forgejo-mcp/openspec/changes/forgejo-code-review-skill/tasks.md

Tasks

1. Command File Scaffold

• 1.1 Create .claude/commands/code-review.md with frontmatter (description, argument-hint, model, allowed-tools) matching existing command patterns (/commit, /catchup)
• 1.2 Write argument parsing section: extract optional PR number and --comment flag from $ARGUMENTS

2. PR Resolution and Pre-screening

• 2.1 Implement PR auto-detection: git branch --show-current → list_repo_pull_requests lookup → prompt if ambiguous
• 2.2 Implement explicit PR number path: parse number from arguments, call get_pull_request_by_index
• 2.3 Add pre-screening gate: skip draft, closed/merged, trivial (<5 lines), and warn on oversized (>500 lines) PRs

3. Context Gathering

• 3.1 Read CLAUDE.md and AGENTS.md from the repository via get_file_content (handle missing files gracefully)
• 3.2 Retrieve PR metadata via get_pull_request_by_index and read each changed file's content via get_file_content

4. Change Summarization

• 4.1 Write the Haiku-model summarization prompt: produce structured summary with file paths, change types (added/modified/deleted), and line number ranges per hunk

5. Parallel Review Agents

• 5.1 Write compliance agent prompt (Sonnet): check changed code against CLAUDE.md/AGENTS.md rules, output findings as JSON with {file, line, rule, description}
• 5.2 Write bug hunter agent prompt (Opus): analyze changed code for bugs, error handling gaps, edge cases, output findings as JSON with {file, line, description}
• 5.3 Write logic/security agent prompt (Opus): analyze changed code for logic errors, security issues, race conditions, output findings as JSON with {file, line, severity, description}
• 5.4 Wire all three agents as parallel Task tool calls with run_in_background: true

6. Confidence Scoring and Filtering

• 6.1 Write the scoring prompt: independently score each finding 0-100, auto-zero pre-existing issues and linter-catchable problems
• 6.2 Add threshold filter: remove findings with confidence < 80

7. Output Formatting

• 7.1 Write terminal output format: group findings by file, show path, line, severity, agent source, confidence, and description
• 7.2 Write clean-result output: "No significant issues found" with filtered count

8. Forgejo Posting

• 8.1 Implement --comment path: build create_pull_review call with state COMMENT, summary body, and inline comments array (path, body, new_position)
• 8.2 Implement clean-review posting: call create_pull_review with "No significant issues found" body when no findings pass threshold

9. Integration and Testing

• 9.1 End-to-end manual test: run /code-review against a real PR in terminal-only mode
• 9.2 End-to-end manual test: run /code-review --comment against a test PR and verify inline comments appear on Forgejo