capabilities.toolcall is populated but never honored — tool_call: false still sends the full tool roster #45327

Description

Setting tool_call: false on a model has no effect. The flag reaches Provider.Model.capabilities.toolcall (packages/opencode/src/provider/provider.ts:1288 and :1519), but nothing in the session, tool, or provider-transform paths ever reads it back — so every request still carries the full native tool roster.

For local models behind OpenAI-compatible endpoints that can't emit native tool calls, the result is a dead-ended turn: the model writes the call as text (<tool_call>{...}</tool_call> or a fenced JSON block), opencode stores it as ordinary assistant text, and the tool never runs. From the user's side it looks like the model simply refused to do the work.

Steps to reproduce

1. Point a provider at a local OpenAI-compatible endpoint (ollama, llama-server, vLLM) and set "tool_call": false on the model in opencode.json
2. Ask for something that needs a tool ("read util.py and describe it")
3. The request still includes the full tool roster; when the model answers with a text-shaped call it is stored as plain text and nothing executes

Expected

The flag should have an observable effect — either withhold the native roster and parse a text protocol back into tool calls, or at minimum surface that the model cannot use tools.

OpenCode version

1.18.23