Implementation

Feature development for IMPLEMENT state.

When to Load

• •State: IMPLEMENT
• •Condition: Pending feature exists in feature-list.json

Core Workflow

Step 2: TEMPLATE- Gate (Critical)

If TEMPLATE-*.sh files exist → STOP

Only proceed when: ls .claude/scripts/TEMPLATE-*.sh returns nothing

See references/template-scripts.md for detailed customization guide.

Step 5: Feature Implementation Sub-Workflow (CRITICAL)

NEVER commit until ALL verification steps pass.

5c: Verification Commands (Run in Order)

Stop at first failure. Fix issue, then restart from top.

Common commands by project type:

Golden Rule: git commit ONLY when all commands exit with code 0.

5d: Manual Testing (When Needed)

Purpose: Confirm implementation works before committing.

Token-Efficient Analysis (Step 3)

Rule: Never load raw logs/code into context.

Exit Criteria (All Must Pass)

• • No TEMPLATE-*.sh files remain
• • Code written with proper types/error handling
• • Tests written (unit tests for happy paths, errors, edge cases)
• • ALL verification commands pass (exit code 0)
  • • Type check passes
  • • All tests pass
  • • Build succeeds
  • • Health check passes (if applicable)
• • Manual testing confirms functionality (if needed)
• • Committed with feature ID in message
• • feature-list.json status updated
• • Session introspection complete (Step 8)
  • • Issues identified and documented
  • • Scripts updated (if issues found)
  • • Script updates committed separately (chore:)

NEVER skip verification. Commit only after ALL checks pass.

Post-Implementation: Store Learnings

After successful feature completion, store key decisions in context-graph:

Example:

context_store_trace(
  decision="Used StateStore for session management with 30min TTL",
  category="architecture",
  outcome="success",
  feature_id="FEATURE-001"
)

Purpose: Future features can query past decisions for consistency.

Step 8: Session Introspection & Script Updates

Purpose: Create continuous improvement feedback loop. Detect friction points and update scripts to prevent future issues.

When to Run

• •After feature marked complete
• •After learnings stored in context-graph
• •Before session ends

Detection Patterns

Detection Methods

Output Format

When issues detected:

🔧 Session Issues Detected:
- [Issue 1]: [Impact]
- [Issue 2]: [Impact]

📝 Recommended Script Updates:
- [script-name]: [Update description] (BEFORE → AFTER)

Apply updates? (y/n)

Implementation Guidelines

Commit Format

git add .claude/scripts/
git commit -m "chore: Update scripts based on [FEATURE-ID] learnings

- [script-1]: [update description]
- [script-2]: [update description]

Prevents: [issue description]"

Example from SDK-BUYER-CART-003

Issues Detected:

• •Rebuilt api-server 5+ times manually
• •Tests in nested packages not discovered
• •New endpoints (cart, checkout) untested

Updates Applied:

1. •restart-servers.sh: Use tsx watch for hot-reload
2. •vitest.config.ts: Added packages/*/*/src/**/*.test.ts
3. •run-tests.sh: Added POST /api/cart, PUT /api/cart/buyer, POST /api/checkout

Result: Future sessions have hot-reload, all tests discovered, endpoints validated

Success Criteria

Scripts Reference

If Scripts Don't Exist

When project scripts are missing → Create minimal versions:

References