docs/TONIGHT-RUN.md

# TONIGHT RUN - get chatting in under 15 minutes

You need: any machine with Python 3.8+, ~1 GB free (Qwen3-0.6B) or ~7 GB free (with Ornith-9B),
a browser (Chrome/Edge recommended), and internet ONCE for the download.

## 1. Download the drive (one command, ~100 MB, no weights in it)

    pip install huggingface_hub
    huggingface-cli download Johnny5b/portable-ai-runner --repo-type dataset --local-dir DRIVE

## 2. One-time online setup (downloads real GGUF weights onto the drive)

    pip install requests
    python setup.py                 # Qwen3-0.6B GGUF (639 MB) + wllama engine assets
    python setup.py --model ornith  # optional: add Ornith-1.5-9B (5.8 GB)

setup.py resumes if interrupted (HTTP Range) and verifies every file's sha256
against the pinned Hub hashes before it says done. A corrupted download cannot pass.

## 3. Launch (offline from here on)

    pip install websockets
    python serve.py

Open http://localhost:8080 in Chrome.

## 4. Chat

1. Models tab -> Scan drive -> Load next to Qwen3-0.6B (reads 639 MB from the drive,
   caches it in the browser; second loads are instant).
2. Chat tab. Type. Tokens stream from llama.cpp compiled to WASM, weights never leave the drive.

Honest expectations (nothing here is a quoted benchmark):
- The 76.7 / 122.7 / 146.3 tok/s numbers in cores.json were measured on GPU server tiers,
  NOT in the browser. Browser WASM speed depends on your machine; a 0.6B model on a
  modern laptop typically gives a few-to-15 tok/s. If Chrome offers WebGPU on your
  machine, wllama 3.6.1 uses it automatically.
- Qwen3-0.6B is a thinking model: it may emit <think>...</think> blocks before the
  answer. That is the model, not a bug. Add /no_think to your message to skip it.
- Set max tokens >= 512 for reasoning-style replies.

## 5. Mesh (optional, same Wi-Fi)

Host machine: Network tab -> Wi-Fi Host. Other devices on the same Wi-Fi open
http://<host-LAN-IP>:8080 -> Wi-Fi Client -> enter ws://<host-LAN-IP>:8081/ws.
The host runs the model; clients are thin.

## Troubleshooting

- single-thread warning in the engine pill: you opened index.html via file:// - use serve.py.
- Load fails with fetch error: models/gguf/ is empty -> run setup.py (step 2).
- Bluetooth tab: Chrome/Edge on desktop/Android only; experimental.
- If serve.py port 8080 is busy: another app owns it - close it or edit the port in serve.py.

## What is on the drive after setup

    models/gguf/Qwen3-0.6B-Q8_0.gguf      639 MB  sha256-verified
    vendor/wllama/esm/...                  llama.cpp WASM engine, vendored
    everything else                        docs, symbols, modules, vault, memories


## Ornith-1.5-9B tonight (the flagship drive tier)

    python setup.py --model ornith

- Pulls ornith-ai/Ornith-1.5-9B-GGUF (5.8 GB, Q4_K_M) with resume + sha256
  verification against the pinned Hub hash.
- RAM honesty: loading 5.8 GB into the browser needs a 16 GB RAM machine. On an
  8 GB Chromebook expect failure or swapping - use Qwen3-0.6B there, or run
  Ornith on the host with llama-server (CPU) instead of the browser.
- Browser tok/s is UNMEASURED (the 76.7 number is GPU llama.cpp, do not cross-quote).
  The app now measures itself: every reply shows a trace line
  [TRACE:AWAKE] symbols=... files_read=... lookup_ms=... | [TRACE:GEN] tok=... tok_s=...
  and appends it to memories/audit_ledger.jsonl via the drive server.
  Your tonight numbers are the first browser-tier datapoints in the whole project.