fix(databricks): accept Reasoning(effort=..., summary=...) dict for reasoning_effort

[hclsys]

Third + final adapter in the #28196 lane — direct Anthropic path was fixed in #25359, Bedrock Converse companion is #29329, this is Databricks Claude.

Same root cause: map_openai_params passed the raw reasoning_effort straight to AnthropicConfig._map_reasoning_effort, which compares with string equality (reasoning_effort == "low"), so the OpenAI Responses Reasoning(effort, summary) dict shape returned None → no thinking / output_config set. The downstream isinstance(reasoning_effort_value, str) adaptive-Claude check would then short-circuit, surfacing a misleading invalid reasoning_effort from _raise_invalid_reasoning_effort.

fix

One-liner coercion before any downstream branch — matches what bedrock + anthropic already do:

if isinstance(reasoning_effort_value, dict):
    reasoning_effort_value = reasoning_effort_value.get("effort")

proof

$ .venv/bin/python -m pytest tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py -k reasoning_effort -v
[...]
test_reasoning_effort_accepts_dict_shape[databricks/databricks-claude-3-7-sonnet] PASSED
test_reasoning_effort_accepts_dict_shape[databricks/databricks-claude-opus-4-7] PASSED
test_reasoning_effort_bare_string_still_works PASSED
======================= 3 passed, 11 deselected in 0.09s =======================

sibling tests in the same file (11 pre-existing) still pass.

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

Fixes the Databricks Claude adapter to accept reasoning_effort as either a plain string ("low") or the OpenAI Responses dict shape ({"effort": "low", "summary": "concise"}), by coercing the dict to its "effort" value before passing it to AnthropicConfig._map_reasoning_effort. This is the third of three adapters patched in the #28196 lane, matching identical coercions already present in the direct Anthropic and Bedrock Converse adapters.

• transformation.py: Adds a two-line isinstance(dict) guard that extracts effort from the dict before the downstream string-comparison branches run, preventing None being forwarded and thinking/output_config being silently dropped.
• test_databricks_chat_transformation.py: Adds three new mock-only unit tests — parametrised over claude-3-7-sonnet and claude-opus-4-7 for the dict path, and a separate bare-string regression guard — with no modifications to existing tests.

Confidence Score: 5/5

Safe to merge — the change is a minimal, well-tested one-liner that mirrors an identical fix already applied to two sibling adapters.

The coercion is isolated to a local variable, does not touch non_default_params, and the post-coercion path (string equality checks, adaptive-model output_config branch) is identical to what worked before for bare-string callers. Three new mock-only unit tests cover both the dict and string input shapes across the affected models, and no existing tests were modified.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/databricks/chat/transformation.py	Adds dict-to-string coercion for reasoning_effort before passing to AnthropicConfig._map_reasoning_effort, matching the pattern already used in the direct Anthropic and Bedrock Converse adapters.
tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py	Adds three new unit tests covering the dict-shape and bare-string paths for reasoning_effort; all tests are mocked with no real network calls, and no existing tests are modified.

_{Reviews (1): Last reviewed commit: "fix(databricks): accept Reasoning(effort..." | Re-trigger Greptile}

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing hclsys:fix/databricks-reasoning-effort-dict-coerce (c2b8927) with litellm_internal_staging (39c01fe)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[krrish-berri-2]

@hclsys — could you add a screenshot or short video showing a Databricks request with a Reasoning(effort=..., summary=...) dict being accepted end-to-end? Visual proof really helps reviewers verify the fix quickly. Thanks!

[hclsys]

companion PR #29329 (same one-line dict coercion for the Bedrock Converse adapter) — both bedrock + databricks have the same #28196 root cause that the direct Anthropic adapter already fixed via #25359. Same coercion, sequential audit of the 3 anthropic-via-X adapters in litellm.

[hclsys]

@krrish-berri-2 before/after exercised against DatabricksConfig.map_openai_params for adaptive Claude 4.7:

from litellm.llms.databricks.chat.transformation import DatabricksConfig
cfg = DatabricksConfig()
op = cfg.map_openai_params(
    non_default_params={"reasoning_effort": {"effort": "low", "summary": "concise"}},
    optional_params={},
    model="databricks/databricks-claude-opus-4-7",
    drop_params=False,
    replace_max_completion_tokens_with_max_tokens=False,
)
print(op.get("thinking"), op.get("output_config"))

unpatched main — actually worse than silent-drop here, it raises:

litellm.exceptions.BadRequestError: litellm.BadRequestError: Invalid reasoning_effort:
{'effort': 'low', 'summary': 'concise'}. Must be one of: 'minimal', 'low', 'medium', 'high', 'xhigh', 'max', 'none'

this PR:

{'type': 'adaptive'} {'effort': 'low'}    # ← dict accepted, identical to passing reasoning_effort="low"
match str path: True

[hclsys]

branch now starts from internal_staging head. same one-commit change, just clean base. CI should pass.

[hclsys]

@krrish-berri-2 best i can offer from this env is the test stdout exercising the Reasoning(effort=..., summary=...) dict path end-to-end through DatabricksConfig.map_openai_params:

$ uv run pytest tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py -k reasoning_effort
3 passed in 0.23s

CI is green. happy to add more coverage if you want it.