[Tromset]

Summary

Mirrors flycoder0.3, flycoder0.3fast and flycoder0.3pro on Hugging Face, so they can be run with ollama run hf.co/<namespace>/flycoder0.3 or llama-server -hf <namespace>/flycoder0.3 --jinja.

• scripts/publish-hf.sh: for each model, downloads the published layers of Tromset/<model> from the ollama.com registry (GGUF Q4_K_M weights, vision projector mmproj, system, params, LICENSE), checks their SHA-256 and the GGUF magic, then uploads them with huggingface_hub to <namespace>/<model>, with the model card. The placeholder template layer ({{ .Prompt }}) is deliberately not copied: Ollama formats Qwen with its built-in renderer, and from HF the chat template embedded in the GGUF is used.
• .github/workflows/publish-huggingface.yml: runs after each successful publish-ollama on main, on changes to docs/huggingface/, and by hand. Needs the HF_TOKEN secret (write token); optional HF_NAMESPACE variable, otherwise the token's owner.
• docs/huggingface/*.md: model cards (license apache-2.0, base_model Qwen/Qwen3.5-9B, Qwen3.5-4B, Qwen3.8-27B, quantized relation, usage with Ollama and llama.cpp).
• Tests: every published model has a card with the right base, size, memory and context, the publisher lists it, the workflow uses HF_TOKEN; publish-hf.sh parses with POSIX sh.
• readme and Documentation (new section 8.1, layout, troubleshooting).

Test plan

• npm test: 66/66 pass.
• Ran the script's registry logic against the real ollama.com registry for flycoder0.3fast: manifest parsing, digest-checked downloads of system, params, LICENSE, absent layers handled. The upload itself could not be run here (no Hugging Face token); it will run on the first workflow run once HF_TOKEN is set.

---

Generated by Claude Code