Wave Execution Skill

Wave Execution Skill

Purpose: Context-efficient parallel execution with fresh agent contexts.

Implements 4 GSD patterns:

1. Context rot prevention (orchestrator ~15%, agents 100% fresh)
2. Wave-based parallelism (automatic dependency grouping)
3. Structured state (STATE.md <100 lines)
4. Atomic commits (1 task = 1 commit)

---

Quick Start

/wave:init          # Initialize STATE.md for current project
/wave:plan <tasks>  # Analyze tasks, group into waves
/wave:execute       # Execute waves with parallel agents
/wave:status        # Show current state

---

1. Context Rot Prevention

The Problem

As context fills, Claude's output quality degrades. Long sessions = worse code.

The Solution

Orchestrator: ~15% context (~30K on 200K model)
  └── Spawns fresh agents per task
      └── Each agent: 100% fresh 200K context

Orchestrator Rules

LOAD ONLY:

• STATE.md (current position, <100 lines)
• Task definitions (what to execute)
• Dependency graph (what order)

NEVER LOAD:

• Full codebase
• Historical context
• Previous execution details

DELEGATE EVERYTHING:

• Code reading → subagent
• Implementation → subagent
• Testing → subagent
• Each gets fresh context

Agent Spawn Pattern

// Orchestrator spawns with MINIMAL context
Task({
  prompt: `
    <objective>Execute: ${task.name}</objective>
    <context>
      @STATE.md
      @${task.planFile}
    </context>
    <success_criteria>${task.criteria}</success_criteria>
    <on_complete>
      1. Commit changes: git commit -m "${task.commitMessage}"
      2. Return: TASK COMPLETE + files changed
    </on_complete>
  `,
  subagent_type: task.agentType || "senior-fullstack-developer"
})

---

2. Wave-Based Parallelism

Concept

Group independent tasks into waves. Execute wave in parallel. Next wave after completion.

Wave 1: [Task A, Task B, Task C]  ← No dependencies, run parallel
         ↓ wait for all
Wave 2: [Task D, Task E]          ← Depend on Wave 1, run parallel
         ↓ wait for all
Wave 3: [Task F]                  ← Depends on D and E

Dependency Analysis

tasks:
  - id: setup-db
    wave: 1
    depends: []

  - id: create-api
    wave: 1
    depends: []

  - id: add-auth
    wave: 2
    depends: [setup-db, create-api]

  - id: write-tests
    wave: 3
    depends: [add-auth]

Parallel Execution Pattern

// Execute all Wave 1 tasks in SINGLE message with multiple Task calls
// This runs them truly parallel

for (const wave of waves) {
  // Spawn ALL tasks in wave simultaneously
  const tasks = wave.tasks.map(task =>
    Task({
      prompt: fillTemplate(task),
      subagent_type: selectAgent(task)
    })
  );

  // All tasks in same message = parallel execution
  // Wait for all to complete before next wave
}

Wave Assignment Rules

Condition	Wave
No dependencies	Wave 1
Depends on Wave N tasks	Wave N+1
Depends on multiple waves	Max(dependency waves) + 1

---

3. Structured State (STATE.md)

Purpose

Session continuity without context bloat. Always <100 lines.

Template

# Project State

## Quick Reference
**Project:** [name]
**Core goal:** [one line]
**Current:** Phase [X], Task [Y]
**Last:** [YYYY-MM-DD] - [what happened]

## Position
- Phase: [X] of [N] - [phase name]
- Task: [Y] of [M] - [task name]
- Status: [planning|executing|blocked|complete]
- Progress: [████░░░░░░] 40%

## Active Context
**Current wave:** [N]
**In progress:** [task names]
**Blocked by:** [none | blocker description]

## Decisions (recent 5)
| Date | Decision | Rationale |
|------|----------|-----------|
| ... | ... | ... |

## Deferred
- [ ] [ISS-001] [description] (from phase X)

## Resume
**Last session:** [timestamp]
**Stopped at:** [description]
**Continue with:** [next action]

Size Rules

Section	Max Lines
Quick Reference	5
Position	6
Active Context	5
Decisions	10 (keep recent 5)
Deferred	10 (reference ISSUES.md if more)
Resume	5
TOTAL	<50 lines ideal, <100 max

Update Triggers

Event	Update
Task complete	Position, Progress
Decision made	Add to Decisions (trim old)
Issue found	Add to Deferred
Session end	Resume section
Phase complete	Clear completed, update Position

---

4. Atomic Commits

Rule

1 Task = 1 Commit. No exceptions.

Commit Format

<type>(<scope>): <description>

- [x] <task that was completed>

Files: <list of changed files>

Types

Type	When
feat	New functionality
fix	Bug fix
refactor	Code restructure
test	Add/update tests
docs	Documentation
chore	Maintenance

Examples

# Good - atomic
git commit -m "feat(auth): add JWT token validation

- [x] Implement token verification middleware

Files: src/middleware/auth.ts, src/utils/jwt.ts"

# Bad - multiple tasks
git commit -m "add auth and fix bugs and update tests"

Agent Commit Instructions

Include in every agent prompt:

<on_complete>
After completing the task:
1. Stage only files for THIS task
2. Commit with format: <type>(<scope>): <task description>
3. Include task checkbox and file list in commit body
4. Return commit hash in completion message
</on_complete>

---

Integration with Existing Setup

With Decision Logs (GitHub Issues)

STATE.md = Fast session state (local, <100 lines)
Decision Log = Permanent record (GitHub, detailed)

Flow:
1. Session start → Read STATE.md (instant context)
2. Session work → Update STATE.md after each task
3. Session end → Sync important decisions to GitHub log

With agent-context-config.json

{
  "wave-orchestrator": {
    "defaultContext": "32K",      // Stays lean
    "maxAutoEscalation": "64K",   // Never needs more
    "specialRules": {
      "orchestration": "32K"       // Force minimal
    }
  }
}

With Specialized Agents

Wave execution uses YOUR agent types:

• senior-fullstack-developer for implementation
• qa-testing-engineer for tests
• backend-specialist for API work
• etc.

NOT generic "general-purpose" like GSD.

---

Commands

/wave:init

Initialize STATE.md for current project.

/wave:plan

/wave:plan "
  1. Setup database schema
  2. Create API endpoints
  3. Add authentication (needs 1,2)
  4. Write integration tests (needs 3)
"

Output: Wave assignment + dependency graph

/wave:execute

Execute planned waves with parallel agents.

/wave:status

Show STATE.md content + execution progress.

---

Example Workflow

# 1. Initialize state
/wave:init

# 2. Plan tasks with dependencies
/wave:plan "
  - Create user table [db]
  - Create product table [db]
  - User API endpoints [api, needs: user table]
  - Product API endpoints [api, needs: product table]
  - Auth middleware [auth, needs: user api]
  - Integration tests [test, needs: auth]
"

# Output:
# Wave 1: [user-table, product-table]     ← parallel
# Wave 2: [user-api, product-api]         ← parallel
# Wave 3: [auth-middleware]
# Wave 4: [integration-tests]

# 3. Execute
/wave:execute

# Orchestrator spawns Wave 1 tasks parallel
# Waits for completion
# Spawns Wave 2 tasks parallel
# ... continues until done

---

5. Two-Stage Review (NEW)

Every task completion goes through TWO review stages before proceeding.

Stage 1: Spec Compliance Review

Purpose: Catch over-building and spec misses.

Questions:

• Does the code do EXACTLY what was specified?
• Is there code that wasn't asked for?
• Are there missing requirements?

Reviewer prompt:

You are a SKEPTICAL spec compliance reviewer.

Compare the implementation against the original task spec:
- Task: {task_description}
- Files changed: {files}

Check for:
1. MISSING: Required functionality not implemented
2. EXTRA: Code added that wasn't requested (YAGNI violation)
3. WRONG: Implementation doesn't match spec

Be STRICT. If anything is off, report it.

Respond with:
- PASS: Spec fully met, nothing extra
- FAIL: [List specific issues]

Stage 2: Code Quality Review

Purpose: Ensure code is maintainable and tested.

Only runs if Stage 1 passes.

Reviewer prompt:

You are a SENIOR code quality reviewer.

Review for:
1. CODE CLARITY: Readable? Good naming? Comments where needed?
2. TEST COVERAGE: Are there tests? Do they test behavior not implementation?
3. ERROR HANDLING: Graceful failures? No silent errors?
4. PERFORMANCE: Any obvious bottlenecks?
5. SECURITY: Input validation? No hardcoded secrets?

Respond with:
- PASS: Code is production-ready
- FAIL: [List specific issues with severity]

Review Loop

Task Complete
    ↓
Stage 1: Spec Compliance
    ↓
  FAIL? → Return to implementer → Fix → Re-review
    ↓
  PASS
    ↓
Stage 2: Code Quality
    ↓
  FAIL? → Return to implementer → Fix → Re-review
    ↓
  PASS
    ↓
Commit & Proceed to Next Task

Implementation in Wave Execution

// After task agent returns
const specReview = await Task({
  prompt: `Review for spec compliance: ${task.spec} vs ${result.changes}`,
  subagent_type: "qa-testing-engineer"
});

if (specReview.verdict !== 'PASS') {
  // Send back to implementer with feedback
  return retryTask(task, specReview.feedback);
}

const qualityReview = await Task({
  prompt: `Review code quality: ${result.changes}`,
  subagent_type: "qa-testing-engineer"
});

if (qualityReview.verdict !== 'PASS') {
  // Send back to implementer with feedback
  return retryTask(task, qualityReview.feedback);
}

// Both passed - commit and continue
await commitTask(task, result);

Why Two Stages?

Stage 1 catches:

• Feature creep
• Missing requirements
• Wrong implementations
• Scope changes

Stage 2 catches:

• Poor code quality
• Missing tests
• Security issues
• Performance problems

Single reviewer misses things. Two specialized reviewers catch more.

---

6. Quality Gates Summary

Gate	When	What It Checks
TDD	Before implementation	Test exists and fails
Spec Review	After implementation	Matches requirements exactly
Quality Review	After spec approval	Code is production-ready
Commit	After both reviews	Atomic, conventional format

All gates must pass. No shortcuts.