| name | autodev |
| description | Use when the user asks to take a task end-to-end autonomously (e.g. "autodev", "/autodev", "build this on autopilot"). Runs the full chain — plan → implement → review → PR → converge — stopping only at real decisions. |
| metadata | {"short-description":"Autopilot — plan, build, review, PR, converge"} |
Autodev (Codex)
Full-pipeline autopilot, Codex-side. Same workflow as Claude's /autodev; this
skill adapts it for a Codex cockpit.
Canonical procedure: read ~/.codex/skills/skyhook-skills-commands/autodev.md and follow it
(chain, modes, plan gate, the --auto ceiling, hand-back). Apply these Codex
translations wherever it names a Claude command:
| Claude command | Codex equivalent |
|---|
/plan-loop | the plan-loop skill |
/review-loop | the review-loop skill |
/cross-review (cross-model) | the claude-review skill — from Codex the secondary reviewer stays Claude (it reviews your work) |
/qa | read the repo's .claude/commands/qa.md and follow it (Codex reads it as a file) |
/pr (verify + open/update PR) | inline via git + gh: run /qa, then create/update the PR (stage relevant files only, --force-with-lease). Write the body per ~/.codex/skills/skyhook-skills-commands/pr.md — depth proportional to the change, lead with motivation + design, no review-fix trivia, re-derive don't append. (Codex especially tends to dump 4 terse bullets — don't.) |
/fix-pr-loop (converge) | inline the reactive loop: wait for CI + bots (gh pr checks, gh pr view --json comments,reviews), triage each finding skeptically, fix the real ones, push, repeat until settled or capped. The AI reviewers + failing build/test CI are the primary signal — wait for those, triage them. Don't block on CodeQL >~2 min — it's a secondary scanner; if it's the only laggard, proceed and note CodeQL pending |
/triage-findings, /fix-findings, /simple | no such commands — perform the discipline inline (read the real code, never auto-accept, cite evidence on skips, don't ping-pong, fix confirmed issues) |
Non-negotiable invariants (don't let translation lose them):
- Default mode gates on the plan — stop for sign-off before code.
--auto
logs assumptions to NOTES.md and proceeds, but always stops for the
ceiling: prod/deploy/data-migrations · external-publish/money ·
security/auth/secrets · breaking-public-API/destructive-file-ops.
- Triage every reviewer skeptically; never auto-accept. Cross-review only
when nontrivial.
- Scenario-sensitive work needs a ledger before "done." For user-facing copy,
diagnostics/remediation, detector precision, error classification,
security/permissions, or UI states, list each scenario with expected final
behavior/copy, source-of-truth evidence, self-review verdict, Claude status,
tests/live proof, and open decision. Claude status is either the verdict or
skipped: <reason> when cross-review was intentionally skipped. If review
happened before later fixes, state whether the final head was re-reviewed.
- Risk / blast radius belongs in the hand-back. For nontrivial changes, state
affected surface, likely failure mode, mitigation/test proof, and residual risk.
Keep it proportional; low-risk copy/test-only work can be one sentence.
- Every loop caps and reports — never loop or proceed-past-a-blocker
silently.
- Never merge, deploy, push to
main, hard---force, or stage unrelated
files. PRs are fine; shipping is the user's call.
- Narrate the run with the glyph-tagged phase banners and close with the
end-of-run summary ledger defined in the canonical file (per-step counts, triage
breakdowns, result line,
--auto assumptions, open items).
- Phases are judgment calls, not a fixed sequence — decide which apply each
run, and honor free-text steering in the invocation (e.g.
consult claude,
no PR, plan only, quick, focus on <area>) as an override, per the
canonical file's directives section.
- UI / product-facing features: lean toward running visual-test +
product-review in the review phase, before opening the PR (recommended, not
forced — a skip is fine when small/non-visual, but must be stated). State
visual-test and product-review status in the hand-back (ran/skipped + why),
never silent. Don't open the PR blind to its own rendered result.
