infra_advisorlisted

# infra_advisor — cut token cost beyond the code Most audits only look at code. This looks at the **machine and the local cluster** and recommends concrete hardware/software/infra changes that move work off paid cloud models — then bundles a fast one-pass audit of the project. ## Commands ```bash # Cluster tips + ASCII diagram python -m skills.infra_advisor.cli tips python -m skills.infra_advisor.cli tips --subnet --json # One-pass audit on the project where Botte Secrète is installed python -m skills.infra_advisor.cli auto . python -m skills.infra_advisor.cli auto /path/to/project --json ``` ## Infra tips (rules, prioritized P0→P3) - **Install a local LLM** (LM Studio/Ollama) if none — the foundation for any saving. - **Add/upgrade a GPU** (≥12-16 GB VRAM) → run a 7-14B coder locally. - **Add a Hailo-8 / 8L / 10 NPU** → vision (detection/OCR/PDF) at ~0 tokens via `media_loader`. - **Move the always-on inference node to Linux/WSL2** → no forced Windows reboots/updates killing the endpoint; more free VRAM/RAM. - **Run Qdrant locally** → unlock the semantic response cache (-60% repeats). - **Dedicate the strongest host** as a shared inference node for all projects. - **Wire the MCP server** (`bootstrap`) so the agent actually uses the local tools. Each tip carries `why` and an `impact` (expected token/cost effect). An ASCII cluster diagram is rendered by default (`--json` for machine-readable). ## Auto audit `auto` combines the cheap, reliable passes into one report: