Execute Plan

---
name: execute-plan
description: Plan execution workflow - parallel SC implementation, worktree mode, verification patterns, GPT delegation. Use for executing plans with TDD + Ralph Loop.
---

# SKILL: Execute Plan (Plan Execution Workflow)

> **Purpose**: Execute plans using TDD + Ralph Loop with parallel execution
> **Target**: Coder Agent implementing Success Criteria from plans

**When to Use**: Execute plans from `/01_confirm`, implement SCs with TDD, parallel execution of independent SCs, worktree mode for isolated development

**Quick Reference**:
```bash
PROJECT_ROOT="$(pwd)"
PLAN_PATH="$(find "$PROJECT_ROOT/.pilot/plan/pending" "$PROJECT_ROOT/.pilot/plan/in_progress" -name "*.md" -type f 2>/dev/null | sort | head -1)"
Task: subagent_type: $AGENT_TYPE, prompt: "Execute SC-1 from $PLAN_PATH"
```

**Scope**: Plan detection, SC dependency analysis, parallel execution, worktree mode, GPT delegation. TDD/Ralph/Vibe → separate skills.

---

## Execution Steps

**⚠️ EXECUTION DIRECTIVE**: Execute ALL steps IMMEDIATELY. NEVER move plan to done/ (only `/03_close` has this authority), do NOT call close-plan automatically, plan MUST remain in `.pilot/plan/in_progress/`

---

## ⛔ MAIN ORCHESTRATOR RESTRICTIONS (ABSOLUTE)

**FORBIDDEN** (orchestrator direct use prohibited):
- Edit tool: Code file modifications
- Write tool: Code file creation
- Direct SC implementation without Task tool

**MANDATORY** (must delegate via Task tool):
- Sequential: `Task: subagent_type: coder, prompt: "SC-N..."`
- Parallel: Multiple Task calls in SAME response for ParallelGroup
- Even 1-2 SCs: Always delegate (Context Protection)

**WHY**: Subagent isolation provides 50-80% context savings (CLAUDE.md:58-59)

---

## Step 1: Plan Detection & SC Extraction

**⚠️ CRITICAL**: Always use absolute path based on Claude Code's initial working directory.

**Full Implementation**: See @REFERENCE.md - Step 1: Plan Detection (Full Bash)

---

## Step 2.7: Pre-Execution Confidence

**Purpose**: Evaluate confidence before complex SCs. If < 0.5 → consult GPT Architect (rubric: @.claude/skills/gpt-delegation/SKILL.md).

**Full Implementation**: See @REFERENCE.md - Step 2.7: Pre-Execution Confidence (Full Bash)

---

**⚠️ REMINDER**: All SC execution MUST use Task tool. Main orchestrator NEVER uses Edit/Write directly.

## Step 3: Execute with Ralph Loop (Per-SC Agent Selection)

**Dependency Analysis, File Extraction & Agent Selection** (supports both `### SC-N:` and `- [ ] **SC-N**` formats):

**Full Implementation**: See @REFERENCE.md - Step 3: Dependency Analysis (Full Bash)

**Test Type Detection Rules**: Path-based (`**/e2e/**`, `**/integration/**`, `**/*.e2e.*`) + Keyword-based (e2e, integration, playwright, cypress) + Script-based (package.json script name). **Fail-Safe**: Defaults to `unit` (parallel allowed, safe with `--maxWorkers=50%` from SC-1). **Routing**: `e2e` → Sequential (environment-bound, stateful), `unit` → Parallel allowed.

**Smart Grouping**: When SCs follow Atomic SC Principle (@.claude/skills/spec-driven-workflow/SKILL.md), parallel execution naturally emerges. SCs modifying same file type automatically group for specialized agents (e.g., frontend-engineer for `src/components/*`, backend-engineer for `src/api/*`). Test type detection automatically groups E2E SCs sequentially (path: `**/e2e/**`, `**/integration/**`, `**/*.e2e.*`; keywords: e2e, integration, playwright, cypress).

**Execution Strategies**: Parallel (Independent SCs, 50-70% speedup): `Task: subagent_type: $SC_AGENT, prompt: "Execute SC-{N} from $PLAN_PATH. Skills: tdd, ralph-loop, vibe-coding. Output: <CODER_COMPLETE> or <CODER_BLOCKED>"` | Sequential (Dependent SCs): One agent with all SCs | Single Coder (1-2 SCs): Always delegate

**Note**: `$SC_AGENT` is selected per-SC (not per-plan), so parallel SCs may use different specialized agents (e.g., SC-1 → frontend-engineer, SC-2 → backend-engineer).

**Process Results**: Check `<CODER_COMPLETE>`, run `npm test`, mark complete or retry. Quality Gates: Tests pass, Coverage ≥80%, Type-check clean, Lint clean.

**Handle CODER_BLOCKED**: Delegate to GPT Architect (gpt-5.2, workspace-write, reasoning_effort=medium). Fallback: Continue with Claude.

```bash
if grep -q "<CODER_BLOCKED>" /tmp/coder_output.log 2>/dev/null; then
    if command -v codex &> /dev/null; then
        codex exec -m gpt-5.2 -s workspace-write -c reasoning_effort=medium --json "$ARCHITECT_PROMPT"
    fi
fi
```

**TaskCreate/TaskUpdate Pattern**: Create TaskList entry per SC before execution. Update on completion with `<CODER_COMPLETE>`. Progress tracking: `Task: status=in_progress, id=SC-N` → Execute → `TaskUpdate: status=completed, id=SC-N`.

**No-Excuses Enforcement**: PROHIBITED phrases ("I cannot", "Too complex", "Out of scope"). Required pattern: "To achieve X, I will: [alternative]" or "Breaking down into: [steps]". Exception: User explicit abort only. Details: @REFERENCE.md

**Mandatory Oracle Consultation** (mandatory_oracle_consultation): Per-SC execution (GPT Role: Engineer), Blocker resolution (GPT Role: Problem Solver). Pattern: `codex exec -m gpt-5.2 -s read-only -c reasoning_effort=medium`. Graceful fallback if unavailable. Full table: @REFERENCE.md

**No-Partial-Delivery Policy**: No "simplified version", "basic implementation", "you can extend later". On BLOCKED: 1) GPT delegation, 2) User collaboration (not "give up"). All SCs must be fully completed.

---

## Step 3.5: Per-SC TODO Verification (BLOCKING)

After each SC execution, verify all TODO checkboxes are marked `[x]`. If unchecked items remain, BLOCKING until coder completes them. Full implementation: @REFERENCE.md - Per-SC TODO Verification

---

## Step 3.9: Final TODO Sweep (BLOCKING)

Before E2E verification, sweep entire plan for unchecked TODOs. Max 50 retries with GPT escalation at 10, user escalation at 50. Full implementation: @REFERENCE.md - Final TODO Sweep

---

## Step 4: Completion & E2E Verification

All SCs complete → E2E verification. Detect project type: Web (Chrome in Claude), CLI (exit code/stdout), Library (tests). Retry max 3 → GPT → user. Full: @REFERENCE.md

---

## Further Reading

**Internal**: @.claude/skills/execute-plan/REFERENCE.md | @.claude/skills/tdd/SKILL.md | @.claude/skills/ralph-loop/SKILL.md | @.claude/skills/gpt-delegation/SKILL.md

**External**: [Test-Driven Development](https://www.amazon.com/Test-Driven-Development-Kent-Beck/dp/0321146530) | [Working Effectively with Legacy Code](https://www.amazon.com/Working-Effectively-Legacy-Michael-Feathers/dp/0131177052)

---

**Version**: claude-pilot 4.4.14