epci:implement-auto

Headless EPCI workflow for autonomous feature implementation via claude -p. Executes Explore-Plan-Code-Inspect without interaction. Produces incremental JSON output. 3-level circuit breaker protects a

SKILL.md

GitHub

implement-auto

Headless EPCI skill for autonomous feature implementation without user interaction. Functionally equivalent to /implement — same subagents, stack skills, and review depth — with zero breakpoints.

Quick Start

# With spec (full EPCI workflow, in-place)
claude --dangerously-skip-permissions -p "/implement-auto feature-slug @path/to/spec.md"

# With worktree isolation (isolated git worktree)
claude --dangerously-skip-permissions -p "/implement-auto feature-slug @path/to/spec.md --worktree"

# With plan (skip Explore + Plan phases)
claude --dangerously-skip-permissions -p "/implement-auto feature-slug @.claude/plans/feature-plan.md"

# With all quality gates disabled (fast mode)
claude --dangerously-skip-permissions -p "/implement-auto feature-slug @spec.md --skip-plan-validation --skip-review --skip-security --skip-qa"

Input

Parameter Required Description
feature-slug Yes Kebab-case feature identifier
@spec-path Yes* Path to spec/PRD file (Markdown). *Required unless @plan-path provided
@plan-path No Path to Claude Code plan (.claude/plans/*.md). Skips Explore + Plan phases
--worktree No Enable git worktree isolation. Default: disabled (work in-place)

Input Detection

INPUT
├── @.claude/plans/*.md → PLAN-FIRST workflow (skip E-P)
│   └─ Plan already done, go directly to CODE
├── @docs/specs/*.md or @*.md → SPEC-FIRST workflow (full E-P-C-I)
│   └─ Full Explore + Plan phases
└── No path → ERROR (spec or plan required in headless mode)

Output

JSON file at .implement-auto-output.json in the working directory. See references/output-json-schema.md.

Status values: SUCCESS, PARTIAL, FAILED.

MANDATORY EXECUTION RULES:

  • NEVER call AskUserQuestion (headless mode)
  • NEVER call EnterPlanMode
  • NEVER display ASCII breakpoint boxes
  • NEVER wait for user input
  • ALWAYS write JSON output at each step transition
  • ALWAYS follow step order sequentially
  • ALWAYS enforce TDD per component (RED-GREEN-REFACTOR)
  • ALWAYS apply circuit breaker rules on failure
  • ALWAYS load stack skills per file extension before implementation
  • FOCUS on autonomous execution from start to finish

Workflow

step-00-init-auto     Parse args, complexity detection, branch setup (+ worktree if --worktree), Feature Doc, JSON init
       |
       ├── @plan-path provided? → Skip to step-03-code-auto
       |
step-01-explore-auto  Explore codebase + sanity check (30% threshold)
       |
step-02-plan-auto     @planner + @plan-validator + implementation plan
       |
step-03-code-auto     TDD implementation + stack skills + circuit breaker + background reviewer
       |
step-04-review-auto   Self-review + @code-reviewer + @security-auditor + @qa-reviewer
       |
step-05-document-auto Feature Document completion + executive summary
       |
step-06-finish-auto   Final validation, commit, index.json update, status determination
       |
step-07-output-auto   Final JSON write
       |
step-08-publish-auto  Push branch, create PR, cleanup worktree (if --worktree)

EXECUTION PROTOCOLS:

  1. Load step-00-init-auto.md
  2. Execute current step protocols completely
  3. Write JSON output after step completion
  4. Evaluate circuit breaker conditions
  5. Proceed to next_step or abort if circuit breaker triggers

Subagents

Agent Model Step Default Skip Flag
Explore Haiku 01 Always -
@planner Sonnet 02 Always -
@plan-validator Opus 02 Always --skip-plan-validation
@code-reviewer Opus 03 (background) + 04 Always --skip-review
@security-auditor Opus 04 Conditional* --skip-security
@qa-reviewer Sonnet 04 Conditional** --skip-qa

* Triggered when auth/security patterns detected OR complexity is LARGE. ** Triggered when >3 acceptance criteria OR >5 components OR complexity is LARGE.

Stack Skills

Dynamic per-file loading via domain-mapping. See references/domain-mapping.md.

Extension Stack Skill
*.py python-django
*.php php-symfony
*.java java-springboot
*.tsx, *.jsx, *.ts, *.js javascript-react
*.css, *.scss, *.html frontend-editor

Each stack skill is loaded once per type (cached). Provides architecture patterns, testing conventions, and anti-patterns.

Circuit Breaker (3 Levels)

Level Scope Threshold Action
1 Component 2 retries Mark FAILED, skip to next
2 Task 3 consecutive OR >50% ABORT task, status FAILED
3 Timeout External (orchestrator) Kill process, read partial JSON

See references/circuit-breaker-rules.md.

TDD Rules

Self-contained TDD workflow: RED -> GREEN -> REFACTOR -> VERIFY. Coverage targets adapt to complexity: STANDARD (70%/60%), LARGE (80%/70%). See references/tdd-rules.md.

Review Pipeline

  1. Self-review (always): Automated grep-based checks (tests, code quality, architecture, security basics)
  2. @code-reviewer (default): Deep Opus analysis — quality, tests, performance, architecture, plan alignment
  3. @security-auditor (conditional): OWASP Top 10 — triggered on auth/security patterns or LARGE complexity
  4. @qa-reviewer (conditional): Acceptance criteria, edge cases, error handling — triggered on >3 AC or >5 components

See references/review-checklist.md.

Steps

Step Name Phase Description
00 init-auto - Parse args, complexity detection, branch setup (+ worktree if --worktree), Feature Doc, JSON
01 explore-auto [E] Explore + sanity check (skippable via @plan-path)
02 plan-auto [P] @planner + @plan-validator (skippable via @plan-path)
03 code-auto [C] TDD + stack skills + circuit breaker + background reviewer
04 review-auto [I] Self-review + @code-reviewer + @security-auditor + @qa-reviewer
05 document-auto - Feature Document + summary
06 finish-auto - Finalization + commit + index.json update
07 output-auto - Final JSON output
08 publish-auto - Push, PR, worktree cleanup (if --worktree)

Step Files

Reference Files

Flags

Flag Default Effect
--worktree Off Enable git worktree isolation. Without this flag, works in-place on current directory
--skip-plan-validation Off Skip @plan-validator invocation (plan still created by @planner)
--skip-review Off Skip @code-reviewer, keep self-review only
--skip-security Off Skip @security-auditor even if patterns detected
--skip-qa Off Skip @qa-reviewer even if threshold met
--skip-publish Off Skip push, PR creation, and worktree cleanup
--auto-merge Off Squash-merge PR immediately (SUCCESS only, PARTIAL keeps draft)

Conventions

The skill loads stack skills dynamically based on file extensions for architecture patterns and testing conventions. It also uses the target project's CLAUDE.md and .claude/rules/ for project-specific conventions.

Limitations

  • No interactive breakpoints (by design)
  • No team mode / multi-agent orchestration (background reviewer serves similar purpose)
  • Timeout managed externally by orchestrator
  • Auto-merge requires gh CLI installed + merge permissions