Claude 19af4ae540 feat: integrate AirLLM as optional high-performance backend ...

Adds the `bigbrain` optional dependency group (airllm>=2.9.0) and a
complete second inference path that runs 8B / 70B / 405B Llama models
locally via layer-by-layer loading — no GPU required, no cloud, fully
sovereign.

Key changes:
- src/timmy/backends.py   — TimmyAirLLMAgent (same print_response interface
                            as Agno Agent); auto-selects AirLLMMLX on Apple
                            Silicon, AutoModel (PyTorch) everywhere else
- src/timmy/agent.py      — _resolve_backend() routing with explicit override,
                            env-config, and 'auto' Apple-Silicon detection
- src/timmy/cli.py        — --backend / --model-size flags on all commands
- src/config.py           — timmy_model_backend + airllm_model_size settings
- src/timmy/prompts.py    — mentions AirLLM "even bigger brains, still fully
                            sovereign"
- pyproject.toml          — bigbrain optional dep; wheel includes updated
- .env.example            — TIMMY_MODEL_BACKEND + AIRLLM_MODEL_SIZE docs
- tests/conftest.py       — stubs 'airllm' module so tests run without GPU
- tests/test_backends.py  — 13 new tests covering helpers + TimmyAirLLMAgent
- tests/test_agent.py     — 7 new tests for backend routing
- README.md               — Big Brain section with one-line install
- activate_self_tdd.sh    — bootstrap script (venv + install + tests +
                            watchdog + dashboard); --big-brain flag

All 61 tests pass. Self-TDD watchdog unaffected.

https://claude.ai/code/session_01DMjQ5qMZ8iHeyix1j3GS7c

2026-02-21 16:53:16 +00:00

952 B

Raw Blame History

View Raw