| name | autocoder |
| description | Use this skill for autonomous GitHub issue resolution, issue triage, regression-test failure handling, proposal review, blocked-issue review, worker monitoring, and continuous fix loops in repositories that use the autocoder workflow. |
Autocoder
Use this skill when the user wants Codex to operate the repo's autonomous issue-resolution or worker-coordination workflow.
Start Here
- Read
AGENTS.md first for repo rules.
- Use the workflow order in references/workflow-map.md.
- Use references/command-mapping.md when translating legacy Claude plugin commands.
- Prefer existing scripts over restating shell procedures.
Operating Rules
- CONTEXT COMPACTION IS MANDATORY (NON-NEGOTIABLE). Before starting work on ANY new issue in a fix loop, you MUST clear/compact the session context (the Codex equivalent of Claude's
/compact). Every issue, every time, no exceptions. Each issue starts from a fresh, compacted context; never carry over detailed investigation notes from a previous issue. Skipping this WILL exhaust the context window and crash the worker mid-issue. See "Context Management (MANDATORY)" in plugins/autocoder/commands/fix.md.
- Prioritize: triage -> bugs -> regression failures -> approved enhancements -> proposals.
- Respect blocking labels:
needs-approval, needs-design, needs-clarification, too-complex.
- Use
plugins/autocoder/scripts/regression-test.sh for full regression runs when appropriate.
- Use
scripts/append-to-history.sh to log significant work.
- Treat
AGENTS.md as the primary configuration source; use CLAUDE.md only as legacy fallback.
When To Read More
