Release v1.3.0 — lifecycle correctness - #7
Merged
Merged
Conversation
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…rn failure, fileChange guard - An `error` notification without `willRetry: true` now ends the turn as failed instead of waiting for a turn/completed that never comes (openai#698). - A task whose turn fails without an error payload records an errorMessage and a summary naming the turn status, not the first line of rawOutput (openai#757). The job status report shows the error for failed jobs. - `fileChange` started items without `changes` no longer crash progress (openai#775). - Test helper `run()` now forwards `timeout` to spawnSync. Co-authored-by: ALV0612 <ALV0612@users.noreply.github.com> Co-authored-by: Soumya95 <Soumya95@users.noreply.github.com> Co-authored-by: kevin9327 <kevin9327@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…or only when it adds information Co-authored-by: ALV0612 <ALV0612@users.noreply.github.com> Co-authored-by: Soumya95 <Soumya95@users.noreply.github.com> Co-authored-by: kevin9327 <kevin9327@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: ALV0612 <ALV0612@users.noreply.github.com> Co-authored-by: Soumya95 <Soumya95@users.noreply.github.com> Co-authored-by: kevin9327 <kevin9327@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…turn id A turn/start response without turn.id left state.turnId null, so every notification stayed buffered and the capture never completed (openai#781). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A broker socket stuck in `connecting` fires neither connect nor error, so waitForBrokerEndpoint and the broker client's initialize could hang forever (openai#773). Each probe attempt is now bounded by min(500 ms, remaining budget) and destroys its socket on expiry; the broker client rejects with ETIMEDOUT after connectTimeoutMs (default 2000), and withAppServer falls back to a direct app-server on ETIMEDOUT when a broker was requested. `status <id> --wait` reported a timed-out wait with exit 0 and no hint (openai#774). It now prints "Timed out after <N>s while the job was still running." and exits 1; the --json snapshot is unchanged but exits 1 too. Co-authored-by: kevin9327 <kevin9327@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…s probe, never signal a stale pid ensureBrokerSession now defaults killProcess to terminateProcessTree, so a replaced or never-ready broker is no longer leaked (openai#753/openai#762). A live, owned broker that misses the 150 ms probe gets a full retry window before it is treated as wedged (openai#768). A dead or foreign pid from a stale record is never signalled; only its files are cleared (openai#749). A probe socket that connected but closes slowly now reads as ready. Co-authored-by: Soumya95 <Soumya95@users.noreply.github.com> Co-authored-by: mzl9039 <mzl9039@users.noreply.github.com> Co-authored-by: sylvesterkaczmarek <sylvesterkaczmarek@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…al and escape hatch; drop hooks.json description - setup --review-gate-model/--review-gate-effort (inherit clears); the stop gate forwards them to its review task (openai#769) - CODEX_REVIEW_GATE_MAX_ROUNDS defaults to 3; explicit 0 keeps unbounded (openai#548) - failure reasons name the kill signal and end with the disable hint (openai#589, openai#483) - hooks.json drops the top-level description key (openai#459) Co-authored-by: mittalpk <mittalpk@users.noreply.github.com> Co-authored-by: SomSamantray <SomSamantray@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
/codex:transfer hardcoded ~/.claude/projects as the only accepted transcript root, so a Claude Code install using CLAUDE_CONFIG_DIR to relocate its config directory could never pass the source-path check (openai#721). resolveClaudeProjectsDir(env) now derives the projects dir from CLAUDE_CONFIG_DIR (resolved via path.resolve when set) and falls back to ~/.claude/projects otherwise; resolveClaudeSessionPath threads options.env through to it and to the TRANSCRIPT_PATH_ENV lookup, defaulting to process.env. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…model catalogue Model aliases now resolve against the local Codex catalogue ($CODEX_COMPANION_MODEL_CATALOG -> $CODEX_HOME/models_cache.json -> `codex debug models --bundled` -> hardcoded fallback), picking the listed model whose slug ends in -<alias> (lowest priority, newest family on ties). Exact slugs pass through; hidden models never match an alias. --effort is rejected when the catalogued model does not list it. Adds the astra alias. Tests pin the catalogue to a fixture. (openai#468/openai#703/openai#485/openai#128) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…e broker.json before use Co-authored-by: weivwang <weivwang@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: weivwang <weivwang@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…n pid sidecar, job records and broker.json A recorded pid is signalled, reaped as dead, or evicted from the state lock only once its start identity proves it is still the recorded process (openai#743): linux /proc/<pid>/stat starttime, darwin `ps -o lstart=,comm=` pinned to LC_ALL=C/TZ=UTC. Unavailable identity never authorises a kill; win32 reports identity-unavailable (CIM identity is v1.4.0); posix records without an identity fall back to the command-line check. runCommand reports a timed-out command as status null instead of 0. Cancel says when it left a worker running, and the SessionEnd teardown line prints its reason. Co-authored-by: sylvesterkaczmarek <sylvesterkaczmarek@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…led regardless of identity The not-ready branch holds the child handle it just spawned, so its pid cannot have been recycled: kill it directly on every platform instead of routing it through the stored-record identity check (which cannot answer on win32). Co-authored-by: sylvesterkaczmarek <sylvesterkaczmarek@users.noreply.github.com> Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Bump package.json, package-lock.json, plugins/codex/.claude-plugin/plugin.json and .claude-plugin/marketplace.json to 1.3.0 via scripts/bump-version.mjs. Add the 1.3.0 CHANGELOG.md section (Fixed/Added/Changed/Known limitations, upstream #/PR citations) and sync plugins/codex/CHANGELOG.md to match. Add a README FAQ note on the Windows kill limitation, resolved in v1.4.0. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ies no id (openai#781) The timeout path could not send turn/interrupt without a turnId. Also correct the stale comments in failTurnOnTimeout and on the error notification. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…cy pid A fresh broker that fails readiness is killed through its child handle, and not at all once it has exited; the numeric process-group kill is only a fallback behind identity or command-line proof. A legacy broker.json pid is re-verified at kill time after the readiness retry, and the record is only cleared while it still names the endpoint we tore down. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A worker cancel may not signal but that is still alive would later overwrite the cancelled record. The job now stays running with its pid sidecar, the log and output say cancellation is not confirmed, and cancel exits 1. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The deadline now starts before the self-probe; every probe gets at most min(500 ms, time left) and is skipped (entry held) under 50 ms, and a failed self-probe is cached per process, so slow ps calls cannot push a lock wait past the SessionEnd budget. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…y write Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…n, triage) The fallback-root limitation now says state is orphaned on every plugin update when CLAUDE_PLUGIN_DATA is unset; the cancel entry describes the pending-cancellation outcome; rescue docs stop hardcoding the spark slug; triage marks the issues and PRs this release closes as fixed-in v1.3.0. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The test relied on cancel recording cancelled after a refused kill, which is the behaviour 41ca814 removed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… record A group kill raises ESRCH for a live pid that leads no process group; fall back to the pid itself before reporting not delivered. Cancel treats an undelivered kill of a live worker as pending, and a worker that outlives an acknowledged cancel no longer overwrites the cancelled record (checked under the state lock). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A broker stuck in connect has no cleanup handlers yet, so signalling only the broker left its app-server child behind. The live, unreaped child is a detached group leader, so its group is killed; the handle is the fallback. An exited child is still never signalled. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A running record without an identity whose pid was recycled by a long-lived unrelated process was never reaped, so a refused cancel stayed pending and the thread stayed in use. A readable command line without codex-companion.mjs now marks the job dead; nothing is signalled, and an unreadable one proves nothing. Test fixtures standing in for live workers now carry a companion command line. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A model-only --review-gate-model change skipped effort validation and could persist a model with an effort it cannot run. The stored effort is now checked against the new model before any write; on failure nothing is written. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…iation Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ion pid Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…no group (H1) terminateProcessTree no longer falls back to kill(pid) after the group kill raised ESRCH; it reports groupGone instead. terminateRecordedProcess re-proves identity (or the legacy command line) immediately before the bare-pid SIGTERM and refuses when the pid was recycled in between. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… (H2) processCommandLine reads /proc/<pid>/cmdline on linux and runs `ps -ww -o command=` with COLUMNS=10000 elsewhere; empty or unreadable output is null (unknown), which the reaper and broker ownership check already treat as no action. Legacy command-line fixtures that faked ps now say platform darwin. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…stop (H3) cleanupSessionJobs now reads each kill outcome and removes a running foreground job's record only when the kill was delivered, the pid is provably gone, or the job has no pid. A refused, undelivered or failed kill, or a job the budget never reached, keeps its record and files, is logged as "[codex] SessionEnd left <id> running: <reason>", and keeps the broker up through activeWorkspaceJobs. Each identity probe now gets half of what is left (worker cleanup) or of the step budget (broker teardown): the H1 re-verify can add a second probe per kill, and both must fit inside the 12 s SessionEnd budget. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Halved SessionEnd probe budgets could be fractional (e.g. 500.5); spawnSync threw ERR_OUT_OF_RANGE, the hook logged kill-failed and kept a live worker. Floor both halving sites, and clamp in runCommand so no caller can pass a fractional, zero (unbounded) or negative timeout. Same clamp on the stop review gate's env override. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A dead legacy worker left as a zombie has an empty /proc/<pid>/cmdline while kill 0 still sees it, so processCommandLine returned null (unknown) and the reaper kept the job running forever. On an empty cmdline read /proc/<pid>/stat: state Z or X yields "<defunct>" (as ps prints it, no companion marker), so the reaper's legacy rule reconciles the job without signalling. Any other state or an unreadable stat stays null. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ring teardown The SessionEnd hook's clearBrokerSession(cwd) guarded the unlink with an existsSync pre-check, but the broker's own SIGTERM handler (clearOwnSessionRecord) can delete the same broker.json concurrently: the check and the unlink are not atomic, so the hook could still hit ENOENT and exit 1. Replace the pre-check with try/unlink/catch-ENOENT in clearBrokerSession and in every unlinkSync site of teardownBrokerSession (pidFile, logFile, the unix socket path) so a concurrently self-cleaning broker can remove any of them without failing the hook. Other error codes (e.g. EPERM) still surface exactly as before. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
teardownBrokerSession's pidFile/logFile unlinks were only tolerant of ENOENT, so a locked or otherwise unremovable file (EPERM, notably reported on Windows upstream in openai#633/openai#626) still escaped and failed the SessionEnd hook. These two unlinks are best-effort cleanup, same as the socket-path unlink beside them (already a catch-all) and rmdirSync below them: swallow every error, not just ENOENT. clearBrokerSession is unchanged and stays ENOENT-only, since it is the one unlink whose target (broker.json) is the ownership record itself. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… subagent labels survive The replay of notifications buffered before the turn/start response ran every message through belongsToTurn, dropping a subagent's thread/started (its thread id is not yet known) and losing the label. Route live and replayed notifications through one function. Adds the FAKE_CODEX_SUBAGENT_EARLY_STARTED fixture knob and a deterministic test. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
CBEPX
added a commit
that referenced
this pull request
Sep 28, 2026
- broker-reuse E2E reads errorMessage from the stored job, not the foreground JSON - explicit fileURLToPath/pathToFileURL import in tests/process.test.mjs - export cleanupSessionJobs and guard main() behind direct execution in the hook - renderCancelPending wired into the cancel caller with diagnostic/json/log split - SURVIVOR-before-finally assertion checks both markers exist first - spec: post-snapshot descendants have no lifetime bound or later cleanup - inline cancel example reads broker.json only on win32 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Release v1.3.0 of the CBEPX fork: lifecycle correctness on posix, stop-gate controls, data-driven model aliases, and two security fixes. Full notes in
CHANGELOG.md(section 1.3.0).Highlights:
errornotifications end the turn instead of hanging (captureTurntreats theerrornotification as non-terminal, so a Codex-side failure hangs the turn forever and wedges the job atstatus: runningopenai/codex-plugin-cc#698, A server-side turn failure that terminates stores noerrorMessage, sostatusreports the reason asSummary: {openai/codex-plugin-cc#757, fix: do not crash when a fileChange start event omits changes openai/codex-plugin-cc#775, captureTurn drops every notification (including turn/completed) when the start response has no turn.id, hanging the job forever openai/codex-plugin-cc#781); subagent errors are scoped to their thread.ETIMEDOUTfallback (fix: bound hung broker connects instead of waiting forever openai/codex-plugin-cc#773);status --waitexits 1 on timeout (fix: surface status --wait timeouts instead of looking successful openai/codex-plugin-cc#774).broker.jsonnames, without checking it is still a broker (pid reuse → SIGTERM to an unrelated process group) openai/codex-plugin-cc#743) withpidIdentityin records, sidecars andbroker.json.--review-gate-model/--review-gate-effort, default 3 rounds, signal + escape hatch in reasons,hooks.jsonwithoutdescription(Stop review gate has no way to pin the model or reasoning effort it reviews with openai/codex-plugin-cc#769, Stop-review gate hook loops until CLAUDE_CODE_STOP_HOOK_BLOCK_CAP (missingstop_hook_activeguard) openai/codex-plugin-cc#548, stop-review-gate: signal-terminated review loses signal metadata in the fail-closed reason openai/codex-plugin-cc#589, stop-review-gate-hook.mjs: fail-closed reason strings do not mention the /codex:setup --disable-review-gate escape valve openai/codex-plugin-cc#483, Remove unsupported top-level description from hooks.json openai/codex-plugin-cc#459).CLAUDE_CONFIG_DIR(/codex:transfer is broken when CLAUDE_CONFIG_DIR is set — Claude transcript root hardcoded to ~/.claude/projects openai/codex-plugin-cc#721); model aliases resolve against the local Codex catalogue (Current Plugin does not support gpt-5.6 model family openai/codex-plugin-cc#468, Skillgpt-5-4-promptingstill targets GPT-5.4, retired from the rate card on 2026-08-31 openai/codex-plugin-cc#703, codex-rescue agent references stale gpt-5-4-prompting skill; effort hint omits max/ultra (default model is now gpt-5.6-sol) openai/codex-plugin-cc#485).broker.jsonvalidated (Predictable os.tmpdir() fallback state dir (0755) + unvalidated broker.json lets a co-located user MITM the Codex IPC and force arbitrary process-kill / file-delete openai/codex-plugin-cc#521, Plugin state dir has no plugin-identity segment: sibling plugins share one jobs array, and pruneJobs deletes the other plugin's records openai/codex-plugin-cc#609).Known limitations (documented in CHANGELOG/README): Windows kills from stored records are refused until process identity lands in v1.4.0.
Validation
npm test: 313 tests, 0 failures, 0 leaked processes;npm run build,check-version,claude plugin validate . --strict,npm audit --omit=dev,npm pack --dry-runall green.Ported with reference to upstream PRs by ALV0612, Soumya95, kevin9327, mzl9039, sylvesterkaczmarek, mittalpk, SomSamantray, weivwang.
🤖 Generated with Claude Code