lab-llm

lab-llm is the browser-side observability and control surface for a local LLM lab.

This repo is intentionally separate from the runner/backend repo. Its job is to:

• consume telemetry from the runner
• visualize runtime behavior
• compare sessions and experiments
• grow toward a shared browser surface for observability first, chat later

Current posture

This repo starts as a dashboard product, not a model runner.

That means:

• no backend execution logic should be duplicated here
• model-runner remains the telemetry producer
• this repo owns ingestion, read models, dashboard APIs, SSE, and browser UX

The current architecture notes live in:

• docs/dashboard_repo_architecture.md
• docs/decisions/
• docs/specs/

Setup

python -m venv .venv
source .venv/bin/activate
pip install ruff pyright pytest python-dotenv
cd frontend && npm install && cd ..

Checks

./scripts/dev.sh
./scripts/check-quick.sh
./scripts/check-full.sh

pyright is intentionally not in the default local loop.

Dev

Backend with fixture replay:

./scripts/dev-backend.sh --fixtures fixtures/telemetry/sample_session.jsonl

Backend ingesting model-runner JSONL telemetry:

./scripts/dev-backend.sh --telemetry-jsonl /path/to/model-runner-telemetry.jsonl

Frontend:

./scripts/dev-frontend.sh

Integrated build served by Python:

cd frontend && npm run build && cd ..
./scripts/dev-backend.sh --no-replay-fixtures