feat(skill): draft truthful replies in pr-feedback-quality-gate (#6317)

The bundled pr-feedback-quality-gate skill inspected PR feedback, applied
review fixes, validated them, and ran a read-only cross-review before pushing
follow-up commits. It had no advisory path for requests like "continue PR;
new responses appeared; tell me how to answer them," which left two risks: the
agent could miss a still-open request after a later approval, or imply a
change had happened when it hadn't.

Add a read-only reply-advice mode: deterministic target resolution, full
feedback-surface inspection (issue comments, review summaries, inline
threads, checks, local worktree), checkpoint/freshness semantics that call
the result "current feedback" rather than "new feedback" without a baseline,
five-way classification (action required / question / acknowledgement /
informational-approval / stale), truthful drafting rules that forbid
claiming performed changes that haven't happened, and the read-only authority
boundary (no edits, comments, thread resolutions, commits, pushes, or check
reruns without separate user authorization). The mutating fix workflow stays
opt-in below reply-advice so advisory intent is never mistaken for
authorization.

Add English (`continue PR`, `reply to PR feedback`, `answer PR feedback`)
and Polish (`kontynuuj PR`, `odpowiedz na uwagi do PR`) continuation/reply
triggers and an `od.example_prompt` that demonstrates a read-only "inspect
and draft, do not post" request, all surfaced through the runtime skill
registry.

Add a focused `listSkills(skillsRoot)` contract test in
`apps/daemon/tests/skills.test.ts` that verifies the triggers and
example_prompt are exposed as runtime metadata and that the composed body
contains the classification rules, truthful-draft markers, full
review-surface inspection, the read-only authority boundary, and the
preserved mutating fix/cross-review workflow. All 22 skills tests pass;
`pnpm guard` stays green.

Closes #6317.

Signed-off-by: xxiaoxiong <2482929840@qq.com>

Author: xxiaoxiong

apps/daemon/tests/skills.test.ts

@@ -298,6 +298,70 @@ describe('listSkills', () => {
     expect(skill.body).toContain('pkill -f -- "--user-data-dir=${CHROME_USER_DATA_DIR}"');
   });
 
+  it('exposes reply-advice mode on the pr-feedback-quality-gate skill (#6317)', async () => {
+    const skills = await listSkills(skillsRoot);
+    const skill = skills.find((entry: { id: string }) => entry.id === 'pr-feedback-quality-gate');
+
+    if (!skill) throw new Error('pr-feedback-quality-gate skill not found');
+
+    // English and Polish continuation/reply triggers surfaced via the runtime
+    // registry so the user-facing skill picker actually offers them.
+    expect(skill.triggers).toContain('continue PR');
+    expect(skill.triggers).toContain('reply to PR feedback');
+    expect(skill.triggers).toContain('answer PR feedback');
+    expect(skill.triggers).toContain('kontynuuj PR');
+    expect(skill.triggers).toContain('odpowiedz na uwagi do PR');
+
+    // An explicit English `od.example_prompt` is advertised as `examplePrompt`
+    // and demonstrates a read-only "inspect and draft, do not post" request.
+    expect(skill.examplePrompt).toBeTruthy();
+    expect(typeof skill.examplePrompt).toBe('string');
+    expect(skill.examplePrompt).toContain('Continue PR');
+    expect(skill.examplePrompt).toContain('draft');
+    expect(skill.examplePrompt).toContain('Do not post');
+
+    // The composed skill body exposes the reply-advice classification rules.
+    const body = skill.body ?? '';
+    expect(body).toContain('Reply-advice mode (read-only)');
+    expect(body).toContain('action required before replying');
+    expect(body).toContain('question requiring an answer');
+    expect(body).toContain('acknowledgement after a completed change');
+    expect(body).toContain('informational / approval / automated — no reply required');
+    expect(body).toContain('stale, outdated, or already resolved');
+
+    // Truthful drafting rules — never claim a change has been made when it
+    // has not; never claim new validation has run when it has not.
+    expect(body).toContain('Never claims a change has been made when it has not');
+    expect(body).toContain('Never claims new validation has run when it has not');
+    expect(body).toContain('current feedback');
+    expect(body).toContain('rather than');
+    expect(body).toMatch(/never\s+`new feedback`|not.*`new feedback`|call the result `current feedback`/i);
+
+    // Full review-surface inspection — every surface enumerated, and a surface
+    // with no items is itself an observation rather than something to skip.
+    expect(body).toContain('Issue (timeline) comments');
+    expect(body).toContain('Review summaries');
+    expect(body).toContain('Inline review threads');
+    expect(body).toContain('Status checks');
+    expect(body).toContain('Local worktree status');
+    expect(body).toContain('A surface with no items is itself an observation');
+
+    // Read-only authority boundary in reply-advice mode — no edits, comments,
+    // thread resolutions, commits, pushes, or check reruns without separate
+    // authorization.
+    expect(body).toContain('Reply-advice mode is the default path');
+    expect(body).toContain('NEVER edits PRs');
+    expect(body).toContain('resolves threads');
+    expect(body).toContain('commits, pushes');
+    expect(body).toContain('reruns checks');
+
+    // Existing minimal-fix, validation, and read-only cross-review rules
+    // remain present in the mutating workflow below reply-advice mode.
+    expect(body).toContain('mutating, opt-in only');
+    expect(body).toContain('read-only cross-review');
+    expect(body).toContain('pnpm guard');
+  });
+
   it('keeps html-ppt PNG export on one managed Chromium screenshot path', async () => {
     const skills = await listSkills(designTemplatesRoot);
     const skill = skills.find((entry: { id: string }) => entry.id === 'html-ppt');

skills/pr-feedback-quality-gate/SKILL.md

@@ -1,24 +1,133 @@
 ---
 name: pr-feedback-quality-gate
 description: |
-  Safely track pull request feedback, resolve review comments or merge conflicts, validate fixes, and use a read-only cross-review before committing or pushing follow-up changes.
+  Safely track pull request feedback, resolve review comments or merge conflicts, validate fixes, and use a read-only cross-review before committing or pushing follow-up changes. Also drafts truthful, ready-to-copy replies when the user asks how to answer new or current PR feedback without authorizing mutations.
 triggers:
   - "PR feedback"
   - "review comments"
   - "merge conflicts"
   - "cross-review"
   - "Claude CLI review"
   - "monitor PR"
+  - "continue PR"
+  - "reply to PR feedback"
+  - "answer PR feedback"
+  - "kontynuuj PR"
+  - "odpowiedz na uwagi do PR"
 od:
   mode: utility
+  example_prompt: "Continue PR #6016 — inspect current feedback, classify each item, and draft truthful copy-paste replies for anything actionable. Do not post, resolve threads, push commits, or edit the PR; only read and advise."
 ---
 
 # PR Feedback Quality Gate
 
 Use this when a PR has review feedback, merge conflicts, pending checks, or
-needs a monitored follow-up after a fix.
+needs a monitored follow-up after a fix, and use this when the user asks how to
+answer PR feedback (continue PR, reply to reviewer, kontynuuj PR, odpowiedz na
+uwagi do PR) without authorizing mutations.
 
-## Workflow
+## Reply-advice mode (read-only)
+
+Reply-advice mode is the default path whenever the user asks to continue a PR,
+review feedback, or draft replies, AND has not separately authorized the
+mutating fix workflow below. Reply-advice NEVER edits PRs, posts comments,
+resolves threads, commits, pushes, or reruns checks. It only inspects and
+advises. The mutating fix workflow runs only on an explicit separate request
+("apply the fix", "commit", "push", "resolve the thread"); until then, stay
+read-only.
+
+### Target resolution precedence
+
+Resolve the target PR deterministically before any inspection:
+
+1. Explicit URL or number supplied by the user.
+2. The PR open on the current git branch (`gh pr view` from the worktree).
+3. A single open PR authored by the current user in the current repository.
+
+If two or more candidates remain, list them with PR number, title, head SHA,
+author, and current review decision; then STOP. Do not pick silently. Let the
+user disambiguate before any inspection.
+
+### Feedback surfaces to inspect
+
+Read every surface that can carry actionable feedback on the resolved PR:
+
+- PR body, current head SHA, and mergeable state.
+- Issue (timeline) comments on the PR.
+- Review summaries (approve / request changes / comment) and reviewer bodies.
+- Inline review threads, including resolved and outdated state per thread.
+- Status checks (`gh pr checks`) and mergeability (`gh pr view --json
+  mergeable,mergeStateStatus`).
+- Local worktree status (`git status`, `git log`) when relevant.
+
+A surface with no items is itself an observation ("no inline review threads
+present on this PR"), not a gap to skip.
+
+### Freshness and checkpoint semantics
+
+Compare against a prior report or conversation checkpoint when one exists in
+this session. Without a checkpoint, call the result `current feedback` — never
+`new feedback`. Saying `new` without a comparison baseline is a fabrication
+risk; only the delta over a recorded observation is `new`.
+
+### Classification
+
+Classify every relevant item on every inspected surface as exactly one of:
+
+- **action required before replying** — a request that must be addressed
+  (code change, PR body update, validation, doc edit) before any truthful
+  reply can claim it is done.
+- **question requiring an answer** — a reviewer question that needs a
+  response, with no implicit code change required.
+- **acknowledgement after a completed change** — a reviewer note reacting to
+  an already-completed fix; reply optionally, briefly.
+- **informational / approval / automated — no reply required** — CI runs,
+  Looper approvals, status updates, automation logs. These are not human
+  feedback. Do not draft a substantive reply; an optional courtesy `thanks for
+  the review and approval` is the maximum allowed.
+- **stale, outdated, or already resolved** — threads marked resolved/outdated
+  by GitHub, or items superseded by a newer change. Acknowledge their state
+  once; do not relitigate.
+
+An item may have a single classification. If two classifications apply, the
+most actionable one wins (a question that is also an approval is still a
+question).
+
+### Truthful drafting rules
+
+For every actionable item, write a ready-to-copy reply that:
+
+- States the smallest required action and why, before any draft reply.
+- Never claims a change has been made when it has not. If the user has not
+  separately run the fix, the draft must say `Once X is done, the suggested
+  reply is:` and present the reply as conditional, not past tense.
+- Never claims new validation has run when it has not. If the user has not run
+  tests, say `after you run \`pnpm X\`` rather than `I ran \`pnpm X\``.
+- Mirrors the reviewer's language register and addresses the reviewer by name
+  when known.
+- Is concise — one short paragraph per actionable item, not a wall of text.
+
+### Output shape
+
+Reply-advice output, in order:
+
+1. PR state and freshness — number, title, head SHA, review decision,
+   mergeable state, last update timestamp, and the comparison baseline (or
+   `no prior checkpoint — current feedback, not new`).
+2. Feedback item — author, source (issue comment / review summary / inline
+   thread / check), timestamp, classification.
+3. Required action and why (when classification is `action required` or
+   `question`).
+4. Suggested reply, or explicit `no reply needed` rationale. Conditional
+   language if the underlying action has not yet been performed.
+5. Checks and mergeability — label unknown causes as `unknown`; do not
+   invent a blocker.
+6. Remaining risk and next step — the smallest action that unblocks the PR.
+
+## Workflow (mutating, opt-in only)
+
+This workflow runs only on an explicit mutating request from the user. It does
+not run inside reply-advice mode.
 
 1. Inspect PR state first: comments, reviews, mergeability, checks, branch, and
    local worktree status. Keep unrelated local changes out of the PR.
@@ -50,4 +159,5 @@ needs a monitored follow-up after a fix.
 
 Always report PR state, actions taken, cross-review verdict, accepted or
 rejected findings, validation run, commits pushed, skipped checks with reasons,
-remaining risks, and next step.
+remaining risks, and next step. In reply-advice mode, replace `actions taken`
+with `actions advised` and `commits pushed` with `no mutations performed`.