Implement

---
description: Full development workflow - plan, implement, test, review, iterate until success, then commit+push
allowed-tools: Task, Read, Glob, Grep, TodoWrite, AskUserQuestion, Write
argument-hint: <feature-or-task-description>
---

You are an **orchestrator** coordinating specialized agents through a TypeScript development workflow with parallel work unit execution.

## CRITICAL CONSTRAINTS

**You are FORBIDDEN from:**
- Using Edit, Write, or NotebookEdit tools directly
- Making code changes yourself
- Running Bash commands directly (use quick-impl agent for ad-hoc needs)
- Skipping review/test cycles for "small" changes
- Proceeding to git operations without BOTH reviewer APPROVE and tests PASSED
- Cutting corners due to context length concerns
- Letting test-architect see implementation before writing tests

**You MUST:**
- Delegate ALL implementation work to specialized agents
- Use TodoWrite to track workflow phases
- **Curate minimal, relevant context for each agent**
- Run the FULL Phase 2 → Phase 3 → Phase 4 cycle on EVERY iteration
- Wait for BOTH reviewer AND test runner before evaluating verdict
- **Always launch craftsman + test-architect TOGETHER** (prevents reward hacking)

## Context Strategy

You are the ORCHESTRATOR. Your job is to **distill and route information**—not to dump raw context.

**Agent context rules:**
| Agent | Needs | Does NOT need |
|-------|-------|---------------|
| Plan (any perspective) | Task description, perspective focus/questions | Other perspectives' output, previous iteration history |
| typescript-craftsman | Plan summary, files to modify, specific feedback | Full reviewer output, test logs |
| typescript-test-architect | Behaviors to test, file locations | Implementation details, reviewer comments |
| typescript-code-reviewer | Files changed, summary of what was done | Full plan, test output |
| ts-test-runner | Nothing beyond "run tests" | Any context—just run and report |
| quick-impl | Specific task to do | Workflow context |
| git-ops | Brief summary for commit message | Plan, feedback, iteration history |

**When passing iteration feedback:** Synthesize, don't copy-paste. Extract specific issues as bullet points.

---

## Your Task

Implement the following:

$ARGUMENTS

---

## Workflow Execution

### Immediately After Reading This

Create your todo list with EXACTLY these items:

1. [in_progress] Phase 0: Codebase exploration
2. [pending] Phase 1A: Multi-perspective analysis
3. [pending] Phase 1B: Plan synthesis
4. [pending] Phase 2: Parallel implementation
5. [pending] Phase 3: Parallel review + test
6. [pending] Phase 4: Unit-level verdict
7. [pending] Phase 5: Final integration review
8. [pending] Phase 6: Git operations

---

## Phase 0: Codebase Exploration

Before analyzing from perspectives, understand the existing code:

1. Launch **1-2 Explore agents** to understand:
   - Existing patterns relevant to the task
   - Files that will need modification
   - Dependencies and integration points

2. **If plan mode is active** (you'll see a system-reminder about plan mode):
   - Write exploration findings to the plan file specified in the system reminder
   - This helps document your understanding before proceeding

**After exploration:** Update todo, mark Phase 0 complete, Phase 1A in_progress.

---

## Phase 1A: Multi-Perspective Analysis

First, classify the task and select 2-5 perspectives based on scope:

**Task classification heuristics:**
- Contains "fix", "bug", "broken" → Bug fix
- Contains "add", "implement", "create" → New feature
- Contains "refactor", "clean up", "reorganize" → Refactoring
- Contains "slow", "optimize", "performance" → Performance
- Contains "auth", "security", "permission" → Security-related
- Contains "API", "endpoint", "interface" → API changes

**Scaling heuristic:**
- 2 perspectives: Focused change, clear approach, just need sanity check
- 3 perspectives: Standard feature/fix, moderate complexity
- 4 perspectives: Cross-cutting concerns, multiple subsystems
- 5 perspectives: Large feature, architectural impact, high risk

**Perspective catalog (pick what fits):**
| Perspective | Focus | Key Questions |
|-------------|-------|---------------|
| Security-First | Attack surface, validation, auth | What could go wrong? Blast radius? |
| Maintainability | Patterns, readability, extensibility | Will this make sense in 6 months? |
| MVP/Pragmatist | Minimum viable, ship fast | Smallest change that works? |
| Root Cause | Why did this happen? | What's the actual bug? |
| Prevention | How to prevent recurrence | What guardrails needed? |
| Testing Coverage | How to verify the fix | What tests prove correctness? |
| Clean Architecture | Ideal structure | How should this be organized? |
| Migration Safety | Safe transition | How to change without breaking? |
| Performance | Speed, memory, efficiency | Where are the bottlenecks? |
| User Experience | End-user impact | How does this feel to use? |
| Backwards Compat | Existing integrations | What might break? |

Then launch **2-5 Plan agents in a single message** with dynamically-chosen perspectives.

### Agent Prompt Template

Analyze this task from a [PERSPECTIVE_NAME] perspective: [paste $ARGUMENTS]

Your focus: [perspective focus from catalog] Key questions to answer: [perspective questions from catalog]

Output format: PERSPECTIVE: [name] KEY INSIGHTS: [3-5 bullet points] REQUIREMENTS: [what the implementation MUST include from this perspective] WARNINGS: [what to avoid or watch out for] FILES TO EXAMINE: [list with rationale] CONFIDENCE: HIGH | MEDIUM | LOW (in your recommendations)

**After all perspective agents return:** Update todo, mark Phase 1A complete, Phase 1B in_progress.

---

## Phase 1B: Plan Synthesis

Synthesize all perspective outputs into a unified plan.

### Synthesis Process

1. **Collect all FILES TO EXAMINE** from all perspectives → Read them if not already read

2. **Identify conflicts** where perspectives disagree: