name	review-pr
description	Process all open review comments on a PR: fix addressable issues, reply in-thread, create follow-up issues for deferred items
disable-model-invocation	true

Review PR — Process Review Feedback

Read all open review comments on a pull request, fix addressable issues, reply to each comment thread with what was done, and create follow-up issues for items that need deferral. Used by both the /forge-loop pipeline and as a standalone skill for interactive review processing.

Input

$ARGUMENTS — optional PR number. Defaults to the PR for the current branch.

/review-pr — process the current branch's PR
/review-pr 97 — process PR #97

Step 0 — Read Configuration

Read forge.toml:

[project]
repo = "org/project"

[gating]
max_copilot_iterations = 3

Read CLAUDE.md for architecture conventions, testing requirements, and coding patterns. All fixes must follow project conventions.

Step 1 — Identify PR

If no PR number provided:

gh pr view --repo $REPO --json number,headRefName

If still ambiguous (multiple PRs for this branch), list them and ask the user to specify.

Step 2 — Fetch All Open Review Feedback

Fetch formal reviews with CHANGES_REQUESTED or COMMENTED state:

gh api "repos/$REPO/pulls/$PR_NUMBER/reviews" \
  --jq '[.[] | select(
    .state == "CHANGES_REQUESTED" or .state == "COMMENTED"
  ) | select(.user.login != "github-actions[bot]")
  | select(.body | length > 0)]
  | sort_by(.submitted_at) | .[-5:]'

Fetch inline comments not already replied to by this agent:

# Get all inline comments
ALL_COMMENTS=$(gh api "repos/$REPO/pulls/$PR_NUMBER/comments" \
  --jq '[.[] | select(.user.login != "github-actions[bot]")]')

# Get IDs of comments this agent has already replied to
REPLIED_TO=$(echo "$ALL_COMMENTS" \
  | jq '[.[] | select(.user.login == "claude[bot]" or .user.login == "github-actions[bot]")
         | .in_reply_to_id] | unique')

# Unresolved: top-level comments not already replied to
UNRESOLVED=$(echo "$ALL_COMMENTS" \
  | jq --argjson replied "$REPLIED_TO" \
    '[.[] | select(.in_reply_to_id == null)
           | select(.id as $id | $replied | index($id) | not)]
     | sort_by(.created_at) | .[-20:]')

Step 3 — Triage Comments

For each comment, classify:

Category	Criteria	Action
Fix now	Bug, incorrect behavior, broken test, security issue, clear code error	Implement fix in Step 4
Discuss	Design disagreement, ambiguous requirement, architectural tradeoff	Reply explaining the reasoning; create follow-up issue if significant
Defer	Scope creep, nice-to-have, out-of-scope for this PR	Reply noting it's deferred; create follow-up issue
Already addressed	The code has already been fixed by a prior commit	Reply with the commit SHA that addressed it

Summarize the triage before fixing:

Found $TOTAL review comments on PR #$PR_NUMBER:
  Fix now:           $N (bugs, broken tests, clear errors)
  Discuss/explain:   $N (design choices to explain)
  Defer:             $N (out-of-scope, follow-up issues)
  Already addressed: $N

Step 4 — Fix Addressable Comments

For each "fix now" comment:

Read the file at the referenced path and line
Understand the issue being raised
Implement the fix (edit the file)
Note the fix for the reply in Step 5

Follow all conventions from CLAUDE.md. If fixing a test, run the tests before proceeding to the next comment to catch cascading failures early.

If a fix proves more complex than expected (requires architecture changes, affects multiple files, or would expand the PR scope significantly), reclassify as "defer" and create a follow-up issue instead.

Step 5 — Run Verification

After all fixes are applied:

$TEST_COMMAND  # from forge.toml [stack]
$BUILD_COMMAND

If tests fail after fixes: investigate and fix. If still failing after 2 attempts, revert the problematic fix and reclassify as "defer" with a note about why the fix was reverted.

Step 6 — Commit Fixes

git add -A
git commit -m "fix: address review feedback on PR #$PR_NUMBER

$(echo "$FIX_SUMMARY")"
git push

Use a concise commit message describing what was addressed in aggregate, not each individual comment.

Step 7 — Reply to Each Comment Thread

For every comment processed, post a reply in the comment thread. This is required — it closes the loop for reviewers and makes it visible what was addressed vs. deferred.

# For inline comments
gh api -X POST \
  "repos/$REPO/pulls/comments/$COMMENT_ID/replies" \
  -f body="$REPLY_BODY"

Reply templates:

Fixed:

Fixed in $COMMIT_SHA — $ONE_LINE_DESCRIPTION_OF_WHAT_CHANGED.

Deferred (follow-up filed):

Good catch — deferred to #$FOLLOWUP_ISSUE_NUMBER to keep this PR focused.

Discussed/explained:

$EXPLANATION_OF_REASONING

Happy to revisit if this doesn't address the concern.

Already addressed:

This was addressed in $COMMIT_SHA (before this review was posted).

Step 8 — Create Follow-up Issues

For each "defer" comment, create a GitHub issue:

gh issue create \
  --repo $REPO \
  --title "follow-up from PR #$PR_NUMBER: $SHORT_DESCRIPTION" \
  --label "backlog/needs-refinement" \
  --body "$(cat <<EOF
## Context

Raised as a review comment on PR #$PR_NUMBER by @$REVIEWER.

## Original comment

> $COMMENT_BODY

(File: $FILE_PATH, line $LINE_NUMBER)

## Why deferred

$DEFERRAL_REASON

## Acceptance criteria

$SUGGESTED_CRITERIA
EOF
)"

Step 9 — Update PR Phase State

If the PR has an autodev-state tracker comment, increment copilot_iterations:

PR_BODY=$(gh pr view $PR_NUMBER --repo $REPO --json body --jq .body)
CURRENT_STATE=$(echo "$PR_BODY" | grep -oP '(?<=<!-- autodev-state: ).*?(?= -->)')
CURRENT_ITERS=$(echo "$CURRENT_STATE" | jq -r '.copilot_iterations // 0')
NEW_ITERS=$(( CURRENT_ITERS + 1 ))

NEW_STATE=$(echo "$CURRENT_STATE" | jq ". + {\"copilot_iterations\": $NEW_ITERS}")
UPDATED_BODY=$(echo "$PR_BODY" | sed "s|<!-- autodev-state: .*-->|<!-- autodev-state: $NEW_STATE -->|")
gh pr edit $PR_NUMBER --repo $REPO --body "$UPDATED_BODY"

Step 10 — Summary Report

Review processing complete for PR #$PR_NUMBER.

  Fixed:    $FIX_COUNT comments addressed and committed ($COMMIT_SHA)
  Replied:  $REPLY_COUNT comment threads replied to
  Deferred: $DEFER_COUNT follow-up issues created (#$N1, #$N2, ...)
  Explained: $DISCUSS_COUNT design decisions explained inline

Iteration: $NEW_ITERS / $MAX_COPILOT_ITERATIONS

If NEW_ITERS >= MAX_COPILOT_ITERATIONS:

Max Copilot iterations reached. If there are still unresolved findings, they will be
handled in the Claude review phase. Run /forge-loop or trigger claude-code-review.yml manually.

Guardrails

Never modify files listed in forge.toml [protected_files] patterns
Never commit without running tests first
Never reply "done" without a specific description of what changed
Never create follow-up issues for already-addressed comments
If the PR has no open comments at all, report that and exit cleanly
Copilot reviews once per PR — it only posts its initial review on the first push. It does NOT re-review after subsequent commits. Never poll for a second Copilot review after pushing fixes; proceed directly to the Claude review pass instead.