thinkpool-pair 0.7.223 → 0.7.225

package/bridge.mjs

@@ -50,7 +50,7 @@ import { validateProviderSwitch, providerSwitchPlan, BUILTIN_PROVIDER } from './
 import { createSdkMcpServer, tool } from '@anthropic-ai/claude-agent-sdk'
 import { z } from 'zod'
 import { startClaudeSession } from './claude-session.mjs'
-import { startCodexSession, readCodexDefaultModel, readCodexModels, codexConfigForMode, defaultModeForRuntime } from './codex-session.mjs'
+import { startCodexSession, readCodexDefaultModel, readCodexModels, codexConfigForMode, codexThreadCanResume, defaultModeForRuntime } from './codex-session.mjs'
 import { canonicalRoomFilePath, waitForNativeImages } from './codex-images.mjs'
 import { createManagedLaneWorktree, removeManagedLaneWorktree } from './lane-worktree.mjs'
 import { commandOnPath } from './agent-detect.mjs'
@@ -103,7 +103,7 @@ const flowRedispatch = new Map()
 // broadcasts; without persistent state the cap can never bite.
 const flowBudgets = new Map()
 import { formatPeek, PEEK, siblingsOf, resolveSibling, crossPostDecision, CROSSPOST, spawnDecision, SPAWN, CROSSROOM, formatPairRoster, crossRoomPostDecision, formatRoomNow, formatClosableHint, laneStatusOf } from './cross-terminal.mjs'
-import { dispatchPendingRecap, recoverInterruptedTurn, recoverMissingResumeOnce, armInterruptedResume, sendInterruptedContinue, supersedeInterruptedResume } from './interrupted-resume.mjs'
+import { dispatchPendingRecap, recoverInterruptedTurn, recoverMissingResumeOnce, sendInterruptedContinue, supersedeInterruptedResume } from './interrupted-resume.mjs'
 import { turnInFlight } from './update-gate.mjs'
 import { saveSession, flushSession, deleteSession, loadAll, canResume, loadPtyId, savePtyId, loadNames, saveNames, appendDurableEvents, seedDurableEvents, readDurablePage, readDurableOldestSeq, hasDurableArchive, servesHistoryPage } from './session-store.mjs'
 import { stampEvent, makeSeqCounter, maxSeq, seqable, capReplayEvents, chunkReplayEvents, boundEventForBroadcast, inlineImageBlocks, usageReportLine, codexUsageReportLine, buildRecapFromLog, RECAP_CAP, trimmedBeforeSeq, firstSeq } from './event-id.mjs'
@@ -1671,8 +1671,11 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
   // `claude-fable-5` → 400 [1211][Unknown Model]. One value, one truth.
   // undefined → omit the SDK option entirely and let the provider env/endpoint default.
   const codexModels = runtime === 'codex' ? readCodexModels() : []
+  const codexRequestedModel = runtime === 'codex' ? (model || readCodexDefaultModel() || undefined) : undefined
   const laneModel = runtime === 'codex'
-    ? resolveCodexModel(model || readCodexDefaultModel() || undefined, codexModels)
+    ? (codexRequestedModel
+      ? resolveCodexModel(codexRequestedModel, codexModels)
+      : flowLaneModelFor({ sliceType: 'feature', runtime: 'codex', catalog: codexModels }))
     : effectiveLaneModel({ provider, model, configuredModel: resolveProviderEnv(provider)?.ANTHROPIC_MODEL })
   // spawnedBy: set when this lane was Dispatched (spawn_terminal). Restored from the
   // session store so the Ensemble flag survives a bridge restart (else a respin
@@ -1698,7 +1701,6 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
     flowReviewRound: Number.isInteger(flowReviewRound) && flowReviewRound >= 0 ? flowReviewRound : 0,
     dispatchBaseSha: dispatchBaseSha || null,
     revertTarget: revertTarget || null,
-    cwd: cwd || null, managedWorktree: managedWorktree || null,
     // Stable creation order — persisted so a bridge restart restores tabs in the SAME
     // order (not readdir/filesystem order). Legacy recs (no openedAt) derive it from the
     // first transcript event ts, so even the first post-fix restart is ordered right.
@@ -3627,7 +3629,13 @@ channel
           all.sort((a, b) => okey(a) - okey(b))
           if (all.length) for (const rec of all) {
             const wasInterrupted = restoredTurnOpen(rec.log || [])
+            // Codex can persist a thread id before its rollout receives the
+            // first metadata row. A stop followed immediately by a restart
+            // leaves that file at zero bytes; both App Server and exec reject
+            // it. Treat only that provably-empty case as non-resumable so the
+            // existing fresh-thread recap recovery handles it.
             const resumable = canResume(rec)
+              && (rec.runtime !== 'codex' || codexThreadCanResume(rec.sessionId))
             const recoveryRecap = wasInterrupted ? buildRecapFromLog(rec.log || [], RECAP_CAP) : ''
             // FL-M6 — restore the flow context (id/role/cwd) so an in-flight flow survives a
             // bridge restart: the conductor keeps its subagent-block + plan interception, and

package/codex-session.mjs

@@ -123,6 +123,22 @@ function codexRolloutPath(sessionId, { home, fsImpl }) {
   } catch { return null }
 }
 
+// A stopped App Server turn can publish its thread id before Codex has written
+// even the session_meta row. If the bridge restarts in that window, both
+// thread/resume and `codex exec resume` reject the zero-byte rollout. Missing
+// files remain "unknown" (Codex may resolve them through another store), but a
+// rollout we can prove is empty is never resumable.
+export function codexThreadCanResume(sessionId, {
+  home = process.env.CODEX_HOME || path.join(os.homedir(), '.codex'),
+  fsImpl = fs,
+} = {}) {
+  if (!sessionId) return false
+  const file = codexRolloutPath(sessionId, { home, fsImpl })
+  if (!file) return true
+  try { return Number(fsImpl.statSync(file).size) > 0 }
+  catch { return true }
+}
+
 // `codex exec --json` exposes lifetime thread totals on `turn.completed`. Those
 // totals are useful for billing, but they are NOT context occupancy: an agentic
 // turn can make hundreds of cached model calls and process many context windows.
@@ -365,8 +381,12 @@ export function startCodexSession({ cwd, model, effort: initialEffort = 'high',
   sandbox = normalizeCodexSandbox(sandbox || modeConfig.sandbox)
   const publishAccess = codexPublishAccess({ cwd, env: { ...process.env, ...(env || {}) } })
   const childEnv = publishAccess.env
-  let sessionId = resume || null
-  let turnNo = resume ? 1 : 0
+  const requestedResume = resume || null
+  const resumeUsable = !requestedResume || codexThreadCanResume(requestedResume, {
+    home: childEnv.CODEX_HOME || path.join(os.homedir(), '.codex'),
+  })
+  let sessionId = resumeUsable ? requestedResume : null
+  let turnNo = sessionId ? 1 : 0
   let child = null
   let aborted = false
   let ended = false
@@ -398,6 +418,9 @@ export function startCodexSession({ cwd, model, effort: initialEffort = 'high',
   })
 
   const note = (text) => { try { onEvent?.({ kind: 'note', text }) } catch { /* noop */ } }
+  if (requestedResume && !resumeUsable) {
+    queueMicrotask(() => note('The stopped Codex turn had no resumable context; continuing in a fresh thread.'))
+  }
   if (!['0', 'false', 'off'].includes(String(childEnv.TP_CODEX_APP_SERVER || '').trim().toLowerCase()) && !appServerGateResult.enabled) {
     queueMicrotask(() => note(`Codex App Server disabled: ${appServerGateResult.reason}`))
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thinkpool-pair",
-  "version": "0.7.223",
+  "version": "0.7.225",
   "description": "Share a local coding-agent CLI (Claude Code, Codex, Gemini, Aider, …) into a ThinkPool Code room, live.",
   "type": "module",
   "bin": {