fetch-pr-feedback

name	fetch-pr-feedback
description	Fetch unresolved review comments from a PR and evaluate with receive-feedback skill
disable-model-invocation	true

Fetch PR Feedback

Fetch review comments from all reviewers on the current PR, format them, and evaluate using the receive-feedback skill. Excludes the PR author and current user by default. Line-specific comments belonging to resolved review threads are also excluded by default.

Usage

Invoke the fetch-pr-feedback skill, optionally passing these flags:

fetch-pr-feedback [--pr <number>] [--include-author] [--include-resolved]

Flags:

• --pr <number> - PR number to target (default: current branch's PR)
• --include-author - Include PR author's own comments (default: excluded)
• --include-resolved - Include line-specific comments from resolved review threads (default: excluded)

Instructions

Gates (sequence; do not skip)

Advance only after each Pass when is satisfied.

1. PR context — Pass when: $PR_NUMBER is set to a positive integer and gh pr view / gh api for that PR completed with exit code 0, or you stop in Get PR Context with only the failure given there (“No PR found for current branch…”).
2. Fetch — Pass when: the resolved-thread GraphQL call (skipped if --include-resolved) and both paginated gh api … | jq -s -f … runs (issue comments + review comments) exit 0 and parse as JSON (empty [] is valid). On non-zero exit or jq error, stop; surface command stderr—do not invent comments.
3. Formatted artifact — Pass when: output is either (a) markdown matching Format Feedback Document (header # PR #$PR_NUMBER Review Feedback, per-reviewer ## Reviewer: … with Summary / Line-Specific sections), or (b) exactly: No review comments found on this PR (excluding PR author, current user, and resolved threads).
4. Load receive-feedback — Pass when: the receive-feedback skill is loaded; only then run that skill’s verify → evaluate → execute loop on that formatted document.

1. Parse Arguments

Extract flags from $ARGUMENTS:

• --pr <number> or detect from current branch
• --include-author flag (boolean, default false)
• --include-resolved flag (boolean, default false)

2. Get PR Context

# If --pr was specified, use that number directly
# Otherwise, get PR for current branch:
gh pr view --json number,headRefName,url,author --jq '{number, headRefName, url, author: .author.login}'

# Get repo owner/name
gh repo view --json owner,name --jq '{owner: .owner.login, name: .name}'

# Get current authenticated user
gh api user --jq '.login'

Store as $PR_NUMBER, $PR_AUTHOR, $OWNER, $REPO, $CURRENT_USER.

Note: $OWNER, $REPO, etc. are placeholders. Substitute actual values from previous steps.

If no PR exists for current branch, fail with: "No PR found for current branch. Use --pr to specify a PR number."

3. Fetch Comments

Fetch both types of comments, excluding $PR_AUTHOR and $CURRENT_USER (unless --include-author is set). Use --paginate with jq -s to combine paginated JSON arrays into one.

Unless --include-resolved is set, first fetch the databaseIds of every review comment that belongs to a resolved review thread. Issue comments (summary/walkthrough) aren't part of review threads, so this only affects line-specific review comments.

Resolved-thread comment IDs (skip this block entirely when --include-resolved is set; set RESOLVED_IDS='[]' instead):

RESOLVED_IDS=$(gh api graphql \
  -F owner="$OWNER" -F repo="$REPO" -F pr="$PR_NUMBER" \
  -f query='
    query($owner: String!, $repo: String!, $pr: Int!) {
      repository(owner: $owner, name: $repo) {
        pullRequest(number: $pr) {
          reviewThreads(first: 100) {
            nodes {
              isResolved
              comments(first: 100) { nodes { databaseId } }
            }
          }
        }
      }
    }
  ' \
  --jq '[.data.repository.pullRequest.reviewThreads.nodes[]
         | select(.isResolved)
         | .comments.nodes[].databaseId
         | select(. != null)]')

The first: 100 limits cover typical PRs. For very large PRs (>100 threads or >100 comments in a single thread), thread/comment-level pagination would need to be added; for now, that's a known limitation.

Write jq filters to temp files using heredocs with single-quoted delimiters (prevents shell escaping issues with !=, regex patterns, and angle brackets):

Issue comments (summary/walkthrough posts):

cat > /tmp/issue_comments.jq << 'JQEOF'
def clean_body:
  gsub("<!-- suggestion_start -->.*?<!-- suggestion_end -->"; ""; "s")
  | gsub("<!--.*?-->"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*🧩 Analysis chain[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*🤖 Prompt for AI Agents[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*📝 Committable suggestion[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>Past reviewee.*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>Recent review details[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*Tips\\b.*?</details>"; ""; "s")
  | gsub("\\n?\\n---\\n+(?=\\s*(?:✨\\s*Finishing Touches|🪧\\s*Tips?|🤖\\s*Generated|<sub>|_Generated by|Generated by Claude|Generated with))[\\s\\S]*$"; ""; "s")
  | gsub("^\\s+|\\s+$"; "")
  | if length > 4000 then .[:4000] + "\n\n[comment truncated]" else . end
;
[(add // []) | .[] | select(
  .user.login != $pr_author and
  .user.login != $current_user
)] |
map({id, user: .user.login, body: (.body | clean_body), created_at})
JQEOF

gh api --paginate "repos/$OWNER/$REPO/issues/$PR_NUMBER/comments" | \
  jq -s --arg pr_author  --arg current_user  \
  -f /tmp/issue_comments.jq

Review comments (line-specific):

cat > /tmp/review_comments.jq << 'JQEOF'
def clean_body:
  gsub("<!-- suggestion_start -->.*?<!-- suggestion_end -->"; ""; "s")
  | gsub("<!--.*?-->"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*🧩 Analysis chain[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*🤖 Prompt for AI Agents[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*📝 Committable suggestion[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>Past reviewee.*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>Recent review details[\\s\\S]*?</details>"; ""; "s")
  | gsub("<details>\\s*<summary>\\s*Tips\\b.*?</details>"; ""; "s")
  | gsub("\\n?\\n---\\n+(?=\\s*(?:✨\\s*Finishing Touches|🪧\\s*Tips?|🤖\\s*Generated|<sub>|_Generated by|Generated by Claude|Generated with))[\\s\\S]*$"; ""; "s")
  | gsub("^\\s+|\\s+$"; "")
  | if length > 4000 then .[:4000] + "\n\n[comment truncated]" else . end
;
[(add // []) | .[] | select(
  .user.login != $pr_author and
  .user.login != $current_user and
  (.id as $comment_id | ($resolved_ids | index($comment_id) | not))
)] |
map({
  id,
  user: .user.login,
  path,
  line_display: (
    .line as  | .start_line as  |
      and  !=   
      end
  ),
  body: (.body | clean_body),
  created_at
})
JQEOF

gh api --paginate  | \
  jq -s --arg pr_author  --arg current_user  \
  --argjson resolved_ids  \
  -f /tmp/review_comments.jq

If --include-author is set, omit the --arg pr_author parameter and the .user.login != $pr_author condition from both jq filter files. Keep the $current_user exclusion either way. The $resolved_ids filter naturally becomes a no-op when RESOLVED_IDS='[]' (the --include-resolved path), so leave it in place.

4. Format Feedback Document

Noise stripping — handled by the clean_body jq function in Step 3. Order matters: <!-- suggestion_start -->...<!-- suggestion_end --> blocks are removed first, then remaining HTML comments, then known-noise <details> blocks (Analysis chain, Prompt for AI Agents, Committable suggestion, Past reviewee, Recent review details, Tips), and finally the --- footer boilerplate. The <details> blocks must be stripped before the --- footer pattern because bot analysis chains contain --- separators that would otherwise truncate the actual finding. Substantive <details> blocks (e.g. "Suggested fix", "Proposed fix") are preserved. Comments exceeding 4000 chars after stripping are truncated with a [comment truncated] marker.

Group by reviewer — organize the formatted output by reviewer username:

# PR #$PR_NUMBER Review Feedback

## Reviewer: coderabbitai[bot]

### Summary Comments
[Issue comments from this reviewer, each separated by ---]

### Line-Specific Comments
[Review comments from this reviewer, each formatted as:]

**File: `path/to/file.ts:42`**
[cleaned comment body]

---

## Reviewer: another-reviewer

### Summary Comments
...

### Line-Specific Comments
...

If no comments found from any reviewer, output: "No review comments found on this PR (excluding PR author, current user, and resolved threads)."

5. Evaluate with receive-feedback

Load the receive-feedback skill.

Then process the formatted feedback document:

1. Parse each actionable item from the formatted document
2. Process each item through verify → evaluate → execute
3. Produce structured response summary

Example

# Fetch unresolved reviewer comments on current branch's PR (default)
fetch-pr-feedback

# Fetch from a specific PR
fetch-pr-feedback --pr 123

# Include PR author's own comments
fetch-pr-feedback --include-author

# Include line-specific comments from resolved review threads
fetch-pr-feedback --include-resolved

# Combined
fetch-pr-feedback --pr 456 --include-author --include-resolved