minnarone-runtime-doctor

name	minnarone-runtime-doctor
description	Perform a read-only Minnarone runtime readiness check for profiles P0-P5, including tools, configuration, pinned model artifacts, digests, and safety gates. Use when asked to diagnose setup readiness or verify a Twitch shadow profile; do not use to download models, mutate config, or enable live sending.

Minnarone Runtime Doctor

Return a deterministic PASS, FAIL, or SKIP report. Do not modify files, install dependencies, download weights, start a long-running capture, or call Twitch live-send APIs.

1. Select one profile

Infer the smallest requested profile and state it before checking:

• P0 chat-only: core install and Twitch chat in shadow;
• P1 capture smoke: Streamlink/FFmpeg and bounded raw chat/audio/video smoke;
• P2 CPU-light audio: P1 plus local ASR and English VoxCeleb CAM++;
• P3 Apple Silicon full: audio/video/VLM/TUI, Qwen2-VL on MPS;
• P4 CUDA full: audio/video/VLM/TUI, Qwen2-VL on CUDA;
• P5 llama.cpp full/local: local GGUF/mmproj and multimodal server.

If intent is ambiguous, default to P0 and mark higher-profile checks SKIP. Use docs/research/runtime-model-profiles.md for profile budgets and docs/runtime-model-manifest.json for artifact pins and SHA-256 digests.

2. Run read-only checks

Check in this order and record the exact command or evidence for each result:

1. Repository and Python: supported Python, uv.lock, importability, and the selected optional extras.
2. External tools: only those required by the profile, such as Streamlink, FFmpeg, or llama-server; record versions.
3. Configuration: load the requested YAML, confirm requested channels and shadow/off send mode, and run uv run python -m minnarone <config> --check. Explain that --check is local/lazy and does not prove live capture or a first VLM inference.
4. Models: treat weights alone as insufficient. Resolve the bundle's pinned source owner and revision, then verify all required bundle files. Each file must have an authenticated integrity entry: either direct byte size and SHA-256 or an artifact reference that resolves to the same filename, owner, revision, byte size, and SHA-256. Missing provenance, an unresolved reference, or any mismatch must be FAIL. Then run the bundle's read-only, offline local loader smoke and record its evidence. Primary-weight digest success without authenticated bundle completeness and a profile-specific loader smoke must be FAIL; never fetch replacements.
5. Speaker contract for Italian/non-Mandarin audio: English VoxCeleb CAM++, dimension 512, threshold 0.5. Treat the old zh-cn 192-dimension model as incompatible with the public profile.
6. Capacity: report CPU, RAM, disk, and GPU/VRAM evidence against the selected profile as measured facts or SKIP, never as invented certainty.
7. Public safety: config remains shadow/off; secrets stay in environment; live would require a dedicated bot account, broadcaster consent, token validation, disclosure choice, attended operation, and immediate stop.

Do not print secret values. It is safe to report only whether expected variable names are set. Do not access the network merely to establish readiness.

3. Classify results

• PASS: directly verified and matches the selected profile.
• FAIL: required and absent, invalid, incompatible, or unsafe.
• SKIP: outside the selected profile or cannot be verified read-only.

Never convert unknowns into PASS. A failed higher-profile check does not make P0 fail when P0 is the selected profile.

4. Report

Use this compact structure:

Profile: Pn — name
Overall: PASS|FAIL

[PASS|FAIL|SKIP] check — evidence

Next actions:
1. smallest actionable remediation

Not performed: downloads, writes, live sends

Make the overall result FAIL when any selected-profile requirement fails; otherwise PASS. Keep skipped optional checks visible.