`ai/functiongemma` tool calls are not parsed into `tool_calls`

[yohasebe]

Summary

ai/functiongemma emits well-formed tool-call tokens in its own format, but Docker Model Runner does not parse them:

• tool_calls stays null and the raw <start_function_call>…<end_function_call> text is returned in message.content;
• <end_function_call> is not treated as a stop token, so the call repeats until max_tokens and finish_reason is "length".

There is no error. To a client, the response looks like the model simply did not use the tool.

Steps to reproduce

docker model pull ai/functiongemma

curl -s http://localhost:12434/engines/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ai/functiongemma",
    "max_tokens": 400,
    "messages": [{"role":"user","content":"What is the current time? Use the get_time tool."}],
    "tools": [{"type":"function","function":{
      "name":"get_time","description":"Get the current time",
      "parameters":{"type":"object","properties":{}}}}]
  }' | jq '.choices[0]'

Actual

tool_calls    : null
finish_reason : "length"
content       : "<start_function_call>call:get_time{}<end_function_call>call:get_time{}<end_function_call>call:get_ti…"

Expected

tool_calls populated, finish_reason: "tool_calls", and no call tokens in content. For comparison, an equivalent request to Ollama returns:

{"tool_calls": [{"id": "call_o3fkhnnc", "function": {"name": "get_time", "arguments": {}}}]}

with content empty.

Suggested resolution

Parse the <start_function_call>…<end_function_call> format into tool_calls and treat <end_function_call> as a stop token.

Related: #1046. /api/show exposes no capabilities, so a client also cannot check for tool support before making the request.

Environment

• Model Runner client v1.2.6, Docker Desktop engine 29.7.2
• Backend llama.cpp b9879-metal, macOS, Apple Silicon
• Model: ai/functiongemma