DevPit

Sequential agent pipeline for AI coding agents

DevPit runs specialized AI agents one at a time on a task. Each agent runs in its own tmux session with full visibility. You can attach and watch any agent work in real-time.

How It Works

dp pipeline "Add a health check endpoint"

DevPit executes a workflow — an ordered list of steps. Each step:

1. Spawns an AI agent in a tmux session
2. Sends the task + context from previous steps
3. Waits for the agent to finish
4. Captures output and passes it to the next step

Steps can have loop-back conditions (e.g., tester fails → jump back to coder, retry up to 3 times).

The default workflow runs: architect → coder → tester ↔ coder (retry) → reviewer → design-qa ↔ coder (retry)

Custom workflows support arbitrary step sequences, context dependencies, and configurable pass/fail markers.

Installation

Prerequisites

• tmux 3.0+ — agents run in tmux sessions
• An AI CLI — Claude Code (default), Gemini CLI, Codex, or others

From source

git clone https://github.com/colbymchenry/devpit.git
cd devpit
make install  # builds and installs to ~/.local/bin/dp

From npm

npm install -g devpit

Quick Start

# 1. Create a workflow (one-time)
dp create --default

# 2. Run the pipeline
dp pipeline "Add a health check endpoint"

dp create spawns Claude to interview you about your project, then generates agent files (.claude/agents/*.md) and a workflow (.claude/workflows/default.yaml). Use --default for the standard template or describe a custom workflow.

Commands

dp (no args)

Launch the interactive TUI dashboard. View running and past pipelines, start new runs, create workflows, and edit workflow configs — all from one interface.

dp pipeline "task"

Run a workflow pipeline. Loads the default workflow from .claude/workflows/default.yaml, or specify a custom one with --workflow.

dp pipeline "Fix the login form validation"
dp pipeline "Refactor auth module" --agent gemini
dp pipeline "Optimize performance" --workflow optimize

| Flag | Default | Description | |------|---------|-------------| | --agent | claude | AI runtime (claude, gemini, codex, etc.) | | --model | opus[1m] | Model override | | --timeout | 10m | Max time per step | | --retries | 3 | Max loop-back retries | | --workflow | default | Custom workflow name (from .claude/workflows/) |

dp create [prompt]

Create a new workflow interactively. Claude scans your project, interviews you, and generates agent files and a workflow YAML.

dp create                                           # TUI create form
dp create --default                                 # Standard template
dp create "benchmark loop that tests and improves"  # Custom workflow

dp pipeline agent <name> "prompt"

Run a single agent interactively — spawns a tmux session and attaches your terminal.

dp pipeline agent architect "Design a caching layer"
dp pipeline agent coder "Implement the plan" --detach

dp pipeline follow "task"

Queue a follow-up task that reuses the same agent sessions with full context.

dp pipeline follow "Make the button blue instead of green"

dp pipeline status

Show running pipeline sessions with working/idle state.

dp pipeline peek <name>

Read an agent's recent terminal output.

dp pipeline peek coder
dp pipeline peek tester -n 200

dp pipeline stop

Stop all running pipeline agent sessions.

TUI Dashboard

Run dp with no arguments to launch the interactive dashboard:

• Dashboard — view running and past pipeline runs, retry failed ones, kill active sessions
• New run (n) — launch a pipeline with a task, workflow, and agent selection
• Create workflow (c) — generate a new workflow with Claude
• Edit workflow (e) — modify workflow configs: reorder steps, edit fields, add/remove steps
• History (h) — browse past runs with status and details

Custom Workflows

Workflows are YAML files in .claude/workflows/:

name: optimize
description: Iterative benchmark-and-improve loop
steps:
  - name: baseline
    agent: benchmarker
  - name: analyst
    context: [baseline]
  - name: improver
    agent: coder
    context: [analyst]
    directive: "Implement the improvements proposed by the analyst"
  - name: verifier
    agent: benchmarker
    context: [improver]
    loop:
      goto: analyst
      max: 3
      pass: "PIPELINE_RESULT:PASS"
      fail: "PIPELINE_RESULT:FAIL"

Run with dp pipeline "your task" --workflow optimize.

Edit workflows in the TUI with e from the dashboard, or directly in YAML.

Agent Files

Agents are markdown files in .claude/agents/ with YAML frontmatter:

---
name: architect
description: Plans implementation before code gets written
model: opus
tools: Read, Glob, Grep, Bash
effort: high
---

You are the architect. Analyze the task, identify affected files,
plan the implementation, and flag risks...

dp create generates these based on your project type and preferences.

Multi-Runtime Support

DevPit works with multiple AI CLIs. The --agent flag selects the runtime:

dp pipeline "task" --agent claude    # Claude Code (default)
dp pipeline "task" --agent gemini    # Gemini CLI
dp pipeline "task" --agent codex     # OpenAI Codex
dp pipeline "task" --agent copilot   # GitHub Copilot

Each runtime has its own readiness detection, prompt delivery, and startup dialog handling built into the tmux layer.

License

MIT