My Council

my-council

Multi-agent parallel review → Council synthesis → Prioritized fix list.

⚡ Quick Start

When the user asks for a review or changes cross a threshold:

# Check if review threshold is met
find src -type f -newer .git/HEAD -name '*.ts' -o -name '*.tsx' | wc -l
# If >5 files or ~100+ lines changed, trigger review council

Activate the council:

Spawning 5 parallel reviewers → synthesizing findings → presenting prioritized list

After user picks fixes, execute all approved items immediately.

The Pattern

1. Spawn Parallel Reviewers

Spawn 5 (or more) subagents in parallel, each with a different focus dimension:

Each reviewer:

• Receives the same task description + file list
• Focuses only on their dimension
• Writes a structured report to /tmp/review-{dimension}.md

2. Synthesize Reports

After all agents complete:

1. Read all 5 report files
2. Categorize every finding into:
   • Critical — fix now (security, data loss, crash)
   • Warning — fix soon (bug risk, UX barrier, type hole)
   • Suggestion — consider (style, refactor, enhancement)
3. De-duplicate findings that appear in multiple dimensions
4. Present as numbered list with severity

3. Execute Fixes

User says "fix all" or picks a subset:

• Do NOT ask for confirmation on each individual fix
• Do NOT ask "which one next?" — just execute the list
• Fix in order: Critical → Warning → Suggestion
• Run tests after every batch of related edits
• Commit after all fixes pass

Reviewer Prompt Template

Each reviewer receives this prompt structure:

You are a {DIMENSION} reviewer for this project.

Your task: Review the following files and write a report to /tmp/review-{DIMENSION}.md.

Focus ONLY on {DIMENSION} concerns. Do NOT review other dimensions.

Review scope: {FILE_LIST}

Report format:
## Critical (fix now)
- [File:line] — Issue description
- Why: Impact + evidence
- Fix: Before/After code

## Warning (fix soon)
...

## Suggestion (consider)
...

Research Dimension

The Research reviewer is unique — it does not read the code for bugs. Instead, it:

• Researches modern best practices for technologies used in the project
• Looks up API documentation for unfamiliar libraries or patterns
• Checks for known security advisories on dependencies
• Finds alternative approaches that might be simpler or more robust
• Validates that the project's approach aligns with current community standards

Research reviewer prompt:

You are a Research reviewer for this project.

Your task: Research the technologies and patterns used in the following files.
Write a report to /tmp/review-research.md.

Use @skills/my-web-search-kagi to search the web for current best practices,
security advisories, and modern alternatives.

Focus ONLY on:
- Are we using the current best practices for each technology?
- Are there newer, simpler, or more robust alternatives?
- Are there any known security issues or deprecations in our dependencies?
- Are we missing common patterns that similar projects use?

Do NOT review code for bugs. Focus on external research and knowledge.

Research scope: {FILE_LIST} + package.json / Cargo.toml / pyproject.toml

Orchestrator Prompt

After reviewers complete, read all reports and produce:

## Code Review Council: {PROJECT_NAME}

### Critical (fix now)
1. **[File:line]** — Issue (from {reviewer})
2. ...

### Warning (fix soon)
1. ...

### Suggestion (consider)
1. ...

### Estimated effort: {X} hours

Auto-Review Threshold

Do not wait for the user to ask. Auto-invoke review before commit when:

If threshold is met:

1. Run git diff --stat (or equivalent) to confirm scope
2. Invoke @skills/my-council (self-call)
3. Fix Critical findings before committing
4. Note Warnings for follow-up
5. Commit with clean code

Why auto-review: Review is almost always user-initiated. Agents ship multi-file changes without review because there is no automatic trigger. This threshold removes the activation barrier.

Execution Rules

• Fix Critical first — never leave a Critical finding unaddressed
• Batch related edits — fix all items in one file before moving to next
• Run tests after each batch — unit tests + typecheck + e2e
• Commit as a single batch — one commit per dimension batch, or one commit for all
• Remove dead code immediately — if reviewers flag unused code, delete it
• Add missing tests — if reviewers flag uncovered code, write tests

Test Coverage Rule

If the project has tests → new code MUST have tests.

Check first: find . -name '*.test.*' -o -name '*.spec.*' -o -name 'test_*' | head -5

Coverage goal: Meaningful coverage without bloat.

• Test the contract: inputs → outputs, success + failure paths
• Test edge cases: empty, null, extreme values, race conditions
• Don't test: language built-ins, trivial getters, implementation details
• One comprehensive test > three shallow tests

Build vs Buy Analysis

Question: Is this code reinventing something a mature library already solves?

When to flag:

• Common utility functions (date formatting, validation, HTTP retries)
• Security-sensitive operations (auth, crypto, parsing untrusted input)
• Complex state management with async operations
• Features with known edge cases (timezones, unicode, floating point)
• Data transformation with schema requirements

Evaluation criteria:

Red Flags (always catch)

• Unvalidated user input → database
• Secrets hardcoded or unvalidated env
• Async without error handling
• Blocking operations in request handlers
• Race conditions in concurrent code
• Memory leaks in long-running processes
• Sensitive data in logs/errors
• New code without tests (when project has tests)

Principles

Research Sources (priority order)

1. Context7 library docs — Most authoritative for specific libraries
2. Official docs — Framework/language best practices
3. Codebase patterns — grep for consistency with existing code
4. Community guides — With clear attribution

Fix Format

### Before (lines X-Y)
```ts
// problematic

After

// fixed

Why

• [specific benefit]
• [evidence-backed reasoning]

## Common Mistakes

❌ **Wrong:** Reviewing code without researching current best practices first
✅ **Right:** Always consult Context7 or official docs before making recommendations

❌ **Wrong:** Giving vague feedback like "this could be better"
✅ **Right:** Provide specific before/after code with evidence-backed reasoning

❌ **Wrong:** Catching style issues but missing security red flags
✅ **Right:** Check Red Flags first: input validation, secrets, async errors, race conditions

❌ **Wrong:** Suggesting a rewrite when a small fix suffices
✅ **Right:** Prefer minimal, targeted changes. Only refactor when clearly justified

❌ **Wrong:** Reviewing 50+ files as a single monolithic review
✅ **Right:** Use @skills/my-council to spawn parallel reviewers across dimensions

## When Context Is Unclear

Ask, don't guess:
- "What's intended behavior when [edge case]?"
- "Is this performance-sensitive?"
- "What versions are you targeting?"

Then deliver targeted recommendations.

## When to Skip

- Single file changed (< 20 lines)
- Only documentation or config changes
- Only test file changes
- User explicitly says "skip review"

## When to Use Research Standalone

Sometimes the user wants research without a full code review:
- "Should we use X or Y library?"
- "What's the current best practice for Z?"
- "Are there any security issues with our dependencies?"

In these cases, spawn **only the Research reviewer** and present findings directly. No council synthesis needed.

## Post-Review

After fixes are applied:
1. Run full test suite
2. Run typecheck
3. Run build
4. Verify coverage report
5. Commit with message: `review: fix {N} {dimension} findings`
6. Push

If the Research reviewer found actionable improvements (e.g., "use library X instead of Y"), create a follow-up task or implement immediately if the change is small.

## Related Skills

- **@skills/my-tech-stack** — For tool recommendations when reviewers suggest new dependencies
- **@skills/my-workflow** — For commit discipline and worktree naming during review execution
- **@skills/my-vcs-hygiene** — For committing and pushing after review fixes

## Versioning

- **Last updated:** 2026-05-28
- **Version:** 1.0
- **Update notes:** Initial council pattern based on multi-dimension parallel review experience