Skip to content
cccskills
Official · CodexFreePublished on GitHub

implementation-kickoff

Start or resume a requested Guardrails implementation or PR takeover in the selected linked worktree with bounded scope and verification.

The English summary is being prepared. Showing the original description.

See how to install

Files included (2)

  • SKILL.md4.0 KB
  • agents/openai.yaml125 B

SKILL.md (original)

Review the instructions the agent will receive before you install.

Implementation Kickoff

Start from the user's actual implementation request. An investigation or a plan alone is not permission to implement, push, or publish. Preserve authorization already given; this skill adds none. Apply AGENTS.md and implementation-strategy where applicable.

Establish the checkout and scope

Record the original outcome, acceptance criteria, affected paths, non-goals, and intended base. Reuse the current task's selected linked worktree. If running in a primary checkout, follow AGENTS.md and ask to start or hand off to a worktree. Do not automatically create another worktree or change unrelated checkouts.

Refresh remote references and record the full intended starting SHA. For a new independent task use the refreshed default branch unless the user chose another base. For an explicitly requested stack, use the exact previous slice's tip. Verify the checkout equals that SHA before editing a new task. For a resumed task, preserve its existing changes and record HEAD plus the original comparison base instead of resetting it to main. Stop on an unexplained base mismatch.

Inspect tracked and untracked state before edits. Keep task deliverables distinct from existing user work and operational plans/review notes. Use codex/ branch names by default, preserving an explicitly selected branch or user preference.

Take over an existing PR

Read current metadata, author commits, discussions, and the complete PR diff. Confirm the PR is still an appropriate source. Preserve its underlying outcome, but evaluate the proposal against the current architecture and toolchain. Import only the accepted scope; a takeover does not require copying the entire patch or preserving intermediate commits.

Record the source PR and verified author/coauthor identities. Credit adapted work with valid Co-authored-by trailers, without inventing identities. If the identity needed for attribution cannot be verified, resolve it before committing. For a split takeover, describe each slice as part of the replacement effort; do not claim the first slice completes the whole source PR.

Implement and verify

Keep changes within the agreed scope. Preserve user work and avoid automatic stashing, rebasing, or rewriting unrelated branches. Check integration with an advanced target before submission; if an update is needed, handle only this task's branch and rerun checks/review affected by any content change. Do not force a detached checkout, one-commit topology, or history rewrite to manufacture a clean handoff.

Use verification and final review before pushing or creating or updating a PR. Follow the release guide: ordinary user-visible SDK changes need a changeset; internal workflow changes need none. Leave version and changelog generation to the release PR.

Handoff

Use pr-draft-summary for the final description and authorized submission/follow-up.

Inspect status, full diff statistics, and every task commit before staging or reporting. Stage only task-owned deliverables; keep local planning and review artifacts out of the PR. Preserve hooks and inspect the committed result.

Report branch, base, commit, requested outcome, verification, review status, and remaining blockers. When authorized to submit, follow AGENTS.md for current-head CI monitoring, scoped fixes, root-channel review requests, and feedback replies. For a requested stack, create each PR against its predecessor and verify that relationship. Do not assume all workflows run on feature-branch PR targets: inspect triggers and use an authorized manual CI run on each head when needed. Never merge or close the source PR without authorization.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

code-change-verification

Official · CodexFree

Select and run Guardrails verification for code, dependency, test, tooling, documentation, or repository-workflow changes.

The English summary is being prepared. Showing the original description.

openai/openai-guardrails-js110Updated Oct 21, 2026

implementation-final-review

Official · CodexFree

Prepare independent review of a complete Guardrails change before pushing or updating a PR, using the repository adversarial-review procedure.

The English summary is being prepared. Showing the original description.

openai/openai-guardrails-js110Updated Oct 21, 2026

implementation-strategy

Official · CodexFree

Choose bounded implementation scope and existing owning modules before Guardrails runtime, configuration, client, resource, streaming, Agents, or SDK compatibility changes and feedback fixes.

The English summary is being prepared. Showing the original description.

openai/openai-guardrails-js110Updated Oct 21, 2026

pr-draft-summary

Official · CodexFree

Draft a Guardrails PR from its complete final diff and carry out authorized CI and review follow-up, including stacked PRs and takeovers.

The English summary is being prepared. Showing the original description.

openai/openai-guardrails-js110Updated Oct 21, 2026

See all skills by openai

Report a problem with this skill