Release 0.18.5: dead-session replay

CHANGELOG.md

@@ -3,6 +3,13 @@
 All notable changes to `codesema` (the npm package in `packages/cli`) are documented here.
 Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Versioning: [SemVer](https://semver.org).
 
+## [0.18.5] - 2026-08-28
+
+### Fixed
+
+- **A turn no longer dies when its stored agent session has vanished.** The cage home volume can be recycled between turns (it is released when the task first parks), and `claude --resume` then exits with "No conversation found". The runner now drops the dead session and replays the turn once with the full context rebuilt into the prompt.
+- **Agent stderr reaches the exit error.** Fatal reasons (a vanished session, an auth failure) go to stderr, not the JSONL stream; both the host and caged runs now tee stderr live AND carry its tail in the "exited with code N" message.
+
 ## [0.18.4] - 2026-08-28
 
 ### Added

package.json

@@ -1,6 +1,6 @@
 {
   "name": "codesema-tools",
-  "version": "0.18.4",
+  "version": "0.18.5",
   "private": true,
   "type": "module",
   "workspaces": [

packages/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codesema",
-  "version": "0.18.4",
+  "version": "0.18.5",
   "description": "Local merge request review, step by step. Your AI agent reviews, codesema displays.",
   "license": "MIT",
   "author": "Hasan TASKIN",

packages/cli/src/agent.test.ts

@@ -24,6 +24,7 @@
   emitsOpencodeJson,
   hardenedReviewCommand,
   hostPolicyUnsafe,
+  isDeadSessionError,
   knownAgent,
   MAX_TIMER_MS,
   OPENCODE_REVIEW_CONFIG,
@@ -461,7 +462,7 @@
 })
 
 describe('createOpencodeTaskParser', () => {
   const line = (event: unknown) => `${JSON.stringify(event)}\n`
 
   test('decoded JSON is activity, first sessionID fires onInit once', () => {
     let beats = 0
@@ -610,7 +611,7 @@
 })
 
 describe('createClaudeStreamParser', () => {
   const delta = (text: string) =>
     `${JSON.stringify({ type: 'stream_event', event: { type: 'content_block_delta', delta: { type: 'text_delta', text } } })}\n`
 
   test('text_delta accumulated and onText called', () => {
@@ -658,7 +659,7 @@
 })
 
 describe('createClaudeTaskParser', () => {
   const line = (event: unknown) => `${JSON.stringify(event)}\n`
 
   test('captures the session id from the system init event', () => {
     let sessionId = ''
@@ -2184,3 +2185,34 @@
     expect(err.message).toBe('agent command exited with code 127')
   })
 })
+
+describe('agentExitError stderr tail', () => {
+  test('the stderr tail joins the stream detail in the message', () => {
+    const err = agentExitError(
+      1,
+      '{"type":"result","subtype":"error_during_execution"}',
+      'warming up\nNo conversation found with session ID: 0123\n',
+    )
+    expect(err.message).toContain('error_during_execution')
+    expect(err.message).toContain('No conversation found with session ID: 0123')
+  })
+
+  test('stderr alone still carries the reason', () => {
+    const err = agentExitError(1, '', 'FATAL: something broke\n')
+    expect(err.message).toBe('agent command exited with code 1: FATAL: something broke')
+  })
+})
+
+describe('isDeadSessionError', () => {
+  test('matches claude wording for a vanished --resume target, and nothing else', () => {
+    expect(
+      isDeadSessionError(
+        new Error(
+          'agent command exited with code 1: error_during_execution; No conversation found with session ID: 0123',
+        ),
+      ),
+    ).toBe(true)
+    expect(isDeadSessionError(new Error('agent command exited with code 1'))).toBe(false)
+    expect(isDeadSessionError('No conversation found with session ID: 0123')).toBe(false)
+  })
+})

packages/cli/src/agent.ts

@@ -1230,11 +1230,22 @@ export function agentFailureDetail(out: string): string | null {
   return tail.length > 0 ? tail.slice(0, FAILURE_DETAIL_MAX) : null
 }
 
+/** Stderr kept per run for the exit error; the stream stays teed to the real stderr. */
+export const AGENT_STDERR_TAIL_MAX = 8192
+
+/** claude's own wording when a --resume target no longer exists (recycled cage home volume, cleared provider state). */
+export function isDeadSessionError(err: unknown): boolean {
+  return err instanceof Error && err.message.includes('No conversation found with session ID')
+}
+
 /** The exit-code message, carrying the agent's dying words when it left any. */
-export function agentExitError(code: number | null, out: string): Error {
+export function agentExitError(code: number | null, out: string, errTail = ''): Error {
   const base = t('agent.exitCode', { code })
-  const detail = agentFailureDetail(out)
-  return new Error(detail === null ? base : `${base}: ${detail}`)
+  const fromStream = agentFailureDetail(out)
+  const fromStderr =
+    errTail.trim().split('\n').slice(-3).join(' ').trim().slice(0, FAILURE_DETAIL_MAX) || null
+  const detail = [fromStream, fromStderr].filter((part) => part !== null).join('; ')
+  return new Error(detail.length === 0 ? base : `${base}: ${detail}`)
 }
 
 // --- clocks, spawning, and the shape of a kill ------------------------------
@@ -1481,7 +1492,7 @@ export function runAgent(opts: AgentRunOptions): Promise<string> {
     const child = spawnFn(command, {
       shell: true,
       cwd: opts.cwd,
-      stdio: ['pipe', 'pipe', 'inherit'],
+      stdio: ['pipe', 'pipe', 'pipe'],
       detached,
       ...(opts.env !== undefined ? { env: opts.env } : {}),
     })
@@ -1490,6 +1501,13 @@ export function runAgent(opts: AgentRunOptions): Promise<string> {
     // Registered BEFORE anything can close stdin: an agent that crashes closes
     // it early, and without this handler the EPIPE would kill the host process.
     stdin?.on('error', () => {})
+    // Teed, not swallowed: fatal reasons (session gone, auth) go to stderr,
+    // not the JSONL stream — keep them visible live AND in the exit error.
+    let errTail = ''
+    child.stderr?.on('data', (chunk: Buffer) => {
+      process.stderr.write(chunk)
+      errTail = (errTail + chunk.toString('utf8')).slice(-AGENT_STDERR_TAIL_MAX)
+    })
 
     let out = ''
     let capped = false
@@ -1549,7 +1567,7 @@ export function runAgent(opts: AgentRunOptions): Promise<string> {
       } else if (code === 0) {
         resolve(parser ? (parser.finalText() ?? out) : out)
       } else {
-        reject(agentExitError(code, out))
+        reject(agentExitError(code, out, errTail))
       }
     }
 

packages/cli/src/task-isolation.ts

@@ -33,6 +33,7 @@ import { join } from 'node:path'
 import {
   AGENT_KILL_GRACE_MS,
   AGENT_SETTLE_GRACE_MS,
+  AGENT_STDERR_TAIL_MAX,
   AGENT_WATCHDOG_DEFAULTS,
   agentExitError,
   AgentWatchdogError,
@@ -1794,10 +1795,19 @@ export const spawnContainer: ContainerSpawnFn = (opts) =>
   new Promise((resolve, reject) => {
     const clock = opts.clock ?? systemClock
     const spawnProcessFn = opts.spawnProcessFn ?? spawn
-    const child = spawnProcessFn(opts.file, opts.args, { stdio: ['pipe', 'pipe', 'inherit'] })
+    const child = spawnProcessFn(opts.file, opts.args, { stdio: ['pipe', 'pipe', 'pipe'] })
     const stdin = child.stdin
     const stdout = child.stdout
     stdin?.on('error', () => {})
+    // Teed, not swallowed: stderr still reaches the operator's journal live,
+    // AND its tail survives into the exit error — claude prints its fatal
+    // reason ("No conversation found with session ID …") there, not in the
+    // JSONL stream, and 'inherit' was throwing that reason away.
+    let errTail = ''
+    child.stderr?.on('data', (chunk: Buffer) => {
+      process.stderr.write(chunk)
+      errTail = (errTail + chunk.toString('utf8')).slice(-AGENT_STDERR_TAIL_MAX)
+    })
 
     let out = ''
     let capped = false
@@ -1837,7 +1847,7 @@ export const spawnContainer: ContainerSpawnFn = (opts) =>
       } else if (code === 0) {
         resolve(out)
       } else {
-        reject(agentExitError(code, out))
+        reject(agentExitError(code, out, errTail))
       }
     }
     const killClient = (signal: NodeJS.Signals): void => {

packages/cli/src/task-runner.ts

@@ -23,6 +23,7 @@ import {
   emitsClaudeStreamJson,
   emitsOpencodeJson,
   flagPresent,
+  isDeadSessionError,
   knownAgent,
   runAgent,
   type AgentHeartbeat,
@@ -2455,7 +2456,10 @@ export function createTaskRunner(opts: TaskRunnerOptions): TaskRunner {
     const attempt: TurnAttempt = { cost: null, folded: false }
     const checksConfig = opts.getChecksConfig ? opts.getChecksConfig() : opts.checksConfig
     const taskCommand = commandForTask(record, opts.command)
-    return (
+    // Re-read per attempt: the dead-session replay below clears
+    // agent_session_id, and both the prompt (full context vs message-only)
+    // and the session flag are derived from it inside runTaskTurn.
+    const execute = (): Promise<TaskTurnOutcome> =>
       runTaskTurn({
         cwd: record.worktree,
         task: record,
@@ -2477,6 +2481,26 @@ export function createTaskRunner(opts: TaskRunnerOptions): TaskRunner {
         ...(checksConfig !== undefined ? { checksConfig } : {}),
         ...(opts.runContainerTurnFn ? { runContainerTurnFn: opts.runContainerTurnFn } : {}),
       })
+    return (
+      execute()
+        // A stored session can stop existing under the task (the cage home
+        // volume was recycled, the provider state was cleared): claude then
+        // dies at birth. One replay with the session dropped rebuilds the
+        // whole context into the prompt instead of failing the turn.
+        .catch((err: unknown) => {
+          if (controller.signal.aborted || !record.agent_session_id || !isDeadSessionError(err)) {
+            throw err
+          }
+          record.agent_session_id = null
+          persist(record)
+          emit(record.id, {
+            type: 'message',
+            data: {
+              text: 'the stored agent session no longer exists: replaying the turn with rebuilt context',
+            },
+          })
+          return execute()
+        })
         .then((outcome) => {
           // The agent process is done: the task stops being interruptible as a
           // running turn here, even though the review that follows still holds