agentleFS
rdmontgomeryrdmontgomery/eigenform/.claude/skills/codex-worker/SKILL.md

codex-worker

Skill1 starsChanged 43 days agoApache-2.0View on GitHub

Delegate a self-contained coding task to an OpenAI Codex worker (`codex exec`) running in its own git worktree, then verify and correct it — or have Codex review your own diff. Use when the user asks for Codex, a second model, a second opinion, a cross-model review, or parallel workers on independent tasks. Not for small edits you can make directly.

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 2

    1. zhengliangliahu-cloud/skills★ 1same name
    2. vasilievyakov/claude-codex-bridge★ 0same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repositoryOne other file it gives its agents.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 5 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every spawn/resume/review prints codex-thread: <id>.

    line 16
  2. Very specific:

    Default sandbox is workspace-write inside the worktree (.codex-worktrees/<name>, branch codex/<name>).

    line 39
  3. Distinctive:

    Never bypass the sandbox on your own initiative.

    line 42
  4. Very specific:

    Wait without polling: Run codex-worker wait <name> as a Bash call with run_in_background: true.

    line 44
  5. Very specific:

    Don't loop on status with sleep.

    line 46
  6. Very specific:

    Run the acceptance command yourself in the worktree (cd .codex-worktrees/<name>).

    line 52
  7. Distinctive:

    If the worker says "tests pass" and you didn't see them pass, they didn't.

    line 53
  8. Distinctive:

    Before you call your own non-trivial change done, have Codex review it:

    line 70
  9. Distinctive:

    Treat each finding as a bug report to verify, not an order.

    line 77
  10. Distinctive:

    When you and Codex disagree (an approach, a diagnosis, a finding you refuted), say so in your reply to the user.

    line 82

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

codex-worker

Codex is a different foundation model with different blind spots. Use it for two things: work done in parallel, and adversarial checking of your own work. Both only pay off if you verify. A worker's report is a claim, not evidence.

codex-worker is the executable next to this file. Call it by that path, or as codex-worker if it's on PATH (just install-codex-skill in the eigenform repo puts it there and links this skill into ~/.claude/skills). It needs codex on PATH, logged in.

Every spawn/resume/review prints codex-thread: <id>. Leave that line in the tool output: eigenform reads it and nests the worker's live transcript under your Bash call, so the user can watch it and open it with codex resume <id>.

The loop

  1. Write the assignment. Codex has none of your context. It sees the repo and your prompt, nothing else. Put the assignment in a file and pass @file. Include:
    • the goal in one sentence, and why it matters;
    • the files and directories in scope, and what must not be touched;
    • the acceptance command: the exact test/lint/build that must pass;
    • known pitfalls you've already found;
    • "Commit nothing. When done, end with: SUMMARY, FILES CHANGED, VERIFICATION (each command you ran and its result), OPEN QUESTIONS."
  2. Spawn. One worker per independent task, at most 3 at once unless the user says otherwise. Name it after the task.

    codex-worker spawn retry-fix -- @/tmp/retry-fix.md
    codex-worker spawn retry-fix --effort high --base main -- @/tmp/retry-fix.md

    Default sandbox is workspace-write inside the worktree (.codex-worktrees/<name>, branch codex/<name>). Codex's sandbox blocks network by default. If the acceptance command needs the network, ask the user before using --sandbox danger-full-access. Never bypass the sandbox on your own initiative.

  3. Wait without polling. Run codex-worker wait <name> as a Bash call with run_in_background: true. You'll be notified when it exits. Do other work meanwhile. Don't loop on status with sleep.
  4. Verify. This is the job.
    • codex-worker result <name>: read the report, then distrust it.
    • codex-worker diff <name>: read every hunk. Look for scope creep, deleted or weakened tests, swallowed errors, and hardcoded expected values.
    • Run the acceptance command yourself in the worktree (cd .codex-worktrees/<name>). If the worker says "tests pass" and you didn't see them pass, they didn't.
  5. Correct. Give the specific failure, not "try again":

    codex-worker resume retry-fix -- "cargo test -p net retry still fails: <paste>. The backoff must cap at 30s; don't change the test."

    Then go back to step 3. After three correction rounds without convergence, stop. Take the task back or ask the user. Don't spiral.

  6. Integrate or discard. Pull accepted changes into the working branch (copy the diff or git -C .codex-worktrees/<name> commit then cherry-pick), and re-run the checks there. Then codex-worker rm <name> [--force] [--delete-branch].

Cross-model review

Before you call your own non-trivial change done, have Codex review it:

codex-worker review --uncommitted
codex-worker review --base main -- "focus on concurrency and error paths"

Treat each finding as a bug report to verify, not an order. Reproduce it or refute it from the code.

Report the divergence

When you and Codex disagree (an approach, a diagnosis, a finding you refuted), say so in your reply to the user. Give what Codex claimed, what you claimed, and what you verified. That disagreement is the point of running a second model; don't smooth it over.

Reference

codex-worker spawn <name> [--base ref] [--model m] [--effort e] [--sandbox mode] [--schema file] [--here] -- <prompt|@file|->
codex-worker resume <name> -- <message|@file|->
codex-worker wait <name> [--timeout secs]
codex-worker status [<name>]
codex-worker result <name>
codex-worker diff <name>
codex-worker review [--base ref | --uncommitted | --commit sha] [-- instructions]
codex-worker rm <name> [--force] [--delete-branch]

--here runs in the current checkout with a read-only sandbox. Use it for investigation, not edits. State lives in .git/codex-workers/<name>/: the event stream, each prompt sent, the final message, and stderr. Read stderr.log there when a run fails to start.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.