| name | kmaz-review-comments |
| description | Address the comments, questions, or feedback on the open Pull Request or Merge Request, then propagate any durable lessons back into the project's architecture, ADRs, roadmap, or build conventions so the next feature inherits them. Use when there is an open PR/MR with review feedback to work through. |
Review Comments → Fixes + Durable Lessons
Pipeline conventions — READ FIRST: before working the comments, read
../kmaz-pipeline/CONVENTIONS.md in full — it holds the compound
learning loop (propagating durable lessons upstream) and the other shared rules. If you dispatch any
sub-agent, copy in only the rules relevant to its task. This skill is where review feedback enters
that loop.
What this is for
A PR review is two things at once: a list of changes to make to this PR, and a stream
of lessons that should outlive this PR. The default failure mode is to address only
the first and lose the second — every reviewer comment that taught you something becomes
tribal knowledge in this one PR's thread, and the next feature repeats the same mistake.
This skill addresses the comments and captures the durable lessons in the artifacts
that future features and agents will actually read.
Workflow
Step 1 — Pull all comments from the PR/MR
Use the appropriate CLI (gh pr view --comments and gh api repos/<o>/<r>/pulls/<n>/comments
for GitHub; glab mr notes for GitLab) to pull every comment, question, and review thread
on the open PR/MR. Include both top-level review comments and inline file/line comments.
Step 2 — Triage each comment
For each comment, decide one of:
- Fix — the comment identifies a real issue; change the code.
- Respond — the comment is wrong, already addressed, or out of scope; reply with a
clear explanation of why and link to evidence (a test, an ADR, an architecture
section). Don't be defensive; do be specific.
- Defer — the comment identifies a real issue but it's larger than this PR; open a
follow-up issue / new feature file, link it in the response, and explain the defer.
Address comments in a deliberate order: blocking / requested-changes first, then
correctness, then style. Don't batch — work them one at a time so the PR history shows
each thread closing.
Step 3 — Apply fixes and respond
For each fix: make the change, run the relevant tests, and reply on the thread linking
to the commit that addressed it. For each respond/defer: post the reply and explain.
Resolving the threads is your job — it is a core part of this skill, not an optional
courtesy and not something to ask permission for. As each thread is addressed (fixed,
responded, or deferred with a linked follow-up), mark it resolved via the forge API
— for GitLab: glab api --method PUT "projects/<o>%2F<r>/merge_requests/<n>/discussions/<id>?resolved=true"
(fetch discussion ids from .../merge_requests/<n>/discussions); for GitHub: resolve the
review thread. Resolving the reviewer's own threads (including an AI review bot's) is
expected workflow — do not stop to confirm it. The reviewer can always re-open if not
satisfied. A review round is not done until the threads it raised are resolved (or
explicitly left open with a stated reason, e.g. a deferral the reviewer must sign off on).
Step 4 — Capture durable lessons (the Compound step)
This is the step that exists only in this skill, and it's the most important. A PR
review surfaces lessons that no adversarial subagent caught — because reviewers bring
context, taste, and history the agent didn't have. Let those lessons compound.
For each review comment that resulted in a fix or a defer, ask: would the next
feature builder benefit from knowing this? If yes, propagate it now, in this same
session, before the PR merges:
- A reviewer disagreed with a decision and you changed it → write a new ADR
capturing the new decision and superseding the old one (if there was one). Don't
let the reasoning live only in a PR thread.
- A reviewer pointed out a pattern the codebase prefers → add it to the relevant
cross-cutting concerns section of
ROADMAP.md, or — if it's about how to build
features in this repo — propose adding it to CLAUDE.md for the user to accept.
- A reviewer caught a class of bug the adversarial review missed → consider whether
it should become an explicit check the next adversarial review run will catch (e.g.
a project-specific lens to brief the security or robustness reviewer with). Note it
for the user.
- A reviewer surfaced a missing integration / contract / dependency → update
ARCHITECTURE.md (a new component, a new flow) or ROADMAP.md (a missing
integration feature, a new cross-cutting concern). Update the diagrams.
- A reviewer flagged a doc/spec gap → update the relevant feature file, ADR
consequences section, or brand contract so it doesn't recur.
If nothing in the review qualifies — sometimes a review is purely about this PR's
code — say so explicitly in the PR body addendum (below). That's a valid outcome; the
discipline is to ask, not to manufacture lessons.
Step 5 — Update the PR body with what you did
Append (or update) a ## Review response section to the PR/MR body listing:
- Threads addressed (count, with links to the resolving commits).
- Threads where you responded without changing code (with brief rationale).
- Threads deferred (with links to follow-up issues/feature files).
- Propagated lessons: ADRs created/superseded, ARCHITECTURE.md / ROADMAP.md
updates, CLAUDE.md proposals, or "no durable lessons this round."
Step 6 — Push and notify
Push the fix commits and the doc-propagation commits. The propagation commits should
be separate from the code fixes so the diff is readable. Reply on the PR with a
short summary comment naming the propagated lessons (so the reviewer sees their
feedback compounded, not just locally absorbed).
Confirm every addressed thread is marked resolved (Step 3) before you consider the
round done — a green re-review with the threads still flagged "unresolved" leaves the MR
showing blocking discussions. If a force-push/rebase re-triggered the reviewer and it came
back clean, that confirms the fixes landed: resolve the original threads. This is the
agent's responsibility; don't defer it to the user.
Do not re-request review automatically; the reviewer will look when ready.
Judgment notes
- The Compound step (Step 4) is the whole point. A skill that only addresses comments
is just
gh pr review --comment; the value here is making review feedback into
inheritable knowledge.
- If a propagation would touch many files or change the architecture meaningfully,
pause and confirm with the user before making the change. Quick edits (a new ADR,
a new bullet in ROADMAP.md) don't need confirmation; restructuring ARCHITECTURE.md
does.
- Don't propagate trivial lessons. Bar: "would the next feature actually benefit?"
If the answer is "probably not," skip it. Quantity of propagation is not a goal.
- Reviewers' comments are evidence; treat them with the same respect as your own
adversarial subagents' findings. If you disagree, push back with a reason — don't
perform agreement.