Github Pr Review

GitHub PR Review Mining

Harvest code review comments from merged GitHub pull requests, classify them for playbook relevance, and feed actionable findings into the packmind-update-playbook workflow.

Workflow: Phase 0 → 1 → 2 → 3 → 4 → 5. Follow every phase in order.

Prerequisite: The GitHub MCP server must be available. Load references/github-mcp-tools.md for tool signatures and parameters.

Phase 0: Collect Parameters

Gather the required inputs before fetching any data.

Execution Mode Detection

Determine the execution mode before collecting any parameters:

1. Check CI env var: run echo $CI. If the value is true, set non_interactive = true.
2. Check skill arguments: if the skill was invoked with --non-interactive, set non_interactive = true.
3. Parse optional arguments:
   • --days N — override the default time period (default: 7 days)
   • --repo owner/repo — override git remote inference

If neither condition is met, non_interactive = false (interactive mode — original behavior).

Collect Parameters

1. Infer owner and repo from the git remote:

   git remote get-url origin
   

   Parse owner/repo from the URL.

   • Interactive: If the remote is unavailable or ambiguous, ask the user.
   • Non-interactive: If --repo was provided, use it. Otherwise abort with error: "Could not infer owner/repo from git remote. Use --repo owner/repo."

2. Gather time_period:

   • Interactive: Ask the user how far back to look (default: 7 days, max: 90 days).
   • Non-interactive: Use --days value if provided, otherwise default to 7 days silently.

   Compute the cutoff date as today - time_period in YYYY-MM-DD format.

3. Confirm before proceeding:

   • Interactive: Display and BLOCK until the user confirms:

     "Mining PR reviews for owner/repo merged in the last N days (since YYYY-MM-DD). Proceed?"

   • Non-interactive: Log the parameters and proceed automatically:

     "Non-interactive mode: mining PR reviews for owner/repo merged in the last N days (since YYYY-MM-DD)."

Phase 1: Fetch Merged PRs

Use mcp__github__search_pull_requests with query:

is:pr is:merged merged:>YYYY-MM-DD repo:owner/repo

• Paginate at 30 per page, collect all PR numbers, titles, and URLs.
• If total_count is 0, inform the user and stop.
• Display progress: "Found N merged PRs to analyze."

Phase 2: Fetch Review Comments

For each PR, call mcp__github__pull_request_read with method: "get_review_comments".

Filtering

Apply these filters to discard noise:

Low-value filtering — discard comments that:

• Are shorter than 15 characters
• Match low-value patterns: consists only of "LGTM", "+1", thumbs-up emoji variants, "nit", or "thanks"

Thread deduplication — when a comment has in_reply_to_id, keep only the root comment unless the reply adds substantive new feedback.

Display progress: "Collected N review comments from M PRs (after filtering)."

If no comments remain after filtering, inform the user and stop.

Phase 3: Classify Comments

Categorize each remaining comment by playbook relevance:

High relevance (keep)

• convention — Enforces or suggests a naming, structure, or organizational pattern
• best-practice — Recommends a better approach for correctness, performance, or maintainability
• recurring-feedback — Same feedback appears across 2+ different PRs
• bug-prevention — Points out a pattern that could lead to bugs

Low relevance (discard)

• style — Pure formatting preference
• clarification — Asks a question without suggesting a change
• nitpick — Minor, one-off preference

Retain only High-relevance comments.

Recurring Theme Detection

Group retained comments by semantic similarity. A recurring theme is 2+ comments across different PRs that address the same underlying pattern. For each theme:

• Assign a short descriptive label
• List all contributing comments
• Note the occurrence count

Phase 4: Build Findings Report

Choose the output path based on execution mode:

• Interactive: .claude/tmp/pr-review-findings.md
• Non-interactive: .claude/reports/pr-review-findings-YYYY-MM-DD.md (using today's date, timestamped for CI artifact upload)

Create the target directory if needed.

Report Structure

# PR Review Findings Report

**Repository**: owner/repo
**Period**: YYYY-MM-DD to YYYY-MM-DD
**PRs analyzed**: N
**Comments collected**: N (after filtering)
**Actionable findings**: N

---

## Recurring Themes

### Theme: <label>
**Occurrences**: N across N PRs

| PR | File | Author | Comment |
|----|------|--------|---------|
| [#123](url) | path/to/file.ts | @author | Summary of comment |
| [#456](url) | path/to/other.ts | @author | Summary of comment |

**Suggested playbook action**: <Create standard | Update standard X | Create skill | ...>
**Rationale**: <Why this theme warrants a playbook change>

---

## Individual Findings

### Finding: <short description>
- **PR**: [#123](url)
- **File**: path/to/file.ts
- **Diff hunk**: (include relevant diff context)
- **Author**: @author
- **Comment**: Full comment text
- **Category**: convention | best-practice | bug-prevention
- **Suggested playbook action**: <action>

URL Requirements

Every PR reference in the report MUST include the full, navigable GitHub URL (e.g., https://github.com/owner/repo/pull/123). This is critical because packmind-update-playbook extracts source URLs from the findings report for its packmind-cli diff --submit -m source attribution. The expected attribution format is:

<topic>: <summary> (source: PR #N https://github.com/owner/repo/pull/N, PR #M https://github.com/owner/repo/pull/M)

Use the html_url field returned by mcp__github__search_pull_requests for each PR. Never use placeholder (url) — always substitute the real URL.

Phase 5: Present and Hand Off

Interactive mode

1. Display a summary to the user:

   • Number of recurring themes found
   • Number of individual findings
   • List of suggested playbook actions

2. Ask the user:

   "Found N recurring themes and M individual findings. Review the full report at .claude/tmp/pr-review-findings.md. Proceed to update the playbook with these findings?"

3. On confirm: Invoke the packmind-update-playbook skill with the full report content as the intent. This maps to Case B (explicit intent) of that skill's Phase 0. Frame the intent as:

   "Update the Packmind playbook based on the following PR review findings report: "

4. On decline: Inform the user the report is available at .claude/tmp/pr-review-findings.md for manual review.

Non-interactive mode

1. Display a summary (same as interactive: themes count, findings count, suggested actions).

2. Log the report output path:

   "Report written to .claude/reports/pr-review-findings-YYYY-MM-DD.md."

3. Invoke packmind-update-playbook with the full report content as the intent, passing along non-interactive mode. Frame the intent as:

   "Update the Packmind playbook based on the following PR review findings report: "

   The packmind-update-playbook skill will auto-approve all changes and delegate creation of new artifacts to the corresponding creation skills (packmind-create-standard, packmind-create-command, packmind-create-skill), all running in non-interactive mode.

Resources

references/

• github-mcp-tools.md — GitHub MCP tool signatures and parameters used by this skill