Contributing

Contributing

Orrin is a single-developer research project, but issues, forks, experiments, and pull requests are welcome. This page is the wiki-side developer guide; the repo also has CONTRIBUTING.md and CLAUDE.md (an agent-oriented guide).

Setup

git clone https://github.com/ric-massey/orrin_v3.git
cd orrin_v3
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Node.js + npm are only needed to build the frontend from source.

The gate: make verify

This is exactly what CI runs — keep it green before opening a PR:

make verify        # ruff + mypy + pytest + frontend typecheck/build
make test          # just the Python suite
make coverage      # suite under the ratcheted coverage floor

Conventions that matter here

1. Keep the suite green. The suite is hermetic — tests/conftest.py redirects state to a tmp dir and fails the session on any live-state write ("isolation breach"). If you see that, you added a path that bypasses brain/paths.py.
2. Resolve state paths through brain/paths.py constants (DATA_DIR, LOGS_DIR, …). Never hand-built or __file__-relative paths — that's what caused a real CI-reddening bug.
3. Symbolic-first, LLM gated. New code must work with no provider key. LLM calls go through brain/utils/generate_response.py only, and callers must be allow-listed. Fail closed — never fabricate content on error.
4. Prefer operational evidence over anthropomorphic language. Cognitive terms name mechanisms.

Where things live

See Cognition Module and the subsystem pages for the map. Long-running components are *_daemon.py and must be resilient and idempotent. Goal handlers register via goals/registry.py; new cognitive functions go through the registry the bandit selects from (Writing a Custom Cognitive Function).

Opening a PR

The repo has a PR template with a verification checklist. Keep changes small and testable, add tests for new behavior, and note anything non-obvious for the reviewer. Don't commit secrets — keys live in .env or the OS keychain (see Security Model).

Reporting issues

Bug and feature templates are provided. Include the run mode, whether an LLM key was configured, and evidence from brain/logs/ or a run report — with no secrets or private logs.