Revert "studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)"

[danielhanchen]

Reverts #5615. Holding the multi-format parser back until we validate the full safetensors / MLX tool-call healing parity package end-to-end with real models.

[gemini-code-assist]

Code Review

This pull request simplifies the tool-call parser by removing support for several model-specific formats, including Llama-3, Mistral, and Gemma 4, to focus on a unified XML-based approach. Review feedback highlights two regressions: the regex for function and parameter names is now too restrictive, and the JSON parsing logic no longer handles non-dictionary arguments or validates for empty tool names as robustly as the previous implementation.

[gemini-code-assist]

The regular expressions for parsing function and parameter names have been restricted from [\w\.\-]+ to \w+. This is a regression, as it no longer supports names containing dots or hyphens, which was possible before.

Suggested change

	_TC_FUNC_START_RE = re.compile(r"<function=(\w+)>\s*")
	_TC_END_TAG_RE = re.compile(r"</tool_call>")
	_TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
	_TC_PARAM_START_RE = re.compile(r"<parameter=([\w\.\-]+)>\s*")
	_TC_PARAM_START_RE = re.compile(r"<parameter=(\w+)>\s*")
	_TC_FUNC_START_RE = re.compile(r"<function=([\w\.\-]+)>\s*")
	_TC_END_TAG_RE = re.compile(r"</tool_call>")
	_TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
	_TC_PARAM_START_RE = re.compile(r"<parameter=([\w\.\-]+)>\s*")

[gemini-code-assist]

This implementation for parsing JSON tool calls has a couple of regressions compared to the original _parse_tool_call_json function that was removed:

1. Incorrect handling of arguments: It no longer correctly handles arguments that are not a dictionary or a string. The previous logic would wrap other types (like lists or numbers) in a {"value": ...} object and serialize it to a JSON string. The current code passes them through as-is, violating the function's contract that arguments is always a JSON string.
2. Missing name validation: It doesn't check if the tool name is empty. The previous implementation would correctly skip tool calls with no name.

Here's a suggestion to restore the more robust, original behavior.

Suggested change

	obj = json.loads(json_str)
	tc = {
	"id": f"call_{id_offset + len(tool_calls)}",
	"type": "function",
	"function": {
	"name": obj.get("name", ""),
	"arguments": obj.get("arguments", {}),
	},
	}
	if isinstance(tc["function"]["arguments"], dict):
	tc["function"]["arguments"] = json.dumps(
	tc["function"]["arguments"]
	)
	tool_calls.append(tc)
	obj = json.loads(json_str)
	name = obj.get("name", "")
	if not name:
	continue
	args = obj.get("arguments", {})
	if isinstance(args, dict):
	args_str = json.dumps(args)
	elif isinstance(args, str):
	args_str = args
	else:
	args_str = json.dumps({"value": args})
	tool_calls.append(
	{
	"id": f"call_{id_offset + len(tool_calls)}",
	"type": "function",
	"function": {"name": name, "arguments": args_str},
	}
	)

[danielhanchen]

Confirming the revert was correct. #5615 shipped five concrete parser regressions (Mistral nested-JSON truncation, <|python_tag|> stop-on-< markup leak, Llama-3 sentinel single-pass dropping calls, UTF-8 corruption in the Llama-3 KV decoder, and GGUF heal-key always-query breaking python/terminal tools). Details on each are in #5615. Replacement is being validated in #5620 — pytest studio/backend/tests/test_safetensors_tool_loop.py studio/backend/tests/test_safetensors_capability_advertise.py -q on pr-5620 is 110 passed in 3.86s and each of the five regressions has a dedicated unit test (TestLoopRePrompt, TestLoopCanonicalHealKey).