8 views
-/https://github.com/berriai/litellm/issues/34388
GitHub · issue

#34388 LiteLLM forwards unsupported strict tool fields to Bedrock Claude models, causing 400 errors and hanging requests

  • State: open
  • Author: @luiscastejon-aily
  • Labels: bug, proxy, llm translation

### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

When using PydanticAI with MCP tools through LiteLLM’s OpenAI-compatible proxy, some tool definitions contain `strict: true`. LiteLLM forwards `strict` and `additionalProperties` to Bedrock Converse without accounting for the specific Claude model or Bedrock tool type. We have observed different failures across multiple Claude models.

This appears to be a regression introduced in LiteLLM [v1.90.0](http://github.com/BerriAI/litellm/releases/tag/v1.90.0) which began forwarding `strict` and `additionalProperties` to Bedrock Converse. ### Claude Sonnet 4.6 and Claude Haiku 4.5 Bedrock reports that the compiled grammar exceeds its maximum allowed size. The deterministic Bedrock error is not returned promptly to the client. The request remains active for some minutes before nginx returns a 504 Gateway Time-out. The same request and MCP tools succeed through LiteLLM when a PydanticAI `prepare_tools` (docs [here](https://pydantic.dev/docs/ai/capabilities/prepare-tools/)) hook forces `strict=False` on every tool definition. No schemas or application …

Open on GitHub

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (4 events)
#0 of 0 · 31d18h51m52s ago — entered · #import:https:::github.com:berriai:litellm post #1998
-/https://github.com/berriai/litellm/issues/34388⭐4:1-/https://github.com/berriai/litellm/issues/23980
34388 demands broader cross-layer investigation, provider-specific compatibility logic, regression coverage, and validation of failure-handling behavior. 23980 is comparatively localized to model-resolution propagation and targeted tests.
#0 of 0 · 31d18h38m14s ago · #import:https:::github.com:berriai:litellm post #2224
-/https://github.com/berriai/litellm/issues/313853:2-/https://github.com/berriai/litellm/issues/34388⭐
31385 spans multiple streaming architectures and callback-dependent execution paths, requiring coordinated instrumentation and regression tests across APIs. 34388 is narrower provider-specific request normalization with targeted compatibility tests, though it carries model-behavior risk.
#0 of 0 · 31d18h15m51s ago · #import:https:::github.com:berriai:litellm post #2576
-/https://github.com/berriai/litellm/issues/34388⭐3:2-/https://github.com/berriai/litellm/issues/30765
The right issue is harder because it spans provider-specific schema normalization, model and tool-type compatibility, regression analysis, and request error/timeout handling. The left issue is comparatively localized to streaming state management and targeted SSE regression tests.
#0 of 0 · 31d18h7m50s ago — current · #import:https:::github.com:berriai:litellm post #2714
-/https://github.com/berriai/litellm/issues/34388⭐3:1-/https://github.com/berriai/litellm/issues/26097
The right-hand task requires broader provider-specific compatibility logic, regression analysis, and comprehensive testing across model variants and failure paths, while the left is a narrower installation and packaging configuration fix.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search