Upstream reference: waybarrios#216
Problem:
vllm-mlx serve already has parser implementations for harmony and gpt-oss, but the CLI does not expose them consistently through --tool-call-parser. That makes the surface area incomplete for local tool-calling experiments.
Path to completion:
- Rebase the branch if needed and keep the diff limited to CLI exposure + tests.
- Re-run
tests/test_server.py.
- Double-check that help text and parser choice validation match the real parser registry.
- If upstream prefers a different naming or grouping for parser options, adjust the flag surface rather than defending the exact spelling.
- Shepherd the PR as a low-risk CLI-surface fix after the correctness backlog is under control.
Acceptance criteria:
Upstream reference: waybarrios#216
Problem:
vllm-mlx servealready has parser implementations forharmonyandgpt-oss, but the CLI does not expose them consistently through--tool-call-parser. That makes the surface area incomplete for local tool-calling experiments.Path to completion:
tests/test_server.py.Acceptance criteria:
serve --tool-call-parserexposes the already-supported parser choices consistently