feat(proxy): add health_check_reasoning_effort for model health checks

[Sameerlite]

Summary

Adds optional model_info.health_check_reasoning_effort so proxy health checks can set reasoning_effort (e.g. none) on chat-style probes only—useful for reasoning models where probes should minimize cost/latency.

Config

model_info:
  health_check_reasoning_effort: none

Applied when mode is unset (chat), chat, completion, skipped for all other modes for now

Tests

• tests/test_litellm/proxy/test_health_check_max_tokens.py::test_update_litellm_params_health_check_reasoning_effort

Docs: BerriAI/litellm-docs#71

---

Note

Low Risk
Low risk: a small, gated parameter addition limited to health-check request construction, with unit tests covering supported/unsupported modes.

Overview
Adds optional model_info.health_check_reasoning_effort so proxy health checks can forward reasoning_effort on chat-style probe requests (mode None, chat, or completion), while skipping other modes.

Extends the health-check max-tokens test suite with coverage ensuring reasoning_effort is set when configured (including dict values) and omitted for unsupported modes or absent config.

^{Reviewed by Cursor Bugbot for commit 32a5e77. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

This PR adds health_check_reasoning_effort to model_info, letting operators pass a reasoning_effort value (e.g. none) to chat-style health-check probes — useful for cost/latency reduction on reasoning models. The implementation is small and focused: a frozenset gate (None, "chat", "completion") in _update_litellm_params_for_health_check and a corresponding unit test.

Confidence Score: 5/5

Safe to merge — the change is additive, opt-in, and does not affect any request path unless health_check_reasoning_effort is explicitly configured.

No P0 or P1 findings. The only comment is a P2 test-coverage gap for mode="completion". Core logic is correct and well-guarded.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/proxy/health_check.py	Adds _HEALTH_CHECK_MODES_SUPPORTING_REASONING_EFFORT frozenset and applies health_check_reasoning_effort from model_info to litellm params only for None/chat/completion modes; logic is clean and correct.
tests/test_litellm/proxy/test_health_check_max_tokens.py	New test covers mode=None, mode=chat, dict value, excluded mode (embedding), and absent key — but mode="completion" (the third value in the frozenset) has no dedicated test case.

_{Reviews (4): Last reviewed commit: "feat(proxy): add health_check_reasoning_..." | Re-trigger Greptile}

[Sameerlite]

@greptile-apps re review

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Sameerlite]

@greptile-apps re review , my bad, I had added another code. Now it is only health check related code

[Sameerlite]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 32a5e77. Configure here.}

[mateo-berri]

LGTM but a small non blocking thought:

The only comment is a P2 test-coverage gap for mode="completion"

consider fixing Greptile's P2 test gap for mode="completion". Looking at the new test test_update_litellm_params_health_check_reasoning_effort in tests/test_litellm/proxy/test_health_check_max_tokens.py, the cases cover no mode key, mode="chat", no mode key -> dict value, mode="embedding", and empty model_info. mode="completion" is the third allowed value in the frozenset that you add to the gate, but never assert that the gate actually lets it through, so a regression that drops "completion" from the set wouldn't be caught. A quick assert would be nice to add

[Sameerlite]

LGTM but a small non blocking thought:

The only comment is a P2 test-coverage gap for mode="completion"

consider fixing Greptile's P2 test gap for mode="completion". Looking at the new test test_update_litellm_params_health_check_reasoning_effort in tests/test_litellm/proxy/test_health_check_max_tokens.py, the cases cover no mode key, mode="chat", no mode key -> dict value, mode="embedding", and empty model_info. mode="completion" is the third allowed value in the frozenset that you add to the gate, but never assert that the gate actually lets it through, so a regression that drops "completion" from the set wouldn't be caught. A quick assert would be nice to add

Fixed it, thanks!