/arc:implement

<build_agents> Available build agents in agents/build/:

Agent	Model	Use For
`implementer`	opus	General task execution — utilities, services, APIs, business logic
`fixer`	haiku	TypeScript errors, lint issues — fast mechanical fixes
`debugger`	sonnet	Failing tests — systematic root cause analysis
`unit-test-writer`	sonnet	Unit tests (vitest) — pure functions, components
`integration-test-writer`	sonnet	Integration tests (vitest + MSW) — API, auth
`e2e-test-writer`	opus	E2E tests (Playwright) — user journeys
`ui-builder`	opus	UI components from design spec — anti-slop, memorable
`design-specifier`	opus	Design decisions when no spec exists — empty states, visual direction
`figma-builder`	opus	Build UI directly from Figma URL
`test-runner`	haiku	Run vitest, analyze failures
`e2e-runner`	opus	Playwright tests — iterate until green or report blockers
`spec-reviewer`	sonnet	Spec compliance check — nothing missing, nothing extra
`code-reviewer`	haiku	Quick code quality gate — no `any`, proper error handling, tests exist
`plan-completion-reviewer`	sonnet	Whole-plan gate — all tasks built, nothing skipped, no scope creep

Before spawning a build agent:

Read the agent file: agents/build/[agent-name].md
Use the model specified in the agent's frontmatter
Include relevant context from the task

Spawn syntax:

Task [agent-name] model: [model]: "[task description with context]"

</build_agents>

You are here in the arc:

/arc:ideate     → Design doc (on main) ✓
     ↓
/arc:implement  → Plan + Execute ← YOU ARE HERE
     ↓
/arc:review     → Review (optional, can run anytime)

Phase 0: Planning (if no plan exists)

Check for existing implementation plan:

ls docs/arc/plans/*-implementation.md docs/plans/*-implementation.md 2>/dev/null | tail -1

If plan exists: Skip to Phase 1.

If no plan exists: Follow the detail skill to create one:

Read: skills/detail/SKILL.md

The detail skill will:

Load design document
Detect project stack
Find reusable patterns
Break down into TDD tasks
Save implementation plan

After plan is created, strongly recommend review:

AskUserQuestion:
  question: "Implementation plan ready. I strongly recommend reviewing the plan before building — it's much cheaper to catch issues now than after writing code."
  header: "Review Before Building?"
  options:
    - label: "Review first"
      description: "Run /arc:review on the plan before implementing (recommended)"
    - label: "Skip review"
      description: "Start implementing immediately"

If "Review first" → invoke /arc:review, then return here.

Phase 1: Setup

If not already in worktree:

# Check current location
git branch --show-current

# If on main/dev, create worktree
git worktree add .worktrees/<feature-name> -b feature/<feature-name>
cd .worktrees/<feature-name>

Install dependencies:

pnpm install  # or yarn/npm based on lockfile

Verify test infrastructure exists:

# Check for test runner in package.json
grep -E '"vitest"|"jest"|"playwright"' package.json

If no test runner → stop and ask user. Cannot proceed with TDD without a runner.

Verify clean baseline:

pnpm test     # or relevant test command

If tests fail before you start → stop and ask user.

Phase 2: Load Plan and Create Todos

Read implementation plan (created in Phase 0 or pre-existing): docs/arc/plans/YYYY-MM-DD-<topic>-implementation.md (fallback: docs/plans/...)

Parse XML tasks: The plan contains <task> elements with structured fields. For each task:

Extract id, depends, type attributes
Note <read_first> files, <action> instructions, <verify> criteria
Build a dependency graph from depends attributes

Create TodoWrite tasks: One todo per <task> in the plan. Mark first as in_progress.

Before implementation starts, confirm the plan includes:

a file structure section
XML tasks with all required elements (<name>, <files>, <read_first>, <action>, <verify>, <done>, <commit>)
<verify> criteria that are concrete (no "works correctly", "looks good")
checkpoint tasks only where human judgment is required

If any of those are missing, fix the plan before dispatching build agents.

Phase 2b: Plan Test Coverage

Before implementation, identify test needs:

## Test Coverage Plan

### Unit Tests (per task)
| Task | Test File | What to Test |
|------|-----------|--------------|
| Task 1: Create utility | src/utils/x.test.ts | Input/output, edge cases |
| Task 2: Create component | src/components/x.test.tsx | Rendering, props |

### Integration Tests (per feature)
| Feature | Test File | What to Test |
|---------|-----------|--------------|
| Signup form | src/features/auth/signup.integration.test.ts | Form + API + validation |

### E2E Tests (critical flows only)
| Flow | Test File | What to Test |
|------|-----------|--------------|
| User signup → dashboard | tests/signup.spec.ts | Full journey |

Determine auth testing needs:

Uses Clerk? → integration-test-writer with Clerk mocks
Uses WorkOS? → integration-test-writer with WorkOS mocks
Has protected routes? → e2e-test-writer with auth.setup.ts

This plan guides which test agent to spawn for each task.

Phase 2c: Handle Build-Agent Statuses

Build agents must report one of:

DONE
DONE_WITH_CONCERNS
NEEDS_CONTEXT
BLOCKED
AUTH_GATE

Controller behavior:

DONE → continue to review
DONE_WITH_CONCERNS → read concerns, then decide whether to clarify or review
NEEDS_CONTEXT → provide the missing context and re-dispatch
BLOCKED → split the task, upgrade model capability, or escalate to the user
AUTH_GATE → present dynamic CHECKPOINT:ACTION, verify auth, re-dispatch same task

Never silently retry a blocked task without changing the conditions.

AUTH_GATE Handling

When an agent reports AUTH_GATE, the task is NOT skipped. Follow this protocol:

Read the agent's report — it includes: attempted command, error, human action, verify command, retry command
Present CHECKPOINT:ACTION to user:

AskUserQuestion:
  question: "The agent tried `[attempted command]` but hit an auth wall: [error]. Please [human action], then confirm."
  header: "Authentication Required"
  options:
    - label: "Done"
      description: "I've completed the authentication step"
    - label: "Need help"
      description: "I need more guidance"

After user confirms "Done" — run the verify command (e.g., vercel whoami) to confirm auth succeeded
If verify passes — re-dispatch the SAME task to the agent (not the next task)
If verify fails — tell the user what went wrong, ask them to try again

CRITICAL: Never skip a task because of an auth error. Never move to the next task. The whole point of AUTH_GATE is that the task is viable — it just needs a human to unlock a door.

Phase 3: Execute in Batches

Default batch size: 3 tasks

Per-task loop:

┌─────────────────────────────────────────────────────────┐
│  1. CLASSIFY  → what type of task? what test level?     │
│  2. TEST      → spawn test agent (unit/integration/e2e) │
│  3. BUILD     → implementer / ui-builder / specialized  │
│  4. TDD       → run test (fail→impl→pass)               │
│  5. FIX       → fixer (TS/lint cleanup)                 │
│  6. SPEC      → spec-reviewer (matches spec?)           │
│       ↳ issues? → fix → re-review                       │
│  7. QUALITY   → code-reviewer (well-built?)             │
│       ↳ issues? → fix → re-review                       │
│  8. COMMIT    → atomic commit, mark complete            │
└─────────────────────────────────────────────────────────┘

For each task:

Step 1: Mark in_progress

Update TodoWrite.

Step 2: Classify Task Type

Determine which build agent(s) may be needed:

Task Type	Primary Agent	When to Use
General implementation	implementer	Utilities, services, APIs, business logic
Write unit tests	unit-test-writer	Pure functions, components, hooks
Write integration tests	integration-test-writer	API mocking, auth states
Write E2E tests	e2e-test-writer	User journeys, Playwright
Build UI from spec	ui-builder	UI components with existing design direction
Build UI from Figma	figma-builder	Figma URL provided
Design decisions needed	design-specifier	No spec exists (empty states, visual direction)
Fix TS/lint errors	fixer	Mechanical cleanup
Debug failing tests	debugger	Test failures
Run E2E tests	e2e-runner	Playwright test suites
Verify spec compliance	spec-reviewer	After implementation, before code quality

Agent selection flow:

Is this general code (no UI)? → implementer
Is this UI with Figma? → figma-builder
Is this UI with design spec? → ui-builder
Is this UI with no spec? → design-specifier first, then ui-builder
Did something break? → debugger or fixer
Task complete? → spec-reviewer to verify

Step 3: Write Tests First (TDD)

Determine test type based on task:

Task Type	Test Agent	Framework
Pure function/utility	unit-test-writer	vitest
Component with props	unit-test-writer	vitest + testing-library
Component + API/state	integration-test-writer	vitest + MSW
Auth-related feature	integration-test-writer	vitest + Clerk/WorkOS mocks
User flow/journey	e2e-test-writer	Playwright

Spawn appropriate test writer:

For unit tests:

Task [unit-test-writer] model: sonnet: "Write unit tests for [function/component].

Behavior to test:
- [expected behavior from plan]
- [edge cases]
- [error cases]

File to create: [path/to/module.test.ts]
Follow vitest patterns from testing-patterns.md"

For integration tests (API/auth):

Task [integration-test-writer] model: sonnet: "Write integration tests for [feature].

Behavior to test:
- [component + API interaction]
- [auth states: loading, signed in, signed out]
- [error handling]

Auth: [Clerk/WorkOS/none]
API endpoints to mock: [list]
File to create: [path/to/feature.integration.test.ts]"

For E2E tests (critical flows):

Task [e2e-test-writer] model: opus: "Write E2E tests for [user journey].

Flow to test:
- [step 1]
- [step 2]
- [expected outcome]

Auth setup: [Clerk/WorkOS/none]
File to create: [tests/feature.spec.ts]"

Step 4: TDD Cycle

1. Tests written (from Step 3)
2. Run test → verify FAIL
3. Write implementation (copy from plan, adapt as needed)
4. Run test → verify PASS
5. Fix TypeScript + lint (spawn fixer if issues)
6. Commit with message from plan

<continuous_quality> After every implementation, before commit:

TypeScript check:

pnpm tsc --noEmit

Biome lint + format:

pnpm biome check --write .

If issues found — spawn fixer:

Task [fixer] model: haiku: "Fix TypeScript and lint errors.

Files with issues: [list files]
Errors: [paste error output]

Project rules: .ruler/typescript.md, .ruler/code-style.md"

Why continuous:

Catching TS errors early is easier than fixing 20 at once
Biome auto-fix keeps code consistent
Each commit is clean and deployable </continuous_quality>

If test doesn't fail when expected:

Test might be wrong
Implementation might already exist
Stop and ask user

If test doesn't pass after implementation — spawn debugger:

Task [debugger] model: sonnet: "Test failing unexpectedly.

Test file: [path]
Test name: [name]
Error: [paste full error]
Implementation file: [path]

Investigate root cause and fix. See disciplines/systematic-debugging.md"

If debugger can't resolve after one attempt → stop and ask user.

Step 5: Spec Compliance Check

After implementation, spawn spec-reviewer:

Task [spec-reviewer] model: sonnet: "Verify implementation matches spec.

Task spec:
[paste full <task> XML element]

Files created/modified: [list]

Check against <verify> and <done> criteria: nothing missing, nothing extra."

If spec-reviewer finds issues → fix with implementer/fixer → re-run spec-reviewer. If compliant → proceed to code quality.

Step 6: Code Quality Gate

After spec compliance passes, spawn code-reviewer:

Task [code-reviewer] model: haiku: "Quick code quality check.

Files: [list of files created/modified]

Check: no any types, error handling, tests exist, style consistent."

If code-reviewer finds issues → fix with fixer → re-run code-reviewer. If approved → commit and mark complete.

Step 7: Commit and Mark Complete

git add [files]
git commit -m "feat(scope): [description from plan]"

Update TodoWrite to mark task completed.

Step 8: Checkpoint after batch

After every 3 tasks:

Completed:
- Task 1: [description] ✓
- Task 2: [description] ✓
- Task 3: [description] ✓

Tests passing: [X/X]

Ready for feedback before continuing?

Wait for user confirmation or adjustments.

Step 8b: Handle Checkpoint Tasks

If the current task is a checkpoint type ([CHECKPOINT:VERIFY], [CHECKPOINT:DECIDE], [CHECKPOINT:ACTION]):

For VERIFY:

Ensure verification environment is running (dev server started, etc.)
Present what was built and verification steps
Ask for approval:

AskUserQuestion:
  question: "[Summary of what was built and how to verify it]"
  header: "Verify Implementation"
  options:
    - label: "Approved"
      description: "Implementation looks correct, continue to next task"
    - label: "Has issues"
      description: "There are problems that need fixing before continuing"

If "Has issues" → ask for description, fix, then re-present checkpoint
If "Approved" → continue to next task

For DECIDE:

Present options with pros/cons from the plan using AskUserQuestion:

AskUserQuestion:
  question: "[Context and trade-offs for this decision]"
  header: "Decision Required"
  options:
    - label: "[Option A name]"
      description: "[Pros/cons summary for option A]"
    - label: "[Option B name]"
      description: "[Pros/cons summary for option B]"

Record decision in progress journal
Continue implementation using selected option

For ACTION:

Explain what was attempted and what blocked (auth gate, etc.)
Provide exact steps for the manual action
Wait for confirmation:

AskUserQuestion:
  question: "[What manual action is needed and the exact steps to perform it]"
  header: "Manual Action Required"
  options:
    - label: "Done"
      description: "I've completed the manual action"
    - label: "Need help"
      description: "I need more guidance on this step"

If "Done" → verify the action succeeded (e.g., vercel whoami)
If "Need help" → provide additional guidance, then re-present checkpoint
Retry the blocked operation and continue

See references/checkpoint-patterns.md for full protocol.

Phase 4: Quality Checkpoints

Before creating new utility functions or services: Spawn duplicate-detector to check for existing similar functionality:

Task [duplicate-detector] model: sonnet: "Scan for functions similar to what I'm about to create.

New function purpose: [what it does]
Search in: [src/utils/, src/lib/, src/helpers/ or relevant dirs]

Report any semantic duplicates so we can reuse instead of reinvent."

If duplicates found → reuse existing code. Skip creating the new function.

After completing data/types tasks:

Spawn data-engineer (from review agents) for quick review
Present findings as questions

Before starting UI tasks:

If design spec exists — spawn ui-builder:

Read: agents/build/ui-builder.md

If no design spec (empty states, undefined visuals) — spawn design-specifier first:

Task [design-specifier] model: opus: "Create design spec for [component].

Context: [what this is for, user's emotional state]
Existing patterns: [what it should feel like]
Project aesthetic: [tone from design doc]

Output actionable spec for ui-builder to implement."

Then spawn ui-builder with the design-specifier's output.

If Figma URL provided — spawn figma-builder:

Read: agents/build/figma-builder.md
Task [figma-builder] model: opus: "Implement from Figma: [URL]"

**For ui-builder, spawn:

Task [ui-builder] model: opus: "Build UI components for [feature].

Aesthetic Direction (from design doc):
- Tone: [tone]
- Memorable element: [what stands out]
- Typography: [fonts]
- Color strategy: [approach]
- Motion: [philosophy]

Figma: [URL if available]
Files to create: [list from implementation plan]

Interface rules: rules/interface/
Project rules: .ruler/react.md, .ruler/tailwind.md

Apply the aesthetic direction to every decision. Make it memorable, not generic."

Fetch Figma context (if available):

mcp__figma__get_design_context: fileKey, nodeId
mcp__figma__get_screenshot: fileKey, nodeId

After completing ALL UI tasks — spawn designer review:

Task [designer] model: opus: "Review the completed UI implementation.

Aesthetic Direction (from design doc):
- Tone: [tone]
- Memorable element: [what stands out]
- Typography: [fonts]
- Color strategy: [approach]

Files: [list of UI component files]
Figma: [URL if available]

Check for:
- Generic AI aesthetics (Inter, purple gradients, cookie-cutter layouts)
- Deviation from aesthetic direction
- Missing memorable moments
- Inconsistent application of design system
- Accessibility concerns
- Missing states (loading, error, empty)"

Address any review findings before proceeding.

When implementing unfamiliar library APIs:

mcp__context7__resolve-library-id: "[library name]"
mcp__context7__get-library-docs: "[library ID]" topic: "[specific feature]"

Use current documentation to ensure correct API usage.

After completing all tasks:

Run full test suite
Run linting

Phase 5: Final Quality Sweep

Spawn parallel build agents for speed:

Task [fixer] model: haiku: "Run TypeScript check (tsc --noEmit) and fix any errors. Report results."

Task [fixer] model: haiku: "Run Biome check (biome check --write .) and fix any issues. Report results."

Wait for agents to complete. If issues found, fix before proceeding.

Run test suite:

pnpm test

If tests fail, spawn debugger to investigate.

Phase 5.5: Plan Completion Verification

This is the whole-plan gate. Per-task spec reviews catch issues within tasks — this catches tasks that were skipped, partially implemented, or scope that crept in.

Re-read the original implementation plan (the file from Phase 2)
Get the list of all files changed:

git diff --name-only main...HEAD

Spawn plan-completion-reviewer:

Task [plan-completion-reviewer] model: sonnet: "Verify the entire implementation matches the original plan.

ORIGINAL PLAN:
[paste full plan text]

FILES CHANGED:
[paste git diff file list]

TEST RESULTS:
[paste test summary — N passing, N failing]

Read each file referenced in the plan. Verify every task was implemented substantively.
Check for skipped tasks, partial implementations, and scope creep.
See agents/build/plan-completion-reviewer.md"

If plan-completion-reviewer finds issues:

Skipped tasks → implement them now
Partial implementations → complete them
Scope creep → ask user if extras should stay or be removed
Re-run plan-completion-reviewer after fixes

Do NOT proceed to Phase 6 until plan-completion-reviewer passes.

Phase 5b: E2E Tests (If Created)

If e2e tests were created as part of this implementation:

Spawn e2e-runner agent:

Task [e2e-runner] model: opus: "Run E2E tests for the feature we just implemented.

Test files: [list e2e test files]
Feature: [brief description]

Run tests, fix any failures, and iterate until all pass or report blockers.
See agents/build/e2e-runner.md for protocol."

Why a separate agent?

E2E tests produce verbose output (traces, screenshots, DOM snapshots)
Fixing may require multiple iterations
Keeps main conversation context clean

Wait for agent to complete. Review its summary of fixes applied.

Phase 6: Expert Review (Optional)

For significant features, offer parallel review:

"Feature complete. Run expert review before PR?"

If yes, spawn review agents in parallel (all use sonnet):

Task [simplicity-engineer] model: sonnet: "Review implementation for unnecessary complexity.
Files: [list of new/modified files]
See agents/review/simplicity-engineer.md"

Task [architecture-engineer] model: sonnet: "Review implementation for architectural concerns.
Files: [list of new/modified files]
See agents/review/architecture-engineer.md"

Add a conditional third reviewer based on what was built:

If the implementation includes...	Also spawn
Auth, sessions, API keys, user data	security-engineer
Significant UI (components, pages)	senior-engineer
Database migrations, data models	data-engineer

Present findings as Socratic questions (see references/review-patterns.md). Blockers → fix → re-verify (max 2 cycles). Should-fix → fix if quick, otherwise note as follow-up.

Post-Completion: Doc Staleness Check

Before shipping, check if documentation may need updating:

Get modified files from git status
Check for existing docs — Glob for docs/**/*.md, docs/**/*.mdx, content/**/*.md
If docs exist — Search doc files for references to modified file paths, exported function names, or component names
If matches found — Ask user: "These docs reference code you just changed. Update them?"
If user says yes — Read each stale doc and the changed source. Rewrite only affected sections.

Phase 7: Ship

Ensure all tests pass:

pnpm test
pnpm lint

Create PR:

git push -u origin feature/<feature-name>

gh pr create --title "feat: <description>" --body "$(cat <<'EOF'
## Summary
- What was built
- Key decisions

## Testing
- [X] Unit tests added
- [X] E2E tests added (if applicable)
- [X] All tests passing

## Screenshots
[Include if UI changes]

## Design Doc
[Link to design doc]

## Implementation Plan
[Link to implementation plan]
EOF
)"

Report to user:

PR URL
Summary of what was built
Any follow-up items

Cleanup worktree (optional):

cd ..
git worktree remove .worktrees/<feature-name>

Phase 8: Cleanup

Kill orphaned subagent processes:

After spawning multiple build agents, some may not exit cleanly. Run cleanup:

scripts/cleanup-orphaned-agents.sh

This is especially important after parallel agent runs.

<when_to_stop> STOP and ask user when:

Test fails unexpectedly and debugger can't resolve
Implementation doesn't match plan
Stuck after 2 debug attempts
Plan has ambiguity
New requirement discovered
Security concern identified

Don't guess. Ask. </when_to_stop>

<tool_restrictions_reminder> REMINDER: You must NEVER call EnterPlanMode or ExitPlanMode at any point during this skill — not at the start, not after creating the plan, not before implementation, not at the end. This skill manages its own flow. All output goes directly to the user as normal messages. </tool_restrictions_reminder>