Bedrock provider: support adaptive thinking and output_config.effort for Claude 4.6 / 4.7

[elliolso-amazon]

Summary

The Bedrock provider's Claude reasoning configuration only emits the legacy {"type": "enabled", "budget_tokens": N} form. It does not support the adaptive thinking configuration used by Claude Sonnet 4.6 and Claude Opus 4.7, nor the associated output_config.effort field. As a result, GenerateConfig.reasoning_effort and GenerateConfig.effortare silently ignored when targeting Claude models through Bedrock.

Current Behavior

In src/inspect_ai/model/_providers/bedrock.py:521-528, BedrockAPI.reasoning_config() handles Claude as follows:

elif self.is_claude():
  if config.reasoning_tokens is not None:
    return {
      "reasoning_config": {
        "type": "enabled",
        "budget_tokens": config.reasoning_tokens,
      }
    }

This branch:

• Only reacts to config.reasoning_tokens; config.reasoning_effort and config.effort are not read.
• Emits only type: "enabled" with budget_tokens; type: "adaptive" is not supported.
• Does not emit an output_config field.

The gpt-oss and Nova branches in the same method do consume reasoning_effort, so the omission is specific to Claude.

Expected Behavior

A request such as the following should be producible from the provider:

response = client.converse(
  modelId="us.anthropic.claude-opus-4-7",
  messages=[{"role": "user", "content": [{"text": "..."}]}],
  inferenceConfig={"maxTokens": 128000},
  additionalModelRequestFields={
    "thinking": {"type": "adaptive"},
    "output_config": {"effort": "high"},
  },
)

Concretely, the Bedrock provider should:

1. Emit type: "adaptive" (with an output_config.effort) when config.reasoning_effort is set on a Claude model that supports adaptive thinking (4.6 and later).
2. Emit output_config.effort from config.effort independently of thinking mode, with the same version-gated adjustments applied in the native provider (e.g. xhigh -> high pre-4.7 models).
3. Continue to support the existing type: "enabled" + budget_tokens when only config.reasoning_tokens is set.

Reference Implementation

The native Anthropic provider already implements the desired semantics at src/inspect_ai/model/_providers/anthropic.py:783-805:

# effort
if config.effort is not None:
  ...
  params["output_config"] = OutputConfigParam(effort=effort)

# some thinking-only stuff
if self.is_using_thinking(config):
  reasoning_effort = self.effort_from_reasoning_effort(config)
  if reasoning_effort is not None:
    params["thinking"] = dict(type="adaptive", display="summarized")
    params["output_config"] = OutputConfigParam(effort=reasoning_effort)
  else:
    params["thinking"] = dict(
      type="enabled",
      budget_tokens=config.reasoning_tokens,
      display="summarized",
    )

The Bedrock provider can mirror this logic, placing the resulting fields into additionalModelRequestFields rather than the top-level Messages API parameter.

Environment

• inspect_ai main branch (commit d09d608f8 at time of filing)
• Affected file: src/inspect_ai/model/_providers/bedrock.py
• Affected models: us.anthropic.claude-opus-4-7 and us.anthropic.claude-sonnet-4-6