thincoder 0.12.52 → 0.12.54

package/src/session.mjs

@@ -6,283 +6,24 @@
  *
  * File layout: {hash}.json.N (slots), {hash}.json.manifest (slot metadata + active pointer).
  * Legacy {hash}.json is migrated to a slot on first access.
+ *
+ * 2026-08-31 advisor round1 🔴：slot/清单管理拆至 session-slots.mjs（本文件曾超 500 行
+ * 硬限）；本文件只保留核心读写 + re-export 全部 slot 导出（既有 import 路径不变）。
  */
 
-import { createHash } from "node:crypto"
-import { mkdirSync, readFileSync, writeFileSync, renameSync, unlinkSync, existsSync } from "node:fs"
-import { join, dirname } from "node:path"
-import { execSync } from "node:child_process"
-import { configDir } from "./config.mjs"
-import { migrateHashLength } from "./session-migrate.mjs"
-
-let currentSessionId = null
-
-/** Generate unique session ID for this process */
-export function getSessionId() {
-  if (!currentSessionId) {
-    currentSessionId = `${process.pid}-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`
-  }
-  return currentSessionId
-}
-
-/** Normalize cwd for hashing: uppercase Windows drive letter so both ends
- *  (CLI's process.cwd() vs VS Code's uri.fsPath, which lowercases it) agree. */
-export function normalizeCwd(cwd) {
-  return cwd.replace(/^([a-z]):/, (_, d) => d.toUpperCase() + ":")
-}
-
-/** Full sha1 hex (40 chars), not truncated. Shared contract with the VS Code extension. */
-function cwdHash(cwd) {
-  return createHash("sha1").update(normalizeCwd(cwd)).digest("hex")
-}
-
-/** Derive base session path from cwd hash. Migrates legacy short-hash files on first access. */
-export function sessionPath(cwd) {
-  const hash = cwdHash(cwd)
-  migrateHashLength(cwd, hash)
-  return join(configDir, "sessions", `${hash}.json`)
-}
-
-function slotPath(cwd, n) { return sessionPath(cwd) + "." + n }
-export { slotPath }
-function manifestPath(cwd) { return sessionPath(cwd) + ".manifest" }
-
-/** Path to the active slot's file */
-export function activePath(cwd) {
-  return slotPath(cwd, activeSlot(cwd))
-}
-
-/** Atomic write: write to temp file then rename to replace, preventing truncated JSON from mid-write crash. */
-function writeSessionFile(p, data) {
-  mkdirSync(dirname(p), { recursive: true })
-  const tmp = `${p}.tmp`
-  writeFileSync(tmp, JSON.stringify(data), "utf8")
-  try {
-    renameSync(tmp, p)
-  } catch {
-    try { unlinkSync(p) } catch {}
-    try {
-      renameSync(tmp, p)
-      try { unlinkSync(tmp) } catch {}
-    } catch {
-      writeFileSync(p, readFileSync(tmp, "utf8"), "utf8")
-    }
-  }
-}
-
-// ========== slot management ==========
-
-/** Normalize slot metadata: old format is a raw timestamp number; new format is { ts, ... } object */
-function slotMetaTs(v) { return typeof v === "number" ? v : (v?.ts ?? 0) }
-
-function slotCmp(a, b) { return slotMetaTs(a[1]) - slotMetaTs(b[1]) }
-
-/** Detect a genuine user message (excludes system-reminder injected messages) */
-function isRealUserMsg(m) {
-  return m.role === "user" && typeof m.content === "string" && !m.content.startsWith("[System reminder:")
-}
-
-/** Extract slot metadata from history (shared by slotDigest and loadSlotMeta) */
-function extractSlotMeta(history, activeProvider, updatedAt, title = "") {
-  const userMsgs = history.filter(isRealUserMsg)
-  const first = userMsgs[0]?.content ?? ""
-  return {
-    messageCount: history.length,
-    turnCount: userMsgs.length,
-    firstMessage: first.slice(0, 80),
-    activeProvider: activeProvider ?? "",
-    updatedAt: updatedAt ?? Date.now(),
-    title,
-  }
-}
-
-/** Extract preview summary from session data for manifest storage (with current timestamp) */
-function slotDigest(data) {
-  const meta = extractSlotMeta(data.history ?? [], data.activeProvider, data.updatedAt, data.title ?? "")
-  return { ts: Date.now(), ...meta }
-}
-
-function loadManifest(cwd) {
-  try {
-    const p = manifestPath(cwd)
-    if (!existsSync(p)) return { slots: {}, sessionId: null }
-    const m = JSON.parse(readFileSync(p, "utf8"))
-    if (!m.sessionId) m.sessionId = null
-    return m
-  } catch { return { slots: {}, sessionId: null } }
-}
-
-function saveManifest(cwd, m) {
-  m.sessionId = getSessionId()
-  writeSessionFile(manifestPath(cwd), m)
-}
-
-/**
- * Claim a slot for this process and set it as active. Idempotent.
- * Preference order:
- *  1. The current active slot, if it is unowned / ours / its owner is dead — reuse it.
- *  2. Any slot that is unowned or owned by a dead process (reclaim).
- *  3. A brand-new slot when all are owned by live processes.
- * The owner is recorded in m.slotSessions so other processes (CLI ↔ VS Code) can
- * see which slots are taken and avoid them.
- */
-function ensureActive(cwd, m) {
-  const mySessionId = getSessionId()
-  if (!m.slotSessions) m.slotSessions = {}
-
-  // Already own the active slot — nothing to do.
-  if (m.active && m.slotSessions[m.active] === mySessionId) return
-
-  const isFree = (slot) => {
-    const owner = m.slotSessions[slot]
-    if (!owner || owner === mySessionId) return true
-    return !isProcessAlive(parseInt(owner.split('-')[0]))
-  }
-
-  // 1. Prefer the current active slot if we can take it (preserves "resume where you left off").
-  if (m.active && m.slots[m.active] && isFree(m.active)) {
-    m.slotSessions[m.active] = mySessionId
-    saveManifest(cwd, m)
-    return
-  }
-
-  // 2. Otherwise claim the first slot that is free.
-  const allSlots = Object.keys(m.slots).filter(n => /^\d+$/.test(n)).map(Number).sort((a, b) => a - b)
-  for (const slot of allSlots) {
-    if (isFree(slot)) {
-      m.active = slot
-      m.slotSessions[slot] = mySessionId
-      saveManifest(cwd, m)
-      return
-    }
-  }
-
-  // 3. All slots owned by live processes — allocate a new one (no limit).
-  const newSlot = allSlots.length > 0 ? Math.max(...allSlots) + 1 : 1
-  m.active = newSlot
-  m.slotSessions[newSlot] = mySessionId
-  saveManifest(cwd, m)
-}
-
-/**
- * Check if a process with given PID is still alive.
- * Returns false if process doesn't exist or we can't determine.
- */
-function isProcessAlive(pid) {
-  if (!pid || isNaN(pid)) return false
-  try {
-    // On Windows: tasklist /FI "PID eq <pid>" /NH
-    // On Unix: kill(pid, 0) or check /proc/<pid>
-    if (process.platform === 'win32') {
-      const output = execSync(`tasklist /FI "PID eq ${pid}" /NH`, { encoding: 'utf8', stdio: 'pipe' })
-      return output.includes(String(pid))
-    } else {
-      // Unix: try to send signal 0 (doesn't kill, just checks)
-      process.kill(pid, 0)
-      return true
-    }
-  } catch {
-    return false
-  }
-}
-
-/** Return the active slot number for this process, claiming one if necessary */
-export function activeSlot(cwd) {
-  const m = loadManifest(cwd)
-  ensureActive(cwd, m)
-  return m.active
-}
-
-/** Lazy-load slot metadata from slot file (for old-format manifest entries that lack metadata) */
-function loadSlotMeta(cwd, slot, v) {
-  if (typeof v === "object" && v !== null && "ts" in v) return v
-  const ts = typeof v === "number" ? v : 0
-  try {
-    const p = slotPath(cwd, slot)
-    if (!existsSync(p)) return { ts }
-    const data = JSON.parse(readFileSync(p, "utf8"))
-    const history = data.history ?? []
-    const meta = extractSlotMeta(history, data.activeProvider, data.updatedAt ?? ts, data.title ?? "")
-    return { ts, ...meta }
-  } catch {
-    return { ts }
-  }
-}
-
-/** List all slots, newest first. Includes isActive flag. */
-export function listSlots(cwd) {
-  const m = loadManifest(cwd)
-  const active = m.active ?? activeSlot(cwd)
-  return Object.entries(m.slots)
-    .filter(([n]) => /^\d+$/.test(n))
-    .map(([n, v]) => {
-      const meta = loadSlotMeta(cwd, Number(n), v)
-      return {
-        slot: Number(n),
-        isActive: Number(n) === active,
-        timestamp: meta.ts,
-        date: new Date(meta.ts).toLocaleString(),
-        messageCount: meta.messageCount ?? 0,
-        turnCount: meta.turnCount ?? 0,
-        firstMessage: meta.firstMessage ?? "",
-        activeProvider: meta.activeProvider ?? "",
-        updatedAt: meta.updatedAt ?? meta.ts,
-        updatedDate: new Date(meta.updatedAt ?? meta.ts).toLocaleString(),
-        title: meta.title ?? "",
-      }
-    })
-    .sort((a, b) => b.updatedAt - a.updatedAt)
-}
-
-/**
- * Switch active slot. No file copying — just change the pointer in the manifest.
- * Returns the loaded session data (null if slot doesn't exist).
- */
-export function switchToSlot(cwd, slot) {
-  const m = loadManifest(cwd)
-  if (!m.slots[slot]) return null
-  m.active = slot
-  saveManifest(cwd, m)
-  return loadSession(cwd)
-}
-
-/** Delete a slot: remove its file and manifest entry. Deleting the active slot
- *  resets the manifest active pointer (the next claim re-creates one). */
-export function deleteSlot(cwd, slot) {
-  const n = Number(slot)
-  if (!Number.isInteger(n) || n < 1) return false
-  const m = loadManifest(cwd)
-  if (!m.slots[n]) return false
-  delete m.slots[n]
-  delete m.slotSessions?.[n] // orphan session-id entries bloat the manifest forever
-  try { unlinkSync(slotPath(cwd, n)) } catch { /* missing file is fine */ }
-  if (m.active === n) delete m.active
-  saveManifest(cwd, m)
-  return true
-}
-
-/** Rename a slot: update the slot file's title + the manifest metadata (shared with VS Code). */
-export function renameSlot(cwd, slot, title) {
-  const n = Number(slot)
-  if (!Number.isInteger(n) || n < 1) return false
-  const p = slotPath(cwd, n)
-  if (!existsSync(p)) return false
-  let data
-  try {
-    data = JSON.parse(readFileSync(p, "utf8"))
-  } catch {
-    return false
-  }
-  data.title = title
-  writeSessionFile(p, data)
-  const m = loadManifest(cwd)
-  if (m.slots[n]) {
-    m.slots[n] = slotDigest(data)
-    saveManifest(cwd, m)
-  }
-  return true
-}
+import { readFileSync, renameSync, existsSync, statSync } from "node:fs"
+import { basename } from "node:path"
+import {
+  slotPath, writeSessionFile, loadManifest, saveManifest, slotDigest,
+  activeSlot, sessionPath, getSessionId, isProcessAlive,
+} from "./session-slots.mjs"
 
+// re-export slot 管理（保持既有 import session.mjs 的调用点不变）
+export {
+  getSessionId, normalizeCwd, sessionPath, slotPath, manifestPath, activePath,
+  writeSessionFile, slotDigest, loadManifest, saveManifest, activeSlot, listSlots,
+  deleteSlot, renameSlot, isProcessAlive,
+} from "./session-slots.mjs"
 
 // ========== legacy transient prefix cleanup ==========
 
@@ -343,7 +84,10 @@ function slimForDisplay(m) {
 /** Save agent state to the active slot file (atomic write). `display` (the old
  *  WYSIWYG render snapshot) is DEPRECATED — it drifted out of sync with history
  *  whenever VS Code wrote the slot, and the TUI resumed from a stale snapshot.
- *  Restore now always rebuilds from history (lazy, see startup.mjs). */
+ *  Restore now always rebuilds from history (lazy, see startup.mjs).
+ *  2026-08-31 会诊 F1：slot 粘性——首次认领后缓存 agent._slot，永不重跑 ensureActive
+ *  （原实现每次保存重推，manifest active 被并发方翻动时会话静默迁移）。
+ *  返回轮转的 .bak 路径或 null。 */
 export function saveSession(agent) {
   // _fullHistory is written at the source via pushReal — no flush needed here.
   // history        = FULL, never-compacted (human-readable; VS Code panel & CLI resume read this)
@@ -377,8 +121,51 @@ export function saveSession(agent) {
     pendingReminders: agent._pendingReminders ?? [],
     sessionStart: agent._sessionStart ?? null,
   }
-  const slot = activeSlot(agent.cwd)
-  writeSessionFile(slotPath(agent.cwd, slot), data)
+  const slot = agent._slot ??= activeSlot(agent.cwd)
+  const p = slotPath(agent.cwd, slot)
+  // 2026-08-31 会诊 F2 🔴：写前校验磁盘文件的 sessionStart——与本进程会话不符（另一
+  // 进程/会话的现场）→ 先轮转 .bak 保留再写（11311 条历史被新进程覆盖的实锤场景）。
+  // 检查按 mtime 缓存（_slotMtime）：文件自上次检查/自写未变就跳过全量解析——
+  // 每次保存都 readFileSync+JSON.parse 多 MB 会话 → O(n²) 退化。
+  let rotated = null
+  try {
+    if (existsSync(p)) {
+      const st = statSync(p)
+      let disk = null
+      if (st.mtimeMs !== (agent._slotMtime ?? -1)) {
+        disk = JSON.parse(readFileSync(p, "utf8"))
+        // 2026-08-31 会诊 deepseek 🟡：version>2 的新版文件无论 sessionStart 一律轮转
+        // （loadSlotFile 对 v3 返回 null 不动文件——若其 sessionStart 为 null，旧版首次
+        // 保存会静默覆盖；轮转 .bak 保证新版文件保留）。
+        // 2026-09-01 advisor 🔵：磁盘文件 cwd 不匹配（异项目文件误落本路径）同样轮转——
+        // 与 loadSlotFile/legacy 读的"别人的文件不动"原则对齐（否则 sessionStart null +
+        // version≤2 的异项目文件被静默覆盖且无 .bak）。
+        const diskIsNewer = typeof disk?.version === "number" && disk.version > 2
+        const diskStart = disk?.sessionStart ?? null
+        const myStart = agent._sessionStart ?? null
+        const diskForeign = typeof disk?.cwd === "string" && disk.cwd.toLowerCase() !== agent.cwd.toLowerCase()
+        if (diskIsNewer || diskForeign || (diskStart && diskStart !== myStart)) {
+          const bak = `${p}.bak-${Date.now()}`
+          renameSync(p, bak)
+          rotated = bak
+          console.error(`[session] slot ${slot} holds ${diskIsNewer ? `a newer-version file (v${disk.version})` : diskForeign ? `a foreign-cwd file (${disk.cwd})` : `another session (start ${diskStart}, ours ${myStart})`} — preserved as ${basename(bak)}`)
+        }
+        agent._slotMtime = st.mtimeMs
+      }
+    }
+  } catch {
+    // 文件存在但不可读（损坏/半写）：改名 .corrupted 保留现场（2026-09-01 advisor 🔵——
+    // 与自身 loadSlotFile 的 .corrupted 约定 + VS Code saveSessionToSlot 对齐；.bak 保留
+    // 给 F2 轮转路径，损坏现场不再混入轮转后缀，恢复/清理工具按后缀分类不误判）
+    if (existsSync(p)) {
+      try {
+        renameSync(p, `${p}.corrupted`)
+      } catch {}
+    }
+  }
+  writeSessionFile(p, data)
+  // 记录我们刚写的 mtime——下次保存跳过重复解析
+  try { agent._slotMtime = statSync(p).mtimeMs } catch {}
   // Update slot metadata in manifest
   try {
     const m = loadManifest(agent.cwd)
@@ -388,28 +175,40 @@ export function saveSession(agent) {
     // Manifest update failure is non-fatal — data is safe, metadata will lazy-recover on next listSlots
     console.error(`[session] manifest metadata update failed for slot ${slot}: ${e.message}`)
   }
+  return rotated
 }
 
-/** Load session data from the active slot; returns null if missing, corrupted, or version mismatch */
-export function loadSession(cwd) {
-  const tryLoad = (p) => {
+/** Load slot file with shared validation — 2026-08-31 会诊 deepseek 🟡 抽取：
+ *  loadSession（active 槽）/switchToSlot/ACP session/load 共享。无认领副作用。
+ *  version 1/2 + history 数组 + cwd 匹配（2026-09-01 会诊 kimi 🔵：cwd 先行——"别人的
+ *  文件不动"优先于结构校验，异 cwd + 坏 version 不得改名）；结构不符改名 .unreadable、
+ *  解析失败 .tmp 回退成功后提升为正主（损坏主文件改名 .corrupted 保留）、主文件缺失
+ *  时恢复孤儿 .tmp（rename 前崩溃现场）。 */
+export function loadSlotFile(cwd, slot) {
+  const p = slotPath(cwd, slot)
+  const tryLoad = (path) => {
     try {
-      if (!existsSync(p)) return null
-      const data = JSON.parse(readFileSync(p, "utf8"))
-      if (data?.version !== 1 && data?.version !== 2) return null
-      if (!Array.isArray(data.history)) return null
-      if (data.cwd && data.cwd.toLowerCase() !== cwd.toLowerCase()) return null
+      if (!existsSync(path)) return null
+      const data = JSON.parse(readFileSync(path, "utf8"))
+      if (data.cwd && data.cwd.toLowerCase() !== cwd.toLowerCase()) return null // 别人的文件，不动
+      if (data?.version !== 1 && data?.version !== 2) {
+        if (typeof data?.version === "number" && data.version > 2) return null // 新版 CLI 的文件，不动（F2 轮转兜底）
+        try { renameSync(path, `${path}.unreadable`) } catch {}
+        console.error(`[session] unsupported version ${data?.version} at ${path} — preserved as ${basename(path)}.unreadable`)
+        return null
+      }
+      if (!Array.isArray(data.history)) {
+        try { renameSync(path, `${path}.unreadable`) } catch {}
+        console.error(`[session] slot file ${path}: history is not an array — preserved as ${basename(path)}.unreadable`)
+        return null
+      }
       data.history = data.history.filter((m) => !isLegacyTransient(m))
-      data._recovered = false
       return data
     } catch (e) {
       return { _error: e }
     }
   }
 
-  // Try the active slot first (post-migration, legacy file may be stale)
-  const slot = activeSlot(cwd)
-  const p = slotPath(cwd, slot)
   let result = tryLoad(p)
   if (result && !result._error) return result
   if (result?._error) {
@@ -417,36 +216,95 @@ export function loadSession(cwd) {
     const tmpResult = tryLoad(`${p}.tmp`)
     if (tmpResult && !tmpResult._error) {
       console.error(`[session] recovered from .tmp fallback`)
-      tmpResult._recovered = true
+      // 2026-09-01 会诊 🟢：.tmp 提升为正主（损坏主文件改名 .corrupted 保留现场）——
+      // 否则主文件留在原地，每次加载都重复"恢复"，且保存侧 F2 会持续轮转它。
+      try {
+        renameSync(p, `${p}.corrupted`)
+        renameSync(`${p}.tmp`, p)
+      } catch {}
       return tmpResult
     }
     console.error(`[session] .tmp fallback also failed — session lost.`)
     try { renameSync(p, `${p}.corrupted`) } catch {}
+  } else if (!result) {
+    // 2026-09-01 advisor 🔵：主文件缺失但孤儿 .tmp 存在（原子写 rename 前崩溃）——
+    // 恢复并提升为正主（SESSION.md §4 的 .tmp 回退语义应覆盖此场景）。
+    const orphan = tryLoad(`${p}.tmp`)
+    if (orphan && !orphan._error) {
+      console.error(`[session] slot ${slot} main file missing — recovered orphan .tmp`)
+      try { renameSync(`${p}.tmp`, p) } catch {}
+      return orphan
+    }
   }
+  return null
+}
+
+/** Load session data from the active slot; returns null if missing, corrupted, or version mismatch */
+export function loadSession(cwd) {
+  // Try the active slot first (post-migration, legacy file may be stale)
+  const slot = activeSlot(cwd)
+  let result = loadSlotFile(cwd, slot)
+  if (result) return result
 
   // Fallback: try the legacy current file (pre-migration, or migration failed to clean up)
   const legacy = sessionPath(cwd)
-  result = tryLoad(legacy)
-  if (result && !result._error) {
-    return result
-  }
-  if (result?._error) {
-    console.error(`[session] failed to load legacy ${legacy}: ${result._error.message}`)
+  try {
+    if (existsSync(legacy)) {
+      const data = JSON.parse(readFileSync(legacy, "utf8"))
+      // cwd 不匹配是别人的文件——与 loadSlotFile 一致直接 return null 不改名
+      // （"别人的文件不动"原则，2026-08-31 advisor round2 🟡）
+      if (data.cwd && data.cwd.toLowerCase() !== cwd.toLowerCase()) return null
+      if ((data?.version === 1 || data?.version === 2) && Array.isArray(data.history)) {
+        data.history = data.history.filter((m) => !isLegacyTransient(m))
+        return data
+      }
+      // 结构不匹配：保留现场（version>2 的新版文件不动）
+      if (!(typeof data?.version === "number" && data.version > 2)) {
+        try { renameSync(legacy, `${legacy}.unreadable`) } catch {}
+        console.error(`[session] legacy file ${legacy}: invalid structure — preserved as ${basename(legacy)}.unreadable`)
+      }
+    }
+  } catch (e) {
+    console.error(`[session] failed to load legacy ${legacy}: ${e.message}`)
     try { renameSync(legacy, `${legacy}.corrupted`) } catch {}
   }
-
   return null
 }
 
-/** Apply loaded session data onto an agent object; returns true if provider was switched */
+/** slimForDisplay 截断的 arguments 以 U+2026（…）结尾——不是合法 JSON 的完整值。
+ *  v1 老文件回退播种机器线时置为 {}（合法空参数），防止半截 \\uXXXX 毒化发送载荷（会诊 F6）。 */
+function stripTruncatedToolArgs(m) {
+  if (m?.role !== "assistant" || !Array.isArray(m.tool_calls)) return m
+  let changed = false
+  const tool_calls = m.tool_calls.map((tc) => {
+    const args = tc?.function?.arguments
+    if (typeof args === "string" && args.endsWith("…")) {
+      changed = true
+      return { ...tc, function: { ...tc.function, arguments: "{}" } }
+    }
+    return tc
+  })
+  return changed ? { ...m, tool_calls } : m
+}
+
+/** Apply loaded session data onto an agent object; returns true if provider was switched.
+ *  2026-08-31 会诊 F1：清空 _slot/_slotMtime 缓存——切换后下次保存重新认领（F1b 回归：
+ *  switchToSlot 后保存必须落新槽而非旧槽）。 */
 export function applySession(agent, data) {
   // data.history is the FULL never-compacted record (human line); data.contextHistory is the
   // (possibly compacted) machine line. Restore each line from its own source — the machine
   // context keeps its compaction savings across resume. Legacy files without contextHistory
   // fall back to seeding the machine line from the full history (it re-compacts when needed).
+  // 2026-08-31 会诊 deepseek 🟡：机读线必须从 contextHistory 恢复而非从完整 history 重建——
+  // 后者会把已压缩的中间过程塞回上下文（实测 prompt 膨胀到 283%）。compactThresholdAuto 时
+  // 按恢复后的模型重新推导阈值（bin/thincoder.mjs）。v1 老文件（无 contextHistory）回退
+  // 播种时剥离被 slimForDisplay 截断的 tool_calls.arguments（以 … 结尾 → 置 {}；会诊 F6——
+  // 截断可劈断 \\uXXXX 产生 400 毒载荷）。2026-09-01 会诊三家：length > 0 才当机读线
+  // （contextHistory: [] 是"无机读线"而非空机器线——空机器线会静默丢全部上下文）。
   agent.config ??= {} // ACP test mocks may omit config; be defensive like the ??= below
   const full = Array.isArray(data.history) ? data.history : []
-  const machine = Array.isArray(data.contextHistory) ? data.contextHistory : full
+  const ch = data.contextHistory
+  const machine = (Array.isArray(ch) && ch.length > 0) ? ch : full.map(stripTruncatedToolArgs)
   agent._fullHistory = [...full]
   agent.history = [...machine]
   agent.title = data.title ?? ""
@@ -471,6 +329,8 @@ export function applySession(agent, data) {
   agent._compressFailures = 0
   agent._verifyRetries = 0
   agent._verifyPassed = false
+  agent._slot = null // 粘性缓存清空——切换后重新认领（F1b）
+  agent._slotMtime = null
   if (data.activeProvider && data.activeProvider !== agent.activeProvider) {
     const p = agent.providers?.find((pr) => pr.name === data.activeProvider)
     if (p) {
@@ -502,18 +362,119 @@ export function applySession(agent, data) {
  */
 export function newSession(cwd) {
   const m = loadManifest(cwd)
-  // Ensure active is set (migrate if needed)
-  if (!m.active) ensureActive(cwd, m)
 
-  // Find next available slot number
+  // 2026-09-01 advisor 🔵：先清理死主条目（与 ensureActive 分支2 语义一致）——否则
+  // "死主 + 文件缺失"的空槽（m.slots 有条目、无文件、属主已死）永不复用，槽号持续
+  // 增长。死主且无文件 = 该会话从未落盘（进程死了没保存），条目可安全删除回收。
+  // 2026-09-01 会诊 deepseek/glm 🟡：清理必须经 deletions 显式落盘——saveManifest 条目级
+  // 合并会把磁盘死条目从 fresh 复活回写，仅传 m 等于没删（VS Code newSlot 已修，CLI 对称）。
+  const mySessionId = getSessionId()
+  const deadSlots = []
+  for (const [s, owner] of Object.entries(m.slotSessions ?? {})) {
+    if (owner && owner !== mySessionId) {
+      const pid = parseInt(owner.split("-")[0])
+      if (!pid || !isProcessAlive(pid)) {
+        delete m.slotSessions[s]
+        if (!existsSync(slotPath(cwd, Number(s)))) delete m.slots[s]
+        deadSlots.push(s)
+      }
+    }
+  }
+
+  // Find next available slot number — 2026-08-31 会诊 deepseek 🟡：不能只看 manifest
+  // 条目（丢失更新可能让条目消失而文件仍在）——复用该号会直接覆写真实会话
+  // （F2 防护不覆盖 newSession 的空数据直写）。
+  // 2026-08-31 advisor round2 🟡：同时跳过"已被另一活进程认领但尚未落盘"的号
+  // （slotSessions 有条目、slots 无条目、文件不存在）——否则双进程会认领同一号。
+  const liveClaimed = (n) => {
+    const owner = m.slotSessions?.[n]
+    return !!(owner && owner !== mySessionId && isProcessAlive(parseInt(owner.split("-")[0])))
+  }
   let slot = 1
-  while (m.slots[slot]) slot++
+  while (m.slots[slot] || existsSync(slotPath(cwd, slot)) || liveClaimed(slot)) slot++
 
-  // Write empty session
-  const data = { version: 2, cwd, title: "", updatedAt: Date.now(), history: [], tasks: [], goal: null, autoApprove: false, advisor: null, pendingReminders: [], sessionStart: null }
+  // Write empty session — 2026-09-01 advisor 🔵：补 contextHistory/planMode 字段与
+  // VS Code newSlot 对齐（SESSION.md §3 v2 格式双端一致；两端读侧均有兜底，功能等价）
+  const data = { version: 2, cwd, title: "", updatedAt: Date.now(), history: [], contextHistory: [], tasks: [], planMode: false, goal: null, autoApprove: false, advisor: null, pendingReminders: [], sessionStart: null }
   writeSessionFile(slotPath(cwd, slot), data)
   m.slots[slot] = slotDigest(data)
   m.active = slot
-  saveManifest(cwd, m)
+  // 2026-08-31 advisor round1 🟡：与 ensureActive 认领模型一致——立即记录新 slot 所有权，
+  // 否则 /new 后到首次保存之间并发方（VS Code/另一 CLI）会把新 active 槽当"空闲可恢复"
+  // 认领 → 双进程写同一槽（F2 轮转互旋）。
+  m.slotSessions ??= {}
+  m.slotSessions[slot] = mySessionId
+  // 2026-09-01 会诊三家 🟡：显式翻 active 的调用点传 setActive（saveManifest 默认保留
+  // 磁盘 fresh.active，避免把并发方刚翻的指针回滚）
+  // 2026-09-01 会诊 deepseek/glm 🟡：deletions 过滤掉本调用刚重新认领的槽（防删掉自己的
+  // 新属主）——与 ensureActive deadParam / VS Code newSlot 同型。
+  const deletions = deadSlots.length
+    ? {
+        slotSessions: deadSlots.filter((s) => m.slotSessions[s] !== mySessionId),
+        slots: deadSlots.filter((s) => !m.slots[s]),
+      }
+    : null
+  saveManifest(cwd, m, deletions, { setActive: true })
   return slot
 }
+
+/** 清空会话运行态（/new 用，2026-08-31 会诊 F3）：_fullHistory/title/_sessionStart 等
+ *  全部会话级状态 + 一次性注入标志必须全清——否则新会话首次落盘把旧会话完整人类线 +
+ *  旧标题写进新槽（实锤 .19/.3 双副本）；注入标志不清则 /new 后新会话永不注入
+ *  OS/cwd reminder（2026-09-01 会诊 glm 🟡）。autoApprove 是用户偏好，跨会话保留（有意）。 */
+export function resetSessionState(agent) {
+  agent._fullHistory = []
+  agent.history = []
+  agent.title = ""
+  agent.tasks = []
+  agent._sessionStart = null
+  agent._engDesignToken = null
+  agent._compressFailures = 0
+  agent._verifyRetries = 0
+  agent._verifyPassed = undefined
+  agent._runStartHistoryLen = 0
+  agent._lastPromptTokens = null
+  agent._usageAtLen = null
+  agent.planMode = false
+  agent.goal = null
+  agent._pendingReminders = []
+  agent._slot = null
+  agent._slotMtime = null
+  agent._osReminderInjected = false
+  agent._restartReminderInjected = false
+  agent._lastEngState = false
+}
+
+/** Switch the manifest active pointer to a slot. Returns the slot's session data
+ *  (null if the slot doesn't exist / can't be read). 2026-08-31 会诊 deepseek 🔴：
+ *  原实现经 loadSession → activeSlot 有认领副作用——目标槽被活进程占用时 ensureActive
+ *  分支 3 会把 active 拨到新空槽并读回 null + 劫持对方指针。现改为 loadSlotFile 直接读
+ *  （无认领副作用）；目标槽空闲则一并认领、被另一活进程占用则不认领（slotOccupancy——
+ *  下次保存经 activeSlot 自然 fork 到新槽）；只改 manifest 指针（setActive 意图）。
+ *  2026-09-01 会诊三家 🟡：saveManifest 条目级合并 + setActive（不把并发方刚翻的指针回滚）。 */
+export function switchToSlot(cwd, slot) {
+  const m = loadManifest(cwd)
+  if (!m.slots[slot]) return null
+  const data = loadSlotFile(cwd, slot)
+  if (!data) return null
+  m.active = slot
+  const occ = slotOccupancy(cwd, slot)
+  if (!occ.occupied) {
+    m.slotSessions ??= {}
+    m.slotSessions[slot] = getSessionId()
+  }
+  saveManifest(cwd, m, null, { setActive: true })
+  return data
+}
+
+/** 目标槽是否被另一活进程占用（2026-09-01 会诊/advisor 🟡）：排除本进程属主——
+ *  /session 重选当前槽不误报；同进程双会话防护由 ACP sameProcessPinned 承担。 */
+export function slotOccupancy(cwd, slot) {
+  const m = loadManifest(cwd)
+  const owner = m.slotSessions?.[slot]
+  if (!owner) return { occupied: false }
+  if (owner === getSessionId()) return { occupied: false }
+  const pid = parseInt(owner.split("-")[0])
+  if (!pid || !isProcessAlive(pid)) return { occupied: false }
+  return { occupied: true, owner }
+}

package/src/tools/bash.md

@@ -27,9 +27,9 @@ Output format:
 Notes:
 - There is NO TTY — editors, pagers (vim, less), and interactive prompts WILL hang. Always pass non-interactive flags: `git commit -m`, `git --no-pager`, `-y`/`--yes` where applicable
 - The environment sets GIT_PAGER=cat, PAGER=cat, EDITOR=true, TERM=dumb — but still always use non-interactive flags
-- Output is capped at ~200K chars; if you need more, redirect to a file and read it. Truncated output ends with a `[... truncated: N chars omitted]` marker — the missing tail may contain errors.
+- Output is capped at ~200K chars; if you need more, redirect to a file and read it. Truncated output ends with a `[... truncated: N chars omitted]` marker — the missing tail may contain errors, so read the saved log file's tail before trusting success. Use `filter` to narrow instead of hand-piping.
 - Check `[stderr]` for error messages, warnings, and diagnostic output — it is separated from `[stdout]` so you can quickly identify problems.
-- On Windows the shell is **cmd.exe** (NOT Git Bash, NOT PowerShell): `&&`/`||` chaining works, use cmd built-ins (`del`, `dir`, `type`, `findstr`, `tasklist`) and `/dev/null`→`NUL`, `2>/dev/null`→`>nul 2>&1`. Bash-isms (`rm -rf`, `cp -r`, `head`, `$(...)`) FAIL. For complex logic prefer `node -e` over shell gymnastics.
+- On Windows the shell is **cmd.exe** (NOT Git Bash, NOT PowerShell): `&&`/`||` chaining works, use cmd built-ins (`del`, `dir`, `type`, `findstr`, `tasklist`) and `/dev/null`→`NUL`, `2>/dev/null`→`>nul 2>&1`. Bash-isms (`rm -rf`, `cp -r`, `head`, `$(...)`, `${VAR}`, single quotes, `;` separators) FAIL — the tool prepends a hint when it detects POSIX-only syntax, and still executes. For complex logic prefer the execute tool (node) over shell gymnastics.
 - Never use bash to read, copy, or transmit secret files (.env, keys, tokens)
 - Do NOT run destructive commands (rm -rf, force-push, drop table) without explicit user confirmation
 - After commands that change files (git checkout, npm install, etc.), repo_outline and code_search may be stale — re-run them to get current results.