standalone.py

"""Recover gemma-4 coder tool_calls WITHOUT litellm (or any framework).

Call your llama.cpp / OpenAI-compatible endpoint however you like, then run
`recovered_tool_calls()` on the assistant message text. The heavy lifting is in
`gemma_tool_parse.py` (pure stdlib) — this file is just the thin OpenAI-shape glue.

Run me directly for a demo:  python standalone.py
"""

from __future__ import annotations

import json

from gemma_tool_parse import clean_content, find_tool_calls


def recovered_tool_calls(completion_text: str) -> tuple[list[dict], str | None]:
    """Turn a raw gemma-4 completion into OpenAI-style tool_calls.

    Returns (tool_calls, leftover_content). When no call is present, tool_calls is
    empty and leftover_content is the cleaned text (or None if it was pure markup).
    """
    calls, leftover = find_tool_calls(completion_text)
    if not calls:
        return [], (clean_content(completion_text) or None)
    tool_calls = [
        {
            "id": f"call_{i}",
            "type": "function",
            "function": {
                "name": c["name"],
                "arguments": json.dumps(c["arguments"], ensure_ascii=False),
            },
        }
        for i, c in enumerate(calls)
    ]
    return tool_calls, (leftover or None)


if __name__ == "__main__":
    # The model leaks its call into `content` as native markup instead of structured
    # tool_calls; recover it into the shape an OpenAI client expects.
    demo = 'Sure! <|tool_call>call:get_weather{"city": "Paris", "units": "celsius"}'
    tcs, content = recovered_tool_calls(demo)
    print(json.dumps({"tool_calls": tcs, "content": content}, indent=2, ensure_ascii=False))