brichan 0.10.0

Repository-native AI Chief of Staff orchestration

Brichan

Coding agents are very good at the work in front of them and very bad at remembering why it was asked for. Context lives in a chat window, so it is lost when the window closes; decisions are re-litigated, status is reconstructed from memory, and the record of what was actually verified never outlives the session.

Brichan is an AI Chief of Staff that sits above the coding agent rather than inside it. You give it a project goal; it keeps only the context that goal needs, hands bounded tasks to independent workers, checks the evidence they return, and records decisions and status as files in your repository — outside chat, where they survive the session.

Requirements

• Linux or macOS with Python 3.10 or newer. No runtime Python dependencies.
• An existing top-level Git repository to operate on.
• The codex CLI on PATH — the coordinator runtime Brichan launches.
• Herdr, the worker control plane — required as soon as Brichan hands bounded tasks to workers (a single coordinator session runs without it):

$ curl -fsSL https://herdr.dev/install.sh | sh
$ herdr integration install codex

Installation

$ pip install brichan

Basic commands

$ brichan init --project <repo>            # preview the footprint (dry run)
$ brichan init --apply --project <repo>    # create it
$ brichan status --project <repo>          # one-line state verdict
$ brichan doctor --project <repo>          # health summary (--json for CI)
$ brichan run --project <repo> -- <args>   # launch the coordinator

Usage

Initialization defaults to a dry run with zero writes, so you can preview the complete footprint before anything is created. --apply then creates:

• a versioned .brichan/ directory holding managed policy, model routing, Herdr skill resources, and mutable project memory;
• root AGENTS.md and CLAUDE.md pointers, when the repository does not already have them.

Adding --init-agents also exports the Herdr orchestration skill to .agents/skills/, so a codex session started directly in the repository (without brichan run) discovers it too.

Existing files are never modified, and repeating init against healthy state is idempotent.

From inside a healthy initialized repository, bare brichan launches the coordinator directly:

$ brichan

Arguments after -- are forwarded to the runtime. Option-looking text after -- is treated as literal prompt content, so a prompt beginning with a dash cannot silently become a flag:

$ brichan run --project <repo> -- --review the auth module

Feature notes

• Coordinator over workers. You describe the outcome; Brichan reads only the project context that work needs, gives bounded tasks to independent workers through Herdr, verifies the results, and records decisions and status back into the repository.
• Settings-driven model routing. Coordinator defaults and worker routes come from .brichan/config/model-routing.json, so planning, implementation, review, and scan work may each use a different model.
• Guarded launches. Brichan rejects native delegation, permission bypasses, cwd/scope widening, profiles, remote execution, and arbitrary runtime configuration before launching.
• Strict state diagnostics. Malformed, dangling, symlinked, inaccessible, or incompatible .brichan/ state is reported, never silently repaired. Schema v1 has no automatic migration: a package-version change requires deliberate backup and reinitialization.
• Machine-readable health. brichan doctor --json emits one JSON document covering repository, Git, policy, routing, memory, and dependency checks for scripts and CI.

License

Brichan is available under the MIT License.