review-response

name

review-response

description

Resolve the reviewer findings on your own open PR end-to-end — fetch the diff AND every inline comment (human and bot/automated), verify each against current source (file:line), apply the minimum scoped fix with red→green proof, RE-RUN the §7 gate on the fix (maker ≠ checker — a fix is not self-certified), reply to every comment, and confirm the head is green. Never merges — stops for the owner. Pushes fixes and posts replies (the write posture pr-review lacks). Project specifics from the profile. Use when the user says "address the reviewer comments on PR

/review-response — Claude Code binding

The workflow logic is runtime-neutral and lives in .claude/workflow/review-response.md — read that file now and execute it. It composes existing roles (introduces no new binding-contract row): the pr-review read-and-ground reads, the next-task §5 scoped-fix + falsification and §7 re-gate, and the §8 digest / green / no-merge terminal. Mapping of its abstract [roles] and the concrete reads/writes:

Neutral role / step	Claude Code mechanism
[workflow] (this one)	this skill; the PR number/URL arrives in the invocation text
[bulk-read offload]	the Explore subagent (spawn on the [cheap tier] per .claude/MODELS.md) for large diffs/threads
PR head checkout (before fix + re-gate)	gh pr checkout <n> — make the working tree and git diff main..HEAD BE PR #N's head before editing or dispatching the gate (see "Check out the PR head" note)
[reviewer]s / [orchestrated run] (§5 re-gate)	the §7 gate on the fix commit: the Workflow tool running .claude/workflows/gate-loop.js (binding of workflow/gate-loop.md), models resolved from .claude/MODELS.md and passed in args {taskId, strongModel, cheapModel, dispatchContract, dispatchSpec, taskBranch, fix?} — taskBranch is the checked-out PR head branch (from gh pr checkout <n> below), the required review-mode audited ref the gate hard-errors without (gate-loop.js throws unless taskBranch or workspacePath is present — T639/#240); pass workspacePath instead under an engaged isolated autonomous run; or the constitution-auditor / contract-auditor / spec-auditor / spec-quality-auditor subagents (.claude/agents/) dispatched via the Agent tool, read-only, against git diff main..HEAD on the checked-out head
[code-review pass]	/code-review over the checked-out PR head (the §7-step-3 advisory pass; /security-review when the fix touches privacy, credentials, or payments)
PR reads (diff + every comment)	gh pr view <n> --json title,body,state,baseRefName,headRefName,url,comments (the comments field is the timeline / issue-level channel — §2.5 owner steering) and gh pr diff <n>, plus the line-anchored inline findings gh pr view omits: gh api --method GET --paginate repos/{owner}/{repo}/pulls/<n>/comments (inline review comments) and .../pulls/<n>/reviews (review summaries — where some bots/Codex post). Match automated reviewers by their [bot] login suffix, never an exact string (the Codex reviewer posts as chatgpt-codex-connector[bot]); enumerate first, filter second — an empty filtered set is suspicious, never "nothing to address". Where the GitHub MCP server (mcp__github__*) is connected, its PR review-comment / diff tools are a drop-in alternative for these reads
PR writes (fixes + replies)	fixes: commit the scoped change on the PR branch (stage specific files, never git add .; tick any owned tasks.md box in the same commit) and git push to PR #N's head branch. replies: gh api --method POST repos/{owner}/{repo}/pulls/<n>/comments/<comment-id>/replies (threaded reply to an inline finding) or gh pr comment <n> --body-file <tempfile> (timeline reply), each ending with the [comment marker]. Never gh pr merge, never gh pr close
[comment marker]	the footer line defined in .claude/skills/next-task/SKILL.md → "The [comment marker] concrete form" (the single copy) — appended to every reply this workflow posts
[environment block]	.claude/skills/next-task/SKILL.md → "This environment's concrete forms" (the single copy — multi-line bodies via a UTF-8 temp file + --body-file, gh PATH fallback)
[headless run]	claude -p "/review-response <pr>"

Profile read — packet first (spec 007 US3.AC3). The workflow doc's project-specifics read resolves to .claude/PROJECT.compact.md by default — the compact packet of active routing facts (base branch, conventions, required check, review passes, critical invariants + backstops), drift-checked against the full profile by .claude/hooks/compact-packet-drift.sh in CI verify. Escalate to the full .claude/PROJECT.md explicitly — say you are escalating and why — when a fix needs a fact the packet omits (an invariant's full text or auditor rule, architecture boundaries, coverage policy). The full profile stays the source of truth.

Check out the PR head before fixing or re-gating. A fix must be committed onto PR #N's head, and /code-review + the §7 gate read the current branch — so before §4/§5 run gh pr checkout <n> to make the working tree (and git diff main..HEAD) be PR #N's head. Without the checkout you would commit onto the wrong branch and the gate would grade the wrong diff. For a PR not based on main, capture the diff against the PR's actual base (gh pr view <n> --json baseRefName) and pass the auditors that patch. The PR reads work from any branch — only the fix, the lens passes, and the re-gate need the checkout.

Re-gate is not optional (the maker ≠ checker bound). review-response.md §5 requires that any fix be re-graded by the §7 [reviewer]s on the fix commit — you never push a fix as "resolved" under your own say-so. Commit first (the gate audits the committed diff), dispatch the reviewers per gate-loop.md → "The reviewer roster", loop until every dispatched reviewer returns PASS/JUSTIFY, and keep every verdict verbatim (append the gate-run record per telemetry.md, as the next-task [orchestrated run] row does). Non-convergence after two rounds → stop and surface in the reply/PR body (§5).

Write posture — the allowlist. This skill pushes to the PR branch and posts replies, so unlike pr-review (read-then-comment) it needs write grants beyond read-only gh api --method GET reads (which — per T641/#254 — are not pre-approved by a gh api --method GET:* allowlist wildcard, since its trailing wildcard would auto-approve an appended --method PUT merge; they prompt or route through the GitHub MCP server): git push to the PR branch and gh api --method POST .../pulls/<n>/comments/*/replies (threaded replies) or gh pr comment. It still never merges or closes — gh pr merge / gh pr close are out of scope and must not be allowlisted by this skill. The harness cannot grant its own allow rules, so the owner adds the push/reply entries directly (or connects the GitHub MCP server, whose write tools are permission-gated on their own); until then an interactive run prompts on the first write, and a [headless run] degrades per workflow/README.md → "How an adapter degrades gracefully" (surface the pending writes rather than silently skipping them).

No new §7 gate semantics, and no merge. Per review-response.md → "Write posture": this workflow re-runs the existing §7 gate on its fixes (it does not create a second gate or alter the gate's round limits, veto authority, or tier floors), and its terminal state is a green PR with every comment answered, handed to the owner. Merge authorization is the owner's alone, session-explicit (§2.5) — review-response never merges, even when every check is green.