Implement local compatibility APIs

[CobraSoftware]

overview

This is desgined to add more API endpoints for boarder compatiblity.
It was heavily assisted by AI.

Summary

This PR expands the local compatibility surface for the MLX-backed server while keeping inference local-only.

Main additions:

• OpenAI compatibility improvements, including broader response/resource coverage and realtime session support
• Anthropic and Ollama compatibility validation
• LM Studio native API support under /lmstudio/v1/*
• Deepgram-compatible local APIs under /deepgram/vl/*
• Model-family detection for Qwen3 Omni and Voxtral realtime variants
• Additional local implementations for practical OpenAI-style resource APIs such as files and persisted responses retrieval

What Changed

• Added LM Studio endpoints:

  • /lmstudio/v1/models
  • /lmstudio/v1/models/load
  • /lmstudio/v1/models/unload
  • /lmstudio/v1/models/download
  • /lmstudio/v1/models/download/status
  • /lmstudio/v1/chat

• Added Deepgram endpoints:

  • /deepgram/vl/listen
  • /deepgram/vl/speak
  • /deepgram/vl/read
  • /deepgram/vl/models
  • /deepgram/vl/models/{model_id}

• Expanded OpenAI-compatible surface:

  • Realtime session endpoints
  • Realtime client secrets
  • Realtime transcription sessions
  • Local files API
  • Local responses/{id} retrieval, delete, and input item listing
  • Broader spec alias coverage
  • Local-only 501 not_implemented_local fallback responses for remaining unsupported cloud paths

• Improved auth compatibility:

  • Authorization: Bearer <key>
  • Authorization: Token <key>
  • x-api-key
  • api-key

Testing

I tested compelation python3 -m py_compile vmlx_engine/server.py tests/test_deepgram_api.py tests/test_lmstudio_api.py tests/test_openai_spec_surface.py tests/test_realtime_compat.pyand did get all tests to passs but I lack the hardware to do larger in-depth realworld testing with bigger models.

[jjang-ai]

Thanks for the compatibility API work. I am deferring this from the current release-hardening lane.

It is a large API surface PR, currently conflicts with main, and needs a separate current-source design/test pass before merge. Right now I am prioritizing targeted fixes and correctness regressions.

[jjang-ai]

Reviewed for the current release-hardening pass. I am not direct-merging this PR into the immediate release branch because the diff is very broad (server.py + multiple compatibility APIs) and includes LM Studio, Deepgram, realtime, local files, and persisted response/resource endpoints that are outside the current release-fix scope.\n\nWhat current source already covers and I verified today:\n- OpenAI Chat Completions sampling/default propagation;\n- OpenAI Responses sampling/default propagation and output-cap behavior;\n- Anthropic adapter bundle-default behavior;\n- Ollama gateway request translation / streaming done behavior;\n- panel request-builder sampling and output overrides;\n- API/cache interaction gates for DSV4 native cache, ZAYA typed CCA, hybrid SSM, TQ KV, disk roundtrip, and DSV4/DSML tool parsing.\n\nI also found and fixed a current release-gate harness regression while reviewing this PR: tests/cross_matrix/run_api_surface_contract.py and run_cache_architecture_contract.py referenced a missing tests/cross_matrix/run_noheavy_api_cache_contract.py. Restored that runner and pushed commit 4df8b3b (test: restore API cache contract runner).\n\nVerification after the runner fix:\n- run_noheavy_api_cache_contract.py -> status=pass (15 API, 5 scheduler-cache, 32 TQ/MLLM-cache, 21 DSV4/DSML tool tests)\n- run_api_surface_contract.py -> status=pass\n- run_cache_architecture_contract.py -> status=pass\n- API parity/history focused pytest -> 38 passed\n\nLeaving this PR open as a future broader compatibility-API feature lane. Credit to @CobraSoftware for the local compatibility API direction; pieces should be split/landed with isolated endpoint specs and tests rather than merged wholesale into this release.

[CobraSoftware]

Reviewed for the current release-hardening pass. I am not direct-merging this PR into the immediate release branch because the diff is very broad (server.py + multiple compatibility APIs) and includes LM Studio, Deepgram, realtime, local files, and persisted response/resource endpoints that are outside the current release-fix scope.\n\nWhat current source already covers and I verified today:\n- OpenAI Chat Completions sampling/default propagation;\n- OpenAI Responses sampling/default propagation and output-cap behavior;\n- Anthropic adapter bundle-default behavior;\n- Ollama gateway request translation / streaming done behavior;\n- panel request-builder sampling and output overrides;\n- API/cache interaction gates for DSV4 native cache, ZAYA typed CCA, hybrid SSM, TQ KV, disk roundtrip, and DSV4/DSML tool parsing.\n\nI also found and fixed a current release-gate harness regression while reviewing this PR: tests/cross_matrix/run_api_surface_contract.py and run_cache_architecture_contract.py referenced a missing tests/cross_matrix/run_noheavy_api_cache_contract.py. Restored that runner and pushed commit 4df8b3b (test: restore API cache contract runner).\n\nVerification after the runner fix:\n- run_noheavy_api_cache_contract.py -> status=pass (15 API, 5 scheduler-cache, 32 TQ/MLLM-cache, 21 DSV4/DSML tool tests)\n- run_api_surface_contract.py -> status=pass\n- run_cache_architecture_contract.py -> status=pass\n- API parity/history focused pytest -> 38 passed\n\nLeaving this PR open as a future broader compatibility-API feature lane. Credit to @CobraSoftware for the local compatibility API direction; pieces should be split/landed with isolated endpoint specs and tests rather than merged wholesale into this release.

Thanks. I amd happy to continue working on this. If you have priority list let me know other wise I will open a pr for one change at a time.