Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

README.md

Helm

A Claude Code plugin for orchestrating fully autonomous agent sessions that take a task from discovery through to a merged PR without human interaction.

How It Works

Helm has two layers:

  1. The Orchestrator - a persistent Claude Code session you talk to. It receives your intent, builds a context packet, spawns an autonomous session, and tracks its status. It never does implementation work.

  2. Spawned Sessions - fully autonomous Claude Code sessions that run in your project directories. Each one owns the full pipeline: discovery, planning, implementation, testing, and PR creation. Each stage is delegated to a dedicated subagent with scoped tool access. They never interact with you directly.

Pipeline Stages

Every spawned session follows the same five stages in order:

Stage Agent What it does Output
Discovery discovery Explores the codebase read-only .helm/findings.md
Planning planning Produces a step-by-step implementation plan .helm/plan.md
Implementation implementation Executes the plan, commits per step .helm/implementation.md
Testing testing Runs the test suite, fixes failures All tests passing
PR pr Opens a pull request via gh PR URL in .helm/status.json

If any stage fails after retries, the session writes an escalation report and stops. The orchestrator surfaces the escalation to you.

Usage

Point Claude Code at this plugin and describe the work you want done:

"Add email validation to the signup form in /Users/me/projects/myapp"

The orchestrator will ask for any missing context (test command, target branch, constraints) and then spawn the session.

To check on a running session:

"What's the status of the myapp session?"

Context Packet

The orchestrator builds a context packet before spawning. All fields are required:

Field Description
task What needs to be done
project_dir Absolute path to the project
success_criteria What done looks like
constraints What the session must not do
branch_name Feature branch to work on
test_command How to run the test suite
pr_target Branch the PR should target

Project Structure

helm/
├── .claude-plugin/plugin.json    # Plugin manifest
├── CLAUDE.md                     # Orchestrator routing
├── skills/
│   ├── orchestrator/SKILL.md     # Orchestrator behavior
│   ├── bootstrap/SKILL.md        # Spawned session pipeline contract
│   └── escalation/SKILL.md       # Shared escalation format
├── agents/
│   ├── discovery.md              # Codebase exploration (read-only)
│   ├── planning.md               # Implementation planning (read-only)
│   ├── implementation.md         # Plan execution (read/write)
│   ├── testing.md                # Test suite runner
│   └── pr.md                     # PR creation
├── hooks/
│   ├── hooks.json                # Hook configuration
│   └── scripts/post-spawn.sh     # Session logging
├── schemas/context-packet.md     # Context packet field definitions
└── status/                       # Runtime session tracking

What Helm Is Not

  • Not a replacement for motif. Motif augments how you work with Claude Code directly. Helm is for fully autonomous sessions that run without you.
  • Not a custom runtime. All execution goes through Claude Code's native tools and standard CLI commands.
  • Not persistent across sessions. Each spawned session is fresh. Context comes from the context packet and .helm/ handoff artifacts.

License

MIT