Feature Lifecycle

Overview

Turn feature ideas into designs and implementations. Full lifecycle in one skill: brainstorm → design → implement → review → finish.

Do NOT explore the entire codebase upfront. Read architecture docs (if configured in project.yml) for context, then ask which services are affected. Only explore those specific service directories.

Config Loading

Before anything else, load project config:

1. •Check if .project-orchestrator/project.yml exists
2. •If yes: parse and extract services, plans_dir, plans_structure, architecture_docs, implementation, models, brainstorm, review
3. •If no: use defaults (monorepo, single service at root, docs/plans/ flat structure)
4. •Read architecture docs if paths configured: config.architecture_docs.agent, config.architecture_docs.domain
5. •If architecture docs don't exist, skip — proceed without system topology context

Config validation:

• •brainstorm.default_depth must be shallow, medium, or deep — error otherwise
• •models.explorer and models.implementer must be opus, sonnet, or haiku
• •Warning (not error) if brainstorm.perspective_docs key not in brainstorm.designer_perspectives
• •review.max_fix_iterations must be 1-5 — error: "max_fix_iterations must be 1-5"
• •review.fix_timeout_turns must be 5-30 — error: "fix_timeout_turns must be 5-30"

Review config defaults (v2):

Agent Preferences

• •Never run agents in background — always foreground
• •Team naming: brainstorm-{feature-slug} / implement-{feature-slug}
• •No peer messaging — all team members report to lead only

Living State Document

All important state lives on disk at {config.plans_dir}/YYYY-MM-DD-{slug}-design.md. This is the source of truth that survives compaction. Any agent can read it to get full context.

Directory lifecycle (when config.plans_structure is standard):

• •{plans_dir}/ (root) = active only (designing / approved / in-progress)
• •{plans_dir}/completed/ = shipped
• •{plans_dir}/backlog/ = scoped but not scheduled
• •{plans_dir}/ideas/ = explorations, no commitment

When config.plans_structure is flat, all plans stay in {plans_dir}/ regardless of status.

When status changes (standard structure), move the file to the matching directory and update INDEX.md.

Strict template:

# {Feature Name} — Design & Implementation

## Status: brainstorming | designing | approved | implementing | reviewing | complete

## Design

### Feature Type
{endpoint/scraper/UI/cross-service/infra}

### Services Affected
{list of services}

### Design Details
{design content using templates from Design Templates section below}

## Implementation Tasks

| # | Task | Service | Status | Assignee | Spec | Quality | Fix Iterations |
|---|------|---------|--------|----------|------|---------|----------------|
| 1 | ... | backend | pending | — | — | — | — |
| 2 | ... | frontend | in-progress | implement-auth | — | — | — |
| 3 | ... | backend | complete | implement-api | ✅ | ✅ | 0 |

## Implementation Log

### Task N — {title}
- **Files changed:** {service}/{path} — {what changed}
- **Commit:** {short sha}
- **Spec review:** ✅ / ❌ {notes}
- **Quality review:** ✅ / ❌ {notes}
- **Recovery:** NO_CHANGES (attempt 1), STALLED (attempt 2)  ← optional, only when recovery occurred

## Test Plan              ← optional, added during brainstorming if UI feature has browser tests
## Test Results           ← optional, filled by /project:test

## Decisions & Context
{Key decisions, gotchas, patterns chosen — helps fresh agents after compaction}

Update this document at every significant step. It is the coordination artifact.

Phase 1: Brainstorm & Design

1. Minimal context gathering

• •Read architecture docs if configured in project.yml (config.architecture_docs.agent, config.architecture_docs.domain)
• •Include relevant sections in every spawned agent's prompt (subagents don't inherit lead context)
• •Do NOT run broad codebase exploration
• •Ask one question at a time, prefer multiple choice

2. Scope the services FIRST

Which service(s) does this affect? Present the list from config.services[].name. If no services configured, ask the user to describe the affected areas.

3. Classify the feature type

• • New endpoint/API
• • UI feature
• • Cross-service data flow
• • Configuration/infra
• • Other

4. Choose exploration depth

• • Shallow: Read architecture docs, 1-2 key files → simple changes
• • Medium: Explore 1 service directory patterns → typical features
• • Deep: Multiple service exploration, existing impl review → cross-cutting changes

Default: config.brainstorm.default_depth (default: medium). User can always override.

5. Team decision

services <= 2 AND depth = shallow → NO TEAM (regular Task agents)
services <= 2 AND depth = medium  → NO TEAM (regular Task agents)
services >= config.brainstorm.team_threshold (default: 3) OR depth = deep → TEAM (brainstorm-{slug})

6a. Team exploration (when team triggered)

TeamCreate("brainstorm-{slug}")
TaskCreate per affected service
Spawn explorer agent per service (team_name, name: "explore-{service}")
  → Prompt: user's problem/feature description + which service directory to focus on
  → Include architecture doc excerpts relevant to that service
  → Agent explores service dir, existing patterns, recent changes
  → Reports via TaskUpdate (mark completed)
  → No peer messaging
Lead waits for all exploration tasks completed

Each explorer captures:

• •Existing patterns relevant to the feature
• •API endpoints / routes affected
• •Database tables / schemas involved
• •Event pipelines published or consumed
• •Tech debt or gotchas

6b. Simple exploration (no team)

• •Spawn 1-2 explorer agents via Task tool (foreground)
• •Each agent prompt: user's problem/feature description + which service directory to focus on
• •Include architecture doc excerpts if available
• •Only explore specific service directories identified
• •Check recent commits, existing patterns

7. Cross-service synthesis (team path only)

Lead synthesizes after all explorers report:

• •Shared patterns across services
• •API contracts between services
• •Event schemas needing changes
• •Conflicting patterns to resolve
• •Shared database tables
• •Migration ordering

8. Multi-perspective design

Team path: Spawn design agents (opus) as team members — one per perspective from config.brainstorm.designer_perspectives (default: ["simplicity", "scalability"]):

For each perspective:

• •If config.brainstorm.perspective_docs has a mapping for this perspective name, Read the mapped file and inject its content into the designer agent's prompt
• •If no mapping exists, spawn with perspective name only (no doc injection — not an error)

Each produces a proposal via TaskUpdate. Lead synthesizes one recommendation.

Simple path: Propose 2-3 approaches with trade-offs. Lead with recommended option.

8a. Test plan generation (optional)

If the feature includes UI changes, ask: "Do you want to define browser test scenarios for this feature?"

• •If no: skip — user can add test scenarios later or not at all
• •For non-UI features: skip this step (browser tests don't apply)
• •If yes: run the test reconnaissance phase before writing scenarios:

Step 1 — Route discovery: Grep the codebase for route definitions relevant to the designed feature:

• •React Router: path=, Route, createBrowserRouter
• •Next.js: app/ directory structure, page.tsx files
• •Vue Router: routes:, path:
• •Other frameworks: ask user for routing convention

Document discovered routes in ### App Context section of the test plan.

Step 2 — Authentication analysis: Search for auth patterns in the codebase:

• •Login pages/components (grep for "login", "signin", "auth")
• •Auth middleware/guards (session checks, JWT verification, protected routes)
• •Test credentials (seed scripts, fixtures, .env.example, test helpers)
• •Session mechanism (cookies, localStorage tokens, auth headers)

Ask user to confirm findings and provide test credentials if not found in code. Document in ### Authentication section.

Step 3 — Page structure recon (if app is running): Ask user: "Is the app running? I can take snapshots of key pages to understand the UI structure."

• •If yes: use Chrome DevTools MCP to navigate_page + take_snapshot on each relevant route
  • •Capture form fields (input names, labels, placeholders)
  • •Capture navigation elements (links, buttons, menus)
  • •Capture content areas (headings, lists, data displays)
  • •Document element selectors/labels that test scenarios should target
• •If no: derive UI structure from component code (read relevant component files)
  • •Look for form field names, button text, heading content
  • •Less reliable but still useful for writing scenarios

Step 4 — Generate test scenarios: Using the reconnaissance findings, generate concrete scenarios with:

• •Real URLs from route discovery (not placeholder paths)
• •Real form field labels/names from page structure recon
• •Real auth flow from authentication analysis
• •Scenarios: happy path first, then key error cases, then edge cases
• •Each scenario references actual elements discovered (e.g., "Fill the 'Email address' input" not "Fill email field")

Step 5 — Prerequisites and setup: Document what's needed before /project:test can run:

• •App start command (from package.json scripts, Makefile, docker-compose)
• •Seed data requirements (from seed scripts or fixtures found)
• •Environment variables needed (from .env.example or config files)

Write the complete ## Test Plan section in the living state doc using the test plan template (see Design Templates below).

9. Contract verification (cross-service only, team path)

When 2+ services share an API boundary or event:

• •Producer check — verify endpoint/event producer matches proposed contract
• •Consumer check — verify consumer aligns
• •Event schema check — verify topology, payloads

All report via TaskUpdate. Skip for single-service features.

10. Write design to living state document

• •

  Create {config.plans_dir}/YYYY-MM-DD-{slug}-design.md using strict template

• •

  Set status to designing

• •

  Fill in Design section using templates below

• •

  If config.plans_structure is standard: add entry to {config.plans_dir}/INDEX.md under "## Active" Format: - [{Feature Name}]({filename}) — {one-line summary} Example: - [Orchestrator Fixes](2026-02-16-orchestrator-impl-fixes-design.md) — Improve agent completeness gates and task messaging

  If INDEX.md doesn't exist, create it with this structure:

  Plans Index

  Active

  • •{Feature Name} — {one-line summary}

  Completed

  Backlog

  Ideas

  If "## Active" section exists, add the entry at the top of that section (newest first).

• •

  Present design to user

11. Cleanup brainstorm team

• •Team path: TeamDelete("brainstorm-{slug}")
• •Suggest user run /clear to free context
• •Suggest user run /project:implement to start parallel implementation
• •Suggest /project:progress anytime to check status

Design phase ends here. Implementation is handled by the /project:implement command.

Implementation is handled by /project:implement. See that command for details.

Design Templates

For API Features

### Endpoint Design
- **Method + path:** GET/POST/etc /api/v1/...
- **Request payload:** (required/optional fields)
- **Response shape:** { field: type, ... }
- **Error cases:** 400/401/404/500 scenarios
- **Consumers:** Which services call this?

### Data Flow
[ASCII diagram showing service→service flow]

For UI Features

### Component Structure
- **Hierarchy:** Page → Container → Components
- **State management:** local vs context vs query cache
- **API dependencies:** Which endpoints?
- **Error/loading states:** How handled?

Cross-Service Contract (when applicable)

### Service Contracts
| Producer | Consumer | Contract | Notes |
|----------|----------|----------|-------|
| service-a POST /api/v1/... | service-b client.post(...) | { field: type } | ... |

### Event Pipelines
| Event | Exchange | Producer | Consumers | Payload |
|-------|----------|----------|-----------|---------|
| ... | ... | ... | ... | { ... } |

### Migration Order
1. Database migrations first
2. Producers next (API endpoints, event publishers)
3. Consumers last (frontend, admin, downstream services)

For Backend Services

### Backend Design
- **Feature flag:** Which config flag controls this?
- **Async pattern:** Virtual threads / reactive / message queue?
- **Caching:** What cache? TTL? Invalidation trigger?
- **Scheduler:** Needs periodic task? Interval?
- **Config:** Application config changes needed?

Check the target service's CLAUDE.md for stack-specific design patterns and conventions.

For Browser Test Scenarios

#### T{N} — {Scenario title}
**Precondition:** {Prior scenario state or "none — self-contained"}
**Setup:** {Navigation/login steps if self-contained}
**Steps:**
1. {Action — navigate, click, fill, etc.}
2. {Action}

**Expected:**
- {Visible element, text, URL, or state}
- {Another expected outcome}

Checklists

Cross-Service Checklist

• • API contract changes documented?
• • Event schema changes?
• • Shared types need updating?
• • Which service handles errors?
• • Migration/deploy ordering defined?
• • Auto-deploy services (from config.services[].auto_deploy) — coordinate PR merge order

Common Pitfalls

Check the project's CLAUDE.md and service-level CLAUDE.md files for project-specific pitfalls.

• •Verify APIs exist before designing against them

Parallel Task Design — Avoid File Collisions

When breaking work into implementation tasks, check if multiple tasks will edit the same file. Parallel agents editing one file causes race conditions.

During design, annotate shared files in the task table:

| # | Task | Shared Files | Strategy |
|---|------|-------------|----------|
| 5 | Auto-populate UI | operations/index.tsx | isolation |
| 6 | Source picker | operations/index.tsx | isolation |
| 7 | Dry-run button | operations/index.tsx | isolation |
| 8 | Integration | operations/index.tsx | sequenced after 5-7 |

Design Smells (pause if any apply)

Related Skills

Key Principles

• •Lazy exploration — only explore what's needed, when scoped
• •One question at a time
• •Service scope early — know which services before exploring
• •YAGNI ruthlessly — check design smells
• •Check service CLAUDE.md — for stack-specific patterns and conventions
• •Cross-service awareness — document contracts and event schemas
• •Living state on disk — design doc in {plans_dir}/ is source of truth, survives compaction
• •Max parallel implementers — from config.implementation.max_parallel (default 3)
• •Fresh reviewers — spec and quality reviewers are one-shot Task agents, not team members (compaction-safe)
• •Suggest /clear between phases — brainstorm→implement is natural context boundary
• •Teams for breadth — use teams for 3+ services or deep cross-cutting; simple agents otherwise
• •Always clean up teams — TeamDelete after every team phase
• •Design doc is the handoff — the living state doc is the artifact, not the team state