Commits on Oct 10, 2026

1. fix: resume a session without messages after an app-server restart (#604 …

   )
   
   ## Problem
   
   A provider restart (`providers/set`, `providers/disable`) failed as a
   whole with "Internal error" when an open session had no messages yet.
   The provider change itself was applied, and every other session was
   resumed.
   
   Codex writes a thread to disk only with its first message. A session
   that never got one has no rollout, so `thread/resume` in the new
   app-server fails with "no rollout found for thread id …". The old
   app-server is gone, so `thread/read` fails too. The adapter collected
   that failure into an `AggregateError` and failed the provider request.
   
   This is the deterministic failure of the e2e test "does not run an
   untrusted hook, runs it after consent, and restores startup hooks after
   restart" (`e2e/hooks.test.ts`). The test is red on main's E2E workflow
   since #588. It failed the `verify` job of the 2.2.0 release, so
   **`v2.2.0` stays a tag and GitHub release that never reached npm**; this
   PR's release, 2.2.1, is the first published one after 2.1.1.
   
   ## Fix
   
   - The loss of such a session is now a typed
   `UnpersistedSessionLostError`, raised where the resume fails with a
   missing rollout (`AppServerRecovery.resumeError`).
   - The provider update completes, the same way it already does for a
   session that is refused or was replaced meanwhile.
   - The session reports the loss on its next use with the existing
   message: "had no messages yet and was lost when the Codex app-server
   restarted. Start a new session." Crash recovery already behaved this
   way.
   - Other resume failures still fail the provider update.
   
   ### Why the session is not recreated under the same id
   
   Codex can't bring back an unpersisted thread under its id:
   
   - `thread/start` takes no id.
   - `thread/resume` with `history` rejects an empty history, and a
   non-empty one would invent history.
   
   A new Codex thread behind the same ACP session id would only last as
   long as this agent process: ACP clients store the session id and
   `session/load` it later, and that id would then name no Codex thread.
   With `sessionIndex`, the conversation would also be listed under the new
   thread id. The session had no messages, so nothing is lost by starting a
   new one.
   
   ## The hooks e2e test
   
   With the restart fixed, the test still failed at its last step. It
   waited for the marker of a SessionStart hook after a `session/new`
   without a prompt. Codex 0.160.1 runs SessionStart hooks when the first
   turn of a thread starts, not on `thread/start`: checked against a real
   app-server, the marker appears only after `turn/start`. The test's check
   that an untrusted hook does not run also passed without running
   anything.
   
   The test now:
   
   1. prompts an untrusted session and checks that no marker appears;
   2. keeps a session without messages across the provider restart, and
   checks that its next prompt reports the loss;
   3. prompts a session started after the trust, and checks that the marker
   appears.
   
   ## Not changed
   
   The e2e "applies an approved workspace file edit" also failed in that
   `verify` run, and there is no adapter cause. In the failing attempts the
   model ran `apply_patch` through a shell command (`bash -c 'apply_patch
   …'`). Codex asks for that as a command approval, which the test's
   responder cancels by design: it allows only edits and asserts zero
   execute requests. When the model uses the `apply_patch` tool, it gets a
   file-change approval and the test passes. The same test failed on main
   before (`ca1d971`).
   
   ## Test plan
   
   - [x] New unit test: a provider update with one session without messages
   and one persisted session completes. The persisted session runs a turn,
   and the empty one reports the loss on its next prompt. The test fails
   without the fix.
   - [x] `npm run typecheck`, `npm test` (1484 passed), `npm run
   bundle:all`
   - [x] `RUN_E2E_TESTS=true` `e2e/hooks.test.ts` locally (the live-key
   test needs the CI secret)
   - [x] E2E workflow on this branch (run 38012958734): 33 passed,
   including the live-key hooks test and "applies an approved workspace
   file edit"
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)
   
   ---------
   
   Co-authored-by: Claude Opus 5.5 <[email protected]>

   

   nikita-ashihmin and claude authored Oct 10, 2026
   Configuration menu

   • View commit details
   • Copy full SHA for 9ac3254
   • Browse repository at this point

   Copy the full SHA

   9ac3254 View commit details

   Browse the repository at this point in the history

2. chore(main): release 2.2.1 (#605) …

   🤖 I have created a release *beep* *boop*
   ---
   
   
   ##
   [2.2.1](v2.2.0...v2.2.1)
   (2026-10-10)
   
   
   ### Bug Fixes
   
   * resume a session without messages after an app-server restart
   ([#604](#604))
   ([9ac3254](9ac3254))
   
   ---
   This PR was generated with [Release
   Please](https://github.com/googleapis/release-please). See
   [documentation](https://github.com/googleapis/release-please#release-please).
   
   Co-authored-by: acp-release-bot[bot] <246668977+acp-release-bot[bot]@users.noreply.github.com>

   

   acp-release-bot[bot] authored Oct 10, 2026
   Configuration menu

   • View commit details
   • Copy full SHA for a2e74a4
   • Browse repository at this point

   Copy the full SHA

   a2e74a4 View commit details

   Browse the repository at this point in the history