Add Python tool call parser and use it for LFM2 models

[atdrendel]

Proposed changes

LFM2 models write LFM2 writes Pythonic function calls, not JSON tool calls. This pull request adds PythonToolCallParser and uses it for LFM2 models.

Before these changes were applied

Prior to these changes, testLFM2EndToEndToolCallGeneration always succeeded, but it succeeded because toolCalls was always empty since JSONToolCallParser could never parse the tool calls.

// ToolCallIntegrationTests.testLFM2EndToEndToolCallGeneration()
let (result, toolCalls) = try await generateWithTools(
    container: container,
    input: input,
    maxTokens: 100
)

print("LFM2 Output: \(result)")
print("LFM2 Tool Calls: \(toolCalls)")

After these changes are applied

Now, testLFM2EndToEndToolCallGeneration succeeds and does recognize the tool calls correctly. Here's the output of one of the runs I did on my Mac.

Test Suite 'Selected tests' started at 2026-02-05 22:40:59.659.
Test Suite 'MLXLMIntegrationTests.xctest' started at 2026-02-05 22:40:59.659.
Test Suite 'ToolCallIntegrationTests' started at 2026-02-05 22:40:59.659.
Test Case '-[MLXLMIntegrationTests.ToolCallIntegrationTests testLFM2EndToEndToolCallGeneration]' started.
LFM2 Output: 
LFM2 Tool Calls: [MLXLMCommon.ToolCall(function: MLXLMCommon.ToolCall.Function(name: "get_weather", arguments: ["location": MLXLMCommon.JSONValue.string("Tokyo")]))]
Test Case '-[MLXLMIntegrationTests.ToolCallIntegrationTests testLFM2EndToEndToolCallGeneration]' passed (0.317 seconds).
Test Suite 'ToolCallIntegrationTests' passed at 2026-02-05 22:41:05.423.
	 Executed 1 test, with 0 failures (0 unexpected) in 0.317 (5.764) seconds
Test Suite 'MLXLMIntegrationTests.xctest' passed at 2026-02-05 22:41:05.423.
	 Executed 1 test, with 0 failures (0 unexpected) in 0.317 (5.764) seconds
Test Suite 'Selected tests' passed at 2026-02-05 22:41:05.423.
	 Executed 1 test, with 0 failures (0 unexpected) in 0.317 (5.764) seconds

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[atdrendel]

Oh, look at that, I pushed this without realizing that there was a similar pull request already open: #91

I'm fine with either one, of course. Feel free to close this one, if you want.

[Copilot]

Pull request overview

This pull request adds a PythonToolCallParser to correctly parse LFM2 model outputs, which generate Python-style function calls rather than JSON. Previously, LFM2 models were incorrectly configured to use JSONToolCallParser, which couldn't parse the Python-style syntax, resulting in all tool calls being silently ignored.

Changes:

• Implemented PythonToolCallParser to parse Python-style function calls with support for both single and double quotes, escape sequences, numeric types, booleans, and null values
• Updated ToolCallFormat.swift to use PythonToolCallParser instead of JSONToolCallParser for LFM2 format
• Replaced outdated LFM2 tests that verified JSON parsing with comprehensive new tests for Python-style parsing

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
Libraries/MLXLMCommon/Tool/Parsers/PythonToolCallParser.swift	New parser implementation for Python-style function calls with comprehensive string, number, and type handling
Libraries/MLXLMCommon/Tool/ToolCallFormat.swift	Updated LFM2 format to use PythonToolCallParser and corrected comment to reflect Python-style format
Tests/MLXLMTests/ToolTests.swift	Removed obsolete JSON-based LFM2 tests and added comprehensive Python-style parser tests covering edge cases

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[atdrendel]

The problem solved by this pull request has already been addressed by #91. So, I'll close this one.