fix(drivers-prompt): don't force tool_choice for Claude Opus 5.5 on Bedrock and Anthropic

[kimnamu]

Thanks for the version-keyed gate in anthropic_utils from #2251 and #2261; it made this a small change.

• I have read and agree to the contributing guidelines.

Describe your changes

Claude Opus 5.5 and Sonnet 5.5 reject a forced tool_choice (any or a named tool), which the default structured_output_strategy="tool" sends. This adds supports_forced_tool_choice() next to supports_sampling_params(), backed by a FORCED_TOOL_CHOICE_UNSUPPORTED_MIN_VERSIONS table ({"opus": (5, 5), "sonnet": (5, 5)}). Both use the same version parsing, now in a shared _is_at_or_above(). Both drivers skip the any override when it returns False, so the configured tool_choice (default auto) is sent. Sonnet 5.5 reached Bedrock on 2026-09-28 with the same 400, so c2dfddb adds its row.

case (live, us-east-1)	Before	After
AmazonBedrockPromptDriver + output_schema, us./global.anthropic.claude-opus-5-5, stream on/off	❌ ValidationException	✅ {'answer': 'OK'}
AmazonBedrockPromptDriver + output_schema, us./global.anthropic.claude-sonnet-5-5, stream on/off	❌ ValidationException	✅ {'answer': 'OK'}
AnthropicPromptDriver (AnthropicBedrock client) + output_schema, Opus 5.5 and Sonnet 5.5	❌ 400	✅ {'answer': 'OK'}
Opus 5, Sonnet 5: forced any still sent	✅	✅ unchanged
Non-Claude models (ai21.j2), rule strategy	✅	✅ unchanged (existing tests pass)
supports_sampling_params() results	✅	✅ unchanged (existing tests pass)

$ uv run pytest tests/unit/utils/test_anthropic_utils.py tests/unit/drivers/prompt/test_amazon_bedrock_prompt_driver.py tests/unit/drivers/prompt/test_anthropic_prompt_driver.py
207 passed
$ uv run pytest -n auto tests/unit/drivers/prompt tests/unit/utils tests/unit/structures tests/unit/tasks
1446 passed, 1 skipped
$ uv run ruff format --check . && uv run ruff check . && uv run typos
1236 files already formatted
All checks passed!
$ uv run pyright <changed source files>
0 errors, 0 warnings, 0 informations

RED: source files reset to main, new tests kept

12 of the 16 failures are the helper tests; the 4 driver cases still send any.

E       AttributeError: module 'griptape.utils.anthropic_utils' has no attribute 'supports_forced_tool_choice'
E       AssertionError: assert {'any': {}} == {'auto': {}}
E       AssertionError: assert {'type': 'any'} == {'type': 'auto'}
16 failed, 187 passed

Sonnet 5.5 rows (c2dfddb), anthropic_utils.py reset to 98f3e64, new rows kept:

FAILED tests/unit/utils/test_anthropic_utils.py::TestAnthropicUtils::test_supports_forced_tool_choice[us.anthropic.claude-sonnet-5-5-False]
FAILED tests/unit/drivers/prompt/test_amazon_bedrock_prompt_driver.py::TestAmazonBedrockPromptDriver::test_base_params_structured_output_tool_choice[us.anthropic.claude-sonnet-5-5-expected_tool_choice5]
FAILED tests/unit/drivers/prompt/test_anthropic_prompt_driver.py::TestAnthropicPromptDriver::test_try_run_structured_output_tool_choice[True-claude-sonnet-5-5-expected_tool_choice3]
FAILED tests/unit/drivers/prompt/test_anthropic_prompt_driver.py::TestAnthropicPromptDriver::test_try_run_structured_output_tool_choice[False-claude-sonnet-5-5-expected_tool_choice3]
4 failed, 23 passed

Live, Agent(output_schema=Schema({"answer": str})).run("Reply with answer=OK"), before (main):

us.anthropic.claude-opus-5-5       stream=False -> ErrorArtifact: An error occurred (ValidationException) when calling the Converse operation: The model returned the following errors: tool_choice: type "tool" and "any" are not supported for this model.
us.anthropic.claude-opus-5-5       stream=True  -> ErrorArtifact: An error occurred (ValidationException) when calling the ConverseStream operation: The model returned the following errors: tool_choice: type "tool" and "any" are not supported for this model.
global.anthropic.claude-opus-5-5   stream=False -> ErrorArtifact: An error occurred (ValidationException) when calling the Converse operation: The model returned the following errors: tool_choice: type "tool" and "any" are not supported for this model.
global.anthropic.claude-opus-5-5   stream=True  -> ErrorArtifact: An error occurred (ValidationException) when calling the ConverseStream operation: The model returned the following errors: tool_choice: type "tool" and "any" are not supported for this model.
us.anthropic.claude-opus-5         stream=False -> JsonArtifact: {'answer': 'OK'}
us.anthropic.claude-opus-5         stream=True  -> JsonArtifact: {'answer': 'OK'}
AnthropicPromptDriver us.anthropic.claude-opus-5-5   -> ErrorArtifact: Error code: 400 - {'message': 'tool_choice: type "tool" and "any" are not supported for this model.'}
AnthropicPromptDriver us.anthropic.claude-opus-5     -> JsonArtifact: {'answer': 'OK'}

After (this branch):

us.anthropic.claude-opus-5-5       stream=False -> JsonArtifact: {'answer': 'OK'}
us.anthropic.claude-opus-5-5       stream=True  -> JsonArtifact: {'answer': 'OK'}
global.anthropic.claude-opus-5-5   stream=False -> JsonArtifact: {'answer': 'OK'}
global.anthropic.claude-opus-5-5   stream=True  -> JsonArtifact: {'answer': 'OK'}
us.anthropic.claude-opus-5         stream=False -> JsonArtifact: {'answer': 'OK'}
us.anthropic.claude-opus-5         stream=True  -> JsonArtifact: {'answer': 'OK'}
global.anthropic.claude-sonnet-5   stream=False -> JsonArtifact: {'answer': 'OK'}
global.anthropic.claude-sonnet-5   stream=True  -> JsonArtifact: {'answer': 'OK'}
AnthropicPromptDriver us.anthropic.claude-opus-5-5   -> JsonArtifact: {'answer': 'OK'}
AnthropicPromptDriver us.anthropic.claude-opus-5     -> JsonArtifact: {'answer': 'OK'}

Sonnet 5.5, before (98f3e64) and after (c2dfddb):

before us.anthropic.claude-sonnet-5-5       stream=False -> ErrorArtifact: An error occurred (ValidationException) when calling the Converse operation: The model returned the following errors: tool_choice: type "tool" and "any" are not supported for this model.
before us.anthropic.claude-sonnet-5-5       stream=True  -> ErrorArtifact: An error occurred (ValidationException) when calling the ConverseStream operation: The model returned the following errors: tool_choice: type "tool" and "any" are not supported for this model.
before global.anthropic.claude-sonnet-5-5   stream=False/True -> same ValidationException
before AnthropicPromptDriver us.anthropic.claude-sonnet-5-5 -> ErrorArtifact: Error code: 400 - {'message': 'tool_choice: type "tool" and "any" are not supported for this model.'}
after  us.anthropic.claude-sonnet-5-5       stream=False -> JsonArtifact: {'answer': 'OK'}
after  us.anthropic.claude-sonnet-5-5       stream=True  -> JsonArtifact: {'answer': 'OK'}
after  global.anthropic.claude-sonnet-5-5   stream=False -> JsonArtifact: {'answer': 'OK'}
after  global.anthropic.claude-sonnet-5-5   stream=True  -> JsonArtifact: {'answer': 'OK'}
after  us.anthropic.claude-sonnet-5         stream=False -> JsonArtifact: {'answer': 'OK'}
after  us.anthropic.claude-sonnet-5         stream=True  -> JsonArtifact: {'answer': 'OK'}
after  AnthropicPromptDriver us.anthropic.claude-sonnet-5-5 -> JsonArtifact: {'answer': 'OK'}

With auto the model could answer without calling the tool. Opus 5.5 and Sonnet 5.5 each called it for all 5 prompts I tried, so I left that alone here.

Issue ticket number and link

Fixes #2351

References: announcement · migration guide

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[collindutter]

Thanks for the PR @kimnamu! Opened #2370 which adds on a few more fixes. I preserved your commits.