address-pr-review-locally

Address PR Review Locally

End-to-end workflow for reading a GitHub pull request review, triaging its inline comments, implementing fixes locally, and verifying correctness — without switching to the browser.

When to Use

• A Copilot, human, or bot reviewer has left inline comments on a PR
• You want to address review feedback without leaving the editor
• A review is in COMMENTED or CHANGES_REQUESTED state with unresolved threads

Procedure

1. Fetch the Review

Use mcp_github_pull_request_read twice in parallel:

method: get_reviews       → list all review objects; note state + review ID
method: get_review_comments → full list of review threads with is_resolved/is_outdated

Focus on threads where is_resolved: false and is_outdated: false. Skip is_outdated: true unless the underlying concern still applies to the current HEAD.

2. Triage Comments

Group unresolved threads into three buckets before writing any code:

Priority	Trigger	Action
P1 — Correctness/Safety	Bug, data loss, security, destructive overwrite without guard	Fix first
P2 — Correctness/Accuracy	Misleading docs, wrong test label, logic gap	Fix second
P3 — Style/Improvement	Refactoring suggestion, docblock wording, minor naming	Fix last or defer

If a comment includes a suggestion block (GitHub's diff suggestion format), use it as the basis for the fix — but verify it compiles and passes tests before accepting verbatim.

3. Apply Fixes Locally

For each P1 → P2 → P3 comment:

1. Open the file at the line referenced in comment.path + comment.line
2. Read surrounding context (±20 lines) to understand the full function/block
3. Apply the fix — prefer the smallest change that addresses the reviewer's concern
4. Note the discussion thread ID (comment.html_url) for later reply

Common fix patterns from this project:

• Misleading doc comment → update the comment only, not the implementation
• YAML frontmatter without quoting → wrap the interpolated value in YAML-safe quotes
• Hardcoded platform path (e.g. "/") → use path.parse(current).root instead
• Overwrite-without-merge → read existing JSON, deep-merge, then write back
• Guard-bypass with --yes flag → keep the existing guard, require --force for destructive overwrite
• Non-monotonic progress counter → use a shared atomic counter incremented after each task resolves
• TTY check via exception → check process.stdout.isTTY or spinner.isEnabled explicitly
• Docs referencing files not in the npm package → add a [!NOTE] block explaining availability