bender-pipeline

"I'm gonna build my own pipeline, with blackjack and code review!"

Parallel story development pipelines with Dev+Review agent pairs for Claude Code.

Bender orchestrates up to 4 concurrent pipelines, each pairing a Dev agent with a Review agent. It handles git worktree isolation, automated code review, fix-plan approval loops, merge verification, and configurable notifications — all fully unattended.

How It Works

Stories ──> Bender (team-lead) ──> Dev Agent 1 ──> Review Agent 1 ──> Merge
                                   Dev Agent 2 ──> Review Agent 2 ──> Merge
                                   Dev Agent 3 ──> Review Agent 3 ──> Merge
                                   Dev Agent 4 ──> Review Agent 4 ──> Merge

1. You provide fully written stories (with acceptance criteria, tasks, affected repos)
2. Bender spawns Dev+Review agent pairs as Claude Code subagents
3. Each pipeline gets an isolated git worktree (no branch conflicts between parallel work)
4. Dev agents implement stories, Review agents perform adversarial code review
5. Fix cycles iterate until review passes (or escalation threshold)
6. Bender merges all completed stories, runs tests, and pushes

Prerequisites

Required

• Claude Code with Agent Teams enabled:

  // .claude/settings.json
  { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

• Fully written stories before running the pipeline. Stories must contain:

  • Acceptance Criteria (numbered ACs)
  • Tasks/Subtasks (checkbox items - [ ])
  • Affected Repos
  • Test requirements

  The pipeline implements stories — it does NOT create them. Use BMAD Method or your own process.

For cross-LLM review (recommended)

• Codex CLI installed and in PATH
• BMAD Method installed in BOTH:
  • Claude Code (for the dev agent's workflow engine)
  • Codex (for the review agent's review prompt)

Without cross-LLM review (zero-dependency setup)

• Use the built-in simple dev workflow + claude-self-review review tool
• No Codex, no BMAD needed — works out of the box

Install

npm

npm install bender-pipeline --save-dev

The postinstall script creates a symlink at .claude/skills/bender/, making /bender available in Claude Code.

git clone

git clone https://github.com/Duatadmin/bender-pipeline.git .claude/skills/bender

Quick Start (Zero Dependencies)

1. Install bender-pipeline
2. Open Claude Code in your project
3. Run /bender
4. On first run, the setup wizard creates .claude/bender.yaml:
   • Auto-detects your git repos
   • Uses simple dev workflow (no BMAD needed)
   • Uses claude-self-review (no Codex needed)
5. Point it at your stories: /bender docs/stories/epic-1/

Full Setup (Cross-LLM Review)

For production use with adversarial cross-LLM code review:

1. Install BMAD Method

# In your project root
git clone https://github.com/bmadcode/BMAD-METHOD.git _bmad

2. Install Codex CLI

npm install -g @openai/codex

3. Set up BMAD for Codex

Copy the BMAD code-review prompt to your Codex prompts directory:

mkdir -p .codex/prompts/
# Create or copy your code-review prompt to:
# .codex/prompts/code-review.md

4. Configure .claude/bender.yaml

version: 1

project:
  stories_dir: "docs/stories/"
  story_pattern: "*.story.md"

repos:
  - name: my-backend
    path: "./my-backend"
    default_branch: main
    test_command: "pytest tests/ -v --tb=short"
    test_timeout: 120
  - name: my-frontend
    path: "./my-frontend"
    default_branch: main
    test_command: "npx vitest run"
    test_timeout: 120

pipeline:
  max_pipelines: 4
  max_review_iterations: 3
  branch_prefix: "story/"

dev_workflow:
  adapter: bmad
  dev_agent_command: ".claude/commands/bmad-agent-bmm-dev.md"
  workflow_engine: "_bmad/core/tasks/workflow.xml"
  workflow_config: "_bmad/bmm/workflows/4-implementation/dev-story/workflow.yaml"

review_tool:
  adapter: codex-cli
  review_prompt: ".codex/prompts/code-review.md"
  timeout: 900
  retry_timeout: 1200

notifications:
  adapter: telegram  # or: slack, webhook, none

5. Run

claude
> /bender docs/stories/epic-1/

Configuration Reference

See references/config-schema.md for the full .claude/bender.yaml schema.

Adapters

Dev Workflows

| Adapter | Dependencies | Description | |---------|-------------|-------------| | simple | None | Built-in TDD workflow: read story, implement tasks, test, commit | | bmad | _bmad/ directory | BMAD Method workflow engine with dev persona | | custom | Your prompt file | Bring your own dev workflow |

Review Tools

| Adapter | Dependencies | Description | |---------|-------------|-------------| | claude-self-review | None | Adversarial self-review using Claude's own context | | codex-cli | Codex CLI | Cross-LLM review via Codex (different LLM catches different bugs) | | custom | Your prompt file | Bring your own review tool |

Notifications

| Adapter | Dependencies | Description | |---------|-------------|-------------| | none | None | Silent (default) | | telegram | Bot token + chat ID | Telegram messages on escalations | | slack | Webhook URL | Slack messages on escalations | | webhook | Webhook URL | Generic HTTP POST on escalations |

Architecture

bender-pipeline/
├── SKILL.md                    # Orchestration engine (the team-lead brain)
├── references/                 # Prompt templates, state schema, handler logic
├── adapters/
│   ├── dev-workflows/          # simple | bmad | custom
│   ├── review-tools/           # claude-self-review | codex-cli | custom
│   └── notifications/          # telegram | slack | webhook
├── hooks/                      # Pre-commit test gating
└── scripts/                    # npm postinstall

The pipeline runs as an 8-step process: 0. Load config + pre-run cleanup

1. Discover & select stories
2. Spawn Dev+Review agent teams
3. Create isolated git worktrees per pipeline
4. Dispatch stories to dev agents
5. Monitor: implementation → review → fix cycles
6. Merge completed stories with test gating
7. Report + auto-commit/push
8. Shutdown team

Production Notes

From running this on real projects:

• Git worktree isolation is critical — without it, parallel pipelines corrupt each other's branches
• Cross-LLM review catches real bugs — Codex found JWT validation gaps that Claude missed
• Codex CLI has ~50% failure rate — the retry mechanism handles most failures
• Dev agents spawn research subagents — 80+ total agents per run is normal
• Context compactions happen — the persistent state file survives compaction; always read it first
• 3 review iterations is usually enough — after that, remaining findings are auto-deferred

License

MIT