메뉴
Walk through staged PR reviews from the task store conversationally — APPROVEs first, then COMMENTs, smallest diff to largest within each. For each PR, present a concise summary and accept owner direction (post, skip, draft, discuss). Use when asked to "drain the staged reviews", "walk through staged reviews", or "what's staged".
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | review-staged |
| description | Walk through staged PR reviews from the task store conversationally — APPROVEs first, then COMMENTs, smallest diff to largest within each. For each PR, present a concise summary and accept owner direction (post, skip, draft, discuss). Use when asked to "drain the staged reviews", "walk through staged reviews", or "what's staged". |
Conversational walkthrough of staged reviews. The owner sees one PR at a time and steers each with a verb. No reviews are posted without an explicit "post it".
This skill assumes /shipwright:review has already staged the reviews into the
task store (staged: true) and written review files to state/reviews/PR_REVIEW_{pr}.md
and state/reviews/pr_review_{pr}.json. It does not review new PRs and does
not merge anything — merging is handled exclusively by /shipwright:deploy.
Fetch staged PR records from the task store for each configured repo:
curl -sf -H "Authorization: Bearer $SHIPWRIGHT_TASK_STORE_TOKEN" \
"$SHIPWRIGHT_TASK_STORE_URL/prs?repo={org}/{repo}&staged=true" | jq -c '.prs[]'
For each record, fetch current GitHub metadata (title, additions, deletions, headRefOid, url):
gh pr view {prNumber} --repo {org}/{repo} \
--json number,title,author,headRefName,baseRefName,headRefOid,isDraft,reviews,commits,additions,deletions,url
diffSize = additions + deletions. Treat a record with reviewState: "approved" as an
APPROVE verdict; reviewState: "posted" as a COMMENT verdict. reviewState: "in_progress"
also maps to COMMENT — backwards compatibility for records staged before the verdict-persist
fix (those records had their verdict omitted from the staging PATCH).
If no staged records exist across all repos: print No staged reviews. Run /shipwright:review to stage some. and stop.
Resolve the current GitHub user once and remember it for the rest of the run:
gh api /user -q '.login'
Sort the staged entries:
reviewState: "approved"), then COMMENTs (reviewState: "posted"
or reviewState: "in_progress" — the latter for backwards-compat with records staged
before the verdict-persist fix)diffSize to largestPrint the queue header so the owner sees what's coming:
Staged reviews ({N}) — APPROVEs first, smallest diff first within each group.
| # | PR | Repo | Title | Verdict | Diff |
|---|----|------|-------|---------|------|
| 1 | #123 | example-repo | Add feature X | APPROVE | +45/-12 (57) |
| 2 | #456 | example-repo | Fix bug Y | COMMENT | +120/-30 (150) |
Walk the queue in order. For each entry:
gh pr view {prNumber} --repo {org}/{repo} \
--json number,title,author,headRefName,baseRefName,headRefOid,isDraft,reviews,commits,additions,deletions,url
isDraft == true) — skip with: Skipping #{pr} — PR is now a draft.savepoint, wip, or WIP (case-insensitive), skip with:
Skipping #{pr} — savepoint/WIP commit on branch ({sha}: "{message}").Skipping #{pr} — @{login} requested changes on {date} and there are no commits since.
(Use the same teammate-comment logic from /shipwright:review Step 3: not a bot,
not CURRENT_USER, no commits pushed after the review.)record.commitSha differs from the current headRefOid,
skip with: Skipping #{pr} — new commits since review was staged. Re-run /shipwright:review {org}/{repo}#{pr}.When a PR is skipped for any reason, leave its task store record untouched
(staged: true stays set). CHANGES_REQUESTED state is re-derived from GitHub on every run.
If author.login == "dependabot[bot]" (or the entry has author: "app/dependabot"):
Read state/dependabot-reviews.json. If an entry exists for this PR, capture
{status, recommendation} and surface it in the presentation block (Step 3d) as
a *Dependabot:* line.
Read state/reviews/PR_REVIEW_{pr}.md and extract:
Fetch CI status:
gh api "repos/{org}/{repo}/actions/runs?branch={branch}&per_page=5" \
-q '.workflow_runs[] | "\(.name): \(.status) \(.conclusion)"'
Print this block (and only this block — no extra commentary):
*PR #{pr}: {title}* — {url}
@{author} · {branch} → {base} · +{additions}/-{deletions} ({diffSize})
*Why:* {motivation}
*Changes*
- Web view: {value or "none"}
- API: {value or "none"}
- Database: {value or "none"}
- Breaking: {value or "none"}
- Arch: {value or "none"}
*CI:* {one-line summary — e.g. "all green" / "lint failing" / "1 failure: build"}
{if dependabot: "*Dependabot:* tracked in dependabot-review ({status}, recommendation: {recommendation})"}
*Concerns ({count})*
- {Critical/Important findings, each on one line with file:line}
- ({N} suggestions/nits — show on request)
*Verdict:* {APPROVE | COMMENT} — {one-line reasoning from review file}
Then prompt:
What do you want to do? (post it / skip / next / draft it / discuss)
Stop and wait for the owner's response.
Match the owner's response:
post itSubmit the review:
gh api -X POST /repos/{org}/{repo}/pulls/{pr}/reviews \
--input state/reviews/pr_review_{pr}.json
Capture html_url from the response. The teammate-feedback and staleness gates in
step 3b already ran immediately before this PR was presented, so there's no separate
re-fetch here — this skill owns staged-review posting end to end.
Clear the staged flag:
curl -sf -X PATCH \
-H "Authorization: Bearer $SHIPWRIGHT_TASK_STORE_TOKEN" \
-H "Content-Type: application/json" \
"$SHIPWRIGHT_TASK_STORE_URL/prs/{record.id}" \
-d '{"staged": false}' >/dev/null
Run /shipwright:review Step 11b with PR_RECORD_ID = {record.id} and
{verdict} = APPROVE if record.reviewState == "approved" else COMMENT — it marks
the record complete and reasserts agentId/reviewState. Posting-then-completing a
review is identical whether it happened immediately or was staged first; no need for
a second copy of that logic here.
Print the posted review URL.
Move to the next PR.
skipLeave the task store record as staged: true. Do nothing. Move to the next PR.
nextSame as skip — move on without action.
draft itPrint the full review body and inline comments from state/reviews/pr_review_{pr}.json
(the JSON destined for the GitHub API). Format the inline comments as a list:
{path}:{line} — {body}. Do not post. Re-prompt for an action verb.
Treat the response as a question or pushback. Engage conversationally — explain a
finding, look up code, run git show or gh pr diff as needed. Do not post.
After the discussion settles, re-prompt with the action verbs.
When CI is failing on a staged PR, classify the failure:
If trivial, offer:
CI is failing on {workflow}. Looks like a {format/lint} fix — want me to patch it?
(yes / no — I'll just flag it in the review)
If yes:
~/worktrees/{repo}-{branch-slug})bunx biome check --write,
cargo fmt, golangci-lint run --fix, etc.)fix(ci): apply lint/format auto-fixPushed lint/format fix to {branch}. CI should re-run shortly. I'll wait.If no or non-trivial: include the CI failure as a finding and continue with the
normal action prompt.
When all staged entries have been processed (acted on or skipped):
Done. {posted}/{N} posted, {skipped}/{N} skipped.
/shipwright:deploy.headRefOid differs from
record.commitSha), it skips and points to /shipwright:review {org}/{repo}#{pr}.gh CLI for all GitHub interactions. Respect the active GH_TOKEN / gh auth context.state/reviews.json reads or
writes occur in this skill.Query and manage Shipwright agents via the admin API — cron jobs, env vars, tool permissions, API tokens, and plugins. Use to configure the running agent, manage its schedules, inspect another agent's config, or provision a new agent.
Orchestrates the full test-readiness pipeline (phases 1–5), once per qualifying repo under repos/, each in its own worktree + branch. Runs test-inventory → test-design → test-migration → test-roadmap → test-fix in sequence per repo, starting from the first stale artifact. Invoked by the `shipwright-test-readiness` cron on a daily schedule; exits early for a repo when all of its phase artifacts are fresh.
Scan Sentry for new/regressed unresolved issues, dynamically map them to repos. Report only — no code changes.
Cross-cutting contract for the test-readiness pipeline. Specifies the GitHub repository configuration that makes "tests pass" structurally meaningful instead of advisory — branch protection requiring CI to pass, required secrets, environments, deploy hooks, and PR templates. Referenced by `test-design` (Phase 2 emits the repo-config plan) and `test-roadmap` (Phase 4 auto-pairs workflow tasks with branch-protection tasks). Without these, the verification commands in each task body are honor-system; with them, the workflow is the gate.
Phase 4 of the test-readiness pipeline. Synthesizes the three prior artifacts (inventory, system design, migration) into a single executable roadmap `test-readiness-plan.md` with five sequenced milestones and an agent-executable task list. Includes mandatory sections on where we are now, where we want to be, the gap, the speed delta, and open risks. Output is suitable for handoff to an engineer or to `shipwright /dev-task`. Invoke when the `/test-roadmap` command runs.
Query and update the Shipwright task store — pick the next ready task, mark status transitions, and append new tasks. Use whenever you need to read from or write to the task queue. Calls the task store HTTP API directly via curl.