No local model can be pinned for a tool-using build role. Nothing records the weight file's sha256 or Hugging Face revision, the quantisation, the runtime and its version, or the served context length, and the known traps go unchecked:

No local model can be pinned for a tool-using build role. Nothing records the weight file's sha256 or Hugging Face revision, the quantisation, the runtime and its version, or the served context length, and the known traps go unchecked:

• llama-server (b9860 here) calls tools through the chat template, and a q4 KV cache can degrade tool calling.
• Ollama (0.14.2 here) sets context length server-wide (OLLAMA_CONTEXT_LENGTH, default 4096) or per model in a Modelfile, never per request, so a long prompt is truncated silently at the default.
• mlx_lm.server (0.31.3 here) accepts tools and a seed, but its tool parsing is per model family and fragile.

As with OpenRouter, a tool-using role reaches a local server through the opencode runner, whose provider configuration is the person's own. A model-matrix run (itd-2610161733316518) needs local cells it can name exactly. Versions checked locally on 2026-10-16; see research/notes/2026-10-16-model-matrix-sota.md.