本文へ移動
cccskills
無料GitHub で公開

coco-ship

Use when finished work needs to reach a coco-managed repository (coco-research/coco or similar) — verifies branch state, commits, pushes through the corporate DLP guard only when actually needed, opens a PR against main, waits for CI, and reports the result. Never merges, never pushes to main directly.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md7.1 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Shipping to a CoCo-managed Repo

Overview

Formalizes the branch → commit → guard-aware push → PR → CI-wait → report flow used for shipping work to coco-research/coco (or any repo carrying the same corporate git-guard pre-push hook), so it doesn't have to be re-derived by hand each time.

Core principle: Verify state → commit → detect whether the guard actually applies → confirm → push → open PR → wait for real CI → report. Never merge. Never push to main directly, regardless of how it's asked.

Announce at start: "I'm using the coco-ship skill to ship this branch."

This is the direct formalization of Option 2 ("Push and Create PR") in finishing-a-development-branch, specialized for repos that carry the corporate git-guard. It assumes implementation is already done and reviewed — it doesn't replace planning skills like brainstorming, writing-plans, or GSD's phase workflow, and it doesn't run tests itself (that's finishing-a-development-branch's Step 1, or the project's own CI).

The Process

Step 1: Preflight

git status -sb
git branch --show-current

If the current branch is main or master: stop. Do not commit or push here. Offer to create a feature branch first:

git checkout -b <feature-branch-name>

If there's nothing to commit and the branch isn't already ahead of its upstream: stop and say so — there's nothing to ship.

Step 2: Commit

Use the standard git-commit protocol already in effect for this harness — this isn't a new invention, just invoked from inside the skill:

  1. In parallel: git status, git diff (staged + unstaged), git log (recent messages, for style).
  2. Draft a commit message from the actual diff — why, not just what — matching the repo's existing style.
  3. Stage named files explicitly. Never git add -A or git add ..
  4. Commit with a Co-Authored-By trailer, via a heredoc so formatting survives.
  5. git status again to confirm a clean tree.

If a pre-commit hook fails, fix the issue, re-stage, and create a new commit — never --amend a commit that a failed hook means never actually landed.

Step 3: Detect whether the guard bypass is actually needed

Don't assume — check:

git remote -v
gh repo view <owner>/<repo> --json visibility,owner -q '"\(.visibility) \(.owner.login)"'
git config --show-origin --get core.hooksPath

A GIT_GUARD_REASON bypass is only needed when the repo is PUBLIC and the owner isn't on the guard's allowlist — check the actual guard config or ask the user which org(s) it allows, rather than assuming. If core.hooksPath doesn't point at the guard, don't assume a bypass is relevant on this machine at all — say so.

Step 4: Confirm — every time, no exceptions

This is an outward-facing action (push + open a PR), so it gets the same explicit go-ahead as any other push or PR in this session, regardless of how directly coco-ship was invoked. State plainly:

  • Target repo and branch
  • Whether a guard bypass will be used, and the exact reason string (not a placeholder)
  • That this opens a PR against main — it does not push to main directly

Wait for a clear yes before Step 5. Invoking this skill is not itself that approval.

Step 5: Push

git push -u origin <branch>

Only when Step 3 determined a bypass is needed:

GIT_GUARD_REASON="<specific, real reason>" git push -u origin <branch>

Never --no-verify. Never a plain --force. A --force-with-lease update to an already-open PR's branch is allowed, but still gets its own confirmation, separate from Step 4's.

Step 6: Open the PR

gh pr create --base main --head <branch> --title "<title>" --body "$(cat <<'EOF'
## Summary
<2-3 bullets of what changed and why>

## Test plan
<what was verified, or a checklist of what to verify>
EOF
)"

Keep the title under 70 characters. Put detail in the body, not the title.

Step 7: Wait for CI

gh pr checks <n> --watch

Don't report done while checks are still pending. If a check fails, say so plainly with the failure, don't retry blindly.

Step 8: Report

State the PR URL, the CI result, and the mergeable/review-required status. Say plainly that merging is the user's call — this skill never merges and never self-approves. If the PR author can't self-approve (most repos), say that explicitly rather than implying the PR is ready to merge on its own.

Quick Reference

StepActionConfirmation needed?
1Preflight (refuse on main)—
2Commit—
3Detect guard need—
4State plan, askYes — always
5Push(covered by Step 4)
6Open PR(covered by Step 4)
7Wait for CI—
8Report—

Common Mistakes

Assuming the guard bypass is needed without checking

  • Problem: Sets GIT_GUARD_REASON on every push, even to private or allowlisted repos, or misses that the guard isn't even installed on this machine.
  • Fix: Always run Step 3's checks; only set the env var when the repo is genuinely public and non-allowlisted.

Treating skill invocation as approval

  • Problem: Runs Steps 5–6 immediately because the user said "ship this," without a distinct confirmation of the actual push/PR plan.
  • Fix: Step 4 is a separate, explicit ask every time — state the repo, branch, guard reason (if any), and that a PR (not a direct push to main) will follow.

Reporting success before CI resolves

  • Problem: Says "done" right after gh pr create, while checks are still pending.
  • Fix: Step 7 watches until every check resolves, then Step 8 reports the real result.

Red Flags

Never:

  • Push to main or master directly, no matter how it's requested
  • Use --no-verify or a plain --force
  • Auto-merge or self-approve a PR
  • Fabricate a GIT_GUARD_REASON instead of a specific, true one
  • Skip Step 4's confirmation because the work "seems done"

Always:

  • Refuse and redirect if invoked while on main/master
  • Check whether the guard bypass is actually needed rather than assuming
  • Get explicit confirmation before the push/PR step
  • Wait for and report real CI status, not an assumed one

Integration

Formalizes: Option 2 of finishing-a-development-branch ("Push and Create PR"), specialized for repos carrying the corporate git-guard.

Distinct from: commands/team/ship.md (/team:ship, a full idea-to-product pipeline) and systems/gsd/skills/gsd-ship (GSD milestone-scoped). Neither overlaps with this skill's narrower job — shipping already-finished work through the guard and into a PR.

Pairs with: using-git-worktrees for cleanup after the PR merges (not handled by this skill).

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Train and optimize AI agents using Microsoft's Agent Lightning framework with reinforcement learning. Use when setting up agent training, instrumenting agents with tracing, configuring LightningStore, implementing reward functions, or optimizing prompts with RL/APO algorithms.

日本語の概要は準備中です。原文の説明を表示しています。

coco-research/coco5462026年10月11日 更新

Post-run self-evaluation system that scores agent output on correctness, clarity, actionability, and conciseness. Use after /team runs, skill executions, or when explicitly asked to evaluate output quality.

日本語の概要は準備中です。原文の説明を表示しています。

coco-research/coco5462026年10月11日 更新

Create AI marketing videos for ads, promos, product launches, and brand content. Models: Veo, Seedance, Wan, FLUX for visuals, Kokoro for voiceover. Types: product demos, testimonials, explainers, social ads, brand videos. Use for: Facebook ads, YouTube ads, product launches, brand awareness. Triggers: marketing video, ad video, promo video, commercial, brand video, product video, explainer video, ad creative, video ad, facebook ad video, youtube ad, instagram ad, tiktok ad, promotional video, launch video

日本語の概要は準備中です。原文の説明を表示しています。

coco-research/coco5462026年10月11日 更新

Use when building AI features into a product: LLM integration, RAG pipelines, guardrails, streaming, AI UX, prompt engineering, or AI cost control. Treats prompts as code and validates every model output.

日本語の概要は準備中です。原文の説明を表示しています。

coco-research/coco5462026年10月11日 更新

Your AI research and engineering brain trust. 59 named personas across 8 cells covering frontier labs, applied product, model architecture, reasoning/RL/agents, alignment and interpretability, theory and science of DL, multimodal and…

日本語の概要は準備中です。原文の説明を表示しています。

coco-research/coco5462026年10月11日 更新

Use when designing a new REST or GraphQL API, reviewing an API spec before implementation, setting team API standards, or migrating REST to GraphQL. Covers resources, HTTP semantics, pagination, error handling, and pitfalls.

日本語の概要は準備中です。原文の説明を表示しています。

coco-research/coco5462026年10月11日 更新

coco-research のスキルをすべて見る

このスキルの問題を報告する