Add Harmony format parsers for GPT-OSS models

[waybarrios]

Implements support for GPT-OSS models that use the Harmony response format (closes #38).

Harmony is a multi-channel format with control tokens that separates reasoning, tool calls, and user-facing responses into distinct channels:

• analysis channel for internal reasoning
• commentary channel for tool calls (addressed with to=functions.{name})
• final channel for the user response

Example of a tool call in Harmony format:

<|channel|>commentary to=functions.get_weather
<|constrain|>json
<|message|>{"location": "San Francisco", "unit": "celsius"}
<|call|>

What was added:

• HarmonyToolParser (registered as "harmony" and "gpt-oss") that extracts tool calls from commentary channel blocks and final content from the final channel. Supports both non-streaming and streaming modes.

• HarmonyReasoningParser (registered as "harmony") that extracts reasoning from analysis channel blocks. Multiple analysis blocks are concatenated. Streaming tracks channel transitions and emits reasoning vs content deltas accordingly.

• convert_tools_to_typescript() utility that converts OpenAI JSON Schema tool definitions to the TypeScript namespace format expected by Harmony models:

namespace functions {
  // Get weather for a location
  type get_weather = (_: {
    location: string,
    unit?: "celsius" | "fahrenheit"
  }) => any;
}

• 41 tests covering tool parsing, reasoning parsing, TypeScript conversion, edge cases (unicode, malformed JSON, incomplete calls, mixed channels), and streaming behavior.

• CI updated to include the new tests in the test-matrix job.

New files:

• vllm_mlx/api/harmony_tools.py
• vllm_mlx/tool_parsers/harmony_tool_parser.py
• vllm_mlx/reasoning/harmony_parser.py
• tests/test_harmony_parsers.py

Modified files:

• vllm_mlx/tool_parsers/init.py
• vllm_mlx/reasoning/init.py
• .github/workflows/ci.yml

Usage:

vllm-mlx serve <model> --enable-auto-tool-choice --tool-call-parser harmony --reasoning-parser harmony

Test plan:

• All 41 new tests pass locally
• All 120 existing parser tests still pass (no regressions)
• ruff and black checks pass
• Imports verified

[TomLucidor]

Update with the latest code: vllm-mlx serve: error: argument --tool-call-parser: invalid choice: 'harmony' (choose from auto, mistral, qwen, llama, hermes, deepseek, kimi, granite, nemotron, xlam, functionary, glm47)