bmad-automator

Story Automator

Portable BMAD bmad-story-automator skill/plugin bundle. This repo packages:

• skills/bmad-story-automator
• skills/bmad-story-automator-review
• the Python helper runtime inside skills/bmad-story-automator

This is the Python port of bma-d/bmad-story-automator-go. The Go README is the stylistic and operator-facing reference; this repo now documents the Python implementation in the same spirit, but with Python-specific behavior and Codex child-session support.

The root skills/ folder follows the Claude skill convention: each skill is a directory with its own SKILL.md. You can copy either skill folder directly into .claude/skills/. The repo also includes .claude-plugin/plugin.json, so the same root layout can be loaded as a Claude Code plugin with claude --plugin-dir ..

Claude Plugin Layout

This repo follows the Anthropic Claude Code plugin layout:

.
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── skills/
│   ├── bmad-story-automator/
│   │   └── SKILL.md
│   └── bmad-story-automator-review/
│       └── SKILL.md
├── bin/
└── README.md

• .claude-plugin/plugin.json is the plugin manifest.
• .claude-plugin/marketplace.json is the marketplace catalog entry.
• skills/ stays at the plugin root, not inside .claude-plugin/.
• bin/ stays at the plugin root so plugin executables can be added to Claude Code's Bash path.

Local plugin test:

claude --plugin-dir .

Local marketplace test:

/plugin marketplace add .
/plugin install bmad-automator@bmad-plugins

Quickstart

Install into the target BMAD project:

cd /absolute/path/to/your-bmad-project
npx bmad-story-automator

Or install from anywhere:

npx bmad-story-automator /absolute/path/to/your-bmad-project

Then run the installed skill from Claude:

Use the bmad-story-automator skill.

Manual skill copy:

cp -a skills/bmad-story-automator /absolute/path/to/project/.claude/skills/
cp -a skills/bmad-story-automator-review /absolute/path/to/project/.claude/skills/

Expectations

• This is an orchestrator, not a correctness guarantee. Bad planning artifacts still produce bad implementation runs.
• The orchestrator itself is a Claude skill. Child sessions can use Claude or Codex depending on agent configuration.
• Retrospectives are Claude-only.
• The automator expects sprint planning to be complete before it starts.
• Review completion is gated by verification, not by child-session exit alone.
• If the optional QA automate skill is missing, install still succeeds, but runs should use Skip Automate = true.

What This Is

Story Automator automates the BMAD implementation loop for one or more stories:

1. create story
2. implement story
3. optionally run automate/test generation
4. run adversarial code review with retries
5. commit verified work
6. trigger retrospective when an epic is fully complete

The core runtime model is:

• one orchestrator session
• one markdown state document
• many short-lived tmux child sessions
• one marker file guarding against accidental stop
• sprint-status.yaml plus story files as the source of workflow truth

How It Works

flowchart TD
    A["Install into BMAD project<br/>npx bmad-story-automator"] --> B["Run bmad-story-automator skill in Claude"]
    B --> C["Load BMAD config and determine mode"]
    C --> D{"Mode"}
    D -->|Create| E["Init -> Preflight -> Configure -> Finalize"]
    D -->|Resume| F["Load state -> compare sprint status -> inspect sessions"]
    D -->|Validate| G["State audit -> session audit -> progress audit"]
    D -->|Edit| H["Load state -> modify config -> save"]
    E --> I["Execution loop"]
    F --> I
    I --> J["Wrap up and remove marker"]

sequenceDiagram
    autonumber
    participant O as Orchestrator
    participant S as State Doc
    participant T as tmux Child
    participant P as Sprint Status
    participant R as Review Skill

    O->>S: Create or load orchestration state
    O->>T: Spawn create/dev/auto child session
    T-->>O: monitor-session result
    O->>P: Verify source of truth
    O->>R: Run code-review loop until verified
    R-->>P: Sync story status
    O->>S: Update progress, current step, action log

Practical shape:

• create, resume, validate, and edit are first-class modes
• preflight complexity scoring happens before agent selection
• done is gated by review verification
• retrospectives fire inside the execution loop, per epic, not only at the very end

Docs Map