// Fetch PR comments and reviews, classify them, and produce a response plan that the originating loop can execute. Internal — invoked by /ev-loop-* skills when feedback arrives on a project PR. Produces a plan; does not execute it. Does not write code or reply on GitHub.
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | trout-pr-respond |
| description | Fetch PR comments and reviews, classify them, and produce a response plan that the originating loop can execute. Internal — invoked by /ev-loop-* skills when feedback arrives on a project PR. Produces a plan; does not execute it. Does not write code or reply on GitHub. |
| argument-hint | <project-slug-or-path> <pr-number> |
| user-invocable | false |
| allowed-tools | Read, Bash(node .claude/scripts/trout/*) |
Turn a pile of PR feedback into a structured response plan. Does not modify code or reply; its output is consumed by the originating loop.
The gh CLI plumbing (PR fetch, item normalization, response-NN.md
write) lives in .claude/scripts/trout/pr-respond-plumbing.ts —
verbs fetch and write-plan. This skill orchestrates: invoke
fetch, classify per the taxonomy, author the plan, invoke
write-plan.
<project-slug-or-path> — resolved like .claude/scripts/trout/autosave.ts.<pr-number> — the PR. Must belong to the project; the script's
cross-project guard refuses if the PR's branch has no
checkins/<branch>/ directory under the resolved project.Bash("node .claude/scripts/trout/pr-respond-plumbing.ts fetch <slug> <pr>").
The script emits one JSON document with pr (number, url, state,
branch, title), items[] (kind ∈ issue-comment | review |
review-comment | ci-failure; source, body, location, url),
next_response_number, and response_path. Items are emitted in
deterministic order: issue-comments → reviews → review-comments →
ci-failures. Only failure-class CI conclusions (FAILURE / CANCELLED /
TIMED_OUT) appear; SUCCESS and NEUTRAL are filtered out.
| Class | Meaning | Response style |
|---|---|---|
| Blocker | Reviewer says "change X before merge" or CI is red | Fix in new checkin; do not merge |
| Suggestion | Reviewer offers an alternative or improvement | Evaluate; adopt, counter, or decline with reason |
| Question | Reviewer is asking for clarification | Reply; no code change unless the answer implies one |
| Nit | Style or micro-polish | Apply if trivial; otherwise batch and decline politely |
| Praise / ack | Positive, informational | No action |
| Off-topic | Not about this PR's scope | Park for follow-up project or defer |
| CI failure | A required check is failing | Treat as Blocker; one item per failing check |
When the same issue is raised by multiple reviewers, collapse to one
row and note the multiplicity. kind: "ci-failure" items are
auto-Blockers; the LLM still chooses how to address them.
3a (LLM). Author content using the template below. Write to a
temp file at /tmp/pr-respond-plan-<slug>-<pr>-<timestamp>.md.
# PR #<N> response plan — generated <YYYY-MM-DD HH:MM>
## Items
### Item 1 — <short label> [Blocker | Suggestion | Question | Nit | Off-topic]
- **Source**: <reviewer handle, comment link if available>
- **Summary**: <1–2 sentences>
- **Proposed action**: <Fix in code | Reply only | Decline with reason | Defer>
- **If code change**: <files, approach>
- **If reply only**: <draft reply, 1–3 sentences>
### Item 2 — …
End the plan with a "Recommended next unit" line.
3b (script). Bash("node .claude/scripts/trout/pr-respond-plumbing.ts write-plan <slug> <pr> --content-file=<temp-file>").
The script computes the next NN (max across <NN>.md and
response-<NN>.md; defaults to 01) and writes to
projects/<slug>/checkins/<branch>/response-<NN>.md. The
response-<NN>.md file is intentionally distinct from the eventual
<NN>.md checkin — plan and executed work share the number, separate
files. The loop writes the matching <NN>.md after executing.
Do not execute or post replies. Report the plan path and recommended next unit to the caller. The router or originating loop picks it up.
items[]. No entries uncategorized.checkins/<branch>/ directory → script refuses; stop and report.write-plan fails → temp file is preserved at the path the skill
authored; user can recover.