From d034fb681ca296a0f082a224043eda62a2ce82d4 Mon Sep 17 00:00:00 2001 From: Nate Armstrong Date: Sun, 15 Nov 2026 19:24:06 -0800 Subject: [PATCH] fix(messages): clamp bedrock effort before the drop_params effort gate (#51972) --- .../crates/inference-messages/src/handler.rs | 67 ++++++++++++++++++- .../llms/src/azure_ai/messages/AGENTS.md | 2 +- .../src/base_llm/messages/transformation.rs | 10 ++- .../anthropic_claude_transformation.rs | 11 ++- .../invoke_transformations/request.rs | 2 +- 5 files changed, 86 insertions(+), 6 deletions(-) diff --git a/litellm/rust/crates/inference-messages/src/handler.rs b/litellm/rust/crates/inference-messages/src/handler.rs index f44ae935da3..3138806c196 100644 --- a/litellm/rust/crates/inference-messages/src/handler.rs +++ b/litellm/rust/crates/inference-messages/src/handler.rs @@ -540,9 +540,20 @@ fn requested_optional_params( provider: MessagesProvider, context: &MessagesTransformContext, ) -> Result { - let request = litellm_llms::anthropic::messages::transformation::drop_unsupported_params( - request, context, + let requested_effort = request.params.reasoning_effort.clone(); + let clamped = provider + .config() + .clamp_requested_effort(request, &context.thinking.capabilities)?; + let gated = litellm_llms::anthropic::messages::transformation::drop_unsupported_params( + clamped, context, )?; + let request = MessagesRequest { + params: MessagesOptionalParams { + reasoning_effort: gated.params.reasoning_effort.and(requested_effort), + ..gated.params + }, + ..gated + }; if provider.accepts_service_tier() { return Ok(request); } @@ -1260,6 +1271,58 @@ mod tests { MessagesShaping::default() } + fn opus_4_6_on_bedrock() -> MessagesTransformContext { + use litellm_llms::base_llm::messages::context::{ + BedrockCapabilities, MessagesModelCapabilities, SupportedEffortTiers, + }; + MessagesTransformContext::new( + MessagesModelCapabilities { + supports_reasoning: true, + supports_adaptive_thinking: true, + supports_output_config: true, + effort_tiers: SupportedEffortTiers { + low: true, + medium: true, + high: true, + max: true, + ..SupportedEffortTiers::default() + }, + bedrock: BedrockCapabilities { + output_config_effort_ceiling: Some( + litellm_llms_types::formats::messages::EffortLevel::Max, + ), + ..BedrockCapabilities::default() + }, + ..MessagesModelCapabilities::default() + }, + true, + ) + } + + #[rstest] + #[case::bedrock_clamps_before_the_gate("bedrock/anthropic.claude-opus-4-6-v1", Some(json!("xhigh")))] + #[case::anthropic_has_no_clamp("anthropic/claude-opus-4-6", None)] + fn drop_params_keeps_an_effort_the_bedrock_ceiling_clamps( + #[case] model: &str, + #[case] kept: Option, + ) { + let resolved = resolve_llm_provider(model, None, "messages").unwrap(); + let provider = messages_provider(resolved.provider, resolved.model).unwrap(); + let requested = requested_optional_params( + body(json!({"model": resolved.model, "max_tokens": 4096, "messages": [{"role": "user", "content": "hi"}], "reasoning_effort": "xhigh"})), + provider, + &opus_4_6_on_bedrock(), + ) + .unwrap(); + assert_eq!( + serde_json::to_value(&requested.params) + .unwrap() + .get("reasoning_effort") + .cloned(), + kept + ); + } + fn body(value: Value) -> MessagesRequest { serde_json::from_value(value).unwrap() } diff --git a/litellm/rust/crates/llms/src/azure_ai/messages/AGENTS.md b/litellm/rust/crates/llms/src/azure_ai/messages/AGENTS.md index 85070710755..6553973ec42 100644 --- a/litellm/rust/crates/llms/src/azure_ai/messages/AGENTS.md +++ b/litellm/rust/crates/llms/src/azure_ai/messages/AGENTS.md @@ -8,7 +8,7 @@ - Caller credentials go out exactly as Python's dict and client send them (owner ruling, 2026-11-13). A caller `api-key` spelled exactly so replaces the deployment key and the Entra ID path. Otherwise the deployment key is added as `api-key`, or the Entra ID token as `Authorization` (replacing only that exact key), beside the caller's headers. `api-key` is renamed to `x-api-key` unless `x-api-key` is spelled exactly so, and names that differ only in case are joined into one header with `, ` - `anthropic-beta` goes through the `azure_ai` row of `anthropic_beta_headers_config.json` as Python's handler filters it: unknown and unsupported betas are dropped, the caller's and the ones the request needs alike - The beta-gated system field strip, system-role normalization and the `cache_control.scope` strip run in `finalize_request`, after the betas are picked from the request as sent, in the order Python's transform runs them -- This config alone reads messages and system prompts Rust could not type as Python does (`UnparsedMessages::Read`): the sanitizer, advisor and encrypted-reasoning strips rewrite them and their beta triggers count. Every other provider, `anthropic` included by the owner's ruling, keeps main's behaviour: such messages go as sent, and a system prompt that does not parse is main's request error (`messages_body`) +- Every native Messages provider (anthropic, azure_ai, bedrock invoke) reads messages and system prompts Rust could not type as Python does: the sanitizer, advisor and encrypted-reasoning strips rewrite them, their beta triggers count, and a system prompt that does not parse is kept. There is no per-provider switch for this # references diff --git a/litellm/rust/crates/llms/src/base_llm/messages/transformation.rs b/litellm/rust/crates/llms/src/base_llm/messages/transformation.rs index 74fa13d293c..54c3a971e36 100644 --- a/litellm/rust/crates/llms/src/base_llm/messages/transformation.rs +++ b/litellm/rust/crates/llms/src/base_llm/messages/transformation.rs @@ -7,7 +7,7 @@ use litellm_secret_value::SecretValue; use serde_json::Value; use super::{ - context::{MessagesShapeOptions, MessagesTransformContext}, + context::{MessagesModelCapabilities, MessagesShapeOptions, MessagesTransformContext}, tools::ToolCustomField, }; @@ -49,6 +49,14 @@ pub trait BaseMessagesConfig: Sync { Ok(request) } + fn clamp_requested_effort( + &self, + request: MessagesRequest, + _capabilities: &MessagesModelCapabilities, + ) -> Result { + Ok(request) + } + fn native_output_fallback( &self, _request: &MessagesRequest, diff --git a/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/anthropic_claude_transformation.rs b/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/anthropic_claude_transformation.rs index b0cba0c9d3a..94b06b70f88 100644 --- a/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/anthropic_claude_transformation.rs +++ b/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/anthropic_claude_transformation.rs @@ -4,7 +4,8 @@ use litellm_host::interceptors::RetriedRequest; use crate::{ anthropic::messages::handler::shape_anthropic_messages_request, base_llm::messages::context::{ - MessagesShapeOptions, MessagesTransformContext, NativeOutputSupport, + MessagesModelCapabilities, MessagesShapeOptions, MessagesTransformContext, + NativeOutputSupport, }, }; use litellm_auth::CredentialPlacement; @@ -232,6 +233,14 @@ impl BaseMessagesConfig for AmazonAnthropicClaudeMessagesConfig { shape_anthropic_messages_request(request, options) } + fn clamp_requested_effort( + &self, + request: MessagesRequest, + capabilities: &MessagesModelCapabilities, + ) -> Result { + request::clamp_adaptive_reasoning_effort_for_bedrock(request, capabilities) + } + fn get_complete_url( &self, api_base: Option<&str>, diff --git a/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/request.rs b/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/request.rs index d357bfd2a5d..9373bdd4acf 100644 --- a/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/request.rs +++ b/litellm/rust/crates/llms/src/bedrock/messages/invoke_transformations/request.rs @@ -97,7 +97,7 @@ fn get_bedrock_output_config_effort_ceiling( } /// Port of Python's `_clamp_adaptive_reasoning_effort_for_bedrock`. -fn clamp_adaptive_reasoning_effort_for_bedrock( +pub(super) fn clamp_adaptive_reasoning_effort_for_bedrock( request: MessagesRequest, capabilities: &MessagesModelCapabilities, ) -> Result {