fix(bedrock-converse): accept Reasoning(effort=..., summary=...) dict for reasoning_effort

[hclsys]

Companion to #28196 — the direct Anthropic path was already fixed (litellm/llms/anthropic/chat/transformation.py accepts both string and dict reasoning_effort, see #25359), but Bedrock Converse's map_openai_params still had the old isinstance(value, str) guard, silently dropping the dict shape OpenAI Responses callers send ({"effort": "low", "summary": "concise"}).

Symptom: reasoning_tokens == 0 on Bedrock Claude after upgrading; same as the direct path before fix.

fix

Same shape coercion as the Anthropic adapter — pull value["effort"] out before dispatch:

elif param == "reasoning_effort" and value is not None:
    effort_value: Any = value
    if isinstance(effort_value, dict):
        effort_value = effort_value.get("effort")
    if not isinstance(effort_value, str):
        continue
    self._handle_reasoning_effort_parameter(...)

proof

$ .venv/bin/python -m pytest tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py -k reasoning_effort -v
[...]
test_reasoning_effort_accepts_dict_shape_on_bedrock_converse[bedrock/converse/us.anthropic.claude-opus-4-7] PASSED
test_reasoning_effort_accepts_dict_shape_on_bedrock_converse[bedrock/converse/us.anthropic.claude-sonnet-4-6] PASSED
test_reasoning_effort_invalid_dict_does_not_crash_or_emit_thinking PASSED
(+ 15 pre-existing reasoning_effort tests)
====================== 18 passed, 113 deselected in 0.21s ======================

scope

Bedrock Converse only. Vertex/Databricks Anthropic-via-partner paths use the direct AnthropicConfig path (which is already fixed), so they pick this up transitively — but I haven't audited the partner-model dispatch on those providers, leaving that as a separate triage if reports come in.

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR fixes Bedrock Converse's map_openai_params to accept reasoning_effort values shaped as a dict ({"effort": "low", "summary": "concise"}), matching the coercion already present in the direct Anthropic path. Previously the isinstance(value, str) guard silently dropped dict-shaped inputs, causing reasoning_tokens == 0 on Bedrock Claude for callers using the OpenAI Responses bridge.

• The guard is changed to value is not None with an inline dict-to-string extraction (effort_value.get("effort")); any non-string result after coercion is skipped with continue, which is safe because the subsequent if blocks in the loop check other param names and would not match "reasoning_effort" regardless.
• Two new pure-transformation unit tests are added: a parametrized happy-path test for adaptive Claude 4.6/4.7 models and a defensive test for malformed dicts (missing "effort" key).

Confidence Score: 5/5

Safe to merge — the change is a minimal guard relaxation with matching tests, and the coercion logic is identical to the already-merged Anthropic direct-path fix.

The fix is narrow: one guard expression replaced with an equivalent dict-unpacking pattern lifted directly from litellm/llms/anthropic/chat/transformation.py. The continue in the inner branch only skips param-specific checks that would not match "reasoning_effort" anyway, so no existing code path is affected. New tests cover both the happy path and the malformed-dict edge case, and no existing tests were weakened.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/bedrock/chat/converse_transformation.py	Relaxes the reasoning_effort guard from isinstance(value, str) to value is not None, adding dict-to-string coercion that mirrors the existing Anthropic direct-path handler; logic is correct and minimal.
tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py	Adds two new unit tests: one parametrized happy-path test for dict-shaped reasoning_effort on adaptive Claude 4.6/4.7 models, and one defensive test for a malformed dict (missing "effort" key). No existing tests modified, no network calls made.

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

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing hclsys:fix/bedrock-reasoning-effort-dict-coerce (afc5085) with main (a021a5b)}

[codecov]

Codecov Report

❌ Patch coverage is 88.88889% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...tellm/llms/bedrock/chat/converse_transformation.py	88.88%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[krrish-berri-2]

@hclsys — two quick things before this gets reviewed:

1. The LiteLLM Linting / lint check is failing, and it looks related to the Python changes in this PR. Could you take a look and fix the lint issues?
2. Could you also add a screenshot or short video showing a Bedrock Converse request with a Reasoning(effort=..., summary=...) dict being accepted end-to-end? Visual proof really speeds up review.

Thanks!

[hclsys]

companion PR #29330 (same one-line dict coercion for the Databricks Claude 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. Either can merge first; the other is a stand-alone identical pattern in a different file.

[hclsys]

@krrish-berri-2 quick on both:

lint is now green — your earlier comment was on the pre-ruff-fix commit; commit afc508570 extracted _coerce_reasoning_effort to keep map_openai_params under PLR0915, ruff check is passing now (see lint row in current checks).

before/after exercised directly against AmazonConverseConfig.map_openai_params for adaptive Claude 4.7:

from litellm.llms.bedrock.chat.converse_transformation import AmazonConverseConfig
cfg = AmazonConverseConfig()
non_default = {"reasoning_effort": {"effort": "low", "summary": "concise"}}
op = cfg.map_openai_params(non_default_params=non_default, optional_params={},
    model="bedrock/converse/us.anthropic.claude-opus-4-7", drop_params=False)
print(op.get("thinking"), op.get("output_config"))

unpatched main:

None None    # ← dict shape silently dropped, thinking never gets set, reasoning_tokens=0 at runtime

this PR:

{'type': 'adaptive'} {'effort': 'low'}    # ← identical to passing reasoning_effort="low"

and the bare string still works (regression-guarded by test_reasoning_effort_bare_string_still_works in #29330, same pattern):

PATCHED dict: {'thinking': {'type': 'adaptive'}, 'output_config': {'effort': 'low'}}
PATCHED str:  {'thinking': {'type': 'adaptive'}, 'output_config': {'effort': 'low'}}
match: True

[hclsys]

redid the branch on top of internal_staging — three commits squashed into the three logic commits, no force-push of new content. waiting on CI.

[kimnamu]

Thanks for closing the Bedrock Converse gap here — I ran the branch locally (923a6ccb) and read through _coerce_reasoning_effort, and it checks out cleanly.

• Sibling consistency is the strongest part. The helper normalizes to the same bare-string shape the direct Anthropic adapter expects, and keeps the silent-drop-on-malformed semantics identical (if dict → .get("effort"), then drop if not a string). Mirroring the already-merged Anthropic path (rather than introducing a fail-fast that diverges across providers) is the right call — and the docstring spelling that out is a nice touch.
• Verified the edge cases directly against the helper logic: "low"→low, {"effort":"low","summary":"concise"}→low, {"summary":"concise"} (no effort)→None, {"effort":None}→None, 5→None, {}→None, None→None. No crash on any malformed shape — the OpenAI Responses Reasoning(effort, summary) dict and the defensive cases are all covered.
• The test genuinely catches the bug. Reverting just the guard back to isinstance(value, str) (keeping the new tests) makes the happy-path parametrized case fail (the dict is dropped → thinking never set); restoring the fix → passes. So it's a real regression guard. The walrus-in-elif to stay under PLR0915 is tidy and the skip is safe.

Minimal, well-mirrored, good regression coverage. I'd love to see it land.

(I'm not a maintainer — just a Bedrock user who reviewed and ran this locally. Prepared with the help of an AI agent (Claude Code), human-verified.)