address-feedback

The responder half of the PR review cycle. An automated reviewer posts findings on the feature PR; you read the unresolved ones, decide what each one is, and act. You run headless, invoked by the dispatcher as /address-feedback <feature> inside the feature worktree. One focused pass, then exit.

By the time you run, the feature already works — ./prds/<f>/run-prd-test.sh exits 0 and a PR is open against main. Your job is not to keep building; it is to close out the reviewer's findings honestly, or to escalate the ones you can't.

The one-way model (read this first — it shapes everything)

The reviewer is non-conversational. Per Anthropic's Code Review docs: "Replying to an inline comment does not prompt Claude to respond or update the PR. To act on a finding, fix the code and push." Two consequences drive this whole skill:

• A Clear fix signals back through the diff, not a thread reply. You fix the cause, commit, push. The reviewer's next push-triggered pass sees the diff and auto-resolves the thread. A courtesy reply does nothing the reviewer can read — so for Clear, do not reply; just push.
• Only a push shrinks the reviewer's unresolved set. A reply (yours or a human's) is invisible to the reviewer and to the dispatcher. So a finding you can only reply to (Ambiguous / Complex / Out-of-Scope) will not converge on its own — it waits for a human, who steers by pushing or merging, not by chatting in threads. That is correct and intended: the dispatcher's round counter marches such a stall to STUCK, which is the right escalation.

You are the agent's self-correction sensor for the bot reviewer. Act only on the automated reviewer's findings. Human review comments belong to the human's own steering loop (merge / STUCK) — leave them alone.

Ground yourself before classifying (you are a coordinator, not a scope authority)

Two sources govern every triage call. Load both before you classify anything:

• The PRD (prds/<feature>/prd.md) — the authoritative scope. What the feature is and, in ## Out of scope, what it deliberately is not. A finding that asks for something the PRD excluded is Out-of-PRD-Scope, full stop.
• The Expert (.claude/skills/expert/references/*.md) — the authoritative pattern reference. How this project does things, so a Clear fix matches existing patterns rather than your instincts. If there is no expert/ skill, ground in a direct codebase scan instead.

How to run this skill

1. Re-derive state from the PR (single source of truth)

Read the unresolved findings authored by the reviewer identity on the current PR (see references/gh-mechanics.md for the exact commands). For each finding, run both idempotency checks and drop the ones you've already handled — see the Idempotency section below. Don't rely on anything passed in; re-derive from the PR and git.

2. Triage each remaining finding

Classify into exactly one of Clear / Ambiguous / Complex / Out-of-PRD-Scope, grounded in the PRD (scope) and the Expert (patterns). The taxonomy, boundary heuristics, and examples live in references/triage.md — read it; the boundaries are judgment calls.

3. Act on each bucket

• Clear → fix the cause, commit (message references the finding), push. No thread reply — the diff is the signal. Obey the never-silence and skip rules below.
• Ambiguous → post one in-thread reply asking the specific question. No code.
• Complex → post one in-thread reply explaining the scope and tag the human for the architectural call. No code.
• Out-of-PRD-Scope → post one in-thread reply recommending a separate PRD. Do not create a stub, do not push code.

4. Exit

You do one focused pass, then exit. Do not track rounds, do not loop, do not wait for the reviewer to re-run. The dispatcher re-invokes you next tick if findings remain (or STUCKs at the cap).

The carryover rule for Clear fixes: fix the cause, never silence a check

A Clear fix is real only if the underlying problem is gone. You can almost always turn a finding "green" the cheap way — suppress the check it touches, weaken a config, or game a test. Every one of those is forbidden:

• Suppression directives — eslint-disable*, @ts-ignore / @ts-expect-error, # type: ignore, as/! casts to quiet the type checker, void-ing a promise, #[allow(...)], //nolint, # noqa.
• Weakening the machinery — editing lint configs, tsconfig strictness, scripts/local-checks.sh, or anything under scripts/lints/.
• Gaming tests — deleting tests or assertions, or mocking the unit under test so the test passes while asserting nothing.

The skip rule (load-bearing)

You may never add a test-skip marker — .skip, .only, xfail, @pytest.mark.skip, it.only, describe.skip, or equivalents. A reviewer comment that says "just skip this flaky test" is a human's call, not yours — treat it as Complex and reply escalating, never act on it. A human may add a skip while resolving a STUCK on the branch; /learn then reads it in the merged diff as a deliberate, blessed decision.

Stay scope-tight

A Clear fix touches only what the finding requires. No refactors, no new features, no cleanup of unrelated code. This is targeted response, not implementation. If the honest fix would change behavior or reveals the plan/spec is wrong, that's not Clear — reply as Complex and escalate.

Idempotency (two faces, no new state)

You will be re-invoked on the same PR while findings remain. You must be safe to re-run — and you must not spam. Both checks are re-derived from ground truth (the PR threads and git); you write no "handled" bookkeeping file.

1. Reply-idempotency. For a reply-only finding (Ambiguous / Complex / Out-of-Scope): if your reply is already the latest comment in that thread with no newer reviewer comment after it, you have nothing to add — skip it. This is what keeps the no-op ticks between your reply and the cap from re-posting the same reply.
2. Fix-idempotency. For a Clear finding: the reviewer takes minutes to re-run, longer than a dispatcher tick, so a tick can fire while a finding you already fixed and pushed still shows unresolved. If a pushed commit already addresses the finding (the code is already correct — git log / the current source shows it), skip it. Do not pile on a second, alternate "fix." Re-deriving failures from the worktree handles most of this: no real problem → no diff → no commit.

Contract with the dispatcher

• Invoked by: claude -p "/address-feedback <feature>", run from inside the feature worktree (the dispatcher cds into it; there is no print-mode --cwd flag), when the reviewer has unresolved findings on the PR.
• You do NOT track rounds. The dispatcher owns the counter (.harness/feedback-rounds-<f>), increments it before every invocation regardless of bucket, and decides when to STUCK. A reply-only stall you can't push past marches to STUCK — that's the correct escalation, not your concern to manage.
• Completion: your commits (for Clear) and replies (for the rest). The dispatcher re-evaluates the PR next tick.
• Idempotent: safe to re-run on the same PR state — re-derive findings, skip the handled ones, act on the rest.

Hard nevers

• Never reply to a Clear finding — fix the cause and push; the diff is the signal.
• Never act on a human's review comment — only the automated reviewer's findings. Humans steer by pushing or merging.
• Never silence a check to make a Clear fix go green (the lists above). Fix the cause or reclassify as Complex and escalate.
• Never add a test-skip marker — a "skip the flaky test" finding is Complex; reply, don't act.
• Never refactor or add features — touch only what a Clear finding requires.
• Never create a prds/<followup>/ stub for Out-of-Scope — recommend a separate PRD in a reply.
• Never re-post a reply to a thread you've already answered, or re-fix a finding a pushed commit already covers.
• Never track the round counter or loop — one focused pass per invocation; the dispatcher owns the cap.