sergeiwallace/ai-chassis

ai-chassis is a collection of reusable tools, skills, document templates and Python
modules for working with AI coding agents. It comes from a working setup for running
several coding agents in parallel on one codebase, and it covers the parts that sit
around the model: agent skills, session and context configuration, git and worktree
helpers, review and audit evidence, and the tests for them.
> Status. This repository is a generalized extract that grows in small releases.
> It installs an ai-chassis command and a Python package. It is not on PyPI yet. See
> Get started, Status and the changelog.
## Contents
The idea
What is inside
How the pieces fit
A session, start to finish
Skills in this release
Get started
Status
Provenance
Contribute and report issues
## The idea
A coding agent is only as reliable as the setup around it. Running several of them at
once on one repository adds problems that a single chat window never has: two agents
edit the same files, a session runs out of context and forgets what it was doing, a
change lands without anyone checking it, and the shared instructions drift from
machine to machine.
ai-chassis collects the parts that answer those problems:
Work happens in isolated git worktrees, one per agent, so parallel edits do not collide.
A session writes down its state before its context resets, and the next one resumes from that record.
Instructions are composed from shared modules and per-machine overlays instead of copied by hand.
Changes carry evidence: a validated commit, a review, an audit, a passing test run.
## What is inside
| Path | What it holds |
| skills/ | Agent skills. Each skill is a SKILL.md plus its helper scripts. The list below names the ones that ship a SKILL.md in this release. |
| projects-wide-session-config/ | Shared modules and rules from which per-machine agent context files are composed, with an example config. |
| src/aichassis/ | Python modules that compose and apply that configuration: profiles, schema, settings merge, links and generations. |
| scripts/ | Helpers for session lifecycle, git and worktree safety, review and audit evidence, and quality checks. |
| hooks/, shell/, launchd/ | A Claude Code settings validator, terminal-mode reset helpers, and a macOS launch agent for the job supervisor. |
| docs/ | Procedures, templates for designs, plans, research and audits, and a documentation guide. |
| codex/ | Example role configurations for delegating work to Codex agents. |
| packages/fleet-testkit/ | Test-support code the suite relies on. |
| tests/ | The Python test suite. |
## How the pieces fit
mermaid
flowchart LR
subgraph config["Configuration"]
modules["Shared modules and rules"]
overlay["Per-machine overlay"]
end
modules --> compose["Compose and apply (src/aichassis)"]
overlay --> compose
compose --> context["Agent context files for one machine"]
context --> agent["Coding agent session"]
skills["Skills (skills/)"] --> agent
agent --> work["Isolated git worktree"]
work --> evidence["Evidence commit, review, audit, tests"]
evidence --> main["Merged change"]
The configuration side turns shared modules and a small per-machine overlay into the
context files an agent reads at startup. The work side keeps every agent in its own