pipewright 0.2.1

CLI-first, model-agnostic, plugin-based AI dev workflow automation

Pipewright

CLI-first, plugin-based AI dev workflow automation. Chain AI agents into multi-step pipelines where each step is focused, checkpointed, and cost-optimized.

Install

pip install pipewright

Or from source:

git clone https://github.com/brangi/pipewright.git
cd pipewright
pip install -e ".[dev]"

Requires Python 3.11+ and an Anthropic API key:

echo "ANTHROPIC_API_KEY=sk-..." > .env

Usage

Generate tests for a source file:

pipewright run test-gen ./src/auth.py

Solve a GitHub issue end-to-end (analyze, plan, implement, open PR):

pipewright run issue-solve 42

Review code changes:

pipewright run code-review HEAD~1..HEAD
pipewright run code-review "#2"

Debug an issue systematically:

pipewright run debug "TypeError in auth.py line 42"

Refactor code:

pipewright run refactor ./src/auth.py

Generate documentation:

pipewright run docs-gen ./src/

List all available workflows:

pipewright list

Auto-approve checkpoints for CI/scripted use:

pipewright run test-gen ./src/auth.py -y

Supported Languages

Pipewright works with any programming language. The AI agents read, analyze, and generate code using standard dev tools. These languages have dedicated examples and test framework detection:

Language	Extensions	Default Test Framework	Example
Python	.py	pytest	example/python/utils.py
JavaScript	.js	Jest	example/js/utils.js
TypeScript	.ts, .tsx	Vitest	example/ts/utils.ts
Java	.java	JUnit 5	example/java/Utils.java
Rust	.rs	cargo test	example/rust/src/lib.rs
Go	.go	go test	example/go/utils.go
Ruby	.rb	RSpec	example/ruby/utils.rb

Generate tests for any language:

pipewright run test-gen ./src/utils.js       # JavaScript -> Jest
pipewright run test-gen ./src/lib.rs         # Rust -> cargo test
pipewright run test-gen ./cmd/server.go      # Go -> go test

Create a Plugin

Scaffold a new plugin:

pipewright init my-plugin

Or manually -- create plugins/my_plugin/workflow.py:

from pipewright.workflow import Workflow, Step

class MyPluginWorkflow(Workflow):
    name = "my-plugin"
    description = "Does something useful"
    steps = [
        Step(name="analyze", prompt="Analyze {target}.\n\n{context}",
             tools=["Read", "Glob"], model="haiku"),
        Step(name="execute", prompt="Execute on {target}.\n\n{context}",
             tools=["Read", "Write"], model="sonnet", checkpoint=True),
    ]

Run pipewright list to verify it appears.

Architecture

CLI (Click)
  |
  v
Plugin Loader ---- plugins/*/workflow.py
  |
  v
Engine (async orchestrator)
  |
  +---> Step 1: Agent (Claude SDK) ---> context
  +---> Step 2: Agent (Claude SDK) ---> context
  +---> Step 3: Agent (Claude SDK) ---> context (checkpoint)
  |
  v
Result (with memory persistence via MCP)

Each step runs a Claude agent with a focused prompt. Steps chain via context accumulation. Checkpoints pause for human review. Model tiering (haiku for cheap steps, sonnet for complex) keeps costs low.

Project Structure

src/pipewright/
  cli.py            CLI entry point (Click)
  engine.py         Async orchestrator
  workflow.py       Step, Chain, Workflow dataclasses
  config.py         JSON config (~/.pipewright/config.json)
  plugins/loader.py Plugin discovery
  memory/           Persistent memory (JSON + MCP server)
  observability/    Terminal display and SDK hooks

plugins/
  test_gen/         Generate test suites
  issue_solve/      Solve GitHub issues end-to-end
  code_review/      Review code changes
  refactor/         Refactor code
  docs_gen/         Generate documentation
  debug/            Systematic debugging

Configuration

pipewright config set model sonnet
pipewright config set max_budget_usd 1.00
pipewright config get model

Settings stored in ~/.pipewright/config.json. API keys come from environment variables or .env (never stored in config).

Contributing

See CONTRIBUTING.md for development setup, coding guidelines, and how to add plugins. For the full plugin authoring reference, see docs/PLUGIN_GUIDE.md.

Roadmap

See ROADMAP.md for planned features and milestones.

License

MIT -- see LICENSE.