Unified model catalog across HF, LM Studio, Ollama, llama.cpp + content-based resolution

[AlexCheema]

Summary

• Unified model catalog: pluggable ModelSource registry detects locally-installed models from HuggingFace cache, LM Studio, Ollama, llama.cpp, and exo's own dirs. Each entry surfaces in state.local_models tagged with source + format (safetensors / mlx / gguf). The Models page (renamed from Downloads) shows source badges, a source filter, and an "GGUF · n/a — not yet loadable" badge for entries exo can't run today. Inference path-resolution falls through to external sources so MLX/safetensors models in HF cache or LM Studio load without re-downloading.
• Single-file safetensors models (e.g. Qwen/Qwen3-0.6B): drop the hard requirement on model.safetensors.index.json. fetch_safetensors_size falls back to huggingface_hub.model_info().safetensors.total when the index 404s; _scan_model_directory / is_model_directory_complete recognise a non-.partial model.safetensors as a complete dir.
• Content-based model resolution: resolve_existing_model now does a convention-pass first (unchanged), then falls through to a content-pass that fingerprints architecture-defining keys in config.json and finds a matching dir regardless of folder name. Rename Qwen--Qwen3-0.6B/ to wrong-typo/ and exo still finds it — no redundant re-download. Fingerprint cost is sub-millisecond at realistic model counts (10 μs/dir scanned).

End-to-end verified live on macOS against a real LM Studio install, a real HF cache, and a real Ollama install.

Test plan

• uv run basedpyright — 0 errors
• uv run ruff check — clean
• nix fmt — applied
• uv run pytest — 445 passed (+22 new: 19 source scanners + 4 scanner service + 8 fingerprint + 7 content resolution + 7 single-file)
• npx svelte-check — no new dashboard errors (16 pre-existing, unchanged)
• GET /sources returns the 5 sources with availability flags
• GET /state.localModels populates after the worker's first scan tick
• POST /models/add Qwen/Qwen3-0.6B returns 200 (single-file regression)
• Chat completion against Qwen/Qwen3-0.6B returns tokens
• Mispath: mv ~/.exo/models/Qwen--Qwen3-0.6B ~/.exo/models/wrong-typo, then chat — log shows Resolved … via content fingerprint (folder name 'wrong-typo' != convention 'Qwen--Qwen3-0.6B'), no re-download
• Manual: Models dashboard at /downloads — verify source badges, filter chips, GGUF "not yet loadable" state, and that delete buttons are hidden on non-exo entries
• Manual: confirm a multi-shard MLX model (e.g. mlx-community/Qwen3-30B-A3B-4bit) still downloads and runs — regression check

Out of scope (named so we don't slip)

• A GGUF inference engine for Ollama / llama.cpp / LM Studio GGUF entries. Detect-and-display only; "GGUF · n/a" badge is honest about it.
• Wiring state.local_models into the chat model picker — independent UI follow-up. Today the Models page surfaces them; the picker still searches HF + bundled cards.
• Persisting a fingerprint index to disk. Sub-ms per resolve at realistic scales — premature optimisation.
• Auto-renaming wrong-named folders to canonical. Content resolution makes this unnecessary.

🤖 Generated with Claude Code