CodeTeam - Gamma Release

A team of AI coding agents that ships working code.

One agent writing code is easy. Many agents writing code in the same repository at once is where things break: work collides, tests rot, and nobody can say what actually shipped. codeteam turns that crowd into a team.

It runs a git repository as a standing crew of Claude Code sessions: an orchestrator, a designer (Skippy), a coordinator (Bilby) for each part of the code, a reviewer (Kong) and an integration builder (Nagatha). Each works in its own folder, inside walls the tools enforce. Every change reaches the main branch the same way: tested, reviewed, merged, rechecked and validated, then promoted only when it is proven.

You describe the spec, the Agents build it.

Discovers and Risks are surfaced to the human

Gamma preview. This is parked branch v5.4 (docs/SPEC-branch-flow-v5.4.md),

• Why is it a Gamma release? Because this is completely working version that can people can use and build from.... but it requires a lot of disk space for big projects. A single machine can handle 40-60 parallel worktrees from a mono repo but will be contrained by disk more than compute.
• The next release will be CodeTeam-JJ which will fold in Jujutsu for massive disk savings.
• This release will stand as it is for pure GIT for multi machine and network storage users.
• CodeTeam-JJ will also be release the same way but seperate.
• CTJJ will allow about 30 software factories to run on a single high end Mac Laptop without a problem; your agents will be able to migration from CT to CTJJ.
• Forge will then replace CTJJ and wrap GIT, GITHUB, JJ, Tessera, Mortar, and T3 Code into a full maintained software factory agent migration from version to version.
• This version will receive bug fixes but no enhancements are planned it is feature complete.

Works with but does not require:

• MOOTx01 CE: compatible with the MOOTx01 memory system, so agents keep what they learn across sessions, models and machines.
• CodeGraph Hub: compatible with CodeGraph's pre-indexed code knowledge graph, so agents find callers, impact and tests without grepping the repo. The fork listed here is our custom fork that tracks codegraph but adds a centralized server needed for a software factory work at scale. (mcp.example.json shows the setup).
• If not using Enterprise API pair with Official CLIProxyAPI: or our Codedaptive CLIProxy fork and our Quota forecasting and spend analysis management center fork)
• Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.2
• Astra 6, Sol 6.1, Terra 6.1

All agent files and skills where optimize using hundreds of hours of real python, typescript, swift, rust, agentic transcripts replayed through the models listed and tuned to the published prompting guides for each. If the expectations of future models change automous behavior will degrade; using with other models will require tuning to achieve automous factory production.

Prerequisites

• macOS
• Claude Code
• herdr from Homebrew: brew install herdr
• Python 3.12+
• git 2.54+
• Optional: sccache (shared Rust build cache); the CodeGraph MCP server (code search)

Start here

docs/QUICKSTART.md: install, build the example team's agents, codeteam2 up teams/example.toml, and how landing works.

What is in this repository

Folder	What
bin/	codeteam (team setup, worktrees, channels, roles), codeteam-land (the landing flow), zone-tool and the build queue, wrappers for cargo/swift/xcodebuild, checks
v2/	codeteam2, the runtime on herdr: sessions in panes, message delivery, the loop, the dashboard (codeteam2 deck)
zone/	zone.mk, the make targets every product gets (zone-test-one, check, warm, ...)
share/	Claude Code hooks (zone guard, provenance, worktree hooks), launchd templates
roles/	the role text codeteam appends to each session
catalog/	the agent catalog: roles, skills, a roster and generate.py, which builds a team's agent files; products/example/
teams/example.toml	an example team manifest
docs/	the branch flow spec, new-team checklist, mission template, quickstart
tests/	the test suite (tests/run.sh, tests/unit, tests/integration)
scripts/check-public.sh	a release check: no private names, paths or secrets in the tree

Tests

python3 -m pytest -q tests/unit v2/tests catalog/tests
for f in tests/unit/*.py; do python3 "$f" || echo "FAIL $f"; done

License

MIT; see LICENSE.