[Vl4dk0]

Summary

mlx_lm.server had no harmony tool parser, so gpt-oss returned raw <|channel|> text as content. The checkpoint chat template then rejects that history (You have passed a message containing <|channel|> tags), and the next turn 404s.

Nothing on main parses harmony tool calls. #1867 (closed, withdrawn) only split non-streaming text into content and reasoning, and put the commentary tool payload back into content.

This adds a harmony parser and selects it for gpt-oss:

• analysis channel → reasoning
• commentary to=functions.<name> → a tool call (recipient-first, as the chat template writes it, and channel-first, as the model emits it, including <|constrain|>json)
• final channel and a commentary preamble → content
• <|call|> is an extra eos, so a tool turn stops for the client
• assistant history that replays tool_calls plus reasoning / reasoning_content (as OpenAI-compatible agent clients send it) is rewritten to the template's thinking field. A tool-call turn cannot carry both content and thinking, so a preamble is dropped when analysis is present.

The checkpoint chat template is unchanged.

Test plan

• python -m unittest tests.test_tool_parsing
• Tokenizer from gpt-oss-20b-MXFP4-Q8 selects the parser, round-trips a tool turn through the streaming detokenizer, and renders reasoning + tool_calls without raising
• Multi-turn agent loop against mlx_lm.server with gpt-oss-20b-MXFP4-Q8: a file read, then a shell call, then the final answer, all HTTP 200 with no channel-tag 404
• Eight-call tool loop with consecutive turns, no "does not support tool calling" warning

🤖 Generated with Claude Code