shipwright-cli 3.2.0 → 3.3.0

package/scripts/skills/generated/pattern-matching-similarity-scoring.md

@@ -0,0 +1,195 @@
+# Pattern Matching & Failure Prevention Scoring
+
+## Overview
+
+This skill guides design and implementation of pattern-based proactive failure prevention: matching incoming issues against captured failure patterns, scoring similarity, injecting relevant context, and measuring whether patterns actually prevent repeat failures.
+
+## Similarity Scoring Algorithm (0-100 scale)
+
+For each incoming issue, compute a composite similarity score against each known failure pattern:
+
+### Component 1: Title Similarity (40% weight)
+- Fuzzy string matching using token overlap or Levenshtein distance normalized by string length
+- Captures semantic closeness of the problem description
+- Example: "API timeout on user endpoint" vs "Timeout in auth middleware" → ~0.7 similarity → 28 points
+
+### Component 2: File Overlap (35% weight)
+- Compare changed files in original failure vs incoming issue
+- Score = (overlapping_files / max(original_files, incoming_files)) * 100
+- Files touching the same components are more likely to have similar root causes
+- Example: Both touched `scripts/sw-daemon.sh` and `scripts/lib/daemon-dispatch.sh` → 35 points if full overlap
+
+### Component 3: Error Signature Match (25% weight)
+- Check if error message substrings or error codes appear in both
+- Extract from error-summary.json or stack trace (structured format preferred)
+- Example: Both contain "pipefail" or "ENOENT" → 25 points
+
+**Formula: score = (title_score * 0.4) + (file_score * 0.35) + (error_score * 0.25)**
+
+## Injection Thresholds
+
+- **Below 60**: Pattern not relevant, no injection
+- **60-80**: Inject with confidence tag ("medium confidence match")
+- **80-100**: Inject with high confidence ("strong pattern match")
+- **Configurable threshold**: daemon-config.json `memory_pattern_matching.similarity_threshold`
+
+## Proactive Injection Strategy
+
+When score > threshold (at pipeline spawn time, before plan stage):
+
+1. Extract relevant context from memory pattern:
+   - Root cause description
+   - Applied fix(es)
+   - Environment/version context if present
+   - What worked vs what didn't
+
+2. Inject into pipeline prompt:
+   ```
+   Similar pattern found (confidence: 85%): Issue #123 "API timeout on user endpoint"
+   Root cause: Unbounded goroutine creation in event loop
+   Applied fix: Add semaphore to limit concurrent handlers
+   Files affected: scripts/sw-daemon.sh, internal/loop.go
+   ```
+
+3. Tag injection metadata:
+   - `pattern_id`: which pattern was injected
+   - `injection_score`: 0-100 confidence
+   - `timestamp`: when injected
+   - `source_issue_id`: which failure this pattern came from
+
+4. **Important**: No coercion. Agent can ignore injected pattern if context doesn't match.
+
+## Outcome Tracking
+
+After pipeline completes, record:
+
+```json
+{
+  "pattern_injected": true,
+  "pattern_id": "mem_abc123",
+  "injection_score": 85,
+  "incoming_issue_id": "456",
+  "source_issue_id": "123",
+  "failure_occurred": false,
+  "failure_type": null,
+  "expected_failure_type": "timeout",
+  "failure_type_matched": null,
+  "failure_prevented": true,
+  "confidence_in_prevention": 0.7,
+  "notes": "Applied semaphore fix suggested by pattern; no timeout occurred."
+}
+```
+
+**Caveat**: `failure_prevented` is **inference**, not proof.
+- True positive: pattern injected, agent applied the fix, no failure occurred, failure type matched expected type
+- Could be false positive: pattern coincidentally matched, but agent's own skill prevented failure
+- Always include confidence score; never claim 100% causation
+
+## Effectiveness Metrics (Dashboard)
+
+**Per-pattern metrics:**
+- **Success Rate**: (times_injected AND failure_prevented) / times_injected
+- **Usage Frequency**: times_injected in last 30 days
+- **Confidence Distribution**: histogram of injection_scores
+- **False Positive Rate**: (times_injected AND failure_occurred) / times_injected
+
+**Aggregate metrics:**
+- **Overall Memory Injection ROI**: sum(successful_injections) / sum(total_injections)
+- **Patterns Needing Refinement**: patterns with > 30% usage but < 40% success rate (candidates for root cause re-analysis)
+- **Trending**: success rate on 7-day and 30-day windows; alert if trending down
+- **Pattern Lifecycle**: which patterns are becoming obsolete (< 1% usage in 90 days)?
+
+## Integration Points
+
+1. **sw-memory.sh**
+   - Call `memory_get_patterns()` to retrieve all patterns with timestamps, failure_type, root_cause
+   - Call `memory_add_outcome_tracking()` to record success/failure outcome
+
+2. **sw-intelligence.sh**
+   - Integrate pattern scoring into `intake` stage
+   - Score issue at pipeline spawn time (before plan stage)
+   - Return top 3 matching patterns sorted by score
+
+3. **Pipeline prompt composition**
+   - Add `memory_pattern_context` section to prompt if score > threshold
+   - Include confidence score so agent is aware this is a suggestion, not a fact
+
+4. **Pipeline state tracking**
+   - Add `memory_patterns` section to pipeline-state.md with injected pattern details
+   - Track injection_score, outcome_recorded=true/false
+
+5. **Loop iteration context**
+   - If issue re-runs in build loop, re-score with new error context
+   - Emerging error signatures may match different patterns on retry
+
+## Testing Strategy
+
+**Unit tests:**
+- Similarity scoring against known issue pairs with ground truth
+- Threshold boundary behavior (59, 60, 61)
+- Weight adjustment: verify 0.4 + 0.35 + 0.25 = 1.0
+
+**Edge cases:**
+- Empty pattern database → score undefined, no injection
+- Identical issues with different outcomes → verify both outcomes tracked
+- Pattern with malformed error_signature → graceful fallback
+- Very high similarity (> 95%) → verify no over-confidence
+
+**Integration tests:**
+- Inject pattern, verify it appears in pipeline prompt
+- Run build, record outcome, verify outcome_tracking fires
+- Query dashboard, verify metrics match recorded outcomes
+
+**Effectiveness validation:**
+- Mock a failure type, populate patterns database, run scoring
+- Verify pattern was injected at expected score
+- Mock outcome (failure_prevented=true/false), verify metrics compute correctly
+
+## Configuration Example
+
+```json
+{
+  "memory_pattern_matching": {
+    "enabled": true,
+    "similarity_threshold": 60,
+    "weights": {
+      "title_similarity": 0.4,
+      "file_overlap": 0.35,
+      "error_signature": 0.25
+    },
+    "confidence_tiers": {
+      "high": 80,
+      "medium": 60,
+      "low": 30
+    },
+    "max_patterns_to_inject": 3,
+    "metrics_retention_days": 90,
+    "anomaly_detection_enabled": true
+  }
+}
+```
+
+## Risk Mitigation
+
+**Risk 1: False positive injection**
+- Monitoring: alert if false_positive_rate > 15%
+- Mitigation: lower threshold, disable for specific pattern types, or retire pattern
+
+**Risk 2: Outcome attribution confusion**
+- Always show confidence_in_prevention as a float (0.0-1.0), never binary
+- Document that "prevented" is inferred, not measured
+- Quarterly review of patterns with low confidence
+
+**Risk 3: Circular reasoning**
+- Patterns must capture ROOT CAUSE, not just "solution"
+- Red flag: if pattern root_cause is identical to another pattern → merge
+- Quarterly audit of pattern root_cause quality
+
+**Risk 4: Performance at scale**
+- Scoring 100+ patterns should be < 500ms
+- Use cached similarity scores if possible
+- Parallel scoring if pattern database grows beyond 500
+
+**Risk 5: Stale patterns**
+- Patterns from > 180 days ago with < 5 uses → mark for review
+- Dashboard should surface "patterns never injected" for root cause analysis

package/scripts/skills/generated/test-parallelization-detection.md

@@ -0,0 +1,65 @@
+## Test Parallelization Detection & Coordination
+
+### Problem
+Test parallelization is dangerous: undetected shared state (temp files, global state, database connections) causes race conditions and flaky failures. This skill provides a systematic approach to detect parallelizable test suites and coordinate their execution safely.
+
+### Shared State Detection Heuristics
+
+**Static Analysis (file scanning):**
+- Scan test file imports for singleton patterns (db connections, file handles, global state modules)
+- Detect hardcoded file paths (temp dirs) and network ports — tests using fixed resources conflict
+- Check for `beforeAll`/`afterAll` hooks that modify global state
+- Identify test files importing shared fixtures/setup modules
+
+**Dynamic Analysis (test execution):**
+- Run test suite with `--detectOpenHandles` (Node.js) or equivalent to catch file/port leaks
+- Track temp directory usage per test file — any overlap = unsafe to parallelize
+- Monitor for test isolation violations (tests passing in isolation but failing when run together)
+
+**Safety Levels:**
+- **Green (parallelizable)**: No shared state detected, no fixture conflicts, passes isolation tests
+- **Yellow (conditional)**: Shared fixtures but isolated datasets, parallel execution with coordination (e.g., separate DB schemas)
+- **Red (sequential)**: Database transaction rollback, process spawning, hardware resource contention — must run serially
+
+### Affected-Test Detection via Git Diff
+
+**Module Dependency Tracking:**
+1. Build module-to-test mapping (which tests exercise which modules)
+2. On each commit, run `git diff --name-only HEAD~1` to identify changed modules
+3. Find all tests that import/test those modules
+4. Prioritize affected tests first in execution order (fail-fast on functionality regression)
+5. Cache mapping per commit to avoid re-scanning on retries
+
+**False Negatives to Handle:**
+- Integration tests that cross module boundaries (require broader analysis)
+- Tests that exercise shared utilities or base classes (conservative: mark as affected if any parent module changed)
+- Dynamic imports and string-based test discovery (fallback: scan test code for patterns)
+
+### Parallel Execution Coordination
+
+**Scheduler:**
+- Detect CPU core count, default to `cores - 1` (reserve 1 for OS)
+- Group parallelizable tests into batches, run batches in parallel
+- Within each batch, respect test file order (some test runners depend on execution order)
+- Run non-parallelizable (red) tests serially, either before or after parallel batches (configurable)
+
+**Fast-Fail Policy:**
+- Critical failures: assertion errors, uncaught exceptions → abort immediately
+- Flaky failures: timeout, process exit, known-flaky markers → retry up to N times before aborting
+- Aggregate results across parallel workers before reporting
+- Time tracking: measure wall-clock time for each batch, report parallelization efficiency (theoretical vs actual speedup)
+
+### Dashboard Integration
+
+- Display parallel execution summary: N tests in M workers, X% speedup
+- Visualize test dependency graph (which tests block which)
+- Alert on shared-state violations (test passed alone, failed in parallel)
+- Trend: parallelization efficiency over time (detect regressions where new tests add serial bottlenecks)
+
+### Key Decisions for This Issue
+
+1. **Minimum Parallelization Threshold**: What's the smallest safe granularity? (per file, per suite, per test?)
+2. **Flaky Detection**: How many retries before marking as critical failure? (recommend 3)
+3. **Shared-State Confidence**: Are heuristics sufficient, or require explicit opt-in per test file?
+4. **Fast-Fail Behavior**: Abort on first critical failure globally, or let all workers finish for faster feedback iteration?
+5. **Fallback**: If parallelization detection is uncertain, run serial — safety over speed.

package/scripts/skills/observability.md

@@ -0,0 +1,79 @@
+## Observability: Watch the Deploy Like a Hawk
+
+Post-deploy monitoring catches what tests miss. Real traffic reveals real problems.
+
+### What to Monitor (by Priority)
+
+**P0 — Immediate (first 5 minutes):**
+- Error rate: any increase over baseline?
+- Health check: still returning 200?
+- Latency: p50/p95/p99 within normal range?
+- Memory/CPU: any sudden spikes?
+
+**P1 — Short-term (5-30 minutes):**
+- Business metrics: are users completing key flows?
+- Queue depths: are background jobs processing normally?
+- Connection pools: any exhaustion or leak patterns?
+- Disk usage: any unexpected growth?
+
+**P2 — Medium-term (1-24 hours):**
+- Memory trends: gradual leak over time?
+- Error rate trends: slowly increasing?
+- User-reported issues: any new support tickets?
+- Performance degradation under sustained load?
+
+### Anomaly Detection Patterns
+- **Spike detection**: >2x baseline error rate in any 1-minute window
+- **Trend detection**: steadily increasing error rate over 5-minute window
+- **Absence detection**: expected periodic events stop occurring
+- **Latency shift**: p95 latency increases >50% from baseline
+
+### Log Analysis
+- Search for new ERROR/FATAL/PANIC entries not present before deploy
+- Check for stack traces — they indicate unhandled exceptions
+- Look for retry storms — repeated failed attempts at the same operation
+- Monitor for resource exhaustion messages (OOM, connection refused, disk full)
+
+### Auto-Rollback Triggers
+Automatically rollback if ANY of these occur:
+- Health check fails 3 consecutive times
+- Error rate exceeds threshold for 2+ minutes
+- Critical service dependency becomes unreachable
+- Memory usage exceeds 90% of limit
+
+### Monitoring by Issue Type
+
+**Frontend changes:**
+- JavaScript error rates in browser (if client-side monitoring exists)
+- Asset load failures (404s on new bundles)
+- Core Web Vitals regression (LCP, FID, CLS)
+
+**API changes:**
+- Response status code distribution (2xx vs 4xx vs 5xx)
+- Request throughput — drops indicate client-side breakage
+- Authentication failures — spikes indicate auth regression
+
+**Database changes:**
+- Query latency per endpoint
+- Connection pool utilization
+- Slow query log entries
+- Replication lag (if applicable)
+
+### Incident Escalation
+If monitoring detects issues:
+1. Execute rollback (if auto-rollback enabled)
+2. Create incident issue with monitoring data
+3. Attach relevant logs and metrics
+4. Tag the original issue with `incident` label
+5. Do NOT silence alerts — let them fire
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **Monitoring Checklist**: P0/P1/P2 metrics to watch (error rate, latency, memory, health checks) with specific thresholds
+2. **Anomaly Detection Triggers**: Explicit conditions that trigger alerts (spike detection >2x, trend detection over 5min, absence detection, latency shift >50%)
+3. **Log Analysis**: Search strategy for new ERROR/FATAL entries, stack traces, retry storms, resource exhaustion patterns
+4. **Auto-Rollback Decision Criteria**: Conditions that trigger automatic rollback (health check failures, error rate threshold, critical dependency unreachable, memory exhaustion)
+
+If any section is not applicable, explicitly state why it's skipped.

package/scripts/skills/performance.md

@@ -0,0 +1,48 @@
+## Performance Expertise
+
+Apply these optimization patterns:
+
+### Profiling First
+- Measure before optimizing — identify the actual bottleneck
+- Use profiling tools appropriate to the language/runtime
+- Focus on the critical path — optimize what users experience
+
+### Caching Strategy
+- Cache expensive computations and repeated queries
+- Set appropriate TTLs — stale data vs freshness trade-off
+- Invalidate caches on write operations
+- Use cache layers: in-memory (L1) → distributed (L2) → database (L3)
+
+### Database Performance
+- Add indexes for frequently queried columns (check EXPLAIN plans)
+- Avoid N+1 queries — use batch loading or JOINs
+- Use connection pooling
+- Consider read replicas for read-heavy workloads
+
+### Algorithm Complexity
+- Prefer O(n log n) over O(n²) for sorting/searching
+- Use appropriate data structures (hash maps for lookups, trees for ranges)
+- Avoid unnecessary allocations in hot paths
+- Pre-compute values that are used repeatedly
+
+### Network Optimization
+- Minimize round trips — batch API calls where possible
+- Use compression for large payloads
+- Implement pagination — never return unbounded result sets
+- Use CDNs for static assets
+
+### Benchmarking
+- Include before/after benchmarks for performance changes
+- Test with realistic data volumes (not just unit test fixtures)
+- Measure p50, p95, p99 latencies — not just averages
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **Baseline Metrics**: Current performance metrics before optimization (p50/p95/p99 latency, throughput, resource usage)
+2. **Optimization Targets**: Specific targets (e.g., "reduce p95 latency from 250ms to <100ms") with rationale
+3. **Profiling Strategy**: Tools and methodology to identify bottlenecks (CPU profiler, memory profiler, query analyzer, benchmarks)
+4. **Benchmark Plan**: Before/after benchmarks with realistic data volume and success criteria for each optimization
+
+If any section is not applicable, explicitly state why it's skipped.

package/scripts/skills/pr-quality.md

@@ -0,0 +1,49 @@
+## PR Quality: Ship a Reviewable Pull Request
+
+Write a PR that a reviewer can understand in 5 minutes.
+
+### PR Description Structure
+1. **What** — One sentence: what does this PR do?
+2. **Why** — Link to issue. Why is this change needed?
+3. **How** — Brief technical approach (2-3 sentences max)
+4. **Testing** — What was tested? How to verify?
+5. **Screenshots** — If UI changes, before/after screenshots
+
+### Commit Hygiene
+- Each commit should be a logical unit of work
+- Commit messages: imperative mood, 50-char subject, blank line, body explains WHY
+- No WIP/fixup/squash commits in final PR
+- No merge commits — rebase onto base branch
+- Separate refactoring commits from feature commits
+
+### Diff Quality
+- Remove all debugging artifacts (console.log, print statements, commented-out code)
+- No unrelated formatting changes mixed with logic changes
+- Generated files should be committed separately or excluded
+- File renames should be separate commits (so git tracks them)
+
+### Reviewer Empathy
+- If the diff is >500 lines, add a "Review guide" section explaining the reading order
+- Call out non-obvious decisions with inline comments
+- Flag areas where you're least confident and want careful review
+- If you changed a pattern used elsewhere, note whether existing code needs updating
+
+### Self-Review Checklist
+Before marking as ready:
+- [ ] PR description explains what, why, and how
+- [ ] All CI checks pass
+- [ ] No secrets, credentials, or API keys in diff
+- [ ] No TODO/FIXME comments without issue links
+- [ ] Breaking changes documented in description
+- [ ] Migration steps documented if applicable
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **PR Description**: What (one sentence), Why (issue link), How (2-3 sentence technical approach), Testing (what was tested)
+2. **Commit Hygiene Check**: Verification that each commit is a logical unit, no WIP/fixup/squash, no merge commits
+3. **Diff Review**: Confirmation that all debugging artifacts removed (console.log, commented code), no unrelated formatting changes
+4. **Self-Review Checklist Completion**: All items from checklist checked (secrets scanned, CI green, breaking changes documented)
+
+If any section is not applicable, explicitly state why it's skipped.

package/scripts/skills/product-thinking.md

@@ -0,0 +1,43 @@
+## Product Thinking Expertise
+
+Consider the user perspective in your implementation:
+
+### User Stories
+- Who is the user for this feature?
+- What problem does this solve for them?
+- What is their workflow before and after this change?
+- Define acceptance criteria from the user's perspective
+
+### User Experience
+- What is the simplest interaction that solves the problem?
+- How does the user discover this feature?
+- What happens when things go wrong? (error states, recovery)
+- Is the feature accessible to users with disabilities?
+
+### Edge Cases from User Perspective
+- What if the user has no data yet? (empty state)
+- What if the user has too much data? (pagination, filtering)
+- What if the user makes a mistake? (undo, confirmation)
+- What if the user is on a slow connection? (loading states)
+
+### Progressive Disclosure
+- Show the most important information first
+- Hide complexity behind progressive interactions
+- Don't overwhelm with options — provide sensible defaults
+- Use contextual help instead of documentation
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **User Stories**: In "As a..., I want..., So that..." format with at least one primary and one secondary user story
+2. **Acceptance Criteria**: Given/When/Then format for how to verify the feature works from the user's perspective
+3. **Edge Cases from User Perspective**: At least 3 specific scenarios (empty state, error state, overload state)
+
+If any section is not applicable, explicitly state why it's skipped.
+
+### Feedback & Communication
+- Confirm successful actions immediately
+- Explain errors in plain language — not error codes
+- Show progress for long-running operations
+- Preserve user context across navigation

package/scripts/skills/security-audit.md

@@ -0,0 +1,49 @@
+## Security Audit Expertise
+
+Apply OWASP Top 10 and security best practices:
+
+### Injection Prevention
+- Use parameterized queries for ALL database access
+- Sanitize user input before rendering in HTML/templates
+- Validate and sanitize file paths — prevent directory traversal
+- Never execute user-supplied strings as code or commands
+
+### Authentication
+- Hash passwords with bcrypt/argon2 (never MD5/SHA1)
+- Implement account lockout after failed attempts
+- Use secure session management (HttpOnly, Secure, SameSite cookies)
+- Require re-authentication for sensitive operations
+
+### Authorization
+- Check permissions server-side on EVERY request
+- Use deny-by-default — explicitly grant access
+- Verify resource ownership (user can only access their own data)
+- Log authorization failures for monitoring
+
+### Data Protection
+- Never log sensitive data (passwords, tokens, PII)
+- Encrypt sensitive data at rest
+- Use HTTPS for all communications
+- Set appropriate CORS headers — never use wildcard in production
+
+### Secrets Management
+- Never hardcode secrets in source code
+- Use environment variables or secret managers
+- Rotate secrets regularly
+- Check for accidentally committed secrets (API keys, passwords, tokens)
+
+### Dependency Security
+- Check for known vulnerabilities in dependencies
+- Pin dependency versions to prevent supply chain attacks
+- Review new dependencies before adding them
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **Threat Model (STRIDE)**: Identify threats across Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, Elevation of Privilege
+2. **Auth Flow**: Step-by-step diagram of authentication/authorization flow with session/token handling
+3. **Input Validation Points**: List all places where user input enters the system and how each is validated/sanitized
+4. **Security Checklist**: Items verified (no secrets in code, secrets rotated, HTTPS enforced, CORS configured, rate limiting applied)
+
+If any section is not applicable, explicitly state why it's skipped.

package/scripts/skills/systematic-debugging.md

@@ -0,0 +1,40 @@
+## Systematic Debugging: Root Cause Analysis
+
+A previous attempt at this stage FAILED. Do NOT blindly retry the same approach. Follow this 4-phase investigation:
+
+### Phase 1: Evidence Collection
+- Read the error output from the previous attempt carefully
+- Identify the EXACT line/file where the failure occurred
+- Check if the error is a symptom or the root cause
+- Look for patterns: is this a known error type?
+
+### Phase 2: Hypothesis Formation
+- List 3 possible root causes for this failure
+- For each hypothesis, identify what evidence would confirm or deny it
+- Rank hypotheses by likelihood
+
+### Phase 3: Root Cause Verification
+- Test the most likely hypothesis first
+- Read the relevant source code — don't guess
+- Check if previous artifacts (plan.md, design.md) are correct or flawed
+- If the plan was correct but execution failed, focus on execution
+- If the plan was flawed, document what was wrong
+
+### Phase 4: Targeted Fix
+- Fix the ROOT CAUSE, not the symptom
+- If the previous approach was fundamentally wrong, choose a different approach
+- If it was a minor error, make the minimal fix
+- Document what went wrong and why the new approach is better
+
+IMPORTANT: If you find existing artifacts from a successful previous stage, USE them — don't regenerate from scratch.
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **Root Cause Hypothesis**: List 3 possible root causes ranked by likelihood with specific evidence that would confirm/deny each
+2. **Evidence Gathered**: Exact file:line location of failure, error messages, logs, code examination results, artifact validation (plan.md, design.md correctness)
+3. **Fix Strategy**: Description of the ROOT CAUSE fix (not the symptom), with rationale for why this approach differs from the previous failed attempt
+4. **Verification Plan**: How to verify the fix works (test cases, specific checks, expected behavior confirmation)
+
+If any section is not applicable, explicitly state why it's skipped.

package/scripts/skills/testing-strategy.md

@@ -0,0 +1,47 @@
+## Testing Strategy Expertise
+
+Apply these testing patterns:
+
+### Test Pyramid
+- **Unit tests** (70%): Test individual functions/methods in isolation
+- **Integration tests** (20%): Test component interactions and boundaries
+- **E2E tests** (10%): Test critical user flows end-to-end
+
+### What to Test
+- Happy path: the expected successful flow
+- Error cases: what happens when things go wrong?
+- Edge cases: empty inputs, maximum values, concurrent access
+- Boundary conditions: off-by-one, empty collections, null/undefined
+
+### Test Quality
+- Each test should verify ONE behavior
+- Test names should describe the expected behavior, not the implementation
+- Tests should be independent — no shared mutable state between tests
+- Tests should be deterministic — same result every run
+
+### Coverage Strategy
+- Aim for meaningful coverage, not 100% line coverage
+- Focus coverage on business logic and error handling
+- Don't test framework code or simple getters/setters
+- Cover the branches, not just the lines
+
+### Mocking Guidelines
+- Mock external dependencies (APIs, databases, file system)
+- Don't mock the code under test
+- Use realistic test data — edge cases reveal bugs
+- Verify mock interactions when the side effect IS the behavior
+
+### Regression Testing
+- Write a failing test FIRST that reproduces the bug
+- Then fix the bug and verify the test passes
+- Keep regression tests — they prevent the bug from recurring
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections when this skill is active:
+
+1. **Test Pyramid Breakdown**: Explicit count of unit/integration/E2E tests and their coverage targets (e.g., "70 unit tests covering business logic, 12 integration tests for API boundaries, 3 E2E tests for critical paths")
+2. **Coverage Targets**: Target coverage percentage per layer and which critical paths MUST be tested
+3. **Critical Paths to Test**: Specific test cases for the happy path, 2+ error cases, and 2+ edge cases
+
+If any section is not applicable, explicitly state why it's skipped.

package/scripts/skills/two-stage-review.md

@@ -0,0 +1,52 @@
+## Two-Stage Code Review
+
+This review runs in two passes. Complete Pass 1 fully before starting Pass 2.
+
+### Pass 1: Spec Compliance
+
+Compare the implementation against the plan and issue requirements:
+
+1. **Task Checklist**: Does the code implement every task from plan.md?
+2. **Files Modified**: Were all planned files actually modified?
+3. **Requirements Coverage**: Does the implementation satisfy every requirement from the issue?
+4. **Missing Features**: Is anything from the plan NOT implemented?
+5. **Scope Creep**: Was anything added that WASN'T in the plan?
+
+For each gap found:
+- **[SPEC-GAP]** description — what was planned vs what was implemented
+
+If all requirements are met, write: "Spec compliance: PASS — all planned tasks implemented."
+
+---
+
+### Pass 2: Code Quality
+
+Now review the code for engineering quality:
+
+1. **Logic bugs** — incorrect conditions, off-by-one errors, null handling
+2. **Security** — injection, XSS, auth bypass, secret exposure
+3. **Error handling** — missing catch blocks, silent failures, unclear error messages
+4. **Performance** — unnecessary loops, missing indexes, N+1 queries
+5. **Naming and clarity** — confusing names, missing context, magic numbers
+6. **Test coverage** — are new code paths tested? Edge cases covered?
+
+For each issue found, use format:
+- **[SEVERITY]** file:line — description
+
+Severity: Critical, Bug, Security, Warning, Suggestion
+
+### Required Output (Mandatory)
+
+Your output MUST include these sections for EACH review pass:
+
+**Pass 1 Output:**
+- **Spec Compliance Verdict**: PASS or FAIL with explicit gaps found (if any)
+- **Unimplemented Tasks**: List any planned tasks NOT in the code
+- **Unplanned Code**: List any code added that was NOT in the plan
+
+**Pass 2 Output:**
+- **Code Quality Verdict**: PASS or list all findings by severity
+- **Critical/Security Issues Found**: Explicit count with file:line references
+- **Suggested Improvements**: Optional suggestions that don't block PASS
+
+If any finding is not applicable, explicitly state why it's skipped.