cli: expose harmony and gpt-oss tool parsers

[krystophny]

Summary

• expose harmony and gpt-oss as valid --tool-call-parser values in vllm-mlx serve
• factor CLI parser construction so the flag surface is unit-testable
• add regression coverage for the serve CLI choices

Status

• refreshed onto current upstream main (b4fa030) on 2026-04-09
• no logic changes beyond the base refresh

Main files

• vllm_mlx/cli.py
• tests/test_server.py

Reviewer focus

This is a CLI-surface fix only. The parser implementations already exist; the gap is that the serve CLI does not currently expose them consistently through --tool-call-parser.

Validation

• python -m pytest tests/test_server.py -q -> 35 passed, 3 deselected

[Thump604]

@waybarrios, @krystophny: brief endorsement.

Small CLI-surface fix exposing existing harmony and gpt-oss tool parsers via --tool-call-parser. The parser implementations already exist in the codebase, the gap is just that the serve CLI does not list them as valid choices. The test addition covers the new flag values. Mergeable on current main, low risk.

[krystophny]

Force-pushed a refresh onto current upstream main (b4fa030). No logic change beyond the base refresh. Validation: python -m pytest tests/test_server.py -q -> 35 passed, 3 deselected.

[Thump604]

I rebased this onto current main locally and validated it cleanly, but direct maintainer push back to the original fork failed with 403 from this auth context. To keep it moving, I opened a writable restack PR carrying the same functional diff: I’ll link it here as soon as GitHub returns the new PR URL.

[Thump604]

Replacement restack is up at #284: https://github.com/waybarrios/vllm-mlx/pull/284\n\nThat branch carries the same functional diff rebased on current main and passed local validation here:\n- python -m py_compile vllm_mlx/cli.py tests/test_server.py\n- python -m black --check --fast vllm_mlx/cli.py tests/test_server.py\n- pytest -q tests/test_server.py\n\nI could not push directly back to the original fork because GitHub returned 403 for the computor-org remote from this auth context.

[Thump604]

Closing as superseded by #284, which carries the same functional diff on a writable branch and has now been validated on current main. The original fork branch could not be updated from this auth context because GitHub returned 403 on direct push to the computor-org remote.