Feature Request: Add tool_choice config per model/provider for Ollama tool calling

Summary

When using Ollama models (Mistral, Llama) via openai-completions API, tool calling fails because OpenClaw doesn't send the tool_choice parameter.

Problem

• Ollama's OpenAI-compatible endpoint supports tool calling via /v1/chat/completions
• When tool_choice: "required" or "auto" is passed, models reliably use tools
• Without tool_choice, models like Mistral 7B default to generating text instead of calling tools
• OpenClaw's buildParams() in pi-ai only sends tool_choice when explicitly passed via options
• Sub-agents don't pass this parameter

Proof

# With tool_choice - WORKS ✅
curl http://localhost:11434/v1/chat/completions -d '{
  "model": "mistral:instruct",
  "tool_choice": "required",
  "tools": [{...}]
}'
# Returns: {"tool_calls": [...]}

# Without tool_choice via OpenClaw - FAILS ❌
# Returns: text explaining how to get weather (hallucination)

Proposed Solution

Add tool_choice config option:

1. Per-model in models.providers.*.models[].compat.toolChoice
2. Or per-provider in models.providers.*.compat.toolChoice

Example config:

{
  "models": {
    "providers": {
      "ollama": {
        "models": [{
          "id": "mistral:instruct",
          "compat": {
            "toolChoice": "auto"
          }
        }]
      }
    }
  }
}

Environment

• OpenClaw version: 2026.2.1
• Ollama version: latest
• Models tested: mistral:instruct, llama3.1:8b-instruct-q8_0
• Hardware: RTX 4090, 128GB RAM

Workaround

Currently using Opus for tool-heavy tasks, local Mistral only for text-only tasks (summarization, code review).

Notes

• Llama 3.1 8B appears to use tools without tool_choice — may be model-dependent
• Mistral 7B specifically needs tool_choice to reliably invoke tools