fix(playground): preserve tool choice and strict setting for Anthropic and Bedrock

[adimalkar]

Fixes #15644

The server provider clients drop parts of the tool contract before the request goes out. Because playground calls, evaluator previews, and online evaluator execution all share these clients, the same loss shows up in each.

1. Anthropic tool choice is overwritten

_anthropic_message_params builds tool_choice from tools.tool_choice, then immediately discards it:

                else:
                    assert_never(tc.type)
            if tools.disable_parallel_tool_calls:          # ← runs after, unconditionally
                params["tool_choice"] = ToolChoiceAutoParam(
                    type="auto", disable_parallel_tool_use=True
                )

Every branch of the block above already applies disable_parallel_tool_use, so this overwrite was only ever needed as the fallback for when no explicit choice was given. As written it fires whenever the flag is set, regardless:

tool_choice	disable_parallel_tool_calls	Sent before this PR
specific_function	True	{"type": "auto", ...} — named tool lost
one_or_more	True	{"type": "auto", ...} — "must call a tool" lost
none	True	{"type": "auto", ...} — tool use enabled when it was turned off

The none case is worth calling out separately: it isn't in the issue description, but it inverts the user's intent rather than merely relaxing it — "do not call tools" becomes "call tools freely".

Fix: make the fallback an elif, so an explicit choice wins and the auto default still applies when there is none.

2. Anthropic and Bedrock drop the tool's strict setting

Both function-tool builders pass only name / schema / description, so PromptToolFunctionDefinition.strict never reaches the provider. The OpenAI builder already forwards it:

# OpenAI, existing
strict=f.strict if isinstance(f.strict, bool) else None,

Both target types accept the field, so this is a straightforward omission rather than an unsupported capability:

• anthropic.types.ToolParam → strict
• Bedrock ToolSpecificationTypeDef → strict: NotRequired[bool]

Fix: forward it when it is a bool, matching the adjacent description handling so an unset value is still omitted.

Acceptance criteria

• Anthropic keeps a specific tool choice when parallel tool use is disabled
• Anthropic sends the supported strict tool field from the stored prompt tool
• Bedrock sends the supported strict tool field from the stored prompt tool
• Request-shape tests cover the cases above

Tests

Seven tests added to tests/unit/server/api/helpers/test_playground_clients.py, split intentionally:

Four fail without the source change — the corrected behaviour:

• test_specific_tool_choice_survives_disable_parallel_tool_calls
• test_tool_choice_none_survives_disable_parallel_tool_calls
• TestAnthropicStreamingClient::test_function_tool_strict_is_forwarded
• TestBedrockClient::test_function_tool_strict_is_forwarded

Three pass with and without it — guards that this PR does not change behaviour it shouldn't:

• test_disable_parallel_tool_calls_without_explicit_choice_falls_back_to_auto — the auto fallback still works when no explicit choice is given
• test_function_tool_omits_strict_when_unset (Anthropic and Bedrock) — an unset strict is still omitted, not sent as None

Verified by stashing only the source change and re-running: 4 failed, 3 passed.

Verification

• pytest tests/unit/server/api/helpers/ → 505 passed, 113 skipped
• ruff check and ruff format --check clean on both changed files (both confirmed clean at main first, so no unrelated reformatting is mixed in)
• Source diff is +5 / −1

Note on the local environment

I could not reproduce the CI dependency set exactly, so CI is the real check on this:

• requirements/unit-tests.txt did not resolve for me — litellm>=1.83.14 requires openai>=2.20.0,<3.0.0, while another pinned dependency requires openai>=3.1.0. requirements/type-check.txt hits a related openai conflict.
• I ran the suite with ci.txt plus the test dependencies, excluding litellm and type-check.txt. Nothing in the changed code paths touches litellm.

Happy to adjust if there is a lockfile or resolver setting I missed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[adimalkar]

I have read the CLA Document and I hereby sign the CLA