thincoder 0.12.60 → 0.12.62

package/src/acp.mjs

@@ -19,7 +19,7 @@ import { assembleAgent } from "./cli/make-agent.mjs"
 import { createAcpServer, ACP_ERRORS } from "./acp/transport.mjs"
 import { createAcpSession } from "./acp/session.mjs"
 import { replayHistory } from "./acp/bridge.mjs"
-import { listSlots, applySession, deleteSlot, normalizeCwd, loadSlotFile, slotOccupancy, loadManifest, saveManifest, getSessionId, newSession } from "./session.mjs"
+import { listSlots, applySession, deleteSlot, normalizeCwd, loadSlotFile, slotOccupancy, loadManifest, saveManifest, getSessionId, newSession, slotPath, bindRecordStore } from "./session.mjs"
 import { createCheckpoint, listCheckpoints, rewind, isGitRepo } from "./git/checkpoint.mjs"
 import { createMemory, list as memList, remove as memRemove } from "./memory.mjs"
 
@@ -71,12 +71,17 @@ export function defaultIsConfigured() {
   }
 }
 
+/** 第 27 批 §12.3⑤（R-A1.1）：ACP 会话装配期剔除的工具。ACP 是 headless 通道——
+ *  question 工具需要交互 UI（`src/tools/question.mjs` 无 `ctx.onQuestion` 即 throw），
+ *  模型可见即会调用、调用必错 → 装配期剔除（schema 直接少一条）。 */
+export const ACP_EXCLUDED_TOOLS = ["question"]
+
 /** M1/M2 session factory: one agent per session, built from the process cwd (single-cwd model).
  *  `id` is the ACP session id — it is baked into the callbacks at construction time
  *  (buildAcpCallbacks closure), so it must be known BEFORE createAcpSession runs.
  *  `request` is the transport's reverse-RPC channel (permissions + fs routing). */
 export async function defaultCreateSession({ id, notify, request, log }) {
-  const agent = await assembleAgent()
+  const agent = await assembleAgent({ excludeTools: ACP_EXCLUDED_TOOLS })
   return createAcpSession({ id, agent, notify, request, log })
 }
 
@@ -152,6 +157,9 @@ export function buildAcpHandlers({
           // mySessionId 同进程恒真）→ 第二个会话拿到与第一个相同的槽号 → 双写同槽
           // F2 互旋。getSessionId() 是进程级，_slot 是 agent 级——粒度错配必须在此切断。
           session.agent._slot = newSession(getCwd())
+          // TUI-OOM-ROOTCAUSE（SESSION.md §14.3.4）：新建槽绑定记录存储（baseHistory 空——
+          // identity 待固化）——与 /new 同语义
+          bindRecordStore(session.agent, { slotFile: slotPath(getCwd(), session.agent._slot), identity: session.agent._sessionStart ?? null, baseHistory: [] })
           sessions.set(id, session)
           return { id, configOptions: [{ configId: "model" }, { configId: "thinking" }, { configId: "mode" }] }
         } catch (e) {
@@ -226,7 +234,6 @@ export function buildAcpHandlers({
         try {
           const id = String(nextId++)
           const session = await createSession({ id, notify: notifyRef.current, request: requestRef.current, log })
-          applySession(session.agent, data)
           // 2026-08-31 advisor round2 🟡：钉 _slot 前查活主——目标槽被另一活进程（CLI/另一
           // IDE）占用时不得钉回（双方 sessionStart 一致 → F2 永不轮转 → 同槽 last-write-wins
           // 静默互覆盖）。空闲则认领后钉回；占用则不钉 → 下次保存经 activeSlot 自然 fork
@@ -236,7 +243,11 @@ export function buildAcpHandlers({
           // 同进程防护完全由 sameProcessPinned 承担：本进程另一 session 已钉该槽即视为占用
           // （进程级属主无法区分 agent，双方 sessionStart 相同 → F2 永不触发 → 静默互覆盖）。
           const sameProcessPinned = [...sessions.values()].some((s) => s.agent?._slot === slot)
-          if (!occ.occupied && !sameProcessPinned) {
+          const pinned = !occ.occupied && !sameProcessPinned
+          // TUI-OOM-ROOTCAUSE（SESSION.md §14.3.4）：钉槽（非占用）才绑定记录存储——
+          // 占用/fork 分支模式 F（首保存 fork 新槽后补绑）
+          applySession(session.agent, data, pinned ? { slot } : {})
+          if (pinned) {
             const m = loadManifest(getCwd())
             m.slotSessions ??= {}
             m.slotSessions[slot] = getSessionId()
@@ -273,13 +284,15 @@ export function buildAcpHandlers({
         try {
           const id = String(nextId++)
           const session = await createSession({ id, notify: notifyRef.current, request: requestRef.current, log })
-          applySession(session.agent, data)
           // 2026-08-31 advisor round2 🟡：同 session/load——活主占用的槽不钉回（防同槽双写，
           // 下次保存 fork 新槽）；空闲则认领后钉回。
           const occ = slotOccupancy(getCwd(), slot)
           // 2026-09-01 会诊 kimi/glm 🔴：同 session/load——同进程其他 session 已钉同槽视为占用 → fork
           const sameProcessPinned = [...sessions.values()].some((s) => s.agent?._slot === slot)
-          if (!occ.occupied && !sameProcessPinned) {
+          const pinned = !occ.occupied && !sameProcessPinned
+          // TUI-OOM-ROOTCAUSE（SESSION.md §14.3.4）：同 load——钉槽才绑定（占用/fork 模式 F）
+          applySession(session.agent, data, pinned ? { slot } : {})
+          if (pinned) {
             const m = loadManifest(getCwd())
             m.slotSessions ??= {}
             m.slotSessions[slot] = getSessionId()

package/src/advisor/citations.mjs

@@ -4,9 +4,14 @@
  * a host fact: every `file:line: content` reference in a review is mechanically
  * checked against the CURRENT disk state; mismatches mark the finding
  * unverified and cannot support a push-back.
+ *
+ * 第 11 批（C / F14 / §14.5）：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。失败原因三分（file unreadable /
+ * content mismatch @ path / path traversal）——父侧不再人肉复核。
  */
 import { readFileSync, realpathSync } from "node:fs"
-import { resolve, sep } from "node:path"
+import { resolve, relative, dirname, extname, sep, isAbsolute, join } from "node:path"
 
 // `file:line: content` citations — the file group is narrowed to source/config
 // extensions so URLs (`example.com:8080: …`) don't become false-positive
@@ -25,48 +30,105 @@ export function extractCitations(text) {
   return out
 }
 
+/** 候选解析根（§14.5——纯路径派生，候选顺序 = cwd → 各声明路径的派生根）：
+ *  声明仓根（`cwd/<segs[0]>`）与 声明文件目录 / 声明目录本身。声明在 cwd 之外 → 不派生
+ *  （relative 越出 cwd 或跨盘符——候选与声明脱节即假命中面）。 */
+function citationRoots(cwd, scope) {
+  const base = resolve(cwd)
+  const roots = [base]
+  for (const s of Array.isArray(scope) ? scope : []) {
+    if (typeof s !== "string" || !s.trim()) continue
+    const abs = resolve(base, s)
+    const rel = relative(base, abs)
+    if (!rel || rel.startsWith("..") || isAbsolute(rel)) continue
+    const segs = rel.split(/[\\/]/).filter(Boolean)
+    if (segs.length === 0) continue
+    const repoRoot = join(base, segs[0])
+    const declaredDir = extname(abs) ? dirname(abs) : abs
+    for (const r of [repoRoot, declaredDir]) if (!roots.includes(r)) roots.push(r)
+  }
+  return roots
+}
+
+/** 单引文解析（§14.5）：按候选顺序试 `resolve(root, file)`；三条件全中才算命中。
+ *  @returns {{matched: true, root: string, resolved: string}|{matched: false, reason: string}} */
+function resolveCitation(citation, roots, base) {
+  const fence = base + sep
+  let mismatch = null
+  let traversal = false
+  for (const root of roots) {
+    let real
+    try {
+      // Path confinement: citation paths are LLM-generated — never trust them.
+      // A hallucinated "../config.json" would otherwise read (and leak via the
+      // report) files outside the project, including API-key configs.
+      // realpathSync resolves symlinks too — a link inside the project that
+      // points outside must not pass the prefix check.
+      real = realpathSync(resolve(root, citation.file))
+    } catch {
+      continue // 该候选无此文件（或不可解析）——试下一候选
+    }
+    if (!real.startsWith(fence)) { traversal = true; continue }
+    let line = ""
+    try {
+      line = readFileSync(real, "utf8").split("\n")[citation.line - 1] ?? ""
+    } 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)) {
+      return { matched: true, root, resolved: relative(base, real).split(sep).join("/") }
+    }
+    mismatch ??= real
+  }
+  if (mismatch) return { matched: false, reason: `content mismatch @ ${relative(base, mismatch).split(sep).join("/")}` }
+  if (traversal) return { matched: false, reason: "path traversal" }
+  return { matched: false, reason: "file unreadable" }
+}
+
 /**
  * Mechanically verify citations against the CURRENT file state: read the file,
  * take the exact line, check it CONTAINS the quoted content. Reports
  * N/M matched + the mismatches. Unverified citations cannot support a
  * push-back — the evidence rule becomes a host fact, not a prompt wish.
+ * @param {string} text — the review text
+ * @param {string} cwd — the agent's working directory (workspace root)
+ * @param {{scope?: string[]}} [opts] — scope = 评审对象声明路径（documents + paths）；
+ *   省略 ⇒ 旧行为（仅 cwd 候选——签名向后兼容）。
  */
-export function verifyCitations(text, cwd) {
+export function verifyCitations(text, cwd, opts = {}) {
   const citations = extractCitations(text)
   const matched = []
   const failed = []
-  const root = resolve(cwd) + sep
+  const base = resolve(cwd)
+  const roots = citationRoots(cwd, opts?.scope)
   for (const c of citations) {
-    try {
-      // Path confinement: citation paths are LLM-generated — never trust them.
-      // A hallucinated "../config.json" would otherwise read (and leak via the
-      // report) files outside the project, including API-key configs.
-      // realpathSync resolves symlinks too — a link inside the project that
-      // points outside must not pass the prefix check.
-      const resolved = realpathSync(resolve(cwd, c.file))
-      if (!resolved.startsWith(root)) {
-        failed.push({ ...c, reason: "path traversal" })
-        continue
-      }
-      const line = readFileSync(resolved, "utf8").split("\n")[c.line - 1] ?? ""
-      if (line.includes(c.content)) matched.push(c)
-      else failed.push(c)
-    } catch {
-      failed.push({ ...c, reason: "file unreadable" })
+    const r = resolveCitation(c, roots, base)
+    if (r.matched) {
+      // 命中根记录（§14.5）：经派生根解析（非 cwd 直解）的命中在报告中注明解析路径。
+      matched.push(r.root === base ? c : { ...c, root: r.root, resolved: r.resolved })
+    } else {
+      failed.push({ ...c, reason: r.reason })
     }
   }
   return { total: citations.length, matched, failed }
 }
 
 /** Append the verification report to the review text (visible to the parent agent). */
-export function appendCitationReport(text, cwd) {
-  const { total, matched, failed } = verifyCitations(text, cwd)
+export function appendCitationReport(text, cwd, opts = {}) {
+  const { total, matched, failed } = verifyCitations(text, cwd, opts)
   if (total === 0) return text // no citations — nothing to verify
   const lines = [
     "",
     "---",
     `[host-verified] ${matched.length}/${total} citations match current file state.`,
   ]
+  // 命中根透明（F14/§14.5）：经声明范围派生根解析的命中逐条注明解析路径（cwd 直解的不列
+  // ——零噪音；列的正是修复前会被误报为 unreadable 的裸路径引用）。
+  const derived = matched.filter((c) => c.resolved)
+  if (derived.length > 0) {
+    lines.push("Citations resolved via the declared review scope (bare path — resolved root noted):")
+    for (const c of derived.slice(0, 10)) lines.push(`- ${c.file}:${c.line} → ${c.resolved}`)
+  }
   if (failed.length > 0) {
     lines.push("Citations that do NOT match the current file state (treat their claims as unverified):")
     for (const f of failed.slice(0, 10)) {

package/src/advisor/compaction.mjs

@@ -0,0 +1,174 @@
+/**
+ * advisor/compaction.mjs — advisor review support (split out of advisor/run.mjs,
+ * 第 11 批 — run.mjs was 498/500 硬帽): context trimming + the review's resource
+ * limits + the terminal-state guards + the review-text assembler.
+ *
+ * estimateTokens / compactMessages moved VERBATIM (the only edit is the `pinned`
+ * re-attach — F13/§14.4 #3); renderTimeline moved verbatim too, so the tail
+ * GENERATOR and the tail CLASSIFIER stay in one file with the assembler they
+ * feed (§14.3 谓词 ↔ §14.6 结构化尾——同族单源)。拆分线 = 行数硬帽实测（见批次档 §5）。
+ */
+
+import { providerSpec } from "../config.mjs" // 第 25 批：预算派生（与 loop.mjs:11 同源导入）
+// B4（群 B 批，CLI §18.3——F32）：CJK 加权单源（provider/rate.mjs 叶子向无环）
+import { estimateText } from "../provider/rate.mjs"
+
+export const MAX_ADVISOR_TURNS = 100
+// NOTE: prompts/advisor-round{1,2,3}.md encourage the model to finish within
+// ~30 tool turns — a prompt-level efficiency target, DISTINCT from the
+// 100-turn mechanical hard cap (MAX_ADVISOR_TURNS above; pure runaway-loop
+// guard). They serve different purposes; do NOT synchronize them.
+
+// 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——设计 §16.4）；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
+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）。 */
+export function advisorContextBudget(provider) {
+  const limit = Math.floor(providerSpec(provider).context * CONTEXT_LIMIT_RATIO)
+  return { limit, compactAt: Math.floor(limit * COMPACT_TRIGGER_RATIO) }
+}
+
+export const TOOL_TIMEOUT_MS = 30_000 // single tool timeout
+export const REVIEW_TIMEOUT_MS = 600_000 // whole review timeout (10 minutes)
+export const MAX_RESULT_CHARS = 64 * 1024 // tool result truncation (line-aware; 64K, aligned with main offload limit)
+const MAX_KEY_FILES_IN_COMPACTION = 5 // files named in the compaction summary
+
+/** Estimate token count from messages（B4——群 B 批 CLI §18.3：扁平 chars/4 改 `estimateText`
+ *  加权式——ASCII/4 + 非 ASCII/1；纯 ASCII 与旧式逐值相等；CJK 低估 ~3-4× 修正；
+ *  walker（content / tool_calls 两源）与计数口径零改）。 */
+export function estimateTokens(messages) {
+  return messages.reduce((sum, msg) => {
+    const content = typeof msg.content === "string" ? msg.content : JSON.stringify(msg.content || "")
+    const toolCalls = msg.tool_calls ? JSON.stringify(msg.tool_calls) : ""
+    return sum + estimateText(content + toolCalls)
+  }, 0)
+}
+
+/** Compact early messages when context grows too large — LOCAL trimming only
+ *  (no LLM summarization). MUTATES in place (splice) so the caller's array
+ *  reference stays valid — a reassignment would leave the caller's logging
+ *  (tool-call count, token estimate) reading a stale array. */
+export function compactMessages(messages, pinned = null) {
+  // Keep: system prompt, last 20 messages (≈ 10 assistant+tool exchanges),
+  // user message — the rest is summarized.
+  if (messages.length <= 20) return
+
+  const system = messages[0]
+  const recent = messages.slice(-20)
+  const old = messages.slice(1, -20)
+
+  // Count actual tool messages (old.length counts user/assistant rows too)
+  const toolCount = old.filter((m) => m.role === "tool").length
+  const keyFiles = old
+    .filter((m) => m.role === "tool")
+    .map((m) => m.content?.split("\n")[0]?.slice(0, 50)) // first line of tool results typically names the file that was read/grepped
+    .filter(Boolean)
+    .slice(0, MAX_KEY_FILES_IN_COMPACTION)
+  const filesPart = keyFiles.length > 0 ? ` Key files examined: ${keyFiles.join(", ")}` : ""
+  const summary = `Earlier exploration: ${toolCount} tool calls completed.${filesPart}`
+
+  // F13（第 11 批 §14.4 #3）：压缩丢掉的正是**首条 user 消息**（评审简报，含 token）——
+  // pinned 由评审参数构建（非模型输出），在本次压缩动作内作为一条 user 消息重挂（幂等可读：
+  // 重复压缩允许重复挂回，不做存在性判定）。
+  const pin = pinned ? [{ role: "user", content: pinned }] : []
+  messages.splice(0, messages.length,
+    system,
+    { role: "user", content: `[Context compacted] ${summary}` },
+    ...pin,
+    ...recent)
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 不完整判定族（A / F16 共用单谓词——§14.3；六 kind = 宿主尾族）
+// ─────────────────────────────────────────────────────────────────────────────
+
+/** 宿主尾族六 kind 的块首行逐字前缀（§14.3 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）。 */
+const ADVISOR_INCOMPLETE_PREFIXES = [
+  ["context_limit", "Advisor: context window limit"],
+  ["turn_cap", "Advisor: stopped after"],
+  ["timeout", "Advisor: review timeout"],
+  ["empty", "Advisor: empty response"],
+  ["interrupted", "Advisor: interrupted."],
+  ["review_failed", "Advisor: review failed"],
+]
+
+/** 单谓词（三消费点同源：design 结算 / code 完成守卫 / 报告提示）——**块首行扫描**（按空行
+ *  分块，逐块取首行 trim 后测前缀；时间线与尾以空行相接，六条尾均以块首行形态落地）。
+ *  负向精度（§14.3 修正轮）：引文中同串的**非块首形态**（围栏内行 / 表格行 / 引用行）不判
+ *  incomplete；块首裸行引用同串的残余误报方向安全（fail-closed——多付一轮重跑，如实登记）。
+ *  @returns {string|null} kind 或 null */
+export function advisorIncompleteMarker(text) {
+  for (const block of String(text ?? "").split(/\n\s*\n/)) {
+    const first = block.split("\n").find((l) => l.trim() !== "")
+    if (!first) continue
+    const line = first.trim()
+    for (const [kind, prefix] of ADVISOR_INCOMPLETE_PREFIXES) {
+      if (line.startsWith(prefix)) return kind
+    }
+  }
+  return null
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 预算提示 + 结构化超时尾（D / F15——§14.6 #2/#3）
+// ─────────────────────────────────────────────────────────────────────────────
+
+/** 0.75 一次性预算提示判定（纯函数——阈值两侧可机测；每场评审至多一次）。 */
+export function shouldBudgetNudge(elapsedMs, budgetMs, nudged) {
+  return !nudged && Number.isFinite(budgetMs) && budgetMs > 0 && elapsedMs >= budgetMs * 0.75
+}
+
+/** 预算提示文案（逐字——§14.6 #2；由循环注入一条 user 消息）。 */
+export function budgetNudgeText(elapsedMs, budgetMs) {
+  const secs = (ms) => Math.round(ms / 100) / 10
+  const pct = Math.round((elapsedMs / budgetMs) * 100)
+  return `⏳ review budget: ~${pct}% consumed (${secs(elapsedMs)}s of ${secs(budgetMs)}s). Converge now: emit your findings table for the evidence you have verified, mark anything you could not verify explicitly as \`unverified\` (unverified evidence must not support a pass), and emit your verdict line.`
+}
+
+/** 结构化超时尾（§14.6 #3）：族前缀 `Advisor: review timeout after {S}s.` 逐字保持
+ *  （判定族字面依赖）；其后 = 机读统计（rounds / tool calls / review text produced）
+ *  + 可执行恢复指引（narrower scope / 调预算）。 */
+export function timeoutTail(timeoutMs, rounds, toolCalls, producedText) {
+  const s = Math.round(timeoutMs / 1000)
+  return [
+    `Advisor: review timeout after ${s}s. Review incomplete — the wall-clock budget was exhausted; partial findings (if any) are above.`,
+    `- rounds: ${rounds} · tool calls: ${toolCalls} · review text produced: ${producedText ? "yes" : "no"}`,
+    `- budget: ${s}s (agent.advisor.timeoutMs) — re-run with a narrower scope (split the review across fewer documents) or raise the budget.`,
+  ].join("\n")
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 评审文本装配（loop 的尾经此与时间线合流——与尾族同文件：生成 / 判定 / 装配单源）
+// ─────────────────────────────────────────────────────────────────────────────
+
+// The live "[thinking…]" wait indicator shares its exact text with the TUI
+// cleanup regex (agent-turn.mjs strips it before flushing to history) — keep
+// them in lockstep.
+export const ADVISOR_THINKING_PLACEHOLDER = "\n[thinking…]\n"
+
+/**
+ * Tool-call progress line summary delegates to the single source describeToolArgs
+ * (../tui/tool-args.mjs) — the same function main-agent tool blocks and subagent
+ * blocks use. 2026-08-31: replaced the local picker (action/path/pattern/command-only)
+ * so advisor progress lines show the quoted-path forms everywhere else.
+ */
+/**
+ * Render the ordered review timeline — thinking / tool progress / final text
+ * interleaved EXACTLY as emitted, so the persisted record shows the review
+ * process at its real positions. A summary appended at the end would lose the
+ * order (the user-visible "no tool calls in the advisor record" gap). The
+ * live "[thinking…]" placeholder is stripped (wait indicator, not content).
+ */
+export function renderTimeline(timeline, tail = "") {
+  const body = timeline
+    .map((b) => b.text.replaceAll(ADVISOR_THINKING_PLACEHOLDER, "").trim())
+    .filter(Boolean)
+    .join("\n\n")
+  return [body, tail].filter(Boolean).join("\n\n")
+}