LoadCoach

Turns FreeWeight's measurements (or declared capabilities) into routed, queued, validated inference execution with a fully explainable decision for every job.

Status: 1.0.0 — the M5 release, prepared and not yet published. Every phase of the development plan is built: the registry and task profiles; evidence-weighted routing with a readable explanation for every decision; synchronous and streaming generation with validation and corrective retries; a durable priority queue with leases, ageing, cancellation, recovery and a circuit breaker; FreeWeight evidence import; production feedback, reliability and regression detection; a complete operator UI (dashboard, jobs, live queue, models, reliability, system, settings); and the hardening a LAN bind needs — scopes checked at the route and in the service, per-token rate limits, per-source queue caps, CSRF, Host validation, body limits, content retention.

Installable once tagged. Every runtime dependency is on PyPI (weightsdb 0.2.0 and mirrorwall 0.2.0 landed on 2026-08-30); pip install . in a clean virtualenv resolves entirely from the index, requirements/ci.lock is hash-pinned against it, and every CI job installs from that lock. pip install loadcoach itself waits on the v1.0.0 tag — a human decision.

Part of the Local AI Suite.

Install

pip install loadcoach          # once v1.0.0 is tagged and released
loadcoach serve                # web UI + API on http://127.0.0.1:8766, zero configuration
loadcoach doctor               # every documented failure mode, ✓ / ! / ✗, with what to do

No provider, no GPU, no FreeWeight is required to start; each absence is a documented degraded state. Ollama on 127.0.0.1:11434 is the default provider.

What it does

• Routes explainably. Every decision is persisted in full — candidates, per-capability scores with their source and age, the four adjustment factors with their inputs, every rejection with its numbers — and /jobs/<id> answers why this model? before it shows a table.
• Learns from production. Attempt outcomes and caller feedback become bounded per-model reliability statistics that deprioritize a failing model, exclude it through a circuit breaker with a re-probe, and flag a regression against the model's own history.
• Queues durably. Priority classes, ageing with a proven starvation bound, leases, cancellation within a chunk, VRAM-aware admission, residency management, and recovery after a kill that loses and duplicates nothing.
• Exposes safely. Loopback and open by default; on a LAN, tokens with cumulative scopes checked twice, Host validation before authentication, CSRF on forms, rate limits with Retry-After, and a doctor that names what is wrong.

Documentation

Development

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pre-commit install
ruff format --check . && ruff check . && mypy src tests && lint-imports
pytest -m "not live and not performance"
pytest -m performance            # every spec §15 budget, measured
pytest tests/security            # Security Standards §14, item by item

Licensed under the Apache License 2.0.