feat(bedrock): forward strict and additionalProperties to Converse toolSpec

[mateo-berri]

Relevant issues

A customer reported that strict: true on tools is silently ignored for Claude on Bedrock: give a tool an enum and ask for a value outside it, and GPT and direct-Anthropic stay inside the enum while Claude-on-Bedrock returns the out-of-enum value. This re-applies the fix from #25209, which was correct and approved (Greptile 5/5) but went stale with merge conflicts since April after unrelated refactors landed in _bedrock_tools_pt. The hunks no longer applied, but the change itself was still valid, so this is a clean rebase onto current staging rather than a rewrite. Related: #10062, #16725, #6136

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• 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

CI (LiteLLM team)

• Branch creation CI run
  Link:
• CI run for the last commit
  Link:
• Merge / cherry-pick CI run
  Links:

Screenshots / Proof of Fix

Reproduce against a live proxy hitting real Bedrock (Claude). Define a tool whose only enum value is celsius and ask for Fahrenheit; before this change the call comes back with additionalProperties/strict dropped and the model free to pick an out-of-enum value, after it the schema reaches Bedrock intact.

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    "messages": [{"role": "user", "content": "What is the weather in Dallas in Fahrenheit? Call the tool."}],
    "tool_choice": "required",
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_weather",
        "strict": true,
        "description": "Get the weather for a city",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {"type": "string"},
            "unit": {"type": "string", "enum": ["celsius"]}
          },
          "required": ["city", "unit"],
          "additionalProperties": false
        }
      }
    }]
  }' | jq '.choices[0].message.tool_calls[0].function.arguments'

With the fix the returned unit stays "celsius" (the only enum member) instead of "fahrenheit".

Type

🐛 Bug Fix

Changes

litellm/types/llms/bedrock.py: add additionalProperties: bool to ToolJsonSchemaBlock and strict: bool to ToolSpecBlock (both total=False, so the keys stay optional).

litellm/litellm_core_utils/prompt_templates/factory.py: in _bedrock_tools_pt, forward strict from the OpenAI function and additionalProperties from the schema, each only when present. Bedrock requires additionalProperties: false alongside strict: true or it rejects the request, so both have to travel together.

tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py: add test_bedrock_tools_pt_strict_parameter in the mapped unit-test suite, covering both the pass-through and the absent-by-default cases. It fails (KeyError on strict) without the factory change and passes with it.

[greptile-apps]

Greptile Summary

This PR fixes a silent-drop bug in the Bedrock Converse path: strict on an OpenAI-format tool function and additionalProperties on its schema were both discarded before reaching Bedrock, allowing Claude on Bedrock to ignore enum constraints. The fix gates forwarding of both fields behind a supports_strict_tools check (Anthropic models only), since other Bedrock families reject the strict toolSpec field.

• factory.py: _bedrock_tools_pt now evaluates get_bedrock_base_model(model).startswith("anthropic") once per call, then conditionally writes additionalProperties into the JSON schema block and strict into ToolSpecBlock — each only when the field is present in the caller's input.
• bedrock.py: ToolJsonSchemaBlock and ToolSpecBlock are extended with the two new optional fields so the TypedDict annotations match the emitted payload.
• Tests: test_bedrock_tools_pt_strict_parameter is a pure unit test covering three cases (Claude+strict, Nova+strict-dropped, Claude-no-strict); no network calls are made.

Confidence Score: 5/5

Safe to merge — the change is narrowly scoped to the Bedrock tool-spec transformation and only activates for Anthropic models.

The logic is small, well-guarded by the supports_strict_tools flag, and covered by a new unit test that exercises the positive, negative, and absent-field cases. The only annotation inaccuracy (additionalProperties: bool vs Union[bool, dict]) has no runtime impact.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/types/llms/bedrock.py	Adds additionalProperties: bool to ToolJsonSchemaBlock and strict: bool to ToolSpecBlock; both are total=False so they remain optional. The additionalProperties annotation is narrowed to bool only, but JSON Schema allows a schema-object value too.
litellm/litellm_core_utils/prompt_templates/factory.py	Adds supports_strict_tools guard (Anthropic-only) in _bedrock_tools_pt, then conditionally forwards additionalProperties from the JSON schema and strict from the function spec. Logic is well-isolated and short-circuits on a falsy model.
tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py	Adds test_bedrock_tools_pt_strict_parameter covering Claude with strict=True/additionalProperties=False, Nova (strict rejected), and Claude without strict. Pure unit test, no network calls.

_{Reviews (4): Last reviewed commit: "fix(bedrock): only forward strict tool s..." | Re-trigger Greptile}

[mateo-berri]

@greptileai

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[yassin-berriai]

doesn't this mean that BedrockToolJsonSchemaBlock is incomplete / incorrect? this code basically sidesteps it and mutates the resulting dict directly

[mateo-berri]

I add both additionalProperties and strict to those TypedDicts, so the json_schema["additionalProperties"] = ... assignment is fully type-checked, not a bypass. I intended for this post-construction assignment. The keys are conditional (Claude-only, only when supplied) and should be absent rather than None when they don't apply

[yassin-berriai]

yea, why can't the constructor do this @mateo-berri

[mateo-berri]

it can't conditionally because a kwarg is always passed; the only constructor option is **-unpacking an extras dict, which still needs the same is not None guard and is less type-safe (mypy can check json_schema["key"] = v per-key, but ofc it cannot validate a **dict unpack). So the current post-construction assignment is actually the more type-checked choice