@agent-dev/agentdev
v1.4.1
Published
Hierarchical agent orchestration on herdr — a pi extension (the brain) driving herdr (the body).
Maintainers
Readme
agentdev
A pi extension that runs a small crew of coding agents on top of herdr. You give it a goal. It plans with you, then dispatches workers that build, test and review the work until it is clean and ready for you to approve the commits. Every agent runs in its own herdr pane, so you can watch what is happening and jump in whenever you want.
Quick start
Install the launcher once, then just type agentdev:
npm i -g @agent-dev/agentdev
agentdevThe first launch installs whatever is missing. herdr goes in first, then pi, then the extension itself, and the herdr server is started if it isn't running. After that, pi opens with the crew already on. If pi asks you to log in to a model provider, you do that once and you're set.
No global install? The same flow works through npx:
npx @agent-dev/agentdev install
npx @agent-dev/agentdevOr from a clone:
git clone https://github.com/dozeotarios/agent-dev
cd agentdev
./setup.shWhat happens after you give it a goal
The first part is shared with you. The leader checks the repository (or helps you pick a stack for a new one), builds a small glossary of the domain, runs a risk interview, and asks which project mode you want.
Then the autopilot takes over. A consensus loop with a Planner, an Architect, a Developer and a Critic produces a plan. The plan is sliced into stories, and each story gets its own git worktree and a Subworker. Once the code builds and the tests pass, six reviewers look at it from different angles: code quality, efficiency, security, API shape, reliability and domain. The loop keeps going until nothing actionable is left.
Then everything stops. Nothing gets committed until you confirm, unless you
chose +yolo, which commits on its own.
Commands
Inside pi, with the crew on:
/agentdev onturns the crew on. Every message becomes a goal./agentdev statusshows the goals and where each one is./agentdev confirm <goal>accepts the commit-ready gate, so the commits actually happen./agentdev resumepicks the work back up after a restart.
The launcher also understands agentdev doctor (the state of your setup),
agentdev install, and agentdev version.
Project modes
direct-PR is the default: stop at commit-ready, commit on confirm. The
others are no-mistakes, local-only, and +yolo (auto-commits).
More
Full instructions: INSTRUCTIONS.md
Design and acceptance criteria: ARCHITECTURE.md, specs/acceptance-criteria.md
Development notes: DEVELOPMENT.md
Works on Linux, WSL and Windows.
