start-feature

name: start-feature description: | Start a new feature with multi-agent collaboration (Opus 4.6 + Agent Teams). Phase 1: Codebase understanding (Opus subagent 1M context + Claude user interaction). Phase 2: Parallel research & design (Agent Teams: Researcher + Architect). Phase 3: Plan synthesis & user approval. Implementation is handled separately by /team-implement. metadata: short-description: Feature kickoff with Agent Teams (Plan phase)

Start Feature

Feature kickoff skill leveraging Opus 1M context and Agent Teams.

Preflight: Update CLIs before starting — claude update && npm install -g @openai/codex@latest. Releases drift frequently (model names, flags, sandbox semantics).

Overview

This skill handles the planning phases (Phase 1-3). Implementation is done via /team-implement, and review via /team-review.

/start-feature <feature>    ← This skill (planning)
    ↓ After approval
/team-implement             ← Parallel implementation
    ↓ After completion
/team-review                ← Parallel review

Workflow

Phase 1: UNDERSTAND (Opus 1M context + Claude Lead)
  Opus subagent analyzes the codebase (1M context), Claude interacts with the user
    ↓
Phase 2: RESEARCH & DESIGN (Agent Teams — Parallel)
  Researcher (Opus) ←→ Architect (Codex) communicate bidirectionally for research and design
    ↓
Phase 3: PLAN & APPROVE (Claude Lead + User)
  Integrate research and design, create plan and get user approval

Phase 1: UNDERSTAND (Opus Subagent + Claude Lead)

Analyze the codebase with Opus subagent's 1M context while Claude interacts with the user.

Main orchestrator context is precious. Large-scale codebase analysis is delegated to Opus subagent (1M context).

Step 0: Read PROGRESS.md (always first)

Before anything else, if PROGRESS.md exists at the repository root, read it. It is the rolling summary of the latest 5 checkpoints (maintained by /checkpointing) and carries the most recent session context, in-progress work, and the "将来のアクション" (next actions) from prior sessions. Use it to ground the new feature in what already happened and to avoid re-deciding settled questions. If it is absent (fresh repo), skip this step.

Step 1: Analyze Codebase with Opus Subagent

Use a general-purpose subagent (Opus) to analyze the entire codebase:

# Via general-purpose subagent (recommended)
Task tool:
  subagent_type: "general-purpose"
  prompt: |
    Analyze this codebase comprehensively:
    - Directory structure and organization
    - Key modules and their responsibilities
    - Existing patterns and conventions
    - Dependencies and tech stack
    - Test structure

    Use Glob, Grep, and Read tools to explore the codebase thoroughly.

    Save analysis to .claude/docs/research/{feature}-codebase.md
    Return concise summary (5-7 key findings).

To supplement the subagent's analysis, Claude can use Glob/Grep/Read to inspect specific files.

Step 2: Requirements Gathering

Ask the user questions to clarify requirements:

1. Purpose: What do you want to achieve?
2. Scope: What to include / exclude?
3. Technical requirements: Specific libraries, constraints?
4. Success criteria: How do you determine completion?
5. Final design: What form should it take?

Step 3: Create Project Brief

Combine codebase understanding + requirements into a "Project Brief":

## Project Brief: {feature}

### Current State
- Architecture: {existing architecture summary}
- Relevant code: {key files and modules}
- Patterns: {existing patterns to follow}

### Goal
{User's desired outcome in 1-2 sentences}

### Scope
- Include: {list}
- Exclude: {list}

### Constraints
- {technical constraints}
- {library requirements}

### Success Criteria
- {measurable criteria}

This brief is passed to Phase 2 teammates as shared context.

Phase 2: RESEARCH & DESIGN (Agent Teams — Parallel)

Launch Researcher and Architect in parallel via Agent Teams with bidirectional communication.

Key difference from subagents: Teammates can communicate with each other. Researcher's findings change Architect's design, and Architect's requests trigger new research.

Team Setup

Create an agent team for project planning: {feature}

Spawn two teammates:

1. **Researcher** — Uses WebSearch/WebFetch for external research (Opus 1M context)
   Prompt: "You are the Researcher for project: {feature}.

   Your job: Research external information needed for this project.

   Project Brief:
   {project brief from Phase 1}

   Tasks:
   1. Research libraries and tools: usage patterns, constraints, best practices
   2. Find latest documentation and API specifications
   3. Identify common pitfalls and anti-patterns
   4. Look for similar implementations and reference architectures

   How to research:
   - Use WebSearch for comprehensive research:
     WebSearch: '{topic} best practices constraints recommendations'
   - Use WebFetch for targeted documentation lookup

   Save all findings to .claude/docs/research/{feature}.md
   Save library docs to .claude/docs/libraries/{library}.md

   Communicate with Architect teammate:
   - Share findings that affect design decisions
   - Respond to Architect's research requests
   - Flag constraints that limit implementation options

   IMPORTANT — Work Log:
   When ALL your tasks are complete, write a work log file to:
     .claude/logs/agent-teams/{team-name}/researcher.md

   Use this format:
   # Work Log: Researcher
   ## Summary
   (1-2 sentence summary of what you researched)
   ## Tasks Completed
   - [x] {task}: {brief description of findings}
   ## Sources Consulted
   - {URL or source}: {what was found}
   ## Key Findings
   - {finding}: {relevance to project}
   ## Communication with Teammates
   - → {recipient}: {summary of message sent}
   - ← {sender}: {summary of message received}
   ## Issues Encountered
   - {issue}: {how it was resolved}
   (If none, write 'None')
   "

2. **Architect** — Uses Codex CLI for design and planning
   Prompt: "You are the Architect for project: {feature}.

   Your job: Use Codex CLI to design the architecture and create implementation plan.

   Project Brief:
   {project brief from Phase 1}

   Tasks:
   1. Design architecture (modules, interfaces, data flow)
   2. Select patterns (considering existing codebase conventions)
   3. Create step-by-step implementation plan with dependencies
   4. Identify risks and mitigation strategies

   How to consult Codex:
   codex exec --model "${CODEX_MODEL:-gpt-5.5}" --sandbox read-only "{question}" 2>/dev/null

   Update .claude/docs/DESIGN.md with architecture decisions.

   Communicate with Researcher teammate:
   - Request specific library/tool research
   - Share design constraints that need validation
   - Adjust design based on Researcher's findings

   IMPORTANT — Work Log:
   When ALL your tasks are complete, write a work log file to:
     .claude/logs/agent-teams/{team-name}/architect.md

   Use this format:
   # Work Log: Architect
   ## Summary
   (1-2 sentence summary of what you designed)
   ## Tasks Completed
   - [x] {task}: {brief description of what was done}
   ## Design Decisions
   - {decision}: {rationale}
   ## Codex Consultations
   - {question asked to Codex}: {key insight from response}
   ## Communication with Teammates
   - → {recipient}: {summary of message sent}
   - ← {sender}: {summary of message received}
   ## Issues Encountered
   - {issue}: {how it was resolved}
   (If none, write 'None')
   "

Wait for both teammates to complete their tasks.

Why Bidirectional Communication Matters

Example interaction flow:

Researcher: "httpx has a connection pool limit of 100 by default"
    → Architect: "Need to add connection pool config to design"
    → Architect: "Also research: does httpx support HTTP/2 multiplexing?"
    → Researcher: "Yes, via httpx[http2]. Requires h2 dependency."
    → Architect: "Updated design to use HTTP/2 for the API client module"

Without Agent Teams (old subagent approach), this would require:

1. Researcher subagent finishes → returns summary
2. Claude reads summary → creates new Codex subagent prompt
3. Codex subagent finishes → returns summary
4. If Codex needs more info → another researcher subagent round

Agent Teams collapses this into a single parallel session with real-time interaction.

Phase 3: PLAN & APPROVE (Claude Lead)

Integrate Agent Teams results, create an implementation plan, and request user approval.

Step 1: Synthesize Results

Read outputs from Phase 2:

• .claude/docs/research/{feature}.md — Researcher findings
• .claude/docs/libraries/{library}.md — Library documentation
• .claude/docs/DESIGN.md — Architecture decisions

Step 2: Create Implementation Plan

Create task list using TodoWrite:

{
    "content": "Implement {specific task}",
    "activeForm": "Implementing {specific task}",
    "status": "pending"
}

Task breakdown should follow references/task-patterns.md.

Step 3: Update CLAUDE.md

Append the project context to CLAUDE.md Zone C (the area below the @orchestra:repo-boundary marker box) for cross-session persistence. Never touch Zone A or Zone B.

If @orchestra:repo-boundary is missing, ask the user to run ./scripts/update.sh first — the updater migrates legacy layouts automatically.

---

## Current Project: {feature}

### Context
- Goal: {1-2 sentences}
- Key files: {list}
- Dependencies: {list}

### Architecture
- {Key architecture decisions from Architect}

### Library Constraints
- {Key constraints from Researcher}

### Decisions
- {Decision 1}: {rationale}
- {Decision 2}: {rationale}

Step 4: Present to User

Present the plan to the user:

## Project Plan: {feature}

### Codebase Analysis
{Key findings from Phase 1 — 3-5 bullet points}

### Research Findings (Researcher)
{Key findings — 3-5 bullet points}
{Library constraints and recommendations}

### Design Direction (Architect)
{Architecture overview}
{Key design decisions with rationale}

### Task List ({N} items)
{Task list with dependencies}

### Risks and Considerations
{From Architect's analysis}

### Next Steps
1. Shall we proceed with this plan?
2. After approval, you can start parallel implementation with `/team-implement`
3. After implementation, run parallel review with `/team-review`

---
Shall we proceed with this plan?

Output Files

Tips

• Phase 1: Opus subagent (1M context) analyzes the codebase while Claude interacts with the user
• Phase 2: Agent Teams bidirectional communication allows Researcher (Opus) and Architect (Codex) to influence each other
• Phase 3: After plan approval, proceed to parallel implementation with /team-implement
• Ctrl+T: Toggle task list display
• Shift+Up/Down: Navigate between teammates (when using Agent Teams)