深化**：证据语义（append 缺省 ∥ `replace` 旗）∥ 家族锚解析（`discoverFamily` anchor-local）。
+- **提示词纪律线族**：每个配置项须有配置 UI（用户裁定）∥ 用户视角门（U1/U2 注册）∥ 平实写作 ∥ 禁以用户名义裁定 ∥ 工程判断 ∥ 任务书再派生（#969）∥ 上抛标记（#970）∥ 板-域-文件三层结构（#956）∥ 子代理报告四列表格 ∥ **上抛前先读档**（#1328——上抛须携已查证据行）。
+- provider 配置一致性（核 + CLI 双舱）；gemini-3.8/3.7 flash 规格行（#1198）。
+
+### Changed
+
+- **provider 默认模型清退**（通道模型退役——五舱随动）。
+- gemini 思考签名捕获/传递（#1205——工具环内保真）；代理分块/传输拆分；会话槽面。
+- 普通模式批档作者缝 + 工作管理说课（#1111）。
+- 测试术语统一（unit / integration——单元测试件为开发期工具，永不入任何闸）。
+
+### Fixed
+
+- 主会话异步交付两点（投递 ∥ 消费时点收正）。
+- provider 探测随 per-provider 代理 ∥ loopback 恒直连（#1026）。
+- 上下文预算 = 可用窗口推导（advisor ∥ agent 压实不再越界）。
+- 陈旧注释 ∥ 字符串 ∥ 死导入清扫（#1131）。
+- 台账状态行写后即时刷新（库变更 `fs.watch` 去抖 ⇒ 写后 ≤~0.5s 重扫）。
+
+## [0.10.4] — 2026-10-05
+
+> 0.10.3 → 0.10.4（月内 +1——发布时定号）
+
+### Added
+
+- **零落笔看门狗（#934）**：eng-coder ∥ eng-designer ∥ 评审面连续 50 轮零文本产出 ⇒ 自动上游提醒（阈值单源叶 `zero-write-watch.mjs`）。
+- **工程模式状态档写通路（#941）**：`PROJECT-MANIFEST.json` ∥ `.thincoder/conventions.json` 合规写入面（`state` 第五值——「恰执行一次」语义；`src` 段仍拒）。
+- **空目录清理通路（#943）**：`delete` 空目录臂（非空照拒 ∥ 嵌套空树自底向上——`delete` 描述面同拍）。
+- **台账变体键首启提示（#935）**：盘符大小写变体旧库检测 + `migrate` 指引（不再静默分裂）。
+- **工具面首用可发现性（#938）**：子代理派单六字段骨架进描述面 ∥ batch 头参 ∥ 占位域修正。
+- **`ledger` 统一入口（#923）**：五工具合 1（action 分派；五旧名 = 弃壳零装配）。
+
+### Changed
+
+- **消化账务语义（#930）**：投递-结算两分 ∥ 重试上限 + 升级 ∥ 未结可见。
+- **模型可见面三笔（#950）**：派单六字段骨架 ∥ 消化签收纪律线 ∥ 停摆轮引用清理。
+- **bash 执行器身份（#922）**：运行时 shell 探针 + 注入执行器行（描述与真实执行器不再相左）。
+- **advisor 主循环拆分（#951）**：时间线记录 ∥ 压缩协作外提（行为零变——结构面）。
+
+### Fixed
+
+- **消化账务取件窗（#939）**：多轮 run 首轮签收跨轮存活（假重投根治）。
+- **设计令牌裁决闸（#940）**：评审回执无 `VERDICT: pass` 不签发（机械判 ∥ 回显径）。
+- **评审串台面（#949）**：并发/排队评审按本评审实例定域 + 投递水印（不再借他场轮次）。
+- **引文核验装饰剥离（#928）**：包裹引号 ∥ 行内注记并入捕获串不再误报失配（候选集 c1–c4；零模糊）。
+- **评审 documents 域锚祖先链（#945）**：容器相对形在子仓锚下可解析（受理判据零放宽）。
+- **批次档路径多落点歧义（#942）**：无仓前缀相对形不再静默嵌套 + 歧义显式拒。
+- **docRoot 空值哨兵（#944）**：`null` = 本面无（不再被默认值补回）。
+- **responses 适配面健壮性（#907）**：SSE error 帧双分支 ∥ 错误码映射 ∥ `max_output_tokens` 显式发送。
+- **评审文档路径基面（#921）**：按候选项目根解析 + 诊断拒语 + 七处基面声明在位。
+
 ## [0.10.3] — 2026-10-04
 
 > 0.10.2 → 0.10.3（月内 +1——发布时定号）

package/README.md

@@ -82,7 +82,7 @@ Per-release notes and artifacts: `CHANGELOG.md` in this package and the [GitHub
 The core lives in `thincoder-core/` of the ThinCoder repository; bug reports and questions go to the
 repository issue tracker. Two conventions shape every change: pure ESM `.mjs` with no build step, and
 zero third-party dependencies — if the Node.js standard library can do it, no third-party package is
-allowed. `npm test` runs the repo suite entry (`node test/run.mjs`); after the 2026-09-28 full reset the suite manifest is empty (zero tests = green — collection self-checks still run). Unit tests live as batch-local files under `docs/batches/` in the repository — current state: `docs/core/design/TESTING.md` §2.3. It runs again on release through `prepublishOnly`.
+allowed. `npm test` runs the test entry (`node test/run.mjs`); after the 2026-09-28 full reset the test manifest is empty (zero tests = green — collection self-checks still run). Unit tests live as batch-local files under `docs/batches/` in the repository — current state: `docs/core/design/TESTING.md` §2.3. It runs again on release through `prepublishOnly`.
 Mechanism design documents live in the repository under `docs/core/design/`.
 
 ## License

package/advisor/citations.mjs

@@ -7,8 +7,8 @@
  *
  * 第 11 批（C / F14 / `ADVISOR-GUARDS.md §3`）：解析候选 = cwd + **评审对象声明范围派生根**（声明仓根 /
  * 声明文件目录 / 声明目录——纯路径派生，零扫描、零 git）；命中判据三条件全中（围栏内 ∧
- * 可读 ∧ 目标行含引文内容）⇒ 零新增假命中。失败原因四类（file unreadable / content mismatch @ path /
- * path traversal / 非连续引文（省略号形——形状 vs 内容造假分列））——父侧不再人肉复核。
+ * 可读 ∧ 目标行含**任一内容候选**——c1 原捕获 / c2 剥外层装饰 / c3 首引号后段 / c4 首成对引号内层；
+ * 零模糊、零内容级归一（装饰剥离候选集 = `ADVISOR-GUARDS.md` §3）。失败原因四类（file unreadable / content mismatch @ path / path traversal / 非连续引文（省略号形——形状 vs 内容造假分列））——父侧不再人肉复核。
  */
 import { readFileSync, realpathSync } from "node:fs"
 import { resolve, relative, dirname, extname, sep, isAbsolute, join } from "node:path"
@@ -79,7 +79,7 @@ function resolveCitation(citation, roots, base) {
     } catch {
       continue // 存在但读不了（目录等）——按未命中处理，试下一候选
     }
-    if (line.includes(citation.content)) {
+    if (citationCandidates(citation.content).some((c) => line.includes(c))) {
       return { matched: true, root, resolved: relative(base, real).split(sep).join("/") }
     }
     mismatch ??= real
@@ -147,3 +147,29 @@ export function appendCitationReport(text, cwd, opts = {}) {
   }
   return text + lines.join("\n")
 }
+
+/** 装饰剥离候选集（`ADVISOR-GUARDS.md` §3 · 2026-10-05 批 · 台账 #928）：评审引风常含包裹引号 ∥ 行内
+ *  注记（`file:line: "…引文…" — 注记`）——捕获串带装饰时字节包含判失败 ⇒ 误报 content mismatch。
+ *  候选（任一为所引行**连续子串**即命中——零模糊、零内容级归一）：
+ *  c1 原捕获（现行行为——零回归，不设下限）∥ c2 反复剥「引号族 + 空白」的先导/尾随 ∥ c3 首个引号
+ *  字符后段（剥前注记）∥ c4 首个成对引号段内层文本；派生候选下限 2 字符（空引号 ∥ 单字符垃圾形
+ *  不成候选）。 */
+function citationCandidates(raw) {
+  const out = [raw] // c1——原捕获（零回归）
+  const add = (s) => { if (s.length >= 2 && !out.includes(s)) out.push(s) }
+  const DECOR = /^[\s"'“”「」『』]+|[\s"'“”「」『』]+$/g
+  let s = raw
+  for (;;) { const t = s.replace(DECOR, ""); if (t === s) break; s = t }
+  add(s) // c2——剥离外层装饰（反复至不动点）
+  const qi = raw.search(/["'“”「」『』]/)
+  if (qi >= 0) add(raw.slice(qi + 1).trim()) // c3——首个引号字符后段（剥前注记）
+  const spans = []
+  for (const [open, close] of [['"', '"'], ["“", "”"], ["「", "」"]]) {
+    const a = raw.indexOf(open)
+    if (a < 0) continue
+    const b = raw.indexOf(close, a + 1)
+    if (b > a + 1) spans.push({ a, inner: raw.slice(a + 1, b) })
+  }
+  if (spans.length) add(spans.sort((x, y) => x.a - y.a)[0].inner) // c4——首个成对引号段内层文本
+  return out
+}

package/advisor/compaction.mjs

@@ -7,9 +7,10 @@
  * re-attach — F13/`ADVISOR-GUARDS.md §2` #3); renderTimeline moved verbatim too, so the tail
  * GENERATOR and the tail CLASSIFIER stay in one file with the assembler they
  * feed (`ADVISOR-GUARDS.md §1` 谓词 ↔ §4 结构化尾——同族单源)。拆分线 = 行数硬帽实测（见批次档 §5）。
+ * advisor-loop-split 批：loop 的压缩检查点归位本档（`compactContextIfNeeded`——`context_limit` 尾串生成点同迁；时间线面 ⇒ `advisor/timeline.mjs`）。
  */
 
-import { providerSpec } from "../config.mjs" // 第 25 批：预算派生（与 loop.mjs:11 同源导入）
+import { providerSpec, outputReserve, usableWindow } from "../config.mjs" // 第 25 批：预算派生（与 loop.mjs:12 同源导入）；`outputReserve`/`usableWindow` = 2026-10-06 预留裁定（#975）
 // B4（群 B 批，ADVISOR-GUARDS.md §9——F32）：CJK 加权单源（provider/rate.mjs 叶子向无环）
 import { estimateText } from "../provider/rate.mjs"
 
@@ -21,14 +22,19 @@ export const MAX_ADVISOR_TURNS = 100
 
 // Context window limits
 // 上下文预算（第 25 批——120K 硬编码退场）：预算跟随评审模型窗口（providerSpec：
-// 模型规格表 × provider 级 context 覆盖）。头寸用途 = chars/4 估算误差 + 响应/协议开销
-// （内存不构成约束——`ADVISOR-GUARDS.md §8`）；判死线仍是宿主机自限线，服务端窗口约束不变。
-export const CONTEXT_LIMIT_RATIO = 0.8  // 判死线 = 窗口 × 0.8
+// 模型规格表 × provider 级 context 覆盖）**扣完成预留**（2026-10-06 裁定——预留坐在窗口里 ⇒
+// 压缩先行、不判死；基数 = 可用窗口，见 advisorContextBudget）。头寸用途 = chars/4 估算误差 +
+// 响应/协议开销（内存不构成约束——`ADVISOR-GUARDS.md §8`）；判死线仍是宿主机自限线，服务端窗口约束不变。
+export const CONTEXT_LIMIT_RATIO = 0.8  // 判死线 = 可用窗口 × 0.8
 const COMPACT_TRIGGER_RATIO = 0.8       // 压缩触发 = 判死线 × 0.8（既有关系零改）
 
-/** 评审上下文预算（纯函数——两档阈值可机测；provider 为 null 时退化默认规格）。 */
+/** 评审上下文预算（纯函数——两档阈值可机测；provider 为 null 时退化默认规格）。
+ *  基数 = **可用窗口 = 窗口 − 完成预留**（#975／用户 2026-10-06 裁定——预留坐在窗口里 ⇒
+ *  先压缩、不许 provider 先拒：provider 口径 = 输入 + max_tokens ≤ 窗口；预留 = `outputReserve`）。 */
 export function advisorContextBudget(provider) {
-  const limit = Math.floor(providerSpec(provider).context * CONTEXT_LIMIT_RATIO)
+  const spec = providerSpec(provider)
+  const usable = usableWindow(spec.context, outputReserve(provider, spec))
+  const limit = Math.floor(usable * CONTEXT_LIMIT_RATIO)
   return { limit, compactAt: Math.floor(limit * COMPACT_TRIGGER_RATIO) }
 }
 
@@ -82,13 +88,31 @@ export function compactMessages(messages, pinned = null) {
     ...recent)
 }
 
+/** 压缩检查点（自 advisor/loop.mjs 迁出——advisor-loop-split 批）：超 `budget.compactAt` ⇒ `onText` 通知 +
+ *  `compactMessages` 原地重写（F13 pinned 重挂随它）；重写后仍超 `budget.limit` ⇒ 返回上下文超限尾串
+ *  （调用点收尾——`context_limit` 生成点），否则 `null`。 */
+export function compactContextIfNeeded(messages, budget, pinned, onText) {
+  // Check context window and compact if needed
+  const currentTokens = estimateTokens(messages)
+  if (currentTokens > budget.compactAt) {
+    onText(`\n[Context compacted: ${currentTokens} tokens → reducing to fit window]\n`)
+    compactMessages(messages, pinned)
+    if (estimateTokens(messages) > budget.limit) {
+      // Report the POST-compaction count — the pre-compaction currentTokens
+      // is stale by the time compaction has run.
+      return `Advisor: context window limit reached (${estimateTokens(messages)} tokens). Review incomplete — too many tool calls. Try a narrower scope.`
+    }
+  }
+  return null
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // 不完整判定族（A / F16 共用单谓词——`ADVISOR-GUARDS.md §1`；六 kind = 宿主尾族）
 // ─────────────────────────────────────────────────────────────────────────────
 
 /** 宿主尾族六 kind 的块首行逐字前缀（`ADVISOR-GUARDS.md §1` 表）。变量段（token 数 / 秒数 / 工具轮数）
  *  不入前缀——取各尾的固定字面部分；`review_failed` = run.mjs catch 的字符串 resolve
- *  形态（不 throw），其余五条 = renderTimeline 尾（loop.mjs）。 */
+ *  形态（不 throw）；`context_limit` 生成点 = 本档 `compactContextIfNeeded`，其余四条 = renderTimeline 尾（loop.mjs）。 */
 const ADVISOR_INCOMPLETE_PREFIXES = [
   ["context_limit", "Advisor: context window limit"],
   ["turn_cap", "Advisor: stopped after"],

package/advisor/history.mjs

@@ -17,13 +17,15 @@ const DEFAULT_CRITERIA = `Review the code changes, focusing on:
 /**
  * Extract the agent's response table (| # | Action | Detail |) — the fix-claims
  * reference for convergence rounds.
- * Semantics (decision 2026-08-08): without sinceIdx, scan BACKWARD for the
- * MOST RECENT response table (no prior-table index is carried anymore — the
- * agent response is a focus aid only; format drift falls back to the
- * no-response text and never drives control flow). With sinceIdx, scan
- * FORWARD from it (legacy callers/tests).
+ * Semantics: with sinceIdx, scan FORWARD from it — the review instance passes
+ * its delivery watermark (`agent._advisorResponseAnchor`), so the table taken
+ * is the one written after THIS review's delivery. Without sinceIdx (legacy
+ * direct callers — no watermark), fall back to scanning BACKWARD for the most
+ * recent table. The agent response is a focus aid only — format drift falls
+ * back to the no-response text and never drives control flow (decision
+ * 2026-08-08).
  * @param {Array} history — message history
- * @param {number} [sinceIdx] — legacy: start scanning forward from this index
+ * @param {number} [sinceIdx] — delivery watermark: start scanning forward from this index
  * @returns {string|null} the response table content, or null
  */
 export function extractAgentResponseTable(history, sinceIdx) {

package/advisor/loop.mjs

@@ -2,6 +2,7 @@
  * advisor/loop.mjs — advisor tool loop: chat → execute tools → repeat, plus the
  * review timeline (split out of advisor/run.mjs, 第 11 批 — run.mjs was 498/500
  * 硬帽；拆分保持既有 import 面：run.mjs 继续 re-export 本文件导出）。
+ * 自本档再拆（advisor-loop-split 批）：时间线记录面 ⇒ `advisor/timeline.mjs`；压缩协作面 ⇒ `advisor/compaction.mjs`（`compactContextIfNeeded`）。
  *
  * 第 11 批（F15/`ADVISOR-GUARDS.md §4`）：每次 chat 调用携带硬墙信号（`AbortSignal.any([signal,
  * AbortSignal.timeout(remaining)])`；墙判定绑信号状态——抛错 / partial 两形态同判），
@@ -13,10 +14,11 @@ import { toOpenAISchema } from "../tools/index.mjs"
 import { truncateAdvisorResult } from "./truncate.mjs"
 import { batchTool } from "../agent-tools/batch.mjs"
 import {
-  estimateTokens, compactMessages, shouldBudgetNudge, budgetNudgeText, timeoutTail, renderTimeline,
+  shouldBudgetNudge, budgetNudgeText, timeoutTail, renderTimeline, compactContextIfNeeded,
   MAX_ADVISOR_TURNS, advisorContextBudget, TOOL_TIMEOUT_MS, REVIEW_TIMEOUT_MS, MAX_RESULT_CHARS,
   ADVISOR_THINKING_PLACEHOLDER,
 } from "./compaction.mjs"
+import { createTimelineRecorder } from "./timeline.mjs"
 
 const { readTool, globTool, grepTool, lsTool } = await import("../tools/index.mjs")
 const { lspTool } = await import("../tools/lsp.mjs")
@@ -68,21 +70,7 @@ async function runAdvisorToolLoop(provider, messages, onOutput, signal, agent, c
   // 未注入 ⇒ 不发进度行（核内零端名分支——契约 5 / 10）。端在装配层传入自己的格式化器。
   const describeArgs = seams.describeArgs ?? (() => "")
   // 第 11 批硬墙 / 预算 / 尾：实现注解见下方各点；守卫函数与 renderTimeline 在 compaction.mjs。
-  // Kind-tagged wrappers: the TUI panel colors reasoning / answer / tool progress differently.
-  // Every chunk is ALSO recorded into an ordered timeline — the persisted record
-  // must show the review process (thinking ↔ tool progress ↔ final text) at its
-  // real positions, not a summary appended at the end. Same-kind consecutive
-  // chunks merge (token streams); kind flips start a new entry.
-  const timeline = []
-  const record = (kind, text) => {
-    const last = timeline.at(-1)
-    if (last && last.kind === kind) last.text += text
-    else timeline.push({ kind, text })
-  }
-  const emit = (kind) => (text) => { record(kind, text); onOutput?.({ kind, text }) }
-  const onThink = emit("think")
-  const onText = emit("text")
-  const onTool = emit("tool")
+  const { timeline, onThink, onText, onTool } = createTimelineRecorder(onOutput)
   // toolsOverride = test seam (T-TS8/9): the real advisor tool set, or a mock
   // set with controllable timing/errors.
   const { schemas: toolSchemas, byName: toolByName } = toolsOverride ?? advisorToolsFor(agent, reviewType, batchDoc)
@@ -90,10 +78,20 @@ async function runAdvisorToolLoop(provider, messages, onOutput, signal, agent, c
   let toolCallCount = 0
   let reviewTextProduced = false
   let budgetNudged = false
+  // #934（零落笔看门狗——§6.32.2 ②）：连续无文本产出轮计数（循环局部——单次评审运行内计数，
+  // 修复轮 = 新运行 ⇒ 重新计数）+ 单次运行通报闩（文本产出后不再触发——「连续无产出」条件不回头）
+  // + 本轮有无文本标记（轮顶结上一轮用）。
+  let silentRounds = 0
+  let stallAlerted = false
+  let roundHadText = false
   const startTime = now()
   // 第 25 批（`ADVISOR-GUARDS.md §8`）：上下文预算跟随评审模型窗口——`providerSpec`（模型规格表 × provider 级
   // context 覆盖）派生；函数体内、while 轮次外一次性（provider 全场不变），两档消费见下守卫。
   const budget = advisorContextBudget(provider)
+  // #934（零落笔看门狗——AGENT-LOOP-UPSTREAM.md §6.32.2 ②）：阈值单源（叶档 agent-tools/zero-write-watch.mjs）——
+  // 函数域动态 import：模块域会与既有环（叶 → parent-channel → async-settle ↔ advisor-async → 本档）
+  // 互锁（加载期 TLA 死锁）；函数运行时图已就绪，模块缓存 ⇒ 每次评审一次代价。
+  const { ZERO_WRITE_ALERT_ROUNDS } = await import("../agent-tools/zero-write-watch.mjs")
 
   while (true) {
     // Interrupted (Ctrl+I) — stop immediately instead of spinning a fresh uncancellable signal
@@ -120,19 +118,23 @@ async function runAdvisorToolLoop(provider, messages, onOutput, signal, agent, c
     if (++turns > MAX_ADVISOR_TURNS) {
       return renderTimeline(timeline, "Advisor: stopped after " + MAX_ADVISOR_TURNS + " tool rounds — the review appears to be looping. You may retry with a narrower scope.")
     }
-    
-    // Check context window and compact if needed
-    const currentTokens = estimateTokens(messages)
-    if (currentTokens > budget.compactAt) {
-      onText(`\n[Context compacted: ${currentTokens} tokens → reducing to fit window]\n`)
-      compactMessages(messages, pinned)
-      if (estimateTokens(messages) > budget.limit) {
-        // Report the POST-compaction count — the pre-compaction currentTokens
-        // is stale by the time compaction has run.
-        return renderTimeline(timeline, `Advisor: context window limit reached (${estimateTokens(messages)} tokens). Review incomplete — too many tool calls. Try a narrower scope.`)
+
+    // #934（零落笔看门狗——§6.32.2 ② / §6.32.6 触发接线②）：轮顶结上一轮（首轮无上一轮可结）——
+    // 上一轮无文本 ⇒ silentRounds += 1；越阈且未通报 ⇒ 恰一次通报（可选钩子——缺省零行为；
+    // 钩子恒零抛：抛错不得上抛为评审 failed 面——实现面兜底捕获兑现）。
+    if (turns > 1) {
+      if (!roundHadText) silentRounds += 1
+      roundHadText = false
+      if (!stallAlerted && silentRounds >= ZERO_WRITE_ALERT_ROUNDS) {
+        stallAlerted = true
+        try { seams.onStallRound?.(silentRounds) } catch { /* #934：恒零抛 */ }
       }
     }
     
+    // 压缩协作（迁出 compaction.mjs `compactContextIfNeeded`）：超限尾 ⇒ 收尾
+    const contextTail = compactContextIfNeeded(messages, budget, pinned, onText)
+    if (contextTail) return renderTimeline(timeline, contextTail)
+    
     // LLM generation silence: the reasoning phase produces no SSE bytes for
     // seconds to tens of seconds (server-side prefill on large contexts, per
     // tool-round LLM return). A placeholder keeps the panel visibly working.
@@ -155,7 +157,7 @@ async function runAdvisorToolLoop(provider, messages, onOutput, signal, agent, c
         messages,
         tools: toolSchemas,
         signal: callSignal,
-        onToken: (t) => { if (String(t ?? "").trim()) reviewTextProduced = true; onText(t) },
+        onToken: (t) => { if (String(t ?? "").trim()) { reviewTextProduced = true; silentRounds = 0; roundHadText = true } onText(t) },
         onReasoning: onThink,
         // LOGGING（vscode advisor/run.mjs parity——按 stage 可 grep）
         // TRACES.md §6.1 D-TR4：轨迹元数据增补——kind=advisor（评审独立于子代理——T-TR2）；role

package/advisor/messages.mjs

@@ -53,6 +53,7 @@ export function buildDesignApprovalBlock(designToken, designId) {
   return [
     "## Approval Signal",
     echo,
+    "The verdict line is checked mechanically: without a `VERDICT: pass` line, or with a `VERDICT: changes-required` line, no token is issued even when the token is echoed.",
     "🟡 (Advisory) and 🔵 (Note) findings do NOT block approval — list them if present, but still include the token. If any 🔴 (Critical) issue remains unresolved, do NOT include the token. A reasoned non-fix accepted by the reviewer does not block (🔴 only — 🟡/🔵 never block approval); anything else stays unresolved.",
   ].join("\n")
 }
@@ -195,7 +196,7 @@ function buildAdvisorUserMessageInner(agent, prior, reviewType, designToken = nu
   // conversation can quote literal "\x"/"\u" sequences (server 400 risk).
   if (p && (agent._advisorRound || 0) > 0) {
     const scopeFiles = resolveScopeFiles(agent, paths)
-    const response = extractAgentResponseTable(agent.history)
+    const response = extractAgentResponseTable(agent.history, agent._advisorResponseAnchor ?? undefined)
       || (scopeFiles?.length
         ? "(Agent did not provide a response table — perform a fresh review of: " + scopeFiles.slice(0, 10).join(", ") + ")"
         : "(Agent did not provide a response table — perform a fresh review of the files named in the system prompt context)")

package/advisor/run.mjs

@@ -32,10 +32,10 @@ export function resolveAdvisorProvider(agent) {
       const provider = findProvider(agent.providers?.length ? agent.providers : agent.config?.providersList ?? [agent.provider], cfg.provider)
       // F-2a (MODEL-400-FIX)：无 cfg.model 时渠道克隆须重派生 model——渠道裸克隆会丢 model 键
       // → 无 model 请求 → serde 400。
-      // MODEL-SELECTION v2（M3④）：无 cfg.model → 命中渠道自己的默认模型（`provider.model`
-      // 单值）；渠道无默认模型 → 父 provider 兜底（与 subagent F-2c 同构）——绝不产出静默
-      // undefined-model 请求（最极端两者皆无 → chat 前 assertProviderModel fail-fast）。
-      const result = cfg.model ? { ...provider, model: cfg.model } : { ...provider, model: provider.model ?? agent.provider?.model }
+      // 2026-10-09 清除批（与 CLI `cmd-advisor` 同族收正）：模型面 = **父解析单档**——渠条目不携
+      // 模型（渠道单值模型退场）⇒ 无 cfg.model 取父运行模型 `agent.provider?.model`；两者皆无 →
+      // chat 前 `assertProviderModel` fail-fast（不静默）。
+      const result = cfg.model ? { ...provider, model: cfg.model } : { ...provider, model: agent.provider?.model }
       // 台账 #329 / `MODEL-SPECS.md §15.4-3`：off 形须**活着到达载荷层**——载荷层 off 门首款 = `provider.thinking === null`
       // （`thincoder-core/provider/core.mjs:213-219`）；归一为 `undefined` 会让该门永不开（effort 族 advisor
       // 关思考静默失效）。显式 off 同清**继承档**（渠条目自带 / 主 provider 的 `reasoningEffort` 不得随行——
@@ -140,7 +140,11 @@ export async function runAdvisorReview(agent, reviewType, callbacks, designToken
   const pinned = buildPinnedBrief(reviewType, documents, object, designToken, designId)
 
   try {
-    const result = await runAdvisorToolLoop(provider, messages, onOutput, signal, agent, advisorCwd, null, reviewType, boundBatchDoc, pinned)
+    // #934（零落笔看门狗——§6.32.6 触发接线②）：评审族闸位钩子转发（callbacks 不供 `onStall` ⇒
+    // `onStallRound: undefined` ⇒ 环路零行为——同步径（advisor 工具）不供钩子）。
+    const result = await runAdvisorToolLoop(provider, messages, onOutput, signal, agent, advisorCwd, null, reviewType, boundBatchDoc, pinned, {
+      onStallRound: callbacks?.onStall,
+    })
 
     // Host-verified citations (decision d698434): mechanically check every
     // `file:line: content` reference in the review against the CURRENT file

package/advisor/timeline.mjs

@@ -0,0 +1,26 @@
+/**
+ * advisor/timeline.mjs — 评审环路的有序时间线记录器（自 advisor/loop.mjs 迁出——advisor-loop-split 批）：
+ *  记录序 / 同类合流 / 三 kind 包装一体；下游消费 = renderTimeline（compaction.mjs）。
+ *  零 import 叶（对照 agent/relay-prefix.mjs 先例家族）。
+ *
+ *  @param {Function} [onOutput] — 实时 chunk 出口；缺省 ⇒ 只记录不出口。
+ *  @returns {{ timeline: Array, onThink: Function, onText: Function, onTool: Function }}
+ */
+export function createTimelineRecorder(onOutput) {
+  // Kind-tagged wrappers: the TUI panel colors reasoning / answer / tool progress differently.
+  // Every chunk is ALSO recorded into an ordered timeline — the persisted record
+  // must show the review process (thinking ↔ tool progress ↔ final text) at its
+  // real positions, not a summary appended at the end. Same-kind consecutive
+  // chunks merge (token streams); kind flips start a new entry.
+  const timeline = []
+  const record = (kind, text) => {
+    const last = timeline.at(-1)
+    if (last && last.kind === kind) last.text += text
+    else timeline.push({ kind, text })
+  }
+  const emit = (kind) => (text) => { record(kind, text); onOutput?.({ kind, text }) }
+  const onThink = emit("think")
+  const onText = emit("text")
+  const onTool = emit("tool")
+  return { timeline, onThink, onText, onTool }
+}

package/advisor.mjs

@@ -124,7 +124,9 @@ export function buildAdvisorSystemPrompt(agent, prior, reviewType) {
  * NOTE: the caller (prepareAdvisorMessages) applies escapeLiteralEscapes to
  * the return value — direct callers must do the same (the prior table and
  * agent response can quote literal "\x"/"\u" sequences).
- * @param {Object} agent — the parent agent (history used for the response table)
+ * @param {Object} agent — the parent agent (history used for the response table:
+ *   scanned forward from `agent._advisorResponseAnchor` — this instance's delivery
+ *   watermark; absent → legacy backward scan)
  * @param {Object|null} prior — prior issue table (extracted from history when null)
  * @param {string[]|null} [scopeFiles] — review surface for the no-response fallback (cwd-relative)
  * @param {Object|null} [object] — review-object declaration (AGENT-LOOP-ASYNC-POOL.md §6.18): mechanically
@@ -156,7 +158,7 @@ export function buildAdvisorFollowUp(agent, prior, scopeFiles = null, object = n
   const noResponseFallback = scopeFiles?.length
     ? "(Agent did not provide a response table — perform a fresh review of: " + scopeFiles.slice(0, 10).join(", ") + ")"
     : "(Agent did not provide a response table — perform a fresh full review; the review surface is unknown, ask the user for the file list)"
-  const response = extractAgentResponseTable(agent.history) || noResponseFallback
+  const response = extractAgentResponseTable(agent.history, agent._advisorResponseAnchor ?? undefined) || noResponseFallback
   const round = (agent._advisorRound || 0) + 1
   // Review-object declaration FIRST (T-OA2 — round 2+ stays anchored, no re-archaeology).
   const declaration = buildObjectDeclarationBlock(object)

package/agent/assemble.mjs

@@ -65,7 +65,7 @@ export async function assembleAgent({ cwd, deps = {}, toolsFinalize = null }) {
   const config = D.loadConfig()
   const provider = config.provider
   const providers = config.providersList
-  // 代理注入（double opt-in: provider.proxy + config.proxy.model）；config.provider 是 loadConfig 里的
+  // 代理注入（逐渠独立：provider.proxy ∧ `proxy.uri` 在案 ⇒ 该渠模型请求经代理——无全局闸，2026-10-08）；config.provider 是 loadConfig 里的
   // 独立拷贝 —— 同步注入结果（proxyUri）。
   const injectProxy = D.injectProxy ?? (await import("../proxy.mjs")).injectProxy
   injectProxy(providers, config)

package/agent/chat-call.mjs

@@ -1,7 +1,7 @@
 /**
  * agent/chat-call.mjs — 单回合模型调用（2026-09-29 P2 三拆：agent.mjs runAgent 段 4 逐字迁入 +
  * `streamOutputAllowed` 定义随迁——agent.mjs re-export 保 import 面）。
- * 接口 = callModelTurn(agent, ctx) → response；错误走原样：AbortError-中断臂处理完照旧 `throw e`，
+ * 接口 = callModelTurn(agent, ctx) → response；错误走原样：AbortError-中断分支处理完照旧 `throw e`，
  * 由 runAgent 的 catch/finally 收束。
  */
 import { chat } from "../provider/index.mjs"

package/agent/child-marks.mjs

@@ -3,8 +3,8 @@
  *
  * 收编自 `spawn-child.mjs`（2026-09-20 端差·显示面消差批 §2.2 X6 实现轮）：端壳侧需**静态**
  * 导入锚点做块头注记判定，而 `spawn-child.mjs` 静态链经核 agent 栈可达 `node:sqlite`
- * ⇒ 端壳静态闭包禁达（W8 契约②判据（现载体 = 批件 `docs/batches/2026-09-29-residuals-round2.test.mjs`；
- * 单测树重建时回迁端侧单测档）：fail-closed 红）⇒ 锚点下沉**零依赖叶**（先例 = `relay-prefix.mjs` 同因下沉）。`spawn-child.mjs`
+ * ⇒ 端壳静态闭包禁达（W8 契约②判据（现载体 = 单元测试档 `docs/batches/2026-09-29-residuals-round2.test.mjs`；
+ * 单元测试树重建时回迁端侧单元测试档）：fail-closed 红）⇒ 锚点下沉**零依赖叶**（先例 = `relay-prefix.mjs` 同因下沉）。`spawn-child.mjs`
  * 原样再导出 ⇒ 既有 import 面（核 agent-tools 族 / CLI `tool-events.mjs`）零改。
  *
  * 语义：消费端用 `includes()` 检出锚点 ⇒ 合成块头注记（"work may be partial"）——两端各一处：

package/agent/completion.mjs

@@ -28,10 +28,12 @@ const MAX_EMPTY_RETRIES = 2
  * @param {object} callbacks - { onTurnEnd, onSubTurnBreak }
  */
 export function handleCompletion(agent, response, depth, turn, guardPushbacks, honestReminderInjected, advisorPushbacks, callbacks) {
-  if (!response.content) {
+  if (String(response.content ?? "").trim() === "") {
     // Transient empty response (reasoning exhausted / output truncated): instead of
     // aborting the whole turn, inject a reminder and let the model respond again.
     // Bounded — after MAX_EMPTY_RETRIES consecutive empties, surface the original error.
+    // §6.31.8 F-DA6（#929 同族 · 批 digest-accounting）：空判加宽为**空白判**（纯空白内容
+    // 原过判 truthy ⇒ 静默当正常收尾）；既有重试（≤2）/ 抛错骨架零改。
     const retries = agent._emptyRetries ?? 0
     if (retries < MAX_EMPTY_RETRIES) {
       agent._emptyRetries = retries + 1