Releases: linuxhq/cc-plugin-codex
Release list
v0.2.0
Changes
- Scan uppercase whole words
BLOCKandALLOWthroughout review output, including prose, code examples, and same-line verdicts. A colon is not required for detection. - Any
BLOCKprevents the turn from ending. Passing requires exactly oneALLOWin the supported verdict format. Lowercase words and identifiers such asBLOCK_SIZEdo not count. - Prefer a canonical first- or final-line blocking reason over an earlier quoted example, with the first detected block as fallback.
- Keep README, contributor policy, setup guidance, and gate prompts aligned. Quoted uppercase verdict words can deliberately cause false blocks.
Validation
npm run check: lint, formatting, plugin/skill validation, and all 242 tests passed.- GitHub CI: Ubuntu and macOS with Node 22 and 24, verified for both the release commit and tag before publication.
v0.1.9
Fixes
- Detect interrupted output streams and report supervisor failures while preserving force-kill escalation and cancellation grace.
- Save completed results before final logging. Tolerate a missing final log during concurrent cleanup; other logging errors still propagate.
- Standardize short skill names and description formatting in the skills menu.
- Reorganize README and contributor guidance into shorter bullets and topic sections, preserving documented behavior and adaptation constraints.
- Add regression coverage for completion ordering, failure reporting, helper cleanup, and bounded fake-helper readiness.
Validation
npm run check: lint, formatting, plugin/skill validation, and all 232 tests.- GitHub CI: Ubuntu and macOS, each with Node 22 and 24, verified for the release commit and tag before publication.
v0.1.8
This patch improves upstream parity for repository inspection, progress reporting, and process cleanup.
- Read committed files during branch reviews so dirty working-tree edits do not replace the reviewed code; allow history inspection beyond fifty commits.
- Stream foreground progress and retain activity, command outcomes, and final output in job logs.
- Preserve final results during session cleanup and stop inspection descendants during cancellation or worker shutdown.
- Complete jobs when Claude exits even if helpers retain stdout or stderr, while preserving helper ownership through history pruning and SessionEnd cleanup.
- Simplify README and consolidate contributor guidance in AGENTS.md.
Existing Claude transport and sandbox limitations remain documented in README. The upstream baseline is unchanged.
Validation: local lint, formatting, plugin/skill validation, and all 216 tests passed using the fake Claude CLI. All eight branch and tag CI jobs passed the Ubuntu/macOS matrix on Node 22 and 24.
Commit: 799994b2476b68440ab74d7fab4ff08cbbc9cfc9
v0.1.7
This patch aligns argument parsing, Git evidence collection, and rescue configuration with the pinned upstream baseline, and fixes session cleanup for active workers.
- Match upstream option spellings, unknown task/focus tokens, and inline boolean parsing.
- Restore Git text conversion, untracked symlink target reads, and binary sampling; remove the Git deadline and inspection history cap. Match auto scope for submodules containing only untracked files.
- Let write rescue use normal Claude tools, settings, hooks, skills, and MCP configuration while retaining sandboxed Bash.
- Preserve full provider diagnostics without the adapter's stderr cutoff, and require typed final result events.
- Keep live workers cancellable despite stale heartbeats, and cancel session-owned workers even after their history records are pruned.
The upstream baseline remains db52e28f4d9ded852ab3942cea316258ae4ef346. Documented read-only inspection and transfer limitations, the intentional final-line gate verdict adaptation, and origin-HEAD handling remain in place.
Validation: lint, formatting, plugin/skill validation, and all 200 tests passed locally with the fake Claude CLI. All eight branch and tag CI matrix jobs passed across Ubuntu/macOS and Node 22/24.
v0.1.6
Align job history with the pinned upstream policy and fix cleanup around history eviction.
- Keep the 50 most recently updated jobs across all states. Phase and Claude session changes refresh retention; heartbeat ticks alone do not.
- Separate history from execution controls so pruning does not cancel running Claude jobs. Jobs can return to history on progress changes or completion, and foreground output survives history eviction.
- Reclaim abandoned execution directories and temporary job files when their owning process is known to have exited. Preserve files with live or unverifiable owners.
- Use CODEX_THREAD_ID as the sole environment source for session identity, and clarify upstream-parity rules in AGENTS.md.
Limitations: job history now uses flat .json records. Older /job.json records are not read or migrated. History eviction can temporarily remove a running job from status/result/cancel lookup until its record is saved again. PID reuse or inaccessible processes can prevent abandoned-file cleanup.
Validation: npm run check (lint, formatting, plugin/skill validation, and 188 tests using the fake Claude CLI). Publication requires all four Ubuntu/macOS and Node 22/24 matrix jobs to pass for both the release commit and tag.
v0.1.5
Fix worker retention, session cleanup, and supervisor termination.
- Reclaim interrupted records when their recorded worker PID no longer exists, and prevent stale jobs from crowding out recent results.
- Continue session cleanup when another job disappears or encounters an error.
- Give the supervisor a two-second SIGTERM fallback while preserving the worker's one-second escalation.
- Add regression coverage and make fake Claude request captures atomic.
- Add a repository release skill that requires all branch and tag CI checks on the release commit to pass before publication.
Paused workers and legacy records without verifiable worker ownership remain conservatively retained, as documented in the README.
Validation: npm run check with 178 tests using the fake Claude CLI. GitHub CI passed on both main and the v0.1.5 tag across Ubuntu/macOS and Node 22/24 before publication.
v0.1.4
Fix cancellation and session cleanup for Claude background workers.
- Keep the process supervisor alive during cancellation so force-kill escalation reaches descendants even after Claude exits.
- Preserve unfinished worker records with stale heartbeats during pruning and session cleanup, allowing resumed workers to receive cancellation and finish cleanup.
- Document that workers that never recover can leave records behind, and add regression tests for paused workers and descendant cancellation.
Validation: npm run check passed, including all 172 tests using the fake Claude CLI.
Update the marketplace and reinstall the plugin:
codex plugin marketplace upgrade linuxhq
codex plugin add claude@linuxhqCheck hook trust in /hooks and start a new Codex session to load the updated plugin. Existing review gate settings are preserved.
v0.1.3
Fix automatic review failures when Claude puts explanatory prose before its verdict.
The gate still requests the upstream first-line verdict format. It now also accepts exactly one explicit ALLOW: <reason> or BLOCK: <reason> on the final line after prose. Multiple verdict lines, quoted or fenced fallback verdicts, and missing or incomplete fallback verdicts still block.
Added regression tests and documented the Claude output adaptation. Claude's automatic review gate approved the parser change.
Validation: npm run check passed, including all 170 tests using the fake Claude CLI.
Update the marketplace and reinstall the plugin:
codex plugin marketplace add linuxhq/cc-plugin-codex
codex plugin add claude@linuxhqStart a new Codex session to load the updated plugin.
v0.1.2
Fix three issues found during a full-codebase review:
- Use the remote default branch when origin HEAD is available, allowing reviews without a matching local branch.
- Stop Claude when its owning worker dies in every execution mode, including read-only reviews and rescue.
- Prune interrupted job records and retain the latest finished result even when active workers fill the history budget.
Added four regression tests and documented the behavior in README.
Validation: npm run check passed, including all 168 tests using the fake Claude CLI.
Update the marketplace and reinstall the plugin:
codex plugin marketplace add linuxhq/cc-plugin-codex
codex plugin add claude@linuxhqStart a new Codex session to load the updated plugin.
v0.1.1
Fix the startup warning caused by requesting a 5-second SessionEnd hook timeout. The plugin now requests 3 seconds, matching the limit Codex already enforced.
The README documents this Codex-specific limit, and plugin validation checks the corrected timeout.
Validation: npm run check passed, including all 164 tests using the fake Claude CLI.
Update the marketplace and reinstall the plugin:
codex plugin marketplace add linuxhq/cc-plugin-codex
codex plugin add claude@linuxhqStart a new Codex session to load the updated plugin.