fix(bedrock): forward output_config.effort for adaptive-thinking Claude

[mateo-berri]

Relevant issues

Customer report on the LiteLLM Anthropic-via-Bedrock route: output_config.effort for Claude Sonnet 4.6 was being silently dropped on every Bedrock surface (Converse and Invoke, /v1/messages and /completion). Tested across v1.83.7, v1.83.14, and v1.84.0 and the parameter never appeared in the wire body.

Linear ticket

Resolves LIT-2758

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory
• My PR passes all unit tests on make test-unit (136 bedrock unit tests pass locally)
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

Bedrock exposes Anthropic's effort parameter on three routes:

Route	Wire shape
Bedrock Invoke /v1/messages (Anthropic Messages API)	top-level output_config: {effort: ...}
Bedrock Invoke /completion (legacy chat invoke)	top-level output_config: {effort: ...}
Bedrock Converse	nested additionalModelRequestFields.thinking.effort

Adaptive-thinking models (Claude 4.6 / 4.7) accept effort natively with no beta header. Claude Opus 4.5 accepts it on Bedrock Invoke when the effort-2025-11-24 beta is attached.

LiteLLM was unconditionally stripping output_config on every Bedrock route — a defensive measure originally added because pre-4.6 Claude on Bedrock 400s on the field. That defense was correct for older models but caused effort to vanish for Sonnet 4.6 and friends.

This change splits the strip behavior by model:

Bedrock Invoke /v1/messages (AmazonAnthropicClaudeMessagesConfig)

• Add a model-aware _supports_effort_on_bedrock check (Claude 4.6/4.7 or Opus 4.5).
• For supported models, leave output_config intact; for everyone else, strip it as before.
• Add output_config to the BedrockInvokeAnthropicMessagesRequest TypedDict so the body allowlist preserves it on the supported path. Existing allowlist tests use Haiku and stay green.

Bedrock Invoke /completion (AmazonAnthropicClaudeConfig)

• Mirror the same model-aware strip in _build_bedrock_anthropic_request_base.

Bedrock Converse (AmazonConverseConfig)

• Add _fold_output_config_effort_into_thinking: if the caller supplied output_config.effort on an adaptive-thinking model, fold it into thinking.effort (creating thinking: {type: adaptive, effort: ...} if absent, never overriding an explicit caller value). output_config is then stripped before signing.
• Update _handle_reasoning_effort_parameter so that on adaptive-thinking models, OpenAI reasoning_effort populates thinking.effort (with type=adaptive) instead of the legacy thinking.type=enabled, budget_tokens=... shape.

Beta-headers config

• Map effort-2025-11-24 to itself for bedrock (was null), so the auto-attach in is_effort_used (already triggered by output_config.effort on Opus 4.5) is preserved instead of dropped.

Tests

• test_bedrock_messages_preserves_output_config_for_adaptive_thinking_models — Sonnet 4.6, Opus 4.6, Opus 4.7 (regional + global) all forward output_config, no beta header.
• test_bedrock_messages_preserves_output_config_for_opus_4_5_with_beta — Opus 4.5 forwards output_config and auto-attaches the effort-2025-11-24 beta.
• test_output_config_preserved_for_adaptive_thinking_on_bedrock_invoke and test_output_config_preserved_for_opus_4_5_on_bedrock_invoke — same coverage on /completion.
• test_converse_folds_output_config_effort_into_thinking_for_46, test_converse_folds_output_config_effort_preserves_existing_thinking, test_converse_reasoning_effort_sets_thinking_effort_for_46 — Converse fold path with explicit-thinking precedence.

Existing strip-output_config tests (Haiku, Sonnet 4, Nova) all stay green and still assert removal — the model-aware split keeps behavior unchanged for unsupported models.

Note on the documentation question

The reporter also flagged that on the OpenAI-compatible route, reasoning_effort maps to thinking: {type: "adaptive"} for Claude 4.6+, which was surprising. That mapping is correct (and documented in litellm/llms/anthropic/chat/transformation.py::_map_reasoning_effort and common_utils.py::_is_adaptive_thinking_model) — Anthropic's adaptive thinking is the recommended replacement for budget_tokens on those models. Documentation lives in the separate BerriAI/litellm-docs repo and is not modified here, but the docstring on _map_reasoning_effort already explains the behavior.

Slack Thread

  

[codecov]

Codecov Report

❌ Patch coverage is 93.93939% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...tellm/llms/bedrock/chat/converse_transformation.py	90.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[mateo-berri]

Superceded by #27074