Af Delivery Process

Delivery Process

Quick Reference

The delivery process transforms approved mini-PRDs into working code through iterative Test-Driven Development (TDD). Implementation happens in local development (localhost:3000 + Amplify sandbox) with continuous testing, documentation, and pull request creation.

Typical duration: 4-12 hours per feature End result: Working code, passing tests, complete documentation, ready-for-review PR Scope: Local development only (up to PR creation, not deployment)

When to Use

✅ Use Delivery Process for:

• Implementing approved mini-PRDs from Refinement
• Writing code using TDD (Red → Green → Refactor)
• Running tests iteratively until clean
• Creating documentation and PRs

❌ Not Delivery Phase:

• Defining requirements (use Refinement)
• Deployment to production (separate process)
• Exploratory work without specs (use Discovery)

Git Workflow

Delivery work happens on issue branches (see ADR-009).

# Start implementation (only works if Linear status >= "Approved")
start-work myproject JCN-123    # Creates issue worktree + tmux

# After PR merged, cleanup
stop-work myproject JCN-123     # Removes worktree + tmux

Prerequisites:

• Specs must be PRd from specs branch to develop first
• Linear status must be "Approved" or later
• start-work will warn if feature isn't ready

Issue branches are short-lived - they only exist during implementation. Specs (mini-PRD, scenarios) are already in develop from the Refinement phase.

See Work Management Guide for full branching details.

Entry Criteria

Before starting Delivery, MUST have:

• ✅ Complete mini-PRD from Refinement phase
• ✅ Approved BDD scenarios with glossary compliance
• ✅ Visual specifications (Storybook stories for UI)
• ✅ Linear issue in "In Development" status
• ✅ Development environment ready (localhost + sandbox)
• ✅ Story points set on Linear issue (Refined Estimate preferred)

Do NOT start Delivery without approved Refinement

Estimation check: If the issue has no story points (estimate is null), flag this to the orchestrator before starting implementation. Load af-estimation-expertise and produce at minimum a Discovery Estimate. If only a Discovery Estimate exists (check comments for [Discovery Estimate] tag), consider whether a Refined Estimate is needed given that Refinement is now complete.

Core Workflow

TDD Cycle: Red → Green → Refactor

1. Write failing test (Red)
2. Write minimal code to pass (Green)
3. Refactor for quality
4. Repeat for next requirement

Implementation Pattern

Implement → Test → Fix → Retest → Document → PR

Key characteristic: Iterative with continuous testing

When Tests Are Written

Refinement Phase (ALREADY DONE before Delivery):

• Storybook stories with play functions (primary UI component tests) - PASSING
• RTL tests for non-visual logic (hooks, utils, state management) - PASSING
• Selector contract defined in mini-PRD Section 5
• Markdown scenarios written in mini-PRD Section 4

Delivery Phase (RED):

• AI generates E2E and Integration test files FROM scenarios + selector contract
• Test files go to tests/e2e/ and tests/integration/
• These tests start as test.todo() (no backend/API yet) — visible in output as "todo" count
• Do NOT regenerate RTL or Storybook tests that already exist from Refinement

Delivery Phase (GREEN):

• Implementation makes E2E/Integration tests pass
• Components use selector contract: data-testid={AUTH.signup.email}
• Tests use same contract: [data-testid="${AUTH.signup.email}"]

The selector contract is the bridge between specifications, tests, and implementation. Single source of truth prevents drift.

Step 1: Feature Implementation

A. Start with test generation (RED)

• Read BDD scenarios from mini-PRD Section 4
• Import selector contract from mini-PRD Section 5
• Verify existing tests from Refinement phase still pass (Storybook + RTL)
• Generate NEW test files for E2E and Integration only:
  • E2E → Playwright specs in tests/e2e/
  • Integration → Jest tests in tests/integration/
• New tests start as test.todo('description') (Red state) — never test.skip()

B. Implement minimal code

• Write code to make tests pass (Green state)
• Follow existing project patterns
• Keep changes focused on mini-PRD requirements

C. Refactor for quality

• Improve code structure
• Remove duplication
• Enhance readability
• Maintain all tests passing

Step 2: Iterative Testing

Because dev-test-agent is stateless, use iterative invocation pattern:

❌ Bad Pattern (won't work):

Task → dev-test-agent: "Monitor tests while I fix issues"
# Agent can't maintain state across iterations

✅ Good Pattern (will work):

Task → dev-test-agent: "Run authentication tests"
# Agent reports: 3 failures
# Orchestrator fixes issues
Task → dev-test-agent: "Run authentication tests again"
# Agent reports: 1 failure
# Orchestrator fixes issue
Task → dev-test-agent: "Run authentication tests once more"
# Agent reports: All passing ✅

Testing layers:

1. Unit tests - Individual functions/components
2. Integration tests - API endpoints, data flow
3. E2E tests - Complete user flows (if configured)

Step 3: Background Process Management

Important: Port numbers and backend configuration are defined in your project's CLAUDE.md file.

Frontend (varies by framework):

# Next.js, React, Vue, etc.
npm run dev &  # Start in background (port defined in CLAUDE.md)

# Check status
Task → dev-test-agent: "Check if frontend running on <port-from-CLAUDE.md>"

# Monitor errors
Task → dev-test-agent: "Check frontend console for errors"

Backend (varies by stack):

AWS Amplify (AgentFlow default):

npx ampx sandbox  # Start Amplify sandbox

# Verify APIs
Task → dev-test-agent: "Check if backend APIs responding"

Other backends (Express, Django, Flask, FastAPI, etc.):

• Configuration and startup commands in project's CLAUDE.md
• Each project defines its backend stack and ports
• Follow project-specific setup instructions

Step 4: Documentation Creation

After implementation is stable:

A. API Documentation

Task → technical-writer-agent:
Input: "Create API documentation for auth endpoints"
Output: /docs/api/auth-endpoints.md

B. Code Comments

Task → technical-writer-agent:
Input: "Add JSDoc comments to auth service"
Output: Inline JSDoc in source files

C. User Guides (if needed)

Task → technical-writer-agent:
Input: "Create user guide for authentication flow"
Output: /docs/guides/authentication.md

Step 5: Quality Validation

A. Documentation Validation

Task → docs-quality-agent:
Input: "Validate all documentation created"
Output: Validation report, auto-fixed issues

B. Code Quality Check

Task → quality-agent:
Input: "Check code quality standards"
Output: Code review, improvement suggestions

Step 5.5: Coverage Validation

Before creating PR, verify test coverage:

npm run test:coverage

Check for gaps:

1. Open coverage/lcov-report/index.html
2. Review files with <80% coverage
3. Identify missing test cases:
   • API routes without unit tests
   • Utility functions without unit tests
   • Components without RTL tests
   • Error handling paths not covered

Coverage categories:

If coverage gaps found:

• Write missing unit tests for infrastructure code
• BDD scenarios cover behaviour; infrastructure tests cover plumbing
• Add per-file thresholds for critical modules in jest.config.js

Step 5.6: Local Code Review (Pre-PR)

Before creating a PR, get fresh eyes on the diff:

1. Spawn af-code-quality-agent with the full diff (git diff develop...HEAD)
2. Subagent reviews cold — same criteria as the GitHub CR Action:
   • Code quality and best practices
   • Potential bugs or issues
   • Performance considerations
   • Security concerns
   • Test coverage
3. Read findings, address substantive issues
4. If significant changes made, run the subagent again (max 2 local review cycles)
5. Local CR does NOT create Linear issues — findings go directly to the delivery agent

Why fresh eyes: The delivery agent built the code and has context bias. A subagent reviewing cold catches what the author misses — same principle as the GitHub CR, but faster (no round-trip).

Step 6: Pull Request Creation

Human Approval Gate:

• ✋ Orchestrator presents changes summary
• ✋ Human approves PR creation

PR creation includes:

• Clean commit history
• Descriptive PR description referencing mini-PRD
• Links to Linear issue
• Test summary (all passing)
• Implementation notes

Step 7: Post-PR Code Review Loop

After creating the PR, the delivery agent stays "In Progress" through the full CR cycle:

1. Post Linear comment: [Delivery] PR #X created: <link> (permanent artifact)
2. Post Zulip notification: "PR ready for review" (working conversation)
3. Poll for CI/CR results — use gh pr checks <number> and gh pr reviews <number>
4. When Claude Code Review comments arrive, read via gh api repos/{owner}/{repo}/pulls/{number}/comments
5. Address each comment, push fixes
6. Repeat steps 3-5 (max 3 GitHub CR cycles)
7. When clean → move to "In Review" + "Waiting for Feedback"
8. Agent hibernates — human merges

CR escalation (stop and ask the human via Zulip):

• 3 GitHub CR cycles with unresolved comments
• Same comment recurs after fix attempt
• CR suggests changes beyond original issue scope
• Architectural disagreement
• Cannot understand what CR is asking for

On escalation: post to Zulip with summary of unresolved comments and reason, set "Waiting for Feedback", hibernate.

PR notifications:

• Linear comment: Only the PR link (permanent artifact on the issue)
• Zulip: Everything else (PR created, CI results, CR status, escalation)
• Status transitions: Not notifications — Linear status IS the signal

Scope ends at PR merge - deployment uses label-driven progression (see CLAUDE-agentflow.md)

Human-in-the-Loop Checkpoints

Orchestrator MUST pause for human input at:

1. After initial implementation

   • "Does this approach look correct?"
   • Verify implementation matches mini-PRD

2. When tests fail repeatedly

   • "Need help with this test failure"
   • Get guidance on tricky issues

3. When technical limitation discovered ⚠️

   • STOP immediately - do not work around it
   • "I discovered [TECH] doesn't support [CAPABILITY]"
   • Present options: re-scope, defer, or investigate alternatives
   • Wait for decision before continuing

4. Before major refactoring

   • "Should I refactor this for better structure?"
   • Confirm structural changes

5. Before creating PR

   • "Ready to create PR with these changes?"
   • Final approval gate

Agent Coordination

dev-test-agent (Stateless Testing)

Purpose: Execute tests and report results Pattern: Iterative invocation (call → fix → call again) Limitations: Cannot maintain state across calls Usage: Each test run is independent

technical-writer-agent (Documentation)

Purpose: Create documentation artifacts Outputs:

• API documentation
• Component documentation
• JSDoc/TSDoc comments
• User guides

code-quality-agent (Quality Review)

Purpose: Enforce coding standards Actions:

• Review code patterns
• Suggest improvements
• Verify best practices

docs-quality-agent (Validation)

Purpose: Validate documentation completeness Actions:

• Check frontmatter compliance
• Verify bidirectional links
• Update last_checked dates

Decision Points

Should I refactor now or later?

• Tests passing and code working? → Refactor now (safe)
• Tests failing? → Fix tests first, then refactor
• Major structural change? → Get human approval first

How much testing is enough?

• Minimum: All BDD scenarios passing
• Standard: + Unit tests for business logic
• Comprehensive: + Integration tests + E2E (if configured)

When should I invoke agents?

• Every code change? → NO (too frequent)
• After each test run? → YES (for dev-test-agent)
• After implementation complete? → YES (for documentation/quality agents)

Common Patterns

API-Heavy Features

Focus on:

• GraphQL resolver implementation
• Database query optimization
• Error handling and validation
• API integration tests

UI-Heavy Features

Focus on:

• Component implementation from Storybook stories
• State management
• Event handling
• Responsive design
• Accessibility (ARIA, keyboard nav)

Data-Heavy Features

Focus on:

• Database migrations
• Data validation
• Query performance
• Backup/recovery procedures

Output Artifacts

1. Working Code

• ✅ Feature implementation complete
• ✅ All tests passing (unit, integration, E2E)
• ✅ Code quality validated
• ✅ Follows project patterns
• ✅ No console errors or warnings

2. Documentation

• ✅ API documentation (if backend changes)
• ✅ Component documentation (if UI changes)
• ✅ JSDoc/TSDoc comments (inline)
• ✅ User guides (if complex feature)
• ✅ Architecture diagrams (if new patterns)

3. Pull Request

• ✅ Clean commit history
• ✅ Descriptive PR title and description
• ✅ Links to Linear issue and mini-PRD
• ✅ Test summary included
• ✅ Screenshots/videos (for UI changes)
• ✅ Ready for peer review

Handling Implementation Blockers

When you discover a technical limitation that prevents implementing a feature as specified:

✅ MUST DO

1. STOP immediately - Do not continue implementation
2. Inform the human - Explain the limitation clearly:
   • What you were trying to implement
   • What technical constraint prevents it
   • What the impact is (feature won't work as spec'd)
3. Return to Refinement - The mini-PRD needs to be updated
4. Wait for decision - Human/PM decides:
   • Re-scope the feature (what CAN we build?)
   • Defer the feature (move to backlog)
   • Investigate alternatives (different approach?)

❌ MUST NOT DO

1. Don't build mock/fake implementations

   • "Illustrative" UI that doesn't actually work
   • Hardcoded data pretending to be real
   • Comments saying "this doesn't really work"

2. Don't write tests that pass for broken features

   • Tests should verify BEHAVIOR, not just UI presence
   • toBeVisible() ≠ "feature works"
   • A passing test must mean the feature works

3. Don't make scope decisions alone

   • "I'll just build what I can" - NO
   • "I'll add a note in the code" - NO
   • "The test passes so it's fine" - NO

Why This Matters

Bad outcome (what we're preventing):

Spec: "User can view active sessions"
     ↓
Discovery: Platform doesn't support session listing
     ↓
❌ Dev builds mock UI with fake data
     ↓
❌ Dev writes test: "element is visible" ✓
     ↓
Test passes, feature broken, false confidence

Good outcome (what we want):

Spec: "User can view active sessions"
     ↓
Discovery: Platform doesn't support session listing
     ↓
✅ Agent STOPS and informs human
     ↓
✅ Return to Refinement phase
     ↓
✅ PM decides: Re-scope to "Sign out all devices" only
     ↓
✅ Mini-PRD updated, BDD scenarios updated
     ↓
✅ Implementation matches updated spec
     ↓
Test passes, feature works, real confidence

Example Human Notification

When you hit a blocker, tell the human:

⚠️ Implementation Blocker

I discovered that [PLATFORM/TECH] does not support [CAPABILITY].

Spec says: User can view their active sessions Reality: Cognito API does not provide session listing

Options:

1. Re-scope: Only implement "Sign out all devices" (which IS supported)
2. Defer: Move feature to backlog, investigate custom session tracking later
3. Alternative: Implement custom session tracking (significant extra work)

I have NOT implemented any workarounds. Awaiting your decision.

Common Pitfalls

1. Starting without approved Refinement

   • Always verify mini-PRD is approved
   • Implementation without specs leads to rework

2. Skipping tests

   • TDD is mandatory, not optional
   • Tests document behavior and prevent regressions

3. Not running tests iteratively

   • dev-test-agent is stateless
   • Must call repeatedly after each fix

4. Forgetting documentation

   • Documentation is part of "done"
   • Undocumented code is incomplete

5. Creating PR without human approval

   • Always get approval before PR creation
   • Human validates implementation matches intent

6. Deploying to production

   • Delivery phase ends at PR creation
   • Deployment is separate approval process

7. Working around technical limitations silently

   • Never build mock/fake implementations
   • Never write tests that pass for broken features
   • Always escalate to human immediately

Integration with Other Phases

Before Delivery:

• ← Refinement Phase - Provides approved mini-PRD and BDD scenarios

After Delivery:

• → Code Review - PR reviewed by team
• → Deployment - After PR approved and merged
• → Monitoring - Production monitoring and alerts

Delivery provides:

• Working implementation
• Complete test coverage
• Documentation
• Merge-ready PR

Detailed Reference

For complete delivery documentation:

• Read .claude/docs/guides/delivery-guide.md (full workflow details)

For testing strategies:

• Load af-testing-expertise skill
• Read .claude/docs/guides/testing-guide.md

For agent interactions:

• Read .claude/agents/dev-test-agent.md
• Read .claude/agents/technical-writer-agent.md
• Read .claude/agents/code-quality-agent.md

Success Criteria

Delivery complete when:

• ✅ All BDD scenarios passing (zero test.todo() remaining — all implemented)
• ✅ Unit tests passing (business logic covered)
• ✅ Integration tests passing (API endpoints verified)
• ✅ Zero test.skip() in committed code
• ✅ No console errors or warnings
• ✅ Code quality validated
• ✅ Documentation complete and validated
• ✅ Human has approved implementation
• ✅ PR created and linked to Linear issue
• ✅ Ready for peer review