Skip to content
AlexWahaPublic

About

Claude Code plugin that turns AI agents into a gated dev team: 31 role subagents (architect, devs, QA, reviewer) run a deterministic plan, implement, test, review pipeline. Each stage advances only when your real build, test and lint commands pass. You approve the plan, commit and push.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Agentry - gated pipelines for Claude Code

Python 3.12+ Claude Code MIT license

Agentry makes Claude Code finish work the way a careful team would. Each task walks through fixed stages, and a stage counts as done only when your real build, test or lint command exits clean. Claude proposes the next step. Python hooks decide whether it happens.

Why it exists

Long agent runs fail in three predictable ways.

What goes wrong What Agentry does
The agent drifts away from the plan You approve the plan before any code is written, and hooks deny edits until you do
The agent says "done" and it is not A stage advances only when its exit command passes, and a script records the result
Context is lost between tasks The next task cannot start until the previous one has a written handoff note

What you get

  • Gated stages. plan, implement, test, review, ready, done, where implement, test and review each end on a command you configure.
  • 31 role agents. Architect, backend, frontend, QA, reviewer, security, writer and more. Each has a profile that limits what it may touch.
  • Folder-as-state tasks. A task is in backlog/, active/ or done/. There is no status field to fall out of sync.
  • Three dials you control. Mode (build, plan, talk), approval level (manual, assisted, auto) and swarm (on, off), each set with one slash command.
  • Memory between tasks. Handoff notes and a searchable lesson store are injected into every agent dispatch.

Quick start

You need Claude Code, Python 3.12 or newer and git. Run from your project root:

claude plugin marketplace add AlexWaha/agentry --scope local
claude plugin install agentry@agentry --scope local
claude plugin list

claude plugin list should show agentry@agentry with scope local and status enabled. Restart Claude Code so the hooks load. No authentication is needed. Without a GitHub SSH key, set CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1 before the first command. Tested on Claude Code 2.1.299.

The plugin carries the tools, skills and hooks. Agents, rules and the pipeline config live in your project, and a one-time onboarding sets them up:

  1. Copy .claude/, .agentry/ and CLAUDE.md from a clone of this repository into your project.
  2. Open Claude Code at the project root and paste the prompt from .claude/_init-prompt.md.
  3. Answer its questions. It detects your stack and asks which commands gate each stage.

Step by step, with a first task end to end: Getting started.

How a task flows

Build pipeline in order: plan, plan-approval (you approve the plan), implement, test, review, ready (you approve the commit), done. Bugfix and techdebt tasks skip the plan and start at implement. A rejected plan returns to plan; a rejected commit returns to implement.

Orange stages are yours. At plan-approval you read the plan. At ready you read the diff with /diff and approve the commit, unless the approval level is assisted or auto, which skip that prompt (auto also skips the take step; a push is never automatic). A failed gate sends the task back to the same stage, up to three times, then parks it as blocked and tells you. bugfix and techdebt tasks skip the plan and start at implement. done is reached only when git confirms the branch is merged. The plan pipeline, which turns an idea into a spec and a task backlog, is in Pipeline and approvals.

Commands you will use

Command What it does
/pipeline-show Print the current mode, approval level and swarm setting
/pipeline-plan Switch to plan mode: brief, plan, spec and task files, no product code
/pipeline-build-manual Build mode, you approve every checkpoint including the commit, swarm off
/pipeline-build-assisted Build mode, commits pass without asking, swarm off
/pipeline-talk Talk mode: discuss and prototype, nothing is tracked
/pipeline-swarm-on Run each epic as a parallel agent swarm
new-task skill Ask Claude to file a task into the backlog

The plugin ships 42 skills, 16 of them the /pipeline-* commands. Full list with examples: Commands.

Architecture

Architecture: you talk to the Orchestrator session. The Orchestrator goes through the plugin hooks (PreToolUse, Stop, SubagentStart, SubagentStop) to the pipeline tools (advance.py, approve.py and the gate scripts); the plugin also holds 42 skills and commands. The tools read gates from the .agentry config (pipeline.json, tasks, plans) and record transitions in run.db, events.db and memory.db. The Orchestrator dispatches the 31 agents and rules in your project's .claude folder.

The plugin is installed once and updated on its own schedule. Everything project-specific stays in your repository. Layout and mechanics: design notes.

Documentation

Page Read it for
Getting started Install, onboard a project, run a first task
Commands Every slash command and skill
Pipeline and approvals Stages, checkpoints, modes, approval levels, swarm, worktrees
Agents The roster, when each agent runs, what each may touch
Configuration The pipeline.json keys worth changing
Upgrade, pin and roll back Version control for the plugin
Troubleshooting Hook denies in plain words, Windows notes
Design notes How enforcement works and where it stops

Status and contributing

Current release: 0.2.0. See CHANGELOG.md and the Releases page. Found a bug or a gap in the docs? Open an issue on GitHub.

License

MIT. See LICENSE.


Pixel art of the author at his computer saying: It's a tool, dude.

About

Claude Code plugin that turns AI agents into a gated dev team: 31 role subagents (architect, devs, QA, reviewer) run a deterministic plan, implement, test, review pipeline. Each stage advances only when your real build, test and lint commands pass. You approve the plan, commit and push.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages