Merge remote-tracking branch 'origin/main' into litellm_fix_chat_mcp_stream_billing

AGENTS.md

@@ -31,6 +31,8 @@ A test must only fail when litellm code changes. Never pin facts we don't own (a
 
 End-to-end tests belong in `tests/e2e/`. A rule about one directory's files, like its harness or which CI runs it, goes in that directory's `AGENTS.md`, never only in a README or here. Read a directory's `AGENTS.md` before editing files in it
 
+Before you build a new way to move data between the endpoints, the provider transformations, and cost tracking or logging, read the architecture decision records in `docs/adrs/`, which say why those layers are shaped the way they are
+
 When creating PRs, target the repository's current default branch for both internal and external / OSS contributions. Check it with `python3 scripts/default_branch.py --branch` instead of assuming a branch name or relying on cached `origin/HEAD`
 
 When writing a PR body, treat the comments and imperative instructions inside .github/PULL_REQUEST_TEMPLATE/general.md as rules to follow, not just layout. Agent harnesses may strip HTML comments from copies of that file injected into context, so read .github/PULL_REQUEST_TEMPLATE/general.md from disk before writing a PR body to make sure you see every comment rule. A section you have nothing to put in (Relevant issues, Affected release, Linear ticket, Caveats, QA runbook, and so on) is removed entirely, heading included, never left as an empty title

ARCHITECTURE.md

@@ -2,6 +2,8 @@
 
 This document helps contributors understand where to make changes in LiteLLM.
 
+For why a layer is shaped the way it is, read the architecture decision records in [`docs/adrs/`](docs/adrs/AGENTS.md) (Python SDK and proxy) and [`litellm-rust/docs/adrs/`](litellm-rust/docs/adrs/AGENTS.md) (Rust core)
+
 ---
 
 ## How It Works

CONTRIBUTING.md

@@ -8,6 +8,7 @@ Here are the core requirements for any PR submitted to LiteLLM:
 
 - [ ] **Sign the Contributor License Agreement (CLA)** - [see details](#contributor-license-agreement-cla)
 - [ ] **Keep scope isolated** - Your changes should address 1 specific problem at a time
+- [ ] **Prove it works end to end** - If your change affects what the proxy or SDK does, paste the commands you ran against a live proxy and their output in the PR - [see details](#proving-your-change-works-end-to-end)
 
 #### Proxy (Backend) PRs
 
@@ -248,6 +249,41 @@ If `make test-unit` fails:
 - **Test edge cases**: Don't just test the happy path
 - **Update documentation**: If you change APIs, update docs
 
+## Proving Your Change Works End to End
+
+Tests are required, but they do not prove that your change works for a user. A PR that changes what the proxy or SDK does shows a real run in its "Screenshots / Proof of Fix" section, meaning the exact commands you sent to a live proxy and the output that came back, against real provider APIs and not mocks. `pytest` output does not count as proof, because a passing test only shows that the code does what its mocks expect. A PR that changes only docs or CI shows the rendered page or the CI run instead
+
+A good proof looks like this:
+
+1. Install the proxy dependencies with `make install-proxy-dev`, then start the proxy from your branch with a config that has the model you changed, `uv run litellm --config your_config.yaml --detailed_debug`
+2. Send the request a user would send with `curl`, and paste both the command and the response. When cost or routing is involved, include the response headers, since `x-litellm-response-cost` is what the user sees
+3. For a bug fix, do this twice, once at the merge base with BerriAI/litellm's main (`git merge-base origin/main HEAD`, with your remote for BerriAI/litellm in place of `origin` in a fork clone) to show the bug and once at your branch's latest commit to show the fix. Name both commit hashes
+4. If your change is reachable from more than one endpoint (`/v1/chat/completions`, `/v1/responses`, `/v1/messages`), show each one
+5. If the request can stream, show a `"stream": true` run too, because streaming builds usage and cost on a different path. A stream sends its headers before its usage exists, so show the `usage.cost` on the last chunk, which `/v1/chat/completions` sends when the request has `"stream_options": {"include_usage": true}` and other endpoints send with `include_cost_in_streaming_usage: true` under `litellm_settings`, or, when the proxy has a database, the request's row on the Admin UI's Logs page or `GET /spend/logs`
+6. For UI changes, add before and after screenshots and name the page
+7. If users reach your change through a coding tool such as Claude Code or Codex, drive that tool the way a user does and add before and after screenshots of it, because a `curl` or a headless run alone does not show what the user sees
+
+```bash
+curl -sD - http://localhost:4000/v1/chat/completions \
+  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "your-model", "messages": [{"role": "user", "content": "hi"}]}'
+```
+
+When the provider reports its own usage or spend numbers, compare the proxy's numbers against them in the proof
+
+## Contributing an LLM Provider Integration
+
+Provider work (a new provider, a new parameter, a new usage or cost field) is the most common kind of contribution, and it is where PRs get reworked most often. Follow these rules
+
+**Find the existing machinery before you add plumbing.** LiteLLM already has generic paths that move data between the endpoints, the provider transformations, and cost tracking. A second, hand-rolled path for the same data is a common reason a provider PR gets rewritten before merge. [`ARCHITECTURE.md`](ARCHITECTURE.md) shows where the layers live, and the decision records in [`docs/adrs/`](docs/adrs/AGENTS.md) explain why they are shaped that way. For example, [the ADR on provider usage fields](docs/adrs/adr_001_provider_usage_fields_ride_as_usage_extras.md) covers how provider-specific usage fields reach cost tracking. If you cannot find a mechanism for what you need, say so in the PR or ask in [#pr-review on Slack](https://join.slack.com/t/litellmossslack/shared_invite/zt-3o7nkuyfr-p_kbNJj8taRfXGgQI1~YyA) before you build one
+
+**Bill a new provider usage field from the usage object.** Leave the field on the usage object as an extra and read it in that provider's `cost_calculator.py`. A provider without one gets a new `cost_calculator.py` and an arm for it in `cost_per_token` (`litellm/cost_calculator.py`), which picks the calculator by provider. If the provider's transform builds `Usage` from named fields, as the Anthropic, Gemini, and Bedrock Converse ones do, pass the new field there too. Streamed requests rebuild usage in `ChunkProcessor.calculate_usage` (`litellm/litellm_core_utils/streaming_chunk_builder_utils.py`), which does not copy extras yet the way the Responses-to-chat bridge does (minus a short exclusion list), so until it does, bill the field on streams by naming it there, as Anthropic's `inference_geo` is. A provider that reports its own charge for the request, as OpenRouter does in `usage.cost`, is billed at that charge from `_hidden_params` before any calculator runs, unless the deployment sets its own prices. When the field counts a built-in tool such as web search, add the provider to `get_cost_for_web_search_request` in `litellm/llms/__init__.py`, and add a gate to `litellm/litellm_core_utils/llm_cost_calc/tool_call_cost_tracking.py` only when no existing gate fires. Then show the field's cost in your proof on every endpoint it reaches, streaming included
+
+**Do not change what a caller sees to make internals easier.** Each endpoint keeps the schema of the API it emulates. A `/v1/responses` caller gets `input_tokens` and a `/v1/chat/completions` caller gets `prompt_tokens`, whatever the provider sent and whatever cost tracking needs. A provider usage field that shows up as an extra key is fine, since it only adds a key, but renaming, removing, or reshaping a key the endpoint defines is not. Reshaping a public response to feed an internal consumer breaks every user of that endpoint
+
+If your change makes an architecture decision, such as a new cross-cutting mechanism or a deliberate difference from a provider's API shape, add an ADR to [`docs/adrs/`](docs/adrs/AGENTS.md) in the same PR
+
 ## Building and Running Locally
 
 ### LiteLLM Proxy Server

docs/adrs/AGENTS.md

@@ -0,0 +1,5 @@
+# Architecture decision records for the Python SDK and proxy
+
+This folder records decisions about the Python SDK and proxy the way [`litellm-rust/docs/adrs/`](../../litellm-rust/docs/adrs/AGENTS.md) records them for the Rust core, and that file's process applies here unchanged: one short `adr_NNN_short_title.md` per decision that is hard to reverse or that a new contributor would reasonably question, with Context, Decision, Alternatives Considered, and Consequences, superseded by a new ADR instead of edited once accepted. Each folder numbers its own ADRs, so a reference to an ADR in the other folder names its path
+
+Read the ADRs here before you build a new way to move data between the endpoints, the provider transformations, and cost tracking or logging. A PR that makes a decision of that kind adds its ADR in the same PR

docs/adrs/adr_001_provider_usage_fields_ride_as_usage_extras.md

@@ -0,0 +1,39 @@
+# ADR 001: Provider usage fields ride as extras on the normalized usage object
+
+Status: Accepted
+
+## Context
+
+Providers report usage fields that no other provider has, such as xAI's `usage.server_side_tool_usage_details.web_search_calls` or Anthropic's `usage.inference_geo`, and cost tracking has to read them or that spend never reaches `LiteLLM_SpendLogs`. One provider response can be served on `/v1/chat/completions`, `/v1/responses`, and `/v1/messages`, streaming or not, so a field that only one of those paths carries is billed wrong on the others
+
+[#30817](https://github.com/BerriAI/litellm/pull/30817) fixed unbilled xAI web searches. One draft did the copy in the xAI Responses transform to keep provider code out of shared files, and that swapped `/v1/responses` usage to the chat shape, so callers of that endpoint lost the Responses schema. The merged version keeps the schema, copies unknown usage fields in the shared Responses-to-chat bridge for every provider, and gates on the reported field without naming the provider
+
+In October 2026, `Usage` and `ResponseAPIUsage` kept unknown usage keys as pydantic extras, and that bridge copied them. Paths that build usage from named fields kept only the fields they name: the Anthropic, Gemini, and Bedrock Converse transforms, the stream builder, and the bridges from chat usage to `/v1/responses` and `/v1/messages`
+
+## Decision
+
+Provider-only usage fields stay on the usage object as pydantic extras, and cost tracking reads them there, in that provider's cost calculator or in a shared gate that checks the field instead of the provider. An extra reaches callers as an added usage key. No transform renames, removes, or reshapes a usage field the endpoint's schema defines to feed cost tracking, though it may fill one LiteLLM's usage object already defines, such as `server_tool_use`
+
+A shared path that rebuilds usage, such as the stream builder or a bridge between endpoint shapes, copies the extras instead of naming provider fields. A provider transform that builds usage from named fields names the new field. Neither is bypassed with a side channel. Older code that differs, such as the Azure branches in the shared tracker and the Gemini pricing tier read from `_hidden_params`, predates this ADR
+
+## Alternatives Considered
+
+**Reshape usage in the provider transform**, as a draft of #30817 did. It changes the schema that callers of that endpoint rely on, and every provider and endpoint needs its own copy
+
+**A typed field on `ResponseAPIUsage` and `Usage` for each provider's quirk**. The types grow without bound and every bridge needs an update for each one. A field several providers report, like `server_tool_use`, does get one
+
+**A side channel such as `_hidden_params`**. Cost calculators receive `Usage`, so a side channel makes a second source of truth for the same number
+
+**Provider branches or provider field names in shared code**. They put one provider's logic in a file every provider shares, which is why #30817 replaced its xAI-only gate with one that checks the reported field
+
+## Consequences
+
+Benefits:
+
+- Billing a new provider usage field touches that provider's calculator (plus a `cost_per_token` arm when the calculator is new), its transform when that names fields, and for a built-in tool at most one gate and one `get_cost_for_web_search_request` arm
+
+Costs:
+
+- Extras are untyped, so a reader validates them before use, as `_usage_reports_server_side_web_search_calls` does
+- The stream builder and the bridges from chat usage did not copy extras when this was written, so until they do, a field billed on those paths is named in them, and tests and proof runs cover each endpoint and `"stream": true`
+- This covers usage fields LiteLLM prices, not request-side params or a provider's own charge for the request, which OpenRouter and opencode Go report and `response_cost_calculator` reads from `_hidden_params` first

enterprise/enterprise_hooks/google_text_moderation.py

@@ -15,7 +15,10 @@ import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.proxy._types import UserAPIKeyAuth
-from litellm.proxy.guardrails._content_utils import iter_message_text
+from litellm.proxy.guardrails._content_utils import (
+    is_non_conversational_request,
+    iter_message_text,
+)
 from litellm.types.utils import CallTypesLiteral
 
 _THRESHOLD_VALUE: Final = TypeAdapter(object)
@@ -94,7 +97,7 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
 
     async def async_moderation_hook(
         self,
-        data: dict,
+        data: dict[str, object],
         user_api_key_dict: UserAPIKeyAuth,
         call_type: CallTypesLiteral,
     ):
@@ -102,6 +105,9 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
         - Calls Google's Text Moderation API
         - Rejects request if it fails safety check
         """
+        if is_non_conversational_request(call_type, data):
+            verbose_proxy_logger.debug("Google text moderation: skipping non-conversational call type %s", call_type)
+            return
         # Covers multimodal list content + Responses-API input.
         text = "".join(iter_message_text(data))
         if text:

enterprise/litellm_enterprise/enterprise_callbacks/callback_controls.py

@@ -33,9 +33,8 @@ class EnterpriseCallbackControls:
 
             try:
                 disabled_callbacks = EnterpriseCallbackControls.get_disabled_callbacks(litellm_params, standard_callback_dynamic_params)
-                verbose_logger.debug(f"Dynamically disabled callbacks from {X_LITELLM_DISABLE_CALLBACKS}: {disabled_callbacks}")
-                verbose_logger.debug(f"Checking if {callback} is disabled via headers. Disable callbacks from headers: {disabled_callbacks}")
                 if disabled_callbacks is not None:
+                    verbose_logger.debug(f"Checking if {callback} is in dynamically disabled callbacks: {disabled_callbacks}")
                     #########################################################
                     # premium user check
                     #########################################################

litellm-rust/crates/inference-ocr/tests/ocr/azure_document_intelligence.rs

@@ -12,6 +12,7 @@ use rstest::rstest;
 use super::*;
 
 const MODEL: &str = "azure_ai/doc-intelligence/prebuilt-read";
+const LAYOUT_MODEL: &str = "azure_ai/doc-intelligence/prebuilt-layout";
 
 fn read_request(base: &str, options: Value) -> LiteLLMOcrRequest {
     ocr_request(MODEL, base, options)
@@ -70,6 +71,7 @@ async fn pages_features_and_extra_options_map_to_the_analyze_call() {
 #[case(json!({"pages": "1&&features=bad"}), Error::Pages(PagesFault::Range))]
 #[case(json!({"features": "languages&pages=1"}), Error::Features)]
 #[case(json!({"req_format": "azure"}), Error::RequestFormat)]
+#[case(json!({"outputContentFormat": "html"}), Error::OutputContentFormat)]
 #[tokio::test]
 async fn invalid_pages_features_and_format_are_rejected_before_sending(
     #[case] options: Value,
@@ -130,6 +132,137 @@ async fn an_inline_document_is_sent_as_base64_and_only_page_text_is_kept(#[case]
     assert_eq!(sent.json(), json!({"base64Source": "YWJj"}));
 }
 
+#[rstest]
+#[case::lower_case("markdown")]
+#[case::capitalized("Markdown")]
+#[tokio::test]
+async fn markdown_output_is_requested_on_the_analyze_call_and_split_into_pages(
+    #[case] output_content_format: &str,
+) {
+    let upstream = upstream([json_response(json!({
+        "status": "succeeded",
+        "analyzeResult": {
+            "content": "# Tïtle\n\n<!-- PageBreak -->\n\n| a |",
+            "contentFormat": "markdown",
+            "stringIndexType": "unicodeCodePoint",
+            "pages": [
+                {"pageNumber": 1, "spans": [{"offset": 0, "length": 29}], "lines": [{"content": "Tïtle"}]},
+                {"pageNumber": 2, "spans": [{"offset": 29, "length": 5}], "lines": [{"content": "a"}]}
+            ]
+        }
+    }))])
+    .await;
+
+    let response = perform(ocr_request(
+        LAYOUT_MODEL,
+        &upstream.uri(),
+        json!({"outputContentFormat": output_content_format, "pages": [0, 1]}),
+    ))
+    .await
+    .unwrap();
+
+    let sent = only_request(&upstream).await;
+    assert_eq!(
+        sent.query("outputContentFormat").as_deref(),
+        Some("markdown")
+    );
+    assert_eq!(
+        sent.query("stringIndexType").as_deref(),
+        Some("unicodeCodePoint")
+    );
+    assert_eq!(sent.query("pages").as_deref(), Some("1,2"));
+    assert_eq!(sent.json(), json!({"base64Source": "YWJj"}));
+    assert_eq!(
+        response.content.as_deref(),
+        Some("# Tïtle\n\n<!-- PageBreak -->\n\n| a |")
+    );
+    let pages = response
+        .pages
+        .iter()
+        .map(|page| page.markdown.as_str())
+        .collect::<Vec<_>>();
+    assert_eq!(pages, ["# Tïtle\n\n<!-- PageBreak -->\n\n", "| a |"]);
+    assert_eq!(
+        response.extra_fields.get("stringIndexType"),
+        Some(&json!("unicodeCodePoint"))
+    );
+    assert_eq!(
+        response.usage_info.as_ref().unwrap().pages_processed,
+        Some(2)
+    );
+}
+
+#[rstest]
+#[case::native_markdown(LAYOUT_MODEL, json!({"outputContentFormat": "markdown", "req_format": "native"}), Some("markdown"))]
+#[case::markdown_on_read(MODEL, json!({"outputContentFormat": "markdown"}), Some("markdown"))]
+#[case::text(LAYOUT_MODEL, json!({"outputContentFormat": "text"}), Some("text"))]
+#[case::omitted(LAYOUT_MODEL, json!({}), None)]
+#[tokio::test]
+async fn only_layout_markdown_in_litellm_format_asks_for_code_point_offsets(
+    #[case] model: &str,
+    #[case] options: Value,
+    #[case] output_content_format: Option<&str>,
+) {
+    let upstream = upstream([json_response(json!({
+        "status": "succeeded",
+        "analyzeResult": {
+            "contentFormat": "text",
+            "stringIndexType": "textElements",
+            "pages": [{"pageNumber": 1, "lines": [{"content": "hello"}]}]
+        }
+    }))])
+    .await;
+
+    let response = perform(ocr_request(model, &upstream.uri(), options))
+        .await
+        .unwrap();
+
+    let sent = only_request(&upstream).await;
+    assert_eq!(
+        sent.query("outputContentFormat").as_deref(),
+        output_content_format
+    );
+    assert_eq!(sent.query("stringIndexType"), None);
+    assert_eq!(sent.json(), json!({"base64Source": "YWJj"}));
+    assert_eq!(response.pages[0].markdown, "hello");
+    assert_eq!(response.extra_fields.get("stringIndexType"), None);
+}
+
+#[rstest]
+#[case::snake_case(json!({"output_content_format": "markdown"}), Some("markdown"))]
+#[case::padded(json!({"outputContentFormat": " markdown "}), Some("markdown"))]
+#[case::empty(json!({"outputContentFormat": ""}), None)]
+#[case::empty_camel_case_falls_back_to_snake_case(
+    json!({"outputContentFormat": "", "output_content_format": "markdown"}),
+    Some("markdown")
+)]
+#[case::dropped(json!({"outputContentFormat": "html", "drop_params": true}), None)]
+#[tokio::test]
+async fn the_requested_output_content_format_reaches_azure_in_its_query_spelling(
+    #[case] options: Value,
+    #[case] output_content_format: Option<&str>,
+) {
+    let upstream = upstream([json_response(json!({
+        "status": "succeeded",
+        "analyzeResult": {"pages": [{"pageNumber": 1, "lines": [{"content": "hello"}]}]}
+    }))])
+    .await;
+
+    let response = perform(ocr_request(MODEL, &upstream.uri(), options))
+        .await
+        .unwrap();
+
+    let sent = only_request(&upstream).await;
+    assert_eq!(
+        sent.query("outputContentFormat").as_deref(),
+        output_content_format
+    );
+    assert_eq!(sent.query("output_content_format"), None);
+    assert_eq!(sent.query("drop_params"), None);
+    assert_eq!(sent.json(), json!({"base64Source": "YWJj"}));
+    assert_eq!(response.pages[0].markdown, "hello");
+}
+
 #[tokio::test]
 async fn native_format_normalizes_pages_and_keeps_the_provider_response() {
     let operation = json!({

litellm-rust/crates/llms-types/src/formats/audio_transcription/response.rs

@@ -1,14 +1,31 @@
 use serde_json::Value;
 
+#[macro_rules_attribute::apply(crate::wire_type)]
+pub struct AudioTranscriptionUsage {
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub total_tokens: u64,
+}
+
 #[macro_rules_attribute::apply(crate::wire_type)]
 pub struct AudioTranscriptionResponseData {
     pub text: String,
+    pub usage: Option<AudioTranscriptionUsage>,
 }
 
 impl AudioTranscriptionResponseData {
     pub fn into_json(self) -> Value {
-        serde_json::json!({
-            "text": self.text,
-        })
+        match self.usage {
+            None => serde_json::json!({"text": self.text}),
+            Some(usage) => serde_json::json!({
+                "text": self.text,
+                "usage": {
+                    "type": "tokens",
+                    "input_tokens": usage.input_tokens,
+                    "output_tokens": usage.output_tokens,
+                    "total_tokens": usage.total_tokens,
+                },
+            }),
+        }
     }
 }

litellm-rust/crates/llms/src/azure_ai/ocr/document_intelligence/transformation.rs

@@ -1,4 +1,7 @@
-use std::{collections::BTreeSet, time::Duration};
+use std::{
+    collections::{BTreeMap, BTreeSet},
+    time::Duration,
+};
 
 use base64::{Engine, engine::general_purpose::STANDARD};
 use litellm_auth::{InputSource, Sourced};
@@ -8,7 +11,7 @@ use litellm_llms_types::serde_compat::{FiniteF64, LaxI64};
 use reqwest::Url;
 use serde::{Deserialize, Serialize};
 use serde_json::{Map, Value};
-use serde_with::serde_as;
+use serde_with::{DefaultOnError, serde_as};
 use tokio::time::Instant;
 
 use crate::base_llm::ocr::{
@@ -17,9 +20,10 @@ use crate::base_llm::ocr::{
     handler::{CallHooks, OcrClient, read_json_response},
     settings::OcrSettings,
     transformation::{
-        BaseOcrConfig, DecodedOcrResponse, OCR_INLINE_MAX_BYTES, OCR_POLL_RETRY_SECS,
-        OcrConnection, OcrCredentialInputs, OcrResponseContext, PreparedOcrRequest,
-        ResolvedOcrCredentials, decode_and_normalize_response, decode_response,
+        BaseOcrConfig, DROP_PARAMS_PARAM, DecodedOcrResponse, OCR_INLINE_MAX_BYTES,
+        OCR_POLL_RETRY_SECS, OcrConnection, OcrCredentialInputs, OcrResponseContext,
+        PreparedOcrRequest, ResolvedOcrCredentials, decode_and_normalize_response, decode_response,
+        drops_unsupported_params, response_format,
     },
 };
 use litellm_llms_types::formats::ocr::{
@@ -32,6 +36,30 @@ const AZURE_DI_DEFAULT_HEIGHT: f64 = 11.0;
 
 const AZURE_DI_API_KEY_ENV: &str = "AZURE_DOCUMENT_INTELLIGENCE_API_KEY";
 const AZURE_DI_ENDPOINT_ENV: &str = "AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT";
+const OUTPUT_CONTENT_FORMAT_PARAM: &str = "outputContentFormat";
+const OUTPUT_CONTENT_FORMAT_ALIAS: &str = "output_content_format";
+const CODE_POINT_STRING_INDEX: &str = "unicodeCodePoint";
+const STRING_INDEX_TYPE_FIELD: &str = "stringIndexType";
+const MARKDOWN_MODEL_ID: &str = "prebuilt-layout";
+
+#[derive(
+    Clone,
+    Copy,
+    Debug,
+    PartialEq,
+    strum::EnumString,
+    strum::IntoStaticStr,
+    strum::Display,
+    strum::VariantArray,
+    serde_with::SerializeDisplay,
+)]
+#[strum(ascii_case_insensitive)]
+pub enum OutputContentFormat {
+    #[strum(serialize = "text")]
+    Text,
+    #[strum(serialize = "markdown")]
+    Markdown,
+}
 
 #[derive(Clone, Debug, PartialEq, Serialize)]
 pub struct DocumentIntelligenceParams {
@@ -39,6 +67,13 @@ pub struct DocumentIntelligenceParams {
     pub pages: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub features: Option<String>,
+    #[serde(
+        rename = "outputContentFormat",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub output_content_format: Option<OutputContentFormat>,
+    #[serde(rename = "stringIndexType", skip_serializing_if = "Option::is_none")]
+    pub string_index_type: Option<&'static str>,
 }
 
 #[derive(Clone, Debug, Serialize, Deserialize)]
@@ -80,6 +115,10 @@ pub struct AzureDocumentIntelligenceOperation {
 #[derive(Clone, Debug, Default, Deserialize)]
 struct AzureDocumentIntelligenceAnalyzeResult {
     pub content: Option<String>,
+    #[serde(rename = "contentFormat")]
+    pub content_format: Option<String>,
+    #[serde(rename = "stringIndexType")]
+    pub string_index_type: Option<String>,
     #[serde(default)]
     pub pages: Vec<AzureDocumentIntelligencePage>,
     pub tables: Option<Vec<Map<String, Value>>>,
@@ -100,6 +139,9 @@ struct AzureDocumentIntelligencePage {
     pub unit: Option<String>,
     #[serde(default)]
     pub lines: Vec<AzureDocumentIntelligenceLine>,
+    #[serde(default)]
+    #[serde_as(deserialize_as = "DefaultOnError")]
+    pub spans: Option<Vec<AzureDocumentIntelligenceSpan>>,
 }
 
 #[derive(Clone, Debug, Deserialize)]
@@ -107,6 +149,18 @@ struct AzureDocumentIntelligenceLine {
     pub content: Option<String>,
 }
 
+#[derive(Clone, Debug, Deserialize)]
+struct AzureDocumentIntelligenceSpan {
+    pub offset: usize,
+    pub length: usize,
+}
+
+impl AzureDocumentIntelligenceSpan {
+    fn end(&self) -> usize {
+        self.offset.saturating_add(self.length)
+    }
+}
+
 #[derive(Clone, Debug)]
 pub struct AzureDocumentIntelligenceOcrConfig;
 
@@ -116,7 +170,14 @@ impl BaseOcrConfig for AzureDocumentIntelligenceOcrConfig {
     type Environment = Vec<(String, String)>;
 
     fn get_supported_ocr_params(&self, _model: &str) -> &'static [&'static str] {
-        &["pages", "features", "req_format"]
+        &[
+            "pages",
+            "features",
+            OUTPUT_CONTENT_FORMAT_PARAM,
+            OUTPUT_CONTENT_FORMAT_ALIAS,
+            "req_format",
+            DROP_PARAMS_PARAM,
+        ]
     }
 
     fn get_api_key_env_var(&self) -> Option<&'static str> {
@@ -154,11 +215,17 @@ impl BaseOcrConfig for AzureDocumentIntelligenceOcrConfig {
     fn map_ocr_params(
         &self,
         non_default_params: &CallArguments,
-        _model: &str,
+        model: &str,
     ) -> Result<DocumentIntelligenceParams, Error> {
+        let output_content_format = requested_output_content_format(non_default_params)?;
+        let slices_markdown_pages = output_content_format == Some(OutputContentFormat::Markdown)
+            && response_format(non_default_params)? != OcrResponseFormat::Native
+            && model_id(model).is_ok_and(|id| id == MARKDOWN_MODEL_ID);
         Ok(DocumentIntelligenceParams {
             pages: normalize_pages_param(non_default_params.get("pages"))?,
             features: normalize_features_param(non_default_params.get("features"))?,
+            output_content_format,
+            string_index_type: slices_markdown_pages.then_some(CODE_POINT_STRING_INDEX),
         })
     }
 
@@ -337,6 +404,36 @@ fn normalize_features_param(features: Option<&Value>) -> Result<Option<String>,
     Ok(Some(normalized.join(",")))
 }
 
+fn requested_output_content_format(
+    params: &CallArguments,
+) -> Result<Option<OutputContentFormat>, Error> {
+    let requested = [OUTPUT_CONTENT_FORMAT_PARAM, OUTPUT_CONTENT_FORMAT_ALIAS]
+        .into_iter()
+        .filter_map(|name| params.get(name))
+        .find(|format| !is_unset(format));
+    match requested.map(parse_output_content_format).transpose() {
+        Err(Error::OutputContentFormat) if drops_unsupported_params(params) => Ok(None),
+        parsed => parsed,
+    }
+}
+
+fn is_unset(format: &Value) -> bool {
+    match format {
+        Value::Null => true,
+        Value::String(format) => format.trim().is_empty(),
+        _ => false,
+    }
+}
+
+fn parse_output_content_format(format: &Value) -> Result<OutputContentFormat, Error> {
+    format
+        .as_str()
+        .ok_or(Error::OutputContentFormat)?
+        .trim()
+        .parse()
+        .map_err(|_| Error::OutputContentFormat)
+}
+
 fn build_request(document: OcrDocument) -> Result<DocumentIntelligenceRequest, Error> {
     let source = document.source();
     if source.is_empty() {
@@ -367,10 +464,27 @@ fn transform_completed_response(
         ));
     }
     let result = response.analyze_result.unwrap_or_default();
+    let markdown_content = result
+        .content
+        .as_deref()
+        .filter(|_| {
+            result.content_format.as_deref() == Some(<&str>::from(OutputContentFormat::Markdown))
+                && result.string_index_type.as_deref() == Some(CODE_POINT_STRING_INDEX)
+                && result.pages.iter().all(|page| page.spans.is_some())
+        })
+        .map(|content| {
+            CodePointText::new(
+                content,
+                result
+                    .pages
+                    .iter()
+                    .flat_map(|page| page.spans.iter().flatten()),
+            )
+        });
     let pages = result
         .pages
         .into_iter()
-        .map(|page| transform_azure_page(page, dpi))
+        .map(|page| transform_azure_page(page, markdown_content.as_ref(), dpi))
         .collect::<Result<Vec<_>, _>>()?;
     let pages_processed = i64::try_from(pages.len()).map_err(|_| Error::NumericRange("pages"))?;
     Ok(LiteLLMOcrResponse {
@@ -381,11 +495,20 @@ fn transform_completed_response(
             pages_processed: Some(pages_processed),
             ..Default::default()
         }),
+        extra_fields: result
+            .string_index_type
+            .filter(|unit| unit == CODE_POINT_STRING_INDEX)
+            .map(|unit| Map::from_iter([(STRING_INDEX_TYPE_FIELD.to_owned(), Value::String(unit))]))
+            .unwrap_or_default(),
         ..LiteLLMOcrResponse::new(model, pages)
     })
 }
 
-fn transform_azure_page(page: AzureDocumentIntelligencePage, dpi: i64) -> Result<OcrPage, Error> {
+fn transform_azure_page(
+    page: AzureDocumentIntelligencePage,
+    markdown_content: Option<&CodePointText<'_>>,
+    dpi: i64,
+) -> Result<OcrPage, Error> {
     let index = page
         .page_number
         .unwrap_or(1)
@@ -397,12 +520,20 @@ fn transform_azure_page(page: AzureDocumentIntelligencePage, dpi: i64) -> Result
         page.unit.as_deref().unwrap_or("inch"),
         dpi,
     )?;
-    let markdown = page
-        .lines
-        .iter()
-        .map(|line| line.content.as_deref().unwrap_or_default())
-        .collect::<Vec<_>>()
-        .join("\n");
+    let markdown = match markdown_content {
+        Some(content) => page
+            .spans
+            .iter()
+            .flatten()
+            .map(|span| content.slice(span))
+            .collect::<String>(),
+        None => page
+            .lines
+            .iter()
+            .map(|line| line.content.as_deref().unwrap_or_default())
+            .collect::<Vec<_>>()
+            .join("\n"),
+    };
     Ok(OcrPage {
         index,
         markdown,
@@ -411,6 +542,43 @@ fn transform_azure_page(page: AzureDocumentIntelligencePage, dpi: i64) -> Result
     })
 }
 
+struct CodePointText<'a> {
+    text: &'a str,
+    byte_offsets: BTreeMap<usize, usize>,
+}
+
+impl<'a> CodePointText<'a> {
+    fn new<'s>(
+        text: &'a str,
+        spans: impl Iterator<Item = &'s AzureDocumentIntelligenceSpan>,
+    ) -> Self {
+        let span_bounds = spans
+            .flat_map(|span| [span.offset, span.end()])
+            .collect::<BTreeSet<_>>();
+        Self {
+            text,
+            byte_offsets: text
+                .char_indices()
+                .map(|(byte_offset, _)| byte_offset)
+                .enumerate()
+                .filter(|(code_point, _)| span_bounds.contains(code_point))
+                .collect(),
+        }
+    }
+
+    fn slice(&self, span: &AzureDocumentIntelligenceSpan) -> &'a str {
+        let byte_offset = |code_point: usize| {
+            self.byte_offsets
+                .get(&code_point)
+                .copied()
+                .unwrap_or(self.text.len())
+        };
+        self.text
+            .get(byte_offset(span.offset)..byte_offset(span.end()))
+            .unwrap_or_default()
+    }
+}
+
 fn convert_dimensions(
     width: f64,
     height: f64,
@@ -578,6 +746,16 @@ impl AzureDocumentIntelligenceOcrConfig {
                                 .features
                                 .iter()
                                 .map(|features| ("features", features.as_str())),
+                        )
+                        .chain(
+                            params
+                                .output_content_format
+                                .map(|format| (OUTPUT_CONTENT_FORMAT_PARAM, format.into())),
+                        )
+                        .chain(
+                            params
+                                .string_index_type
+                                .map(|index_type| ("stringIndexType", index_type)),
                         ),
                 )
                 .into_string()
@@ -697,8 +875,12 @@ mod tests {
     }
 
     fn map(value: Value) -> Result<DocumentIntelligenceParams, Error> {
+        map_for_model(value, "azure_ai/doc-intelligence/prebuilt-layout")
+    }
+
+    fn map_for_model(value: Value, model: &str) -> Result<DocumentIntelligenceParams, Error> {
         let arguments = serde_json::from_value(value).unwrap();
-        AzureDocumentIntelligenceOcrConfig.map_ocr_params(&arguments, "model")
+        AzureDocumentIntelligenceOcrConfig.map_ocr_params(&arguments, model)
     }
 
     #[test]
@@ -828,6 +1010,218 @@ mod tests {
         assert_eq!(map(json!({"features": []})).unwrap().features, None);
     }
 
+    #[rstest]
+    #[case::markdown(
+        json!({"outputContentFormat": "markdown"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::markdown_litellm_format(
+        json!({"outputContentFormat": "markdown", "req_format": "litellm"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::markdown_native_format(
+        json!({"outputContentFormat": "markdown", "req_format": "native"}),
+        Some(OutputContentFormat::Markdown),
+        None
+    )]
+    #[case::capitalized_markdown(
+        json!({"outputContentFormat": "Markdown"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::text(json!({"outputContentFormat": "text"}), Some(OutputContentFormat::Text), None)]
+    #[case::upper_case_text(json!({"outputContentFormat": "TEXT"}), Some(OutputContentFormat::Text), None)]
+    #[case::padded_markdown(
+        json!({"outputContentFormat": " markdown\n"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::snake_case_markdown(
+        json!({"output_content_format": "markdown"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::camel_case_wins_over_snake_case(
+        json!({"outputContentFormat": "text", "output_content_format": "markdown"}),
+        Some(OutputContentFormat::Text),
+        None
+    )]
+    #[case::empty_camel_case_falls_back_to_snake_case(
+        json!({"outputContentFormat": "", "output_content_format": "markdown"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::null_camel_case_falls_back_to_snake_case(
+        json!({"outputContentFormat": null, "output_content_format": "markdown"}),
+        Some(OutputContentFormat::Markdown),
+        Some("unicodeCodePoint")
+    )]
+    #[case::empty(json!({"outputContentFormat": ""}), None, None)]
+    #[case::blank(json!({"outputContentFormat": "  "}), None, None)]
+    #[case::null(json!({"outputContentFormat": null}), None, None)]
+    #[case::omitted(json!({}), None, None)]
+    fn output_content_format_maps_to_query_fields(
+        #[case] options: Value,
+        #[case] format: Option<OutputContentFormat>,
+        #[case] string_index_type: Option<&str>,
+    ) {
+        let mapped = map(options).unwrap();
+
+        assert_eq!(mapped.output_content_format, format);
+        assert_eq!(mapped.string_index_type, string_index_type);
+    }
+
+    #[rstest]
+    #[case::layout("azure_ai/doc-intelligence/prebuilt-layout", Some("unicodeCodePoint"))]
+    #[case::bare_layout("prebuilt-layout", Some("unicodeCodePoint"))]
+    #[case::read("azure_ai/doc-intelligence/prebuilt-read", None)]
+    #[case::invoice("azure_ai/doc-intelligence/prebuilt-invoice", None)]
+    #[case::custom("azure_ai/doc-intelligence/my-extraction-model", None)]
+    fn only_the_layout_model_asks_for_code_point_offsets(
+        #[case] model: &str,
+        #[case] string_index_type: Option<&str>,
+    ) {
+        let mapped = map_for_model(json!({"outputContentFormat": "markdown"}), model).unwrap();
+
+        assert_eq!(
+            mapped.output_content_format,
+            Some(OutputContentFormat::Markdown)
+        );
+        assert_eq!(mapped.string_index_type, string_index_type);
+    }
+
+    #[rstest]
+    #[case::unknown(json!({"outputContentFormat": "html"}))]
+    #[case::number(json!({"outputContentFormat": 1}))]
+    #[case::boolean(json!({"outputContentFormat": true}))]
+    #[case::list(json!({"outputContentFormat": ["markdown"]}))]
+    #[case::snake_case_unknown(json!({"output_content_format": "html"}))]
+    #[case::drop_params_off(json!({"outputContentFormat": "html", "drop_params": false}))]
+    fn invalid_output_content_format_is_a_request_error(#[case] options: Value) {
+        let error = map(options).unwrap_err();
+
+        assert!(matches!(error, Error::OutputContentFormat), "{error:?}");
+        assert_eq!(error.http_status_code(), Some(400));
+    }
+
+    #[rstest]
+    #[case::unknown(json!({"outputContentFormat": "html", "drop_params": true, "pages": [0]}))]
+    #[case::number(json!({"outputContentFormat": 1, "drop_params": true, "pages": [0]}))]
+    #[case::snake_case_unknown(
+        json!({"output_content_format": "html", "drop_params": true, "pages": [0]})
+    )]
+    fn drop_params_drops_an_invalid_output_content_format(#[case] options: Value) {
+        let mapped = map(options).unwrap();
+
+        assert_eq!(serde_json::to_value(mapped).unwrap(), json!({"pages": "1"}));
+    }
+
+    #[rstest]
+    #[case::camel_case(json!({"outputContentFormat": "markdown", "drop_params": true}))]
+    #[case::snake_case(json!({"output_content_format": "markdown", "drop_params": true}))]
+    fn drop_params_keeps_a_valid_output_content_format(#[case] options: Value) {
+        let mapped = map(options).unwrap();
+
+        assert_eq!(
+            mapped.output_content_format,
+            Some(OutputContentFormat::Markdown)
+        );
+    }
+
+    fn markdown_pages(analyze_result: Value) -> Vec<String> {
+        let raw =
+            serde_json::to_vec(&json!({"status": "succeeded", "analyzeResult": analyze_result}))
+                .unwrap();
+        AzureDocumentIntelligenceOcrConfig
+            .transform_ocr_response("model", &raw, OcrResponseFormat::Litellm)
+            .unwrap()
+            .pages
+            .into_iter()
+            .map(|page| page.markdown)
+            .collect()
+    }
+
+    #[rstest]
+    fn markdown_pages_are_code_point_slices_of_the_content() {
+        let content = "# Ünïcode क्ष\n\n<!-- PageBreak -->\n\n## Två 🙂\n\n|a|\n\n> tail";
+        let second_page_start = "# Ünïcode क्ष\n\n<!-- PageBreak -->\n\n".chars().count();
+        let pages = markdown_pages(json!({
+            "content": content,
+            "contentFormat": "markdown",
+            "stringIndexType": "unicodeCodePoint",
+            "pages": [
+                {"pageNumber": 1, "spans": [{"offset": 0, "length": second_page_start}],
+                 "lines": [{"content": "Ünïcode क्ष"}]},
+                {"pageNumber": 2, "spans": [
+                    {"offset": second_page_start, "length": 8},
+                    {"offset": second_page_start + 10, "length": 3}
+                ], "lines": [{"content": "Två"}]}
+            ]
+        }));
+
+        assert_eq!(
+            pages,
+            ["# Ünïcode क्ष\n\n<!-- PageBreak -->\n\n", "## Två 🙂|a|"]
+        );
+    }
+
+    #[rstest]
+    fn spans_past_the_content_end_are_clamped() {
+        let pages = markdown_pages(json!({
+            "content": "ab",
+            "contentFormat": "markdown",
+            "stringIndexType": "unicodeCodePoint",
+            "pages": [{"spans": [{"offset": 1, "length": 9}, {"offset": 7, "length": 1}]}]
+        }));
+
+        assert_eq!(pages, ["b"]);
+    }
+
+    #[rstest]
+    #[case::text_content(json!("text"), json!("unicodeCodePoint"))]
+    #[case::grapheme_offsets(json!("markdown"), json!("textElements"))]
+    #[case::no_index_type(json!("markdown"), Value::Null)]
+    #[case::no_content_format(Value::Null, json!("unicodeCodePoint"))]
+    fn other_responses_keep_line_joined_pages(
+        #[case] content_format: Value,
+        #[case] string_index_type: Value,
+    ) {
+        let pages = markdown_pages(json!({
+            "content": "# A\n\nB",
+            "contentFormat": content_format,
+            "stringIndexType": string_index_type,
+            "pages": [{"spans": [{"offset": 0, "length": 6}],
+                       "lines": [{"content": "A"}, {"content": "B"}]}]
+        }));
+
+        assert_eq!(pages, ["A\nB"]);
+    }
+
+    #[rstest]
+    #[case::null_spans(json!("markdown"), json!(null))]
+    #[case::fractional_offset(json!("markdown"), json!([{"offset": 0.5, "length": 1}]))]
+    #[case::missing_length(json!("markdown"), json!([{"offset": 0}]))]
+    #[case::negative_offset(json!("markdown"), json!([{"offset": -1, "length": 1}]))]
+    #[case::text_content(json!("text"), json!([{"offset": "0", "length": 1}]))]
+    fn unusable_page_spans_keep_line_joined_pages(
+        #[case] content_format: Value,
+        #[case] broken_spans: Value,
+    ) {
+        let pages = markdown_pages(json!({
+            "content": "# A\n\nB",
+            "contentFormat": content_format,
+            "stringIndexType": "unicodeCodePoint",
+            "pages": [
+                {"spans": [{"offset": 0, "length": 5}], "lines": [{"content": "A"}]},
+                {"spans": broken_spans, "lines": [{"content": "B"}]}
+            ]
+        }));
+
+        assert_eq!(pages, ["A", "B"]);
+    }
+
     #[tokio::test]
     async fn request_endpoint_cannot_receive_environment_key() {
         let connection = OcrConnection {

litellm-rust/crates/llms/src/base_llm/ocr/error.rs

@@ -59,6 +59,10 @@ pub enum Error {
     Pages(PagesFault),
     #[error("invalid OCR features")]
     Features,
+    #[error(
+        "invalid OCR outputContentFormat, expected text or markdown. To drop unsupported params, set `litellm.drop_params = True`."
+    )]
+    OutputContentFormat,
     #[error("OCR model cannot be a dot segment")]
     DotModel,
     #[error("OCR response exceeds the size limit of {limit} bytes")]
@@ -200,6 +204,7 @@ impl Error {
             | Self::InlineDocumentTooLarge
             | Self::Pages(_)
             | Self::Features
+            | Self::OutputContentFormat
             | Self::DotModel
             | Self::TooLarge { .. }
             | Self::ResponseField { .. }
@@ -242,6 +247,7 @@ impl Error {
             | Self::InlineDocumentTooLarge
             | Self::Pages(_)
             | Self::Features
+            | Self::OutputContentFormat
             | Self::DotModel
             | Self::TooLarge { .. }
             | Self::ResponseField { .. }
@@ -291,6 +297,7 @@ impl Error {
             | Self::InlineDocumentTooLarge
             | Self::Pages(_)
             | Self::Features
+            | Self::OutputContentFormat
             | Self::DotModel
             | Self::InvalidProvider(_)
             | Self::InvalidModel { .. }
@@ -345,6 +352,7 @@ impl Error {
             | Self::InlineDocumentTooLarge
             | Self::Pages(_)
             | Self::Features
+            | Self::OutputContentFormat
             | Self::DotModel
             | Self::InvalidProvider(_)
             | Self::InvalidModel { .. }

litellm-rust/crates/llms/src/base_llm/ocr/transformation.rs

@@ -21,6 +21,7 @@ pub const OCR_RESPONSE_MAX_BYTES: usize = 64 * 1024 * 1024;
 pub const OCR_INLINE_MAX_BYTES: usize = 50 * 1024 * 1024;
 pub const OCR_MAX_FETCH_REDIRECTS: usize = 10;
 pub const OCR_POLL_RETRY_SECS: u64 = 2;
+pub const DROP_PARAMS_PARAM: &str = "drop_params";
 
 #[derive(Clone, Default)]
 pub struct OcrCredentialInputs {
@@ -185,6 +186,10 @@ pub fn response_format(optional_params: &CallArguments) -> Result<OcrResponseFor
         .map(|format| format.unwrap_or_default())
 }
 
+pub fn drops_unsupported_params(optional_params: &CallArguments) -> bool {
+    optional_params.get(DROP_PARAMS_PARAM) == Some(&Value::Bool(true))
+}
+
 #[derive(Debug)]
 pub struct DecodedOcrResponse<T> {
     pub data: T,

litellm-rust/crates/llms/src/bedrock/audio_transcription/mod.rs

@@ -6,8 +6,8 @@ use litellm_auth_aws::{
 };
 use litellm_core_utils::core_helpers::json_type_name;
 use litellm_llms_types::{
-    formats::audio_transcription::AudioTranscriptionResponseData,
-    providers::bedrock::ConverseResponse,
+    formats::audio_transcription::{AudioTranscriptionResponseData, AudioTranscriptionUsage},
+    providers::bedrock::{ConverseResponse, ConverseUsage},
 };
 use serde::Deserialize;
 use serde_json::{Map, Value, json};
@@ -101,10 +101,10 @@ impl BaseAudioTranscriptionConfig for BedrockAudioTranscriptionConfig {
         if let Some(prompt) = optional_string(&optional_params, "prompt") {
             instruction.push_str(&format!(" Additional context: {prompt}"));
         }
-        let mut inference_config = Map::from_iter([("maxTokens".to_string(), json!(4096))]);
-        if let Some(temperature) = optional_params.get("temperature") {
-            inference_config.insert("temperature".to_string(), temperature.clone());
-        }
+        let temperature = optional_params
+            .get("temperature")
+            .cloned()
+            .unwrap_or(json!(0));
         Ok(AudioTranscriptionRequestData {
             body: json!({
                 "messages": [{
@@ -114,8 +114,7 @@ impl BaseAudioTranscriptionConfig for BedrockAudioTranscriptionConfig {
                         {"text": instruction}
                     ]
                 }],
-                "system": [{"text": "You are a transcription assistant."}],
-                "inferenceConfig": inference_config,
+                "inferenceConfig": {"maxTokens": 4096, "temperature": temperature},
             }),
         })
     }
@@ -125,6 +124,7 @@ impl BaseAudioTranscriptionConfig for BedrockAudioTranscriptionConfig {
         _model: &str,
         response_json: Value,
     ) -> Result<AudioTranscriptionResponseData, Error> {
+        let reports_usage = response_json.get("usage").is_some();
         let response: ConverseResponse =
             serde_json::from_value(response_json).map_err(|error| {
                 Error::InvalidResponse(ErrorDetail::invalid("Converse response", error))
@@ -137,6 +137,7 @@ impl BaseAudioTranscriptionConfig for BedrockAudioTranscriptionConfig {
         }
         Ok(AudioTranscriptionResponseData {
             text: response.content_text(),
+            usage: reports_usage.then(|| transcription_usage(&response.usage)),
         })
     }
 
@@ -191,22 +192,41 @@ impl BaseAudioTranscriptionConfig for BedrockAudioTranscriptionConfig {
     }
 }
 
+fn transcription_usage(usage: &ConverseUsage) -> AudioTranscriptionUsage {
+    AudioTranscriptionUsage {
+        input_tokens: usage.input_tokens,
+        output_tokens: usage.output_tokens,
+        total_tokens: usage
+            .total_tokens
+            .unwrap_or(usage.input_tokens + usage.output_tokens),
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
+    use rstest::rstest;
 
     fn no_env(_: &str) -> Option<String> {
         None
     }
 
-    #[test]
-    fn request_matches_python_shape() {
-        let params = Map::from_iter([
-            ("language".to_string(), json!("en")),
-            ("prompt".to_string(), json!("Speaker names")),
-            ("temperature".to_string(), json!(0)),
-            ("timestamp_granularities".to_string(), json!(["word"])),
-        ]);
+    #[rstest]
+    #[case::caller_temperature(Some(json!(0.4)), json!(0.4))]
+    #[case::default_temperature(None, json!(0))]
+    fn request_sends_audio_and_instruction_without_system_block(
+        #[case] temperature: Option<Value>,
+        #[case] expected_temperature: Value,
+    ) {
+        let params = Map::from_iter(
+            [
+                ("language".to_string(), json!("en")),
+                ("prompt".to_string(), json!("Speaker names")),
+                ("timestamp_granularities".to_string(), json!(["word"])),
+            ]
+            .into_iter()
+            .chain(temperature.map(|value| ("temperature".to_string(), value))),
+        );
         let params = BEDROCK_AUDIO_TRANSCRIPTION_CONFIG.map_transcription_params(&params);
         let result = BEDROCK_AUDIO_TRANSCRIPTION_CONFIG
             .transform_audio_transcription_request(
@@ -225,22 +245,39 @@ mod tests {
                         {"text": "Transcribe the audio. Respond with only the transcript. The audio language is en. Additional context: Speaker names"}
                     ]
                 }],
-                "system": [{"text": "You are a transcription assistant."}],
-                "inferenceConfig": {"maxTokens": 4096, "temperature": 0}
+                "inferenceConfig": {"maxTokens": 4096, "temperature": expected_temperature}
             })
         );
     }
 
-    #[test]
-    fn response_concatenates_content_blocks() {
+    #[rstest]
+    #[case::reported_total(
+        json!({"inputTokens": 390, "outputTokens": 60, "totalTokens": 450}),
+        json!({"text": "hello world", "usage": {"type": "tokens", "input_tokens": 390, "output_tokens": 60, "total_tokens": 450}})
+    )]
+    #[case::derived_total(
+        json!({"inputTokens": 1, "outputTokens": 2}),
+        json!({"text": "hello world", "usage": {"type": "tokens", "input_tokens": 1, "output_tokens": 2, "total_tokens": 3}})
+    )]
+    fn response_carries_text_and_converse_usage(#[case] usage: Value, #[case] expected: Value) {
         let result = BEDROCK_AUDIO_TRANSCRIPTION_CONFIG
             .transform_audio_transcription_response(
                 "model",
-                json!({"output": {"message": {"content": [{"text": "hello "}, {"text": "world"}]}}, "usage": {"inputTokens": 1, "outputTokens": 2}}),
+                json!({"output": {"message": {"content": [{"text": "hello "}, {"text": "world"}]}}, "usage": usage}),
             )
             .expect("response");
-        assert_eq!(result.text, "hello world");
-        assert_eq!(result.into_json(), json!({"text": "hello world"}));
+        assert_eq!(result.into_json(), expected);
+    }
+
+    #[rstest]
+    fn response_without_usage_reports_none() {
+        let result = BEDROCK_AUDIO_TRANSCRIPTION_CONFIG
+            .transform_audio_transcription_response(
+                "model",
+                json!({"output": {"message": {"content": [{"text": "hello"}]}}}),
+            )
+            .expect("response");
+        assert_eq!(result.into_json(), json!({"text": "hello"}));
     }
 
     #[test]

litellm-rust/crates/python-bridge/src/routes/ocr/host.rs

@@ -13,7 +13,7 @@ use pyo3::{
 
 use super::{
     errors::route_failure,
-    project::{OcrHostHandles, SETTINGS_ERROR_MARKER, project_request},
+    project::{OcrHostHandles, ROUTE_HOST_MODULE, SETTINGS_ERROR_MARKER, project_request},
 };
 use crate::{marshal::public_response, routes::RequestView};
 
@@ -77,7 +77,7 @@ impl OcrPythonHost {
             _ => "",
         };
         let mapped = py
-            .import("litellm.rust_bridge.host.routes.ocr")
+            .import(ROUTE_HOST_MODULE)
             .and_then(|module| module.getattr("map_failure"))
             .and_then(|map| map.call1((error.value(py), self.view.request(py)?, provider)))
             .and_then(|mapped| mapped.extract::<Py<PyBaseException>>().map_err(PyErr::from));
@@ -115,7 +115,7 @@ impl PythonBinding for OcrPythonHost {
         py: Python<'_>,
         response: LiteLLMOcrResponse,
     ) -> PyResult<Py<PyAny>> {
-        public_response(py, "litellm.rust_bridge.host.routes.ocr", &response)
+        public_response(py, ROUTE_HOST_MODULE, &response)
     }
 
     fn encode_stream_head(

litellm-rust/crates/python-bridge/src/routes/ocr/project.rs

@@ -5,7 +5,7 @@ use litellm_inference_ocr::{
     types::{LiteLLMOcrRequest, OcrDocumentInput},
     wire::{OcrWireRequest, consumed_optional_params, decode_document, decode_request_input},
 };
-use litellm_llms::base_llm::ocr::error::Error;
+use litellm_llms::base_llm::ocr::{error::Error, transformation::DROP_PARAMS_PARAM};
 use pyo3::{prelude::*, types::PyDict};
 use serde_json::{Map, Value};
 
@@ -19,6 +19,8 @@ use crate::{
     python_settings::{PythonSettings, Snapshot},
 };
 
+pub(super) const ROUTE_HOST_MODULE: &str = "litellm.rust_bridge.host.routes.ocr";
+
 /// What the host keeps after projection: the caller's token callable that answers the
 /// token operation, and the provider name the failure mapping reports.
 pub(super) struct OcrHostHandles {
@@ -139,6 +141,15 @@ pub(super) fn project_request(
         Some(defaults) => fold_module_globals(bound.py(), optional_params, &globals, &defaults)?,
         None => optional_params,
     };
+    let optional_params = if names.contains(&DROP_PARAMS_PARAM) {
+        fold_drop_params(
+            bound.py(),
+            optional_params,
+            present(kwargs, bound, DROP_PARAMS_PARAM)?,
+        )?
+    } else {
+        optional_params
+    };
     let input_sources = request_input_sources(
         kwargs,
         names
@@ -216,6 +227,23 @@ fn fold_module_globals(
         .collect())
 }
 
+fn fold_drop_params(
+    py: Python<'_>,
+    optional_params: Map<String, Value>,
+    requested: Option<Bound<'_, PyAny>>,
+) -> PyResult<Map<String, Value>> {
+    let drops: bool = py
+        .import(ROUTE_HOST_MODULE)?
+        .getattr("drops_unsupported_params")?
+        .call1((requested,))?
+        .extract()?;
+    Ok(optional_params
+        .into_iter()
+        .filter(|(name, _)| name != DROP_PARAMS_PARAM)
+        .chain(drops.then(|| (DROP_PARAMS_PARAM.to_string(), Value::Bool(true))))
+        .collect())
+}
+
 /// A misconfigured litellm setting is the operator's error, not the request's or the
 /// provider's, so the host raises it as is instead of mapping it onto a provider failure.
 pub(super) const SETTINGS_ERROR_MARKER: &str = "native_settings_error";

litellm/__init__.py

@@ -681,6 +681,7 @@ moonshot_models: Set = set()
 publicai_models: Set = set()
 darkbloom_models: Set = set()
 openinfer_models: Final[Set[str]] = set()
+streamlake_models: Final[Set[str]] = set()
 zerogpu_models: Final[Set[str]] = set()
 v0_models: Set = set()
 morph_models: Set = set()
@@ -970,6 +971,8 @@ def _populate_provider_model_sets(model_cost_map: Mapping[str, Mapping[str, obje
             darkbloom_models.add(key)
         elif value.get("litellm_provider") == "openinfer":
             openinfer_models.add(key)
+        elif value.get("litellm_provider") == "streamlake":
+            streamlake_models.add(key)
         elif value.get("litellm_provider") == "zerogpu":
             zerogpu_models.add(key)
         elif value.get("litellm_provider") == "v0":
@@ -1145,6 +1148,7 @@ model_list = list(
     | publicai_models
     | darkbloom_models
     | openinfer_models
+    | streamlake_models
     | zerogpu_models
     | v0_models
     | morph_models
@@ -1265,6 +1269,7 @@ def _build_models_by_provider() -> dict:
         "publicai": publicai_models,
         "darkbloom": darkbloom_models,
         "openinfer": openinfer_models,
+        "streamlake": streamlake_models,
         "zerogpu": zerogpu_models,
         "v0": v0_models,
         "morph": morph_models,

litellm/constants.py

@@ -1064,6 +1064,7 @@ openai_compatible_endpoints: Final[list[str]] = [
     "https://api.reka.ai/v1",
     "https://gigachat.devices.sberbank.ru/api/v1",
     "https://api.openinfer.ai/v1",
+    "https://vanchin.streamlake.ai/api/gateway/v1/endpoints",
 ]
 
 
@@ -1145,6 +1146,7 @@ openai_compatible_providers: Final[list[str]] = [
     "prism",
     "sail",
     "reka",
+    "streamlake",
 ]
 
 OPENAI_AUDIO_TRANSCRIPTION_PROVIDERS: Final = frozenset({"openai"} | frozenset(openai_compatible_providers))

litellm/cost_calculator.py

@@ -688,7 +688,7 @@ def cost_per_token(
     elif call_type == "atranscription" or call_type == "transcription":
         if _transcription_usage_has_token_details(usage_block):
             return generic_cost_per_token(
-                model=model_without_prefix,
+                model=model,
                 usage=usage_block,
                 custom_llm_provider=custom_llm_provider,
                 service_tier=service_tier,

litellm/google_genai/adapters/transformation.py

@@ -198,6 +198,11 @@ def _genai_message_part(part: object) -> _GenAIMessagePart | None:
     return ChatCompletionTextObject(type="text", text=text) if isinstance(text, str) else _genai_media_part(fields)
 
 
+def _genai_content_parts(value: object) -> Sequence[object]:
+    single_part: Final = _validated(_CONFIG_FIELDS, value)
+    return (single_part,) if single_part is not None else _validated(_GENAI_CONTENT_PARTS, value) or ()
+
+
 def _genai_message_parts(parts: object) -> tuple[_GenAIMessagePart, ...]:
     items: Final = _validated(_GENAI_CONTENT_PARTS, parts) or ()
     return tuple(converted for part in items if (converted := _genai_message_part(part)) is not None)
@@ -248,7 +253,7 @@ def _genai_content_messages(content: object) -> tuple[AllMessageValues, ...]:
     fields: Final = _validated(_CONFIG_FIELDS, content)
     if fields is None:
         return ()
-    parts: Final = _validated(_GENAI_CONTENT_PARTS, fields.get("parts", ())) or ()
+    parts: Final = _genai_content_parts(fields.get("parts", ()))
     converted: Final = _genai_message_parts(parts)
     role: Final = fields.get("role", "user")
     if role == "user":

litellm/litellm_core_utils/get_litellm_params.py

@@ -77,6 +77,7 @@ OPTIONAL_KWARGS_KEYS: Final = (
             "otpm",
             "tpd",
             "rpd",
+            "daily_reset_time",
             "use_xai_oauth",
             "chatgpt_auth_file",
             "github_copilot_auth_type",

litellm/litellm_core_utils/litellm_logging.py

@@ -1135,9 +1135,6 @@ class Logging(LiteLLMLoggingBaseClass):
         }
         self.litellm_request_debug = litellm_params.get("litellm_request_debug", False)
         self.logger_fn = litellm_params.get("logger_fn", None)
-        if is_debugging_on() or self.litellm_request_debug:
-            verbose_logger.debug("self.optional_params: %s", self.optional_params)
-
         self.model_call_details.update(
             {
                 "model": self.model,

litellm/litellm_core_utils/llm_cost_calc/usage_object_transformation.py

@@ -27,13 +27,18 @@ class TranscriptionUsageObjectTransformation:
         if isinstance(usage_object, TranscriptionUsageDurationObject):
             return None
         elif isinstance(usage_object, TranscriptionUsageTokensObject):
+            input_token_details: Final = usage_object.input_token_details
             return Usage(
                 prompt_tokens=usage_object.input_tokens,
                 completion_tokens=usage_object.output_tokens,
                 total_tokens=usage_object.total_tokens,
-                prompt_tokens_details=PromptTokensDetailsWrapper(
-                    text_tokens=usage_object.input_token_details.text_tokens,
-                    audio_tokens=usage_object.input_token_details.audio_tokens,
+                prompt_tokens_details=(
+                    None
+                    if input_token_details is None
+                    else PromptTokensDetailsWrapper(
+                        text_tokens=input_token_details.text_tokens,
+                        audio_tokens=input_token_details.audio_tokens,
+                    )
                 ),
             )
         return None

litellm/litellm_core_utils/streaming_handler.py

@@ -92,6 +92,18 @@ def _next_sync_or_exhausted(it: Iterator[object]) -> object:
         return _SYNC_ITER_EXHAUSTED
 
 
+def _stamp_usage_cost(
+    usage_chunk: ModelResponse | TextCompletionResponse | ModelResponseStream,
+    complete_response: ModelResponse | TextCompletionResponse | ModelResponseStream,
+) -> None:
+    bill: Final = (get_hidden_params(complete_response) or {}).get("response_cost")
+    usage: Final[object] = getattr(usage_chunk, "usage", None)
+    if usage is None or isinstance(bill, bool) or not isinstance(bill, (int, float)):
+        return
+    if bill > 0 or isinstance(getattr(usage, "cost", None), (int, float)):
+        setattr(usage, "cost", float(bill))  # noqa: B010  # cost is an optional extra on the usage model
+
+
 def _stamp_served_service_tier(response: ModelResponseStream, complete_streaming_response: ModelResponse) -> None:
     served_tier: Final = complete_streaming_response.model_dump().get("service_tier")
     if isinstance(served_tier, str) and served_tier:
@@ -2141,6 +2153,7 @@ class CustomStreamWrapper:
                         getattr(complete_streaming_response, "usage"),
                     )
                     _stamp_served_service_tier(response, complete_streaming_response)
+                    _stamp_usage_cost(response, complete_streaming_response)
                     try:
                         _cache_copy = complete_streaming_response.model_copy(deep=True)
                         _log_copy = complete_streaming_response.model_copy(deep=True)
@@ -2361,6 +2374,7 @@ class CustomStreamWrapper:
                     getattr(complete_streaming_response, "usage"),
                 )
                 _stamp_served_service_tier(response, complete_streaming_response)
+                _stamp_usage_cost(response, complete_streaming_response)
                 try:
                     _copy = complete_streaming_response.model_copy(deep=True)
                 except RuntimeError:

litellm/llms/anthropic/messages/context_management/editors/clear_tool_uses.py

@@ -31,6 +31,7 @@ from ..constants import (
 )
 from ..errors import AnthropicContextManagementError
 from ..placeholders import build_cleared_tool_result_content
+from .compact import count_system_tokens, history_after_client_compaction
 
 if TYPE_CHECKING:
     from litellm.types.llms.openai import ChatCompletionToolParam
@@ -203,9 +204,10 @@ def _trigger_met(
     model: str,
     messages: list[dict[str, object]],
     tools: list[dict[str, object]] | None,
+    system: str | Sequence[Mapping[str, object]] | None,
     counted_tool_uses: int,
 ) -> tuple[bool, int | None]:
-    """Return (trigger_met, input_tokens if counted for reuse)."""
+    """Return (trigger_met, message and tool tokens for reuse); like the API, input_tokens adds the system prompt."""
     trigger_type: Final = trigger.get("type", "input_tokens")
     threshold = trigger.get("value")
 
@@ -216,14 +218,17 @@ def _trigger_met(
 
     if not isinstance(threshold, int):
         threshold = DEFAULT_INPUT_TOKENS_TRIGGER
-    current_tokens: Final = litellm.token_counter(
+    message_tokens: Final = litellm.token_counter(
         model=model,
         messages=messages,
         tools=cast("list[ChatCompletionToolParam] | None", tools),
     )
+    if message_tokens > threshold:
+        return True, message_tokens
+    current_tokens: Final = message_tokens + count_system_tokens(model, system)
     verbose_logger.debug("context_management polyfill: current_tokens: %s", current_tokens)
     verbose_logger.debug("context_management polyfill: threshold: %s", threshold)
-    return current_tokens > threshold, current_tokens
+    return current_tokens > threshold, message_tokens
 
 
 def _resolve_keep_count(keep: dict[str, object]) -> int:
@@ -237,7 +242,7 @@ def _resolve_keep_count(keep: dict[str, object]) -> int:
 
 
 def _last_completed_tool_use_id(
-    messages: list[dict[str, object]],
+    messages: Sequence[Mapping[str, object]],
 ) -> str | None:
     """Latest completed tool_result id; never cleared."""
     last_id: str | None = None
@@ -253,7 +258,7 @@ def _last_completed_tool_use_id(
 
 
 def _clear_tool_results(
-    messages: list[dict[str, object]], ids_to_clear: set[str]
+    messages: list[dict[str, object]], ids_to_clear: Set[str]
 ) -> tuple[list[dict[str, object]], int]:
     """Clear matching tool_result content; return (messages, cleared_count)."""
     cleared = 0
@@ -311,7 +316,11 @@ def apply_clear_tool_uses_20250919(
     system: str | list[dict[str, object]] | None,
     edit_spec: ClearToolUsesEditSpec,
 ) -> tuple[list[dict[str, object]], AppliedEdit | None]:
-    """Apply clear_tool_uses to an edit ``check_clear_tool_uses_spec`` returned; return (messages, AppliedEdit or None)."""
+    """Apply clear_tool_uses to an edit ``check_clear_tool_uses_spec`` returned; return (messages, AppliedEdit or None).
+
+    Like the API, it reads only the history after a client compaction block and counts the block's summary as system.
+    """
+    kept, counted_system = history_after_client_compaction(messages, system)
     fields: Final = _ClearToolUsesFields.model_validate(edit_spec)
     clear_at_least: Final = 0 if fields.clear_at_least is None else fields.clear_at_least.value
     exclude_tools: Final = frozenset(fields.exclude_tools or ())
@@ -326,7 +335,7 @@ def apply_clear_tool_uses_20250919(
         "value": DEFAULT_KEEP_TOOL_USES,
     }
 
-    blocks: Final = _all_blocks(messages)
+    blocks: Final = _all_blocks(kept)
     results: Final = _blocks_by_id(blocks, "tool_result", "tool_use_id")
     uses: Final = _blocks_by_id(blocks, "tool_use", "id")
     tiny: Final = _tiny_result_ids(results)
@@ -335,15 +344,15 @@ def apply_clear_tool_uses_20250919(
         for tool_use_id, name in _tool_uses_in_order(blocks)
         if name not in exclude_tools and (tool_use_id not in tiny or _clears_input(name, input_tools))
     )
-    met, counted_tokens = _trigger_met(trigger, model, messages, tools, len(clearable))
-    if not met:
-        return messages, None
-
     keep_count: Final = _resolve_keep_count(keep)
     if len(clearable) <= keep_count:
         return messages, None
 
-    last_completed_id: Final = _last_completed_tool_use_id(messages)
+    met, counted_tokens = _trigger_met(trigger, model, kept, tools, counted_system, len(clearable))
+    if not met:
+        return messages, None
+
+    last_completed_id: Final = _last_completed_tool_use_id(kept)
     candidates: Final = tuple(
         (tool_use_id, name)
         for tool_use_id, name in clearable[: len(clearable) - keep_count]
@@ -358,7 +367,7 @@ def apply_clear_tool_uses_20250919(
         _select_in_batches(candidate_ids, tokens, clear_at_least) if clear_at_least > 0 else candidate_ids
     )
 
-    edited, cleared_count = _clear_tool_results(messages, set(ids_to_clear))
+    edited, cleared_count = _clear_tool_results(messages, ids_to_clear)
     verbose_logger.debug("context_management polyfill: edited: %s", edited)
     if cleared_count == 0:
         return messages, None
@@ -370,11 +379,12 @@ def apply_clear_tool_uses_20250919(
         counted_tokens
         if counted_tokens is not None
         else litellm.token_counter(
-            model=model, messages=messages, tools=cast("list[ChatCompletionToolParam] | None", tools)
+            model=model, messages=kept, tools=cast("list[ChatCompletionToolParam] | None", tools)
         )
     )
+    kept_after, _ = history_after_client_compaction(final_messages, system)
     tokens_after: Final = litellm.token_counter(
-        model=model, messages=final_messages, tools=cast("list[ChatCompletionToolParam] | None", tools)
+        model=model, messages=kept_after, tools=cast("list[ChatCompletionToolParam] | None", tools)
     )
     cleared_input_tokens: Final = max(tokens_before - tokens_after, 0)
 

litellm/llms/anthropic/messages/context_management/editors/compact.py

@@ -189,6 +189,21 @@ def _strip_compaction_blocks(
     return cleaned
 
 
+def history_after_client_compaction(
+    messages: Sequence[dict[str, object]],
+    system: str | Sequence[Mapping[str, object]] | None,
+) -> tuple[list[dict[str, object]], str | Sequence[Mapping[str, object]] | None]:
+    """Return the history the API keeps after the latest compaction block, without it, and the system with its summary."""
+    sliced, block = _slice_around_compaction_block(messages)
+    if block is None:
+        return list(messages), system
+    tail: Final = _strip_compaction_blocks(sliced)
+    summary: Final = block.get("content")
+    if not isinstance(summary, str) or not summary:
+        return tail, system
+    return tail, _augment_system_with_summary(_system_to_text(system) or None, summary)
+
+
 def _augment_system_with_summary(
     system: str | list[dict[str, object]] | None,
     summary_text: str,
@@ -320,14 +335,16 @@ def _count_effective_tokens(
         content: Final = compaction_block.get("content") or ""
         if content:
             total += litellm.token_counter(model=model, text=content)
+    return total + count_system_tokens(model, system)
+
+
+def count_system_tokens(model: str, system: str | Sequence[Mapping[str, object]] | None) -> int:
     system_text: Final = _system_to_text(system)
-    if system_text:
-        total += litellm.token_counter(model=model, text=system_text)
-    return total
+    return litellm.token_counter(model=model, text=system_text) if system_text else 0
 
 
 def _system_to_text(
-    system: str | list[dict[str, object]] | None,
+    system: str | Sequence[Mapping[str, object]] | None,
 ) -> str:
     """Flatten an Anthropic-style ``system`` value into a single string for
     token counting. Returns ``""`` when ``system`` carries no text."""
@@ -338,7 +355,7 @@ def _system_to_text(
     return "\n".join(
         text
         for block in system
-        if isinstance(block, dict)
+        if isinstance(block, Mapping)
         and block.get("type") == "text"
         and isinstance(text := block.get("text"), str)
         and text

litellm/llms/azure/common_utils.py

@@ -621,7 +621,7 @@ class BaseAzureLLM(BaseOpenAILLM):
 
         # For Azure v1 API, use standard OpenAI client instead of AzureOpenAI
         # See: https://learn.microsoft.com/en-us/azure/ai-services/openai/reference#api-specs
-        if self._is_azure_v1_api_version(api_version):
+        if self.is_azure_v1_api_version(api_version):
             # Extract only params that OpenAI client accepts
             # Always use /openai/v1/ regardless of whether user passed "v1", "latest", or "preview"
             # The OpenAI client accepts a callable for `api_key` and re-invokes it
@@ -922,7 +922,7 @@ class BaseAzureLLM(BaseOpenAILLM):
         else:
             new_url = api_base
 
-        if BaseAzureLLM._is_azure_v1_api_version(api_version):
+        if BaseAzureLLM.is_azure_v1_api_version(api_version):
             # ensure the request go to /openai/v1 and not just /openai
             if "/openai/v1" not in new_url:
                 parsed_url: Final = httpx.URL(new_url)
@@ -952,7 +952,7 @@ class BaseAzureLLM(BaseOpenAILLM):
 
         Returns None when ``api_version`` is a dated one, which still uses the deployment route.
         """
-        if not BaseAzureLLM._is_azure_v1_api_version(api_version):
+        if not BaseAzureLLM.is_azure_v1_api_version(api_version):
             return None
 
         base_url: Final = httpx.URL(api_base)
@@ -970,7 +970,7 @@ class BaseAzureLLM(BaseOpenAILLM):
         )
 
     @staticmethod
-    def _is_azure_v1_api_version(api_version: str | None) -> bool:
+    def is_azure_v1_api_version(api_version: str | None) -> bool:
         if api_version is None:
             return False
         return api_version in {"preview", "latest", "v1"}

litellm/llms/azure/responses/transformation.py

@@ -54,6 +54,16 @@ def _resolve_responses_api_version(api_base: str | None, litellm_params: Mapping
     return api_version
 
 
+def _v1_responses_url(api_base: str) -> httpx.URL:
+    parsed_url: Final = httpx.URL(api_base)
+    path: Final = parsed_url.path.rstrip("/")
+    root: Final = next(
+        (path.removesuffix(route) for route in ("/openai/v1/responses", "/openai/responses") if path.endswith(route)),
+        path,
+    )
+    return parsed_url.copy_with(path=f"{root}/openai/v1/responses", query=None)
+
+
 def _with_responses_api_version(api_base: str | None, litellm_params: Mapping[str, object]) -> Mapping[str, object]:
     return MappingProxyType({**litellm_params, "api_version": _resolve_responses_api_version(api_base, litellm_params)})
 
@@ -132,7 +142,9 @@ class AzureOpenAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
                 return filtered_item
         return item
 
-    def _validate_input_param(self, input: str | ResponseInputParam) -> str | ResponseInputParam:
+    def _validate_input_param(
+        self, input: str | ResponseInputParam, *, drops_set_created_by: bool = False
+    ) -> str | ResponseInputParam:
         """
         Override parent method to also filter out 'status' field from message items.
         Azure OpenAI API does not accept 'status' field in input messages.
@@ -140,7 +152,7 @@ class AzureOpenAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
         from typing import cast
 
         # First call parent's validation
-        validated_input: Final = super()._validate_input_param(input)
+        validated_input: Final = super()._validate_input_param(input, drops_set_created_by=drops_set_created_by)
 
         # Then filter out status from message items
         if isinstance(validated_input, list):
@@ -247,20 +259,20 @@ class AzureOpenAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
         if api_base is None:
             raise ValueError("api_base is required for Azure WebSocket")
 
-        parsed_url: Final = httpx.URL(api_base)
-        path = parsed_url.path.rstrip("/")
-        # Strip existing /openai/responses path if the api_base already contains it
-        for suffix in ("/openai/v1/responses", "/openai/responses"):
-            if path.endswith(suffix):
-                path = path[: -len(suffix)]
-                break
-        scheme: Final = "wss" if parsed_url.scheme == "https" else "ws"
-        return str(parsed_url.copy_with(scheme=scheme, path=f"{path}/openai/v1/responses", query=None))
+        v1_url: Final = _v1_responses_url(api_base)
+        return str(v1_url.copy_with(scheme="wss" if v1_url.scheme == "https" else "ws"))
 
     def model_in_websocket_url(self) -> bool:
         # Azure sends the model in the response.create body, not the URL
         return False
 
+    def compact_url(self, api_base: str) -> str:
+        params: Final = httpx.URL(api_base).params
+        v1_params: Final = (
+            params if BaseAzureLLM.is_azure_v1_api_version(params.get("api-version")) else params.remove("api-version")
+        )
+        return super().compact_url(str(_v1_responses_url(api_base).copy_with(params=v1_params)))
+
     #########################################################
     ########## DELETE RESPONSE API TRANSFORMATION ##############
     #########################################################

litellm/llms/azure_ai/common_utils.py

@@ -140,6 +140,15 @@ def get_azure_ai_auth_headers(
 AZURE_MODEL_ROUTER_SELECTED_MODEL_KEY: Final = "azure_model_router_selected_model"
 
 
+FOUNDRY_CLASSIC_AGENT_ID_PREFIX: Final = "asst_"
+
+
+def is_foundry_project_agent(model: str | None) -> bool:
+    if model is None or AzureFoundryModelInfo.get_azure_ai_route(model) != "agents":
+        return False
+    return not model.split("agents/", 1)[1].startswith(FOUNDRY_CLASSIC_AGENT_ID_PREFIX)
+
+
 def azure_ai_supports_native_responses(model: str | None, api_base: str | None) -> bool:
     resolved_base: Final = AzureFoundryModelInfo.get_api_base(api_base)
     if resolved_base is not None and not is_azure_openai_v1_host(resolved_base):

litellm/llms/azure_ai/responses/transformation.py

@@ -1,26 +1,64 @@
-from typing import Final
+import os
+from collections.abc import Collection, Mapping, Sequence
+from typing import TYPE_CHECKING, Final, NoReturn
 
 import httpx
+from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 
+import litellm
+from litellm.exceptions import UnsupportedParamsError
 from litellm.llms.azure.responses.transformation import AzureOpenAIResponsesAPIConfig
 from litellm.llms.azure_ai.chat.transformation import AzureAIGPT5Config
 from litellm.llms.azure_ai.common_utils import (
+    AZURE_AI_AGENTS_SCOPE,
     AzureFoundryModelInfo,
     api_key_header_for_base,
     get_azure_ai_auth_headers,
 )
+from litellm.types.llms.azure_ai import FoundryAgentReference
+from litellm.types.llms.openai import (
+    ResponseInputParam,
+    ResponsesAPIOptionalRequestParams,
+    ResponsesAPIResponse,
+    ResponsesAPIStreamingResponse,
+)
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import LlmProviders
 
+if TYPE_CHECKING:
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+
 _PROJECT_PATH_PREFIX: Final = ("api", "projects")
 _RESPONSES_PATH: Final = ("openai", "v1", "responses")
+_AGENT_OWNED_PARAMS: Final = frozenset({"temperature", "top_p", "tools", "text", "reasoning"})
+_CALLER_AGENT_KEYS: Final = ("agent", "agent_reference")
+_RESPONSE_FIELDS: Final = TypeAdapter(dict[str, object])
+
+
+def _project_root(base_url: httpx.URL) -> tuple[str, ...]:
+    segments: Final = tuple(segment for segment in base_url.path.split("/") if segment)
+    return segments[:3] if segments[:2] == _PROJECT_PATH_PREFIX else ()
 
 
 def _responses_url(api_base: str) -> str:
     base_url: Final = httpx.URL(api_base)
-    segments: Final = tuple(segment for segment in base_url.path.split("/") if segment)
-    project_root: Final = segments[:3] if segments[:2] == _PROJECT_PATH_PREFIX else ()
-    return str(base_url.copy_with(path="/" + "/".join((*project_root, *_RESPONSES_PATH)), query=None))
+    return str(base_url.copy_with(path="/" + "/".join((*_project_root(base_url), *_RESPONSES_PATH)), query=None))
+
+
+def _without_agent_instructions(response: ResponsesAPIResponse) -> ResponsesAPIResponse:
+    if (response.model_extra or {}).get("agent_reference") is None:
+        return response
+    return response.model_copy(update={"instructions": None})
+
+
+def _event_without_agent_instructions(event: Mapping[str, object]) -> Mapping[str, object]:
+    response: Final = event.get("response")
+    if not isinstance(response, Mapping):
+        return event
+    fields: Final = _RESPONSE_FIELDS.validate_python(response)
+    if fields.get("agent_reference") is None:
+        return event
+    return {**event, "response": {**fields, "instructions": None}}
 
 
 class AzureAIResponsesAPIConfig(AzureOpenAIResponsesAPIConfig):
@@ -42,10 +80,19 @@ class AzureAIResponsesAPIConfig(AzureOpenAIResponsesAPIConfig):
 
     def validate_environment(self, headers: dict, model: str, litellm_params: GenericLiteLLMParams | None) -> dict:
         params: Final = litellm_params or GenericLiteLLMParams()
+        api_base: Final = AzureFoundryModelInfo.get_api_base(params.api_base)
+        on_project: Final = api_base is not None and bool(_project_root(httpx.URL(api_base)))
+        auth_params: Final = (
+            params.model_copy(
+                update={"azure_scope": params.azure_scope or os.getenv("AZURE_SCOPE") or AZURE_AI_AGENTS_SCOPE}
+            )
+            if on_project
+            else params
+        )
         auth_headers: Final = get_azure_ai_auth_headers(
             api_key=AzureFoundryModelInfo.get_api_key(params.api_key),
-            litellm_params=params.model_dump(),
-            api_key_header=api_key_header_for_base(AzureFoundryModelInfo.get_api_base(params.api_base)),
+            litellm_params=auth_params.model_dump(),
+            api_key_header="Authorization" if on_project else api_key_header_for_base(api_base),
         )
         return {
             **headers,
@@ -64,3 +111,207 @@ class AzureAIResponsesAPIConfig(AzureOpenAIResponsesAPIConfig):
                 "Set the api_base parameter or the AZURE_AI_API_BASE environment variable."
             )
         return _responses_url(resolved_base)
+
+    def transform_response_api_response(
+        self, model: str, raw_response: httpx.Response, logging_obj: "LiteLLMLoggingObj"
+    ) -> ResponsesAPIResponse:
+        return _without_agent_instructions(super().transform_response_api_response(model, raw_response, logging_obj))
+
+    def transform_streaming_response(
+        self,
+        model: str,
+        parsed_chunk: Mapping[str, object],
+        logging_obj: "LiteLLMLoggingObj",
+    ) -> ResponsesAPIStreamingResponse:
+        return super().transform_streaming_response(  # pyright: ignore[reportUnknownMemberType]  # parent types parsed_chunk as a bare dict
+            model=model, parsed_chunk=dict(_event_without_agent_instructions(parsed_chunk)), logging_obj=logging_obj
+        )
+
+    def transform_get_response_api_response(
+        self, raw_response: httpx.Response, logging_obj: "LiteLLMLoggingObj"
+    ) -> ResponsesAPIResponse:
+        return _without_agent_instructions(super().transform_get_response_api_response(raw_response, logging_obj))
+
+    def transform_cancel_response_api_response(
+        self, raw_response: httpx.Response, logging_obj: "LiteLLMLoggingObj"
+    ) -> ResponsesAPIResponse:
+        return _without_agent_instructions(super().transform_cancel_response_api_response(raw_response, logging_obj))
+
+
+def _agent_reference(model: str) -> FoundryAgentReference:
+    name, _, version = model.split("agents/", 1)[1].partition(":")
+    if version:
+        return FoundryAgentReference(type="agent_reference", name=name, version=version)
+    return FoundryAgentReference(type="agent_reference", name=name)
+
+
+class _CallerAgentReference(BaseModel):
+    model_config = ConfigDict(frozen=True, coerce_numbers_to_str=True)
+
+    name: str
+    version: str | None = None
+
+
+def _caller_agent_reference(value: object) -> _CallerAgentReference | None:
+    try:
+        return _CallerAgentReference.model_validate(value)
+    except ValidationError:
+        return None
+
+
+def _requested_agent(
+    deployment_agent: FoundryAgentReference, caller_agents: Sequence[_CallerAgentReference | None]
+) -> FoundryAgentReference:
+    requested_versions: Final = frozenset(
+        agent.version
+        for agent in caller_agents
+        if agent is not None and agent.name == deployment_agent["name"] and agent.version is not None
+    )
+    if "version" in deployment_agent or len(requested_versions) != 1:
+        return deployment_agent
+    return FoundryAgentReference(
+        type="agent_reference", name=deployment_agent["name"], version=next(iter(requested_versions))
+    )
+
+
+def _caller_agent_fields(fields: Mapping[str, object]) -> tuple[tuple[str, object], ...]:
+    return tuple((key, fields[key]) for key in _CALLER_AGENT_KEYS if key in fields)
+
+
+def _deployment_agent(
+    model: str, caller_fields: Sequence[tuple[str, object]], litellm_params: GenericLiteLLMParams
+) -> FoundryAgentReference:
+    caller_keys: Final = tuple(key for key, _ in caller_fields)
+    caller_agents: Final = tuple(_caller_agent_reference(value) for _, value in caller_fields)
+    agent: Final = _requested_agent(_agent_reference(model), caller_agents)
+    other_agents: Final = tuple(
+        key
+        for key, caller_agent in zip(caller_keys, caller_agents)
+        if caller_agent is None
+        or caller_agent.name != agent["name"]
+        or caller_agent.version not in (None, agent.get("version"))
+    )
+    if other_agents and not (litellm.drop_params is True or litellm_params.drop_params is True):
+        raise UnsupportedParamsError(
+            message=(
+                f"This model entry always calls Foundry agent {agent['name']!r}, so `{other_agents[0]}` "
+                "in the request cannot pick another agent or version. Remove it, call the model entry for that "
+                "agent, or set `drop_params: true` to drop it"
+            ),
+            llm_provider=LlmProviders.AZURE_AI.value,
+            model=model,
+        )
+    return agent
+
+
+_INPUT_ITEMS: Final = TypeAdapter(list[object])
+
+
+def _with_system_prompt(instructions: str, request_input: object) -> Sequence[object]:
+    conversation: Final = (
+        [{"type": "message", "role": "user", "content": request_input}]
+        if isinstance(request_input, str)
+        else _INPUT_ITEMS.validate_python(request_input)
+    )
+    return [{"type": "message", "role": "system", "content": instructions}, *conversation]
+
+
+class AzureAIAgentResponsesAPIConfig(AzureAIResponsesAPIConfig):
+    def should_fake_stream(
+        self, model: str | None, stream: bool | None, custom_llm_provider: str | None = None
+    ) -> bool:
+        return False
+
+    def get_supported_openai_params(self, model: str) -> list[str]:  # mutable-ok: override contract
+        inherited: Final[list[str]] = super().get_supported_openai_params(model)  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]  # parent returns a bare list  # mutable-ok: same
+        return [param for param in inherited if param not in _AGENT_OWNED_PARAMS]
+
+    def map_openai_params_for_deployment(
+        self,
+        response_api_optional_params: ResponsesAPIOptionalRequestParams,
+        model: str,
+        drop_params: bool,
+        *,
+        declared_reasoning_efforts: Collection[str],
+    ) -> Mapping[str, object]:
+        mapped: Final = super().map_openai_params_for_deployment(
+            response_api_optional_params, model, drop_params, declared_reasoning_efforts=declared_reasoning_efforts
+        )
+        resends_instructions: Final = bool(mapped.get("instructions")) and bool(mapped.get("previous_response_id"))
+        if resends_instructions and not drop_params:
+            raise UnsupportedParamsError(
+                message=(
+                    "azure_ai Foundry agents store a turn's system message with the conversation, so `instructions` "
+                    "with `previous_response_id` would add another copy each turn. Send `instructions` on the first "
+                    "turn only, or set `drop_params: true` to drop it"
+                ),
+                llm_provider=LlmProviders.AZURE_AI.value,
+                model=model,
+            )
+        dropped: Final = _AGENT_OWNED_PARAMS | (
+            frozenset({"instructions"}) if resends_instructions else frozenset[str]()
+        )
+        return {key: value for key, value in mapped.items() if key not in dropped}
+
+    def transform_responses_api_request(
+        self,
+        model: str,
+        input: str | ResponseInputParam,
+        response_api_optional_request_params: dict[str, object],  # mutable-ok: signature fixed by the override contract
+        litellm_params: GenericLiteLLMParams,
+        headers: dict[str, str],  # mutable-ok: same
+    ) -> dict[str, object]:  # mutable-ok: same
+        request: Final[dict[str, object]] = super().transform_responses_api_request(  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]  # parent returns a bare dict  # mutable-ok: same
+            model=model,
+            input=input,
+            response_api_optional_request_params=response_api_optional_request_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+        instructions: Final = request.get("instructions")
+        return {
+            **{key: value for key, value in request.items() if key not in ("model", "instructions")},
+            **(
+                {"input": _with_system_prompt(instructions, request.get("input"))}
+                if isinstance(instructions, str) and instructions
+                else {}
+            ),
+            "agent_reference": _deployment_agent(
+                model, _caller_agent_fields(litellm_params.model_extra or {}), litellm_params
+            ),
+        }
+
+    def transform_extra_body(
+        self,
+        extra_body: Mapping[str, object],
+        request: Mapping[str, object],
+        model: str,
+        litellm_params: GenericLiteLLMParams,
+    ) -> Mapping[str, object]:
+        passthrough: Final = {key: value for key, value in extra_body.items() if key not in _CALLER_AGENT_KEYS}
+        if len(passthrough) == len(extra_body):
+            return passthrough
+        return {
+            **passthrough,
+            "agent_reference": _deployment_agent(
+                model, (*_caller_agent_fields(request), *_caller_agent_fields(extra_body)), litellm_params
+            ),
+        }
+
+    def transform_compact_response_api_request(
+        self,
+        model: str,
+        input: str | ResponseInputParam,
+        response_api_optional_request_params: dict[str, object],  # mutable-ok: signature fixed by the override contract
+        api_base: str,
+        litellm_params: GenericLiteLLMParams,
+        headers: dict[str, str],  # mutable-ok: same
+    ) -> NoReturn:
+        raise UnsupportedParamsError(
+            message=(
+                "azure_ai Foundry agents cannot be compacted, because the project's /responses/compact endpoint "
+                "takes a model deployment, not an agent. Call compact on a model entry for that deployment instead"
+            ),
+            llm_provider=LlmProviders.AZURE_AI.value,
+            model=model,
+        )

litellm/llms/bedrock/audio_transcription/__init__.py

@@ -1,7 +1,11 @@
 import base64
+import math
+from collections.abc import Mapping
+from dataclasses import dataclass
 from typing import Final
 
 import httpx
+from pydantic import BaseModel, ConfigDict
 
 from litellm.litellm_core_utils.audio_utils.utils import process_audio_file
 from litellm.rust_bridge import runtime
@@ -14,7 +18,72 @@ from litellm.rust_bridge.transcription.native import (
     RustAtranscription,
     RustTranscription,
 )
-from litellm.types.utils import FileTypes, TranscriptionResponse
+from litellm.secret_managers.main import get_secret_str
+from litellm.types.utils import FileTypes, TranscriptionResponse, TranscriptionUsageTokensObject
+
+SUPPORTED_OPENAI_TRANSCRIPTION_PARAMS: Final = ("language", "prompt", "temperature", "response_format")
+TEXT_RESPONSE_FORMATS: Final = frozenset({"json", "text"})
+
+
+def _as_float(value: object) -> float | None:
+    if isinstance(value, bool) or not isinstance(value, (int, float, str)):
+        return None
+    try:
+        number: Final = float(value)
+    except ValueError:
+        return None
+    return number if math.isfinite(number) else None
+
+
+def _rejection(key: str, value: object) -> str | None:
+    if key not in SUPPORTED_OPENAI_TRANSCRIPTION_PARAMS:
+        return f"bedrock transcription does not support {key}."
+    if key == "response_format" and value not in TEXT_RESPONSE_FORMATS:
+        return f"bedrock transcription returns only the transcript text, so response_format={value!r} is not supported."
+    if key == "temperature" and _as_float(value) is None:
+        return f"bedrock transcription needs a number for temperature, got {value!r}."
+    return None
+
+
+@dataclass(frozen=True, slots=True)
+class UnsupportedTranscriptionParams:
+    message: str
+
+
+def map_openai_transcription_params(
+    non_default_params: Mapping[str, object], drop_params: bool
+) -> Mapping[str, object] | UnsupportedTranscriptionParams:
+    rejections: Final = {
+        key: reason for key, value in non_default_params.items() if (reason := _rejection(key, value)) is not None
+    }
+    if rejections and not drop_params:
+        return UnsupportedTranscriptionParams(
+            message=f"{' '.join(rejections.values())} To drop it from the call, set `litellm.drop_params = True`."
+        )
+    return {
+        key: _as_float(value) if key == "temperature" else value
+        for key, value in non_default_params.items()
+        if key not in rejections
+    }
+
+
+class _RustTranscriptionResult(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    text: str
+    usage: TranscriptionUsageTokensObject | None = None
+
+
+def _transcription_response(
+    result: Mapping[str, object], optional_params: Mapping[str, object]
+) -> TranscriptionResponse:
+    parsed: Final = _RustTranscriptionResult.model_validate(result)
+    response: Final = TranscriptionResponse(text=parsed.text).model_copy(update={"usage": parsed.usage})
+    region_name: Final = (
+        optional_params.get("aws_region_name") or get_secret_str("AWS_REGION_NAME") or get_secret_str("AWS_REGION")
+    )
+    response.hidden_params = {**response.hidden_params, "region_name": region_name}
+    return response
 
 
 class BedrockAudioTranscriptionRustDispatch:
@@ -64,7 +133,7 @@ class BedrockAudioTranscriptionRustDispatch:
                 "timeout_seconds": timeout_to_seconds(timeout),
             }
             call: Final = NativeCall(args=(), kwargs=fields, base={})
-            return TranscriptionResponse(**rust(call))
+            return _transcription_response(rust(call), optional_params)
 
         return runtime.run(
             RouteContext(Route.TRANSCRIPTION, provider=custom_llm_provider, model=model),
@@ -97,7 +166,7 @@ class BedrockAudioTranscriptionRustDispatch:
                 "timeout_seconds": timeout_to_seconds(timeout),
             }
             call: Final = NativeCall(args=(), kwargs=fields, base={})
-            return TranscriptionResponse(**await rust(call))
+            return _transcription_response(await rust(call), optional_params)
 
         return await runtime.arun(
             RouteContext(Route.TRANSCRIPTION, provider=custom_llm_provider, model=model),

litellm/llms/openai/responses/count_tokens/transformation.py

@@ -22,7 +22,7 @@ from litellm.llms.anthropic.messages.responses_adapters.transformation import (
 )
 from litellm.responses.utils import ResponsesAPIRequestUtils
 from litellm.types.llms.anthropic import AnthropicMessagesRequest, AnthropicOutputSchema
-from litellm.types.llms.openai import ResponsesAPIRequestParams
+from litellm.types.llms.openai import ResponsesAPIRequestParams, replayable_responses_input
 from litellm.types.utils import all_litellm_params
 
 
@@ -387,10 +387,13 @@ def _forwards(field: str, policy: DropPolicy) -> bool:
     return field in _COUNT_REQUEST_FIELDS or (not policy.drop_params and field in _INFERENCE_REQUEST_FIELDS)
 
 
-def _restored(field: str, value: object) -> object:
+def _restored(field: str, value: object, policy: DropPolicy) -> object:
     match field, value:
         case "input", _:
-            return ResponsesAPIRequestUtils.restore_encrypted_content_item_ids_in_input(copy.deepcopy(value))
+            return replayable_responses_input(
+                ResponsesAPIRequestUtils.restore_encrypted_content_item_ids_in_input(copy.deepcopy(value)),
+                drops_set_created_by=policy.drop_params,
+            )
         case "previous_response_id", str():
             return ResponsesAPIRequestUtils.decode_previous_response_id_to_original_previous_response_id(value)
         case _:
@@ -398,5 +401,7 @@ def _restored(field: str, value: object) -> object:
 
 
 def native_count_request(request: Mapping[str, object], model: str, policy: DropPolicy) -> Mapping[str, object]:
-    forwarded: Final = {field: _restored(field, value) for field, value in request.items() if _forwards(field, policy)}
+    forwarded: Final = {
+        field: _restored(field, value, policy) for field, value in request.items() if _forwards(field, policy)
+    }
     return {**forwarded, "model": model}

litellm/llms/openai/responses/guardrail_translation/handler.py

@@ -32,11 +32,11 @@ import copy
 import json
 import time
 import uuid
-from collections.abc import Mapping, MutableSequence, Sequence
+from collections.abc import Iterable, Mapping, MutableSequence, Sequence
 from dataclasses import dataclass
-from itertools import accumulate, chain, repeat
+from itertools import accumulate, chain, groupby, repeat
 from types import MappingProxyType
-from typing import TYPE_CHECKING, Any, Final, NamedTuple, Protocol, Union, cast
+from typing import TYPE_CHECKING, Any, Final, NamedTuple, Protocol, TypeVar, Union, cast
 
 from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 from typing_extensions import ReadOnly, TypedDict
@@ -55,7 +55,9 @@ from litellm.llms.base_llm.guardrail_translation.base_translation import (
 )
 from litellm.llms.base_llm.guardrail_translation.utils import (
     blocked_responses_stream_usage,
+    effective_scan_only_tool_results_for_guardrail,
     effective_skip_system_message_for_guardrail,
+    effective_skip_tool_message_for_guardrail,
     list_field,
     merge_guardrailed_scoped_messages,
     role_out_of_guardrail_scope,
@@ -65,7 +67,10 @@ from litellm.llms.base_llm.guardrail_translation.utils import (
     stream_item_items,
     unappliable_request_rewrite,
 )
-from litellm.llms.openai.responses.guardrail_translation.tool_merge import merge_guardrailed_tools
+from litellm.llms.openai.responses.guardrail_translation.tool_merge import (
+    append_returned_tools,
+    merge_guardrailed_tools,
+)
 from litellm.responses.litellm_completion_transformation.compaction_items import (
     compaction_summary,
     decode_compaction_item,
@@ -74,7 +79,9 @@ from litellm.responses.litellm_completion_transformation.compaction_items import
 from litellm.responses.litellm_completion_transformation.custom_tools import without_malformed_tool_flags
 from litellm.responses.litellm_completion_transformation.tool_search import is_server_tool_search_tool
 from litellm.responses.litellm_completion_transformation.transformation import (
+    TOOL_CALL_OUTPUT_ITEM_TYPES,
     LiteLLMCompletionResponsesConfig,
+    tool_output_to_chat_content,
 )
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.llms.openai import (
@@ -270,6 +277,7 @@ _TOOL_CALL_PAYLOAD_EVENT_TYPES: Final = _TOOL_CALL_PAYLOAD_DELTA_EVENT_TYPES | f
 )
 _OUTPUT_ITEM_EVENT_TYPES: Final = frozenset({"response.output_item.added", "response.output_item.done"})
 _OUTPUT_TEXT_EVENT_TYPES: Final = frozenset({"response.output_text.delta", "response.output_text.done"})
+_TOOL_OUTPUT_TEXT_PART_TYPES: Final = frozenset({"input_text", "output_text", "text"})
 _PATCHABLE_ITEM_FIELDS: Final[Mapping[str, str]] = MappingProxyType(
     {"function_call_output": "output", "custom_tool_call_output": "output", "message": "content"}
 )
@@ -519,10 +527,109 @@ def scannable_instructions(data: Mapping[str, object], *, skip_system: bool = Fa
 
 
 def _input_item_role(item: object) -> str:
+    if stream_item_field(item, "type") in TOOL_CALL_OUTPUT_ITEM_TYPES:
+        return "tool"
     role: Final = item.get("role") if isinstance(item, Mapping) else None
     return role.lower() if isinstance(role, str) else ""
 
 
+def _scanned_text_field(item: Mapping[str, object]) -> str:
+    return "output" if item.get("type") in TOOL_CALL_OUTPUT_ITEM_TYPES else "content"
+
+
+def _tool_output_part_text(part: object) -> str | None:
+    if not isinstance(part, dict):
+        return None
+    fields: Final = _item_fields_adapter.validate_python(part)
+    text: Final = fields.get("text")
+    return text if fields.get("type") in _TOOL_OUTPUT_TEXT_PART_TYPES and isinstance(text, str) else None
+
+
+def tool_output_texts(item: object) -> tuple[tuple[int | None, str], ...]:
+    if not stream_item_field(item, "call_id") or stream_item_field(item, "type") == "computer_call_output":
+        return ()
+    parts: Final = stream_item_items(item, "output")
+    part_texts: Final = tuple(
+        (index, text) for index, part in enumerate(parts) if (text := _tool_output_part_text(part))
+    )
+    if part_texts:
+        return part_texts
+    tool_content: Final = tool_output_to_chat_content(stream_item_field(item, "output"))
+    return ((None, tool_content),) if isinstance(tool_content, str) and tool_content else ()
+
+
+_Grouped: Final = TypeVar("_Grouped")
+
+
+def _pair_call_id(pair: tuple[str, object]) -> str:
+    return pair[0]
+
+
+def _grouped_by_call_id(pairs: Iterable[tuple[str, _Grouped]]) -> Mapping[str, tuple[_Grouped, ...]]:
+    ordered: Final = sorted(pairs, key=_pair_call_id)
+    return MappingProxyType(
+        {call_id: tuple(value for _, value in group) for call_id, group in groupby(ordered, key=_pair_call_id)}
+    )
+
+
+def _texts_by_tool_row(
+    messages: Sequence[Mapping[str, object]], input_data: "str | ResponseInputParam"
+) -> Mapping[int, tuple[str, ...]]:
+    if isinstance(input_data, str):
+        return MappingProxyType({})
+    output_texts_by_call_id: Final = _grouped_by_call_id(
+        (str(stream_item_field(item, "call_id")), tuple(text for _, text in tool_output_texts(item)))
+        for item in input_data
+        if stream_item_field(item, "type") in TOOL_CALL_OUTPUT_ITEM_TYPES
+    )
+    row_indices_by_call_id: Final = _grouped_by_call_id(
+        (str(message.get("tool_call_id")), index)
+        for index, message in enumerate(messages)
+        if message.get("role") == "tool"
+    )
+    return MappingProxyType(
+        dict(
+            chain.from_iterable(
+                zip(indices, output_texts_by_call_id[call_id], strict=True)
+                for call_id, indices in row_indices_by_call_id.items()
+                if len(indices) == len(output_texts_by_call_id.get(call_id, ()))
+            )
+        )
+    )
+
+
+def _is_string_tool_row(message: Mapping[str, object]) -> bool:
+    return message.get("role") == "tool" and isinstance(message.get("content"), str)
+
+
+def _split_into_text_parts(message: AllMessageValues, texts: tuple[str, ...]) -> AllMessageValues:
+    if not _is_string_tool_row(message) or len(texts) < 2:
+        return message
+    return cast(  # cast-ok: a tool row whose content is its text parts as chat text blocks
+        "AllMessageValues", {**message, "content": [{"type": "text", "text": text} for text in texts]}
+    )
+
+
+def _with_split_rows_restored(
+    rows: Sequence[AllMessageValues], split_rows: Sequence[tuple[AllMessageValues, AllMessageValues]]
+) -> Sequence[AllMessageValues]:
+    if not isinstance(rows, list):
+        return rows
+    return [next((original for split, original in split_rows if row == split), row) for row in rows]
+
+
+def _is_textless_tool_row(message: Mapping[str, object], texts: tuple[str, ...] | None) -> bool:
+    return _is_string_tool_row(message) and (not message.get("content") or texts == ())
+
+
+def _rewritten_whole_text(item: object, field: str, guardrail_response: str) -> object | None:
+    if isinstance(stream_item_field(item, field), str):
+        return guardrail_response
+    if tool_output_texts(item) == ((None, guardrail_response),):
+        return None
+    return guardrail_response
+
+
 def _patched_request_fields(
     raw_input: object,
     instructions: object,
@@ -636,17 +743,28 @@ class OpenAIResponsesHandler(BaseTranslation):
 
         Handles both string input and list of message objects.
         """
-        input_data: Final[str | ResponseInputParam | None] = data.get("input")
-        if not isinstance(input_data, (str, list)):
+        raw_input: Final[str | ResponseInputParam | None] = data.get("input")
+        if not isinstance(raw_input, (str, list)):
             return data
+        input_data: Final = cast("str | ResponseInputParam", raw_input)  # cast-ok: fields read via stream_item_field
         skip_system: Final = effective_skip_system_message_for_guardrail(guardrail_to_apply)
+        skip_tool: Final = effective_skip_tool_message_for_guardrail(guardrail_to_apply)
+        scan_only: Final = effective_scan_only_tool_results_for_guardrail(guardrail_to_apply)
         structured_messages: Final = self.get_structured_messages(data)
-        scoped_indices: Final = scoped_structured_message_indices(
-            structured_messages or [], scan_only_tool_results=False, skip_system=skip_system, skip_tool=False
+        all_messages: Final = structured_messages or []
+        texts_by_tool_row: Final = _texts_by_tool_row(all_messages, input_data)
+        scoped_indices: Final = tuple(
+            index
+            for index in scoped_structured_message_indices(
+                all_messages, scan_only_tool_results=scan_only, skip_system=skip_system, skip_tool=skip_tool
+            )
+            if not _is_textless_tool_row(all_messages[index], texts_by_tool_row.get(index))
         )
-        scoped_structured_messages: Final = (
-            [structured_messages[index] for index in scoped_indices] if structured_messages else None
+        split_rows: Final = tuple(
+            (_split_into_text_parts(all_messages[index], texts_by_tool_row.get(index, ())), all_messages[index])
+            for index in scoped_indices
         )
+        scoped_structured_messages: Final = [split for split, _ in split_rows] if structured_messages else None
         raw_tools: Final = data.get("tools")
         original_tools: Final[tuple[Mapping[str, object], ...]] = (
             tuple(raw_tools) if isinstance(raw_tools, list) else ()
@@ -661,7 +779,12 @@ class OpenAIResponsesHandler(BaseTranslation):
             )
         )
         extracted: Final = self._extract_guardrail_inputs(
-            data, input_data, flattened_tool_groups, skip_system=skip_system
+            data,
+            input_data,
+            () if scan_only else flattened_tool_groups,
+            skip_system=skip_system,
+            skip_tool=skip_tool,
+            scan_only=scan_only,
         )
         if not extracted.inputs.get("texts"):
             return data
@@ -674,13 +797,14 @@ class OpenAIResponsesHandler(BaseTranslation):
             logging_obj=litellm_logging_obj,
         )
         self._apply_guardrailed_tools_to_data(
-            data, original_tools, flattened_tool_groups, guardrailed_inputs.get("tools")
+            data, original_tools, flattened_tool_groups, guardrailed_inputs.get("tools"), scan_only=scan_only
         )
         written_back: Final = self._written_back_request_fields(
             data,
             structured_messages or (),
             scoped_indices,
             scoped_structured_messages,
+            tuple((split, original) for split, original in split_rows if split is not original),
             guardrail_to_apply,
             guardrailed_inputs,
         )
@@ -730,8 +854,10 @@ class OpenAIResponsesHandler(BaseTranslation):
         flattened_tool_groups: Sequence[Sequence[Mapping[str, object]]],
         *,
         skip_system: bool = False,
+        skip_tool: bool = False,
+        scan_only: bool = False,
     ) -> _ExtractedInputs:
-        instructions: Final = scannable_instructions(data, skip_system=skip_system)
+        instructions: Final = scannable_instructions(data, skip_system=skip_system or scan_only)
         texts_to_check: Final[list[str]] = [] if instructions is None else [instructions]
         images_to_check: Final[list[str]] = []
         task_mappings: Final[list[tuple[int, int | None]]] = []
@@ -745,11 +871,15 @@ class OpenAIResponsesHandler(BaseTranslation):
             )
         )
         if isinstance(input_data, str):
-            texts_to_check.append(input_data)
+            if not scan_only:
+                texts_to_check.append(input_data)
         else:
             for msg_idx, message in enumerate(input_data):
                 if role_out_of_guardrail_scope(
-                    _input_item_role(message), skip_system_message=skip_system, skip_tool_message=False
+                    _input_item_role(message),
+                    skip_system_message=skip_system,
+                    skip_tool_message=skip_tool,
+                    scan_only_tool_results=scan_only,
                 ):
                     continue
                 self._extract_input_text_and_images(
@@ -775,12 +905,14 @@ class OpenAIResponsesHandler(BaseTranslation):
         structured_messages: Sequence[AllMessageValues],
         scoped_indices: Sequence[int],
         scoped_structured_messages: Sequence[AllMessageValues] | None,
+        split_rows: Sequence[tuple[AllMessageValues, AllMessageValues]],
         guardrail_to_apply: "CustomGuardrail",
         guardrailed_inputs: GenericGuardrailAPIInputs,
     ) -> _RequestFields | None:
-        guardrailed: Final = guardrailed_inputs.get("structured_messages")
-        if guardrailed is None or guardrailed is scoped_structured_messages:
+        returned: Final = guardrailed_inputs.get("structured_messages")
+        if returned is None or returned is scoped_structured_messages:
             return None
+        guardrailed: Final = _with_split_rows_restored(returned, split_rows)
         covers_full_request: Final = len(scoped_indices) == len(structured_messages) or (
             guardrail_to_apply.structured_messages_cover_full_request() and len(guardrailed) == len(structured_messages)
         )
@@ -814,11 +946,14 @@ class OpenAIResponsesHandler(BaseTranslation):
         original_tools: Sequence[Mapping[str, object]],
         flattened_tool_groups: Sequence[Sequence[Mapping[str, object]]],
         guardrailed_tools: Sequence[ChatCompletionToolParam] | None,
+        *,
+        scan_only: bool,
     ) -> None:
         if guardrailed_tools is None:
             return
+        merge: Final = append_returned_tools if scan_only else merge_guardrailed_tools
         data["tools"] = list(  # rebind-ok: in-place request rewrite
-            merge_guardrailed_tools(original_tools, flattened_tool_groups, guardrailed_tools)
+            merge(original_tools, flattened_tool_groups, guardrailed_tools)
         )
 
     def _extract_input_text_and_images(
@@ -839,6 +974,11 @@ class OpenAIResponsesHandler(BaseTranslation):
             texts_to_check.append(summary)
             task_mappings.append((msg_idx, None))
             return
+        if message.get("type") in TOOL_CALL_OUTPUT_ITEM_TYPES:
+            for content_idx, text in tool_output_texts(message):
+                texts_to_check.append(text)
+                task_mappings.append((msg_idx, content_idx))
+            return
         content: Final = message.get("content", None)
         if content is None:
             return
@@ -890,18 +1030,20 @@ class OpenAIResponsesHandler(BaseTranslation):
                 )
                 continue
 
-            content = messages[msg_idx].get("content", None)
+            field = _scanned_text_field(messages[msg_idx])
+            content = messages[msg_idx].get(field, None)
             if content is None:
                 continue
 
-            if isinstance(content, str) and content_idx_optional is None:
-                # Replace string content with guardrail response
-                messages[msg_idx]["content"] = guardrail_response
+            if content_idx_optional is None:
+                rewritten = _rewritten_whole_text(messages[msg_idx], field, guardrail_response)
+                if rewritten is not None:
+                    messages[msg_idx][field] = rewritten
 
-            elif isinstance(content, list) and content_idx_optional is not None:
+            elif isinstance(content, list):
                 # Replace specific text item in list content
-                if isinstance(messages[msg_idx]["content"][content_idx_optional], dict):
-                    messages[msg_idx]["content"][content_idx_optional]["text"] = guardrail_response
+                if isinstance(messages[msg_idx][field][content_idx_optional], dict):
+                    messages[msg_idx][field][content_idx_optional]["text"] = guardrail_response
 
     async def process_output_response(
         self,

litellm/llms/openai/responses/guardrail_translation/tool_merge.py

@@ -6,6 +6,7 @@ from typing import Final, TypeAlias
 from pydantic import BaseModel, TypeAdapter, ValidationError
 
 from litellm._logging import verbose_logger
+from litellm.llms.base_llm.guardrail_translation.utils import merge_returned_tools_into_request_tools, openai_tool_name
 from litellm.responses.litellm_completion_transformation.custom_tools import custom_tool_grammar_suffix
 from litellm.responses.litellm_completion_transformation.transformation import (
     NAMESPACE_DESCRIPTION_SEPARATOR,
@@ -183,3 +184,18 @@ def merge_guardrailed_tools(
         tuple(tool for key, tool in zip(guardrailed_keys, guardrailed) if key not in owned_keys)
     )
     return tuple(chain(merged_originals, appended))
+
+
+def append_returned_tools(
+    original_tools: Sequence[Tool],
+    flattened_groups: Sequence[Sequence[Tool]],
+    returned_tools: Iterable[object],
+) -> tuple[Tool, ...]:
+    request_chat_tools: Final = tuple(chain.from_iterable(flattened_groups))
+    additions: Final = merge_returned_tools_into_request_tools(
+        request_tools=request_chat_tools, returned_tools=_validated_tools(returned_tools), tool_name=openai_tool_name
+    )[len(request_chat_tools) :]
+    return (
+        *original_tools,
+        *LiteLLMCompletionResponsesConfig.transform_chat_completion_tool_params_to_responses_api_tools(additions),
+    )

litellm/llms/openai/responses/transformation.py

@@ -66,6 +66,29 @@ _PROVIDERS_VALIDATING_TOOL_CALL_ITEM_IDS: Final = frozenset({LlmProviders.AZURE,
 _PROVIDERS_REPLAYING_ONLY_THEIR_OWN_REASONING: Final = _PROVIDERS_VALIDATING_TOOL_CALL_ITEM_IDS
 _ROLES_READING_PROMPT_CACHE_BREAKPOINTS: Final = frozenset({"system", "developer", "user"})
 _PART_TYPES_READING_PROMPT_CACHE_BREAKPOINTS: Final = frozenset({"input_text", "input_image", "input_file"})
+_COMPACT_TRACKING_PARAMS: Final = frozenset({"metadata", "user"})
+_COMPACT_REJECTED_PARAMS: Final = frozenset(
+    {
+        "background",
+        "context_management",
+        "max_output_tokens",
+        "max_tool_calls",
+        "moderation",
+        "partial_images",
+        "prompt",
+        "safety_identifier",
+        "store",
+        "stream_options",
+        "temperature",
+        "tool_choice",
+        "top_logprobs",
+        "top_p",
+        "truncation",
+    }
+)
+_PROVIDERS_WITH_STRICT_COMPACT_PARAMS: Final = frozenset(
+    {LlmProviders.OPENAI, LlmProviders.AZURE, LlmProviders.AZURE_AI}
+)
 _INPUT_ITEM: Final[TypeAdapter[Mapping[str, object]]] = TypeAdapter(Mapping[str, object])
 _CONTENT_PARTS: Final[TypeAdapter[Sequence[Mapping[str, object]]]] = TypeAdapter(Sequence[Mapping[str, object]])
 
@@ -85,7 +108,7 @@ class _DropParamsSetting(BaseModel):
     drop_params: object = None
 
 
-def _drops_params(litellm_params: GenericLiteLLMParams) -> bool:
+def drops_params(litellm_params: GenericLiteLLMParams) -> bool:
     setting: Final = _DropParamsSetting.model_validate(litellm_params)
     return litellm.drop_params is True or normalize_drop_params(setting.drop_params) is True
 
@@ -518,7 +541,9 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         litellm_params: GenericLiteLLMParams,
         prompt_cache_options: object = None,
     ) -> tuple[str | ResponseInputParam, Sequence[ALL_RESPONSES_API_TOOL_PARAMS] | None]:
-        validated_input: Final = self._validate_input_param(self._drop_bridge_minted_reasoning_items(input))
+        validated_input: Final = self._validate_input_param(
+            self._drop_bridge_minted_reasoning_items(input), drops_set_created_by=drops_params(litellm_params)
+        )
         marked_input: Final = (
             self._input_with_prompt_cache_breakpoints(validated_input)
             if self._reads_prompt_cache_breakpoints(
@@ -559,7 +584,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         if tools is None or not any(is_anthropic_tool_search_chat_tool(tool) for tool in tools):
             return tools
         callers_own_search: Final = any(is_openai_tool_search_tool(tool) for tool in tools)
-        if callers_own_search or not _drops_params(litellm_params) or _runs_tool_search(model, tools):
+        if callers_own_search or not drops_params(litellm_params) or _runs_tool_search(model, tools):
             return tools_with_anthropic_tool_search_mapped(tools)
         verbose_logger.warning(
             "Dropping tool search under drop_params: OpenAI model '%s' cannot run it or no tool is deferred, so every "
@@ -752,7 +777,9 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         ]
         return tools if all(new is old for new, old in zip(sanitized, tools, strict=True)) else sanitized
 
-    def _validate_input_param(self, input: str | ResponseInputParam) -> str | ResponseInputParam:
+    def _validate_input_param(
+        self, input: str | ResponseInputParam, *, drops_set_created_by: bool = False
+    ) -> str | ResponseInputParam:
         """
         Ensure all input fields if pydantic are converted to dict
 
@@ -764,18 +791,20 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             for item in input:
                 # if it's pydantic, convert to dict
                 if isinstance(item, BaseModel):
-                    validated_input.append(item.model_dump(exclude_none=True))
+                    validated_input.append(
+                        replayable_responses_item(
+                            item.model_dump(exclude_none=True), drops_set_created_by=drops_set_created_by
+                        )
+                    )
                 elif isinstance(item, dict):
-                    # Handle reasoning items specifically to filter out status=None
-                    if item.get("type") == "reasoning":
-                        verbose_logger.debug("Handling reasoning item: %s", item)
-                        # Type assertion since we know it's a dict at this point
-                        dict_item = cast(dict[str, object], item)
-                        filtered_item = self._handle_reasoning_item(dict_item)
-                    else:
-                        # For other dict items, just pass through
-                        filtered_item = cast(dict[str, object], item)
-                    validated_input.append(filtered_item)
+                    validated_input.append(
+                        replayable_responses_item(
+                            self._handle_reasoning_item(cast(dict[str, object], item))
+                            if item.get("type") == "reasoning"
+                            else item,
+                            drops_set_created_by=drops_set_created_by,
+                        )
+                    )
                 else:
                     validated_input.append(item)
             return validated_input
@@ -1182,6 +1211,17 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     #########################################################
     ########## COMPACT RESPONSE API TRANSFORMATION ##########
     #########################################################
+    def compact_url(self, api_base: str) -> str:
+        parsed_url: Final = httpx.URL(api_base)
+        return str(parsed_url.copy_with(path=parsed_url.path.rstrip("/") + "/compact"))
+
+    def compact_dropped_params(self, litellm_params: GenericLiteLLMParams) -> frozenset[str]:
+        if self.custom_llm_provider not in _PROVIDERS_WITH_STRICT_COMPACT_PARAMS:
+            return frozenset()
+        if _drops_params(litellm_params):
+            return _COMPACT_TRACKING_PARAMS | _COMPACT_REJECTED_PARAMS
+        return _COMPACT_TRACKING_PARAMS
+
     def transform_compact_response_api_request(
         self,
         model: str,
@@ -1197,10 +1237,8 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         OpenAI API expects the following request
         - POST /v1/responses/compact
         """
-        # Preserve query params (e.g., api-version) while appending /compact.
-        parsed_url: Final = httpx.URL(api_base)
-        compact_path: Final = parsed_url.path.rstrip("/") + "/compact"
-        url: Final = str(parsed_url.copy_with(path=compact_path))
+        url: Final = self.compact_url(api_base)
+        dropped_params: Final = self.compact_dropped_params(litellm_params)
 
         request_params: Final = cast(  # cast-ok: the base signature leaves the params bag untyped
             "Mapping[str, object]", response_api_optional_request_params
@@ -1214,14 +1252,11 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         )
         if sanitized_tools is not None:
             response_api_optional_request_params["tools"] = sanitized_tools
-        data: Final = cast(  # cast-ok: typed dict spread is inferred as Unknown by basedpyright
-            dict[str, object],
-            {
-                "model": model,
-                "input": replay_safe_input,
-                **response_api_optional_request_params,
-            },
-        )
+        data: Final = {
+            "model": model,
+            "input": replay_safe_input,
+            **{key: value for key, value in request_params.items() if key not in dropped_params},
+        }
 
         return url, data
 

litellm/llms/openai_like/dynamic_config.py

@@ -55,6 +55,15 @@ def create_config_class(provider: SimpleProviderConfig):
             else:
                 return super()._transform_messages(messages=messages, model=model, is_async=False)
 
+        def _should_preserve_cache_control_for_endpoint(
+            self,
+            custom_llm_provider: str | None,
+            api_base: str | None,
+        ) -> bool:
+            return provider.special_handling.get(
+                "preserve_cache_control", False
+            ) or super()._should_preserve_cache_control_for_endpoint(custom_llm_provider, api_base)
+
         def _get_openai_compatible_provider_info(
             self, api_base: str | None, api_key: str | None
         ) -> tuple[str | None, str | None]:
@@ -177,7 +186,7 @@ def create_config_class(provider: SimpleProviderConfig):
     return JSONProviderConfig
 
 
-_responses_config_cache: Final[dict] = {}
+_responses_config_cache: Final[dict[str, "type[OpenAILikeResponsesConfig]"]] = {}
 
 
 def create_responses_config_class(provider: SimpleProviderConfig) -> "type[OpenAILikeResponsesConfig]":

litellm/llms/openai_like/json_loader.py

@@ -3,6 +3,7 @@ JSON-based provider configuration loader for OpenAI-compatible providers.
 """
 
 import json
+from collections.abc import Mapping
 from pathlib import Path
 from typing import Final
 
@@ -12,6 +13,7 @@ from litellm._logging import verbose_logger
 from litellm.types.utils import ProviderSpecificModelInfo
 
 _PROVIDER_INFO_ADAPTER: Final = TypeAdapter(ProviderSpecificModelInfo)
+_SPECIAL_HANDLING_ADAPTER: Final = TypeAdapter(Mapping[str, bool])
 
 
 class SimpleProviderConfig:
@@ -20,12 +22,14 @@ class SimpleProviderConfig:
     def __init__(self, slug: str, data: dict):
         self.slug = slug
         self.base_url = data["base_url"]
-        self.api_key_env = data["api_key_env"]
-        self.api_base_env = data.get("api_base_env")
+        self.api_key_env: str = data["api_key_env"]
+        self.api_base_env: str | None = data.get("api_base_env")
         self.base_class = data.get("base_class", "openai_gpt")
         self.param_mappings = data.get("param_mappings", {})
         self.constraints = data.get("constraints", {})
-        self.special_handling = data.get("special_handling", {})
+        self.special_handling: Mapping[str, bool] = _SPECIAL_HANDLING_ADAPTER.validate_python(
+            data.get("special_handling", {})
+        )
         self.supported_endpoints = data.get("supported_endpoints", [])
         self.provider_info: ProviderSpecificModelInfo = (
             _PROVIDER_INFO_ADAPTER.validate_python(data["provider_info"]) if "provider_info" in data else {}

litellm/llms/openai_like/providers.json

@@ -269,5 +269,21 @@
       "/v1/chat/completions",
       "/v1/responses"
     ]
+  },
+  "streamlake": {
+    "base_url": "https://vanchin.streamlake.ai/api/gateway/v1/endpoints",
+    "api_key_env": "STREAMLAKE_API_KEY",
+    "api_base_env": "STREAMLAKE_API_BASE",
+    "param_mappings": {
+      "max_tokens": "max_completion_tokens"
+    },
+    "provider_info": {
+      "supports_function_calling": true,
+      "supports_reasoning": true
+    },
+    "special_handling": {
+      "preserve_cache_control": true
+    },
+    "supported_endpoints": ["/v1/chat/completions"]
   }
 }

litellm/llms/perplexity/responses/transformation.py

@@ -17,7 +17,7 @@ from pydantic import TypeAdapter
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
-from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
+from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig, drops_params
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import ResponseInputParam, ResponsesAPIResponse
 from litellm.types.router import GenericLiteLLMParams
@@ -84,7 +84,7 @@ class PerplexityResponsesConfig(OpenAIResponsesAPIConfig):
         """Handle preset/ model prefix: send as {"preset": name} instead of {"model": name}."""
         input = self._ensure_message_type(input)
         if model.startswith("preset/"):
-            input = self._validate_input_param(input)
+            input = self._validate_input_param(input, drops_set_created_by=drops_params(litellm_params))
             data: Final[dict] = {
                 "preset": model[len("preset/") :],
                 "input": input,

litellm/model_prices_and_context_window_backup.json

@@ -41939,6 +41939,120 @@
         "supports_system_messages": true,
         "supports_native_structured_output": true
     },
+    "bedrock/ap-southeast-2/mistral.voxtral-mini-3b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 4.12e-08,
+        "input_cost_per_token_batches": 2.06e-08,
+        "input_cost_per_token_flex": 2.06e-08,
+        "input_cost_per_token_priority": 7.21e-08,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 4.12e-08,
+        "output_cost_per_token_batches": 2.06e-08,
+        "output_cost_per_token_flex": 2.06e-08,
+        "output_cost_per_token_priority": 7.21e-08,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-southeast-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/ap-southeast-2/mistral.voxtral-small-24b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 1.03e-07,
+        "input_cost_per_token_batches": 5.15e-08,
+        "input_cost_per_token_flex": 5.15e-08,
+        "input_cost_per_token_priority": 1.8025e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 3.09e-07,
+        "output_cost_per_token_batches": 1.545e-07,
+        "output_cost_per_token_flex": 1.545e-07,
+        "output_cost_per_token_priority": 5.4075e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-southeast-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/ap-northeast-1/mistral.voxtral-mini-3b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 5e-08,
+        "input_cost_per_token_batches": 3e-08,
+        "input_cost_per_token_flex": 3e-08,
+        "input_cost_per_token_priority": 9e-08,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 5e-08,
+        "output_cost_per_token_batches": 3e-08,
+        "output_cost_per_token_flex": 3e-08,
+        "output_cost_per_token_priority": 9e-08,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-northeast-1/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/ap-northeast-1/mistral.voxtral-small-24b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 1.2e-07,
+        "input_cost_per_token_batches": 6e-08,
+        "input_cost_per_token_flex": 6e-08,
+        "input_cost_per_token_priority": 2.1e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 3.6e-07,
+        "output_cost_per_token_batches": 1.8e-07,
+        "output_cost_per_token_flex": 1.8e-07,
+        "output_cost_per_token_priority": 6.3e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-northeast-1/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/eu-west-2/mistral.voxtral-mini-3b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 6e-08,
+        "input_cost_per_token_batches": 3e-08,
+        "input_cost_per_token_flex": 3e-08,
+        "input_cost_per_token_priority": 1.1e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 6e-08,
+        "output_cost_per_token_batches": 3e-08,
+        "output_cost_per_token_flex": 3e-08,
+        "output_cost_per_token_priority": 1.1e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/eu-west-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/eu-west-2/mistral.voxtral-small-24b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 1.6e-07,
+        "input_cost_per_token_batches": 8e-08,
+        "input_cost_per_token_flex": 8e-08,
+        "input_cost_per_token_priority": 2.8e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 4.7e-07,
+        "output_cost_per_token_batches": 2.3e-07,
+        "output_cost_per_token_flex": 2.3e-07,
+        "output_cost_per_token_priority": 8.2e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/eu-west-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
     "mistral/codestral-2508": {
         "cache_read_input_token_cost": 3e-08,
         "input_cost_per_token": 3e-07,
@@ -61275,6 +61389,359 @@
         "supports_tool_choice": true,
         "supports_vision": true
     },
+    "streamlake/deepseek-v4-flash-0731": {
+        "cache_read_input_token_cost": 1.4e-08,
+        "input_cost_per_token": 4.4e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 1.32e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/DeepSeek-V4-Pro-0813": {
+        "cache_read_input_token_cost": 4.4e-08,
+        "input_cost_per_token": 1.32e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048566,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 3.96e-06,
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/DeepSeek-V3": {
+        "cache_read_input_token_cost": 1.14e-07,
+        "input_cost_per_token": 2.86e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 56000,
+        "max_output_tokens": 8000,
+        "mode": "chat",
+        "output_cost_per_token": 1.143e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": false,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/GLM-5.3": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_assistant_prefill": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/GLM-5.3-Flash": {
+        "cache_read_input_token_cost": 3e-08,
+        "input_cost_per_token": 1.5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 5e-07,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_vision": true,
+        "supports_assistant_prefill": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/Kimi-K2-Instruct-0905": {
+        "cache_read_input_token_cost": 1.43e-07,
+        "input_cost_per_token": 5.71e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 224000,
+        "max_output_tokens": 32000,
+        "mode": "chat",
+        "output_cost_per_token": 2.286e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": false,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-r1": {
+        "cache_read_input_token_cost": 2.29e-07,
+        "input_cost_per_token": 5.71e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 98304,
+        "max_output_tokens": 16000,
+        "mode": "chat",
+        "output_cost_per_token": 2.286e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v3.1": {
+        "cache_read_input_token_cost": 2.29e-07,
+        "input_cost_per_token": 5.71e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 98304,
+        "max_output_tokens": 32000,
+        "mode": "chat",
+        "output_cost_per_token": 1.714e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4-flash": {
+        "cache_read_input_token_cost": 2.8e-08,
+        "input_cost_per_token": 1.4e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 2.8e-07,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4-pro": {
+        "cache_read_input_token_cost": 1.45e-07,
+        "input_cost_per_token": 1.74e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 3.48e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4.1-flash-1017": {
+        "cache_read_input_token_cost": 6e-09,
+        "input_cost_per_token": 3e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 1.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4.1-flash-260910": {
+        "cache_read_input_token_cost": 6e-09,
+        "input_cost_per_token": 3e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048568,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 1.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/glm-5": {
+        "cache_read_input_token_cost": 2e-07,
+        "input_cost_per_token": 1e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 202752,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/glm-5.1": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 202752,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/glm-5.2": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048566,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/kimi-k2.6": {
+        "cache_read_input_token_cost": 1.6e-07,
+        "input_cost_per_token": 9.5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 262144,
+        "max_output_tokens": 256000,
+        "mode": "chat",
+        "output_cost_per_token": 4e-06,
+        "supports_function_calling": true,
+        "supports_vision": true,
+        "supports_reasoning": false,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/kimi-k2.7-code": {
+        "cache_read_input_token_cost": 1.9e-07,
+        "input_cost_per_token": 9.5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 262144,
+        "max_output_tokens": 32000,
+        "mode": "chat",
+        "output_cost_per_token": 4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/mimo-v2.6-flash": {
+        "cache_read_input_token_cost": 2.8e-09,
+        "input_cost_per_token": 1.4e-07,
+        "litellm_provider": "streamlake",
+        "mode": "chat",
+        "output_cost_per_token": 2.8e-07,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/mimo-v2.6-pro": {
+        "cache_read_input_token_cost": 3.6e-09,
+        "input_cost_per_token": 4.35e-07,
+        "litellm_provider": "streamlake",
+        "mode": "chat",
+        "output_cost_per_token": 8.7e-07,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/minimax-m2.5": {
+        "cache_creation_input_token_cost": 3.75e-07,
+        "cache_read_input_token_cost": 3e-08,
+        "input_cost_per_token": 3e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 196608,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 1.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/minimax-m3": {
+        "cache_read_input_token_cost": 1.2e-07,
+        "cache_read_input_token_cost_above_512k_tokens": 2.4e-07,
+        "input_cost_per_token": 6e-07,
+        "input_cost_per_token_above_512k_tokens": 1.2e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 512000,
+        "mode": "chat",
+        "output_cost_per_token": 2.4e-06,
+        "output_cost_per_token_above_512k_tokens": 4.8e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.5-397b-a17b": {
+        "cache_read_input_token_cost": 1.2e-07,
+        "input_cost_per_token": 6e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 258048,
+        "max_output_tokens": 64000,
+        "mode": "chat",
+        "output_cost_per_token": 3.6e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.6-27b": {
+        "input_cost_per_token": 6e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 229376,
+        "max_output_tokens": 64000,
+        "mode": "chat",
+        "output_cost_per_token": 3.6e-06,
+        "supports_function_calling": true,
+        "supports_vision": true,
+        "supports_reasoning": false,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.6-35b-a3b": {
+        "input_cost_per_token": 3.75e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 229376,
+        "max_output_tokens": 64000,
+        "mode": "chat",
+        "output_cost_per_token": 2.25e-06,
+        "supports_function_calling": true,
+        "supports_vision": true,
+        "supports_reasoning": false,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.8-2.4t-a95b": {
+        "cache_read_input_token_cost": 2.5e-07,
+        "input_cost_per_token": 2e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 983616,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 6e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "cache_creation_input_token_cost": 2.5e-06,
+        "explicit_cache_read_input_token_cost": 1.7e-07,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.8-27b": {
+        "cache_read_input_token_cost": 1e-07,
+        "input_cost_per_token": 5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 983616,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "cache_creation_input_token_cost": 6.25e-07,
+        "explicit_cache_read_input_token_cost": 5e-08,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
     "deepseek-v4-flash": {
         "cache_creation_input_token_cost": 0.0,
         "cache_read_input_token_cost": 6e-09,
@@ -62590,6 +63057,14 @@
                     "litellm_provider": "anthropic"
                 }
             },
+            {
+                "name": "typesafe-jev-ids",
+                "pattern": "^jev-(?:latest|preview|1\\.13\\.0|1\\.14\\.0)$",
+                "description": "A bare Jev id that has a typesafe/ price entry (jev-latest, jev-preview, jev-1.13.0, jev-1.14.0), the name the /typesafe pass-through forwards to TypeSafe. Listed by name, matched case-insensitively like every rule here, so foo/jev-latest or a model group named like a Jev id TypeSafe does not serve, such as jev-1.0 or jev-internal-router, never resolves to typesafe. A new Jev version goes here together with its price entry. Provider inference feeds the proxy's provider/* wildcard access checks, so this rule is what lets typesafe/* on a key, team or project allowlist cover a jev-latest request.",
+                "model_info": {
+                    "litellm_provider": "typesafe"
+                }
+            },
             {
                 "name": "claude-family-baseline",
                 "pattern": "claude-[a-z]+-\\d+(?:[-.]\\d+)?(?:-\\d{8})?",

litellm/provider_endpoints_support_backup.json

@@ -3785,6 +3785,32 @@
         "generateContent": true
       }
     },
+    "streamlake": {
+      "display_name": "StreamLake (`streamlake`)",
+      "url": "https://docs.litellm.ai/docs/providers/openai_compatible",
+      "endpoints": {
+        "chat_completions": true,
+        "responses": true,
+        "text_completion": true,
+        "embeddings": false,
+        "image_generations": false,
+        "image_edits": false,
+        "audio_transcriptions": false,
+        "audio_speech": false,
+        "rerank": false,
+        "ocr": false,
+        "search": false,
+        "moderations": false,
+        "realtime": false,
+        "video_generations": false,
+        "batches": false,
+        "messages": true,
+        "interactions": true,
+        "systemone": false,
+        "a2a": true,
+        "generateContent": true
+      }
+    },
     "tencent": {
       "display_name": "Tencent TokenHub (`tencent`)",
       "url": "https://docs.litellm.ai/docs/providers/tencent",

litellm/proxy/_types.py

@@ -3816,6 +3816,16 @@ class ManagementEndpointLoggingPayload(LiteLLMPydanticObjectBase):
     end_time: datetime | None = None
 
 
+def proxy_exception_status_code(message: str, code: str) -> str:
+    # Litellm router.py returns "No healthy deployment available" when there are no deployments available
+    # Should map to 429 errors https://github.com/BerriAI/litellm/issues/2487
+    if "No healthy deployment available" in message or "No deployments available" in message:
+        return "429"
+    if RouterErrors.no_deployments_with_tag_routing.value in message:
+        return "401"
+    return code
+
+
 class ProxyException(Exception):
     # NOTE: DO NOT MODIFY THIS
     # This is used to map exactly to OPENAI Exceptions
@@ -3844,13 +3854,7 @@ class ProxyException(Exception):
                     headers[k] = str(v)
         self.headers = headers or {}
         self.provider_specific_fields = provider_specific_fields
-        # rules for proxyExceptions
-        # Litellm router.py returns "No healthy deployment available" when there are no deployments available
-        # Should map to 429 errors https://github.com/BerriAI/litellm/issues/2487
-        if "No healthy deployment available" in self.message or "No deployments available" in self.message:
-            self.code = "429"
-        elif RouterErrors.no_deployments_with_tag_routing.value in self.message:
-            self.code = "401"
+        self.code = proxy_exception_status_code(self.message, self.code)
 
     def to_dict(self) -> dict:
         """Converts the ProxyException instance to a dictionary."""

litellm/proxy/auth/auth_checks.py

@@ -1063,6 +1063,15 @@ def is_mcp_discovery_request(route: str, request_body: Mapping[str, object]) ->
     return isinstance(method, str) and method in MCP_ZERO_SPEND_JSONRPC_METHODS
 
 
+class ModelMaxBudgetHolder(Protocol):
+    @property
+    def model_max_budget(self) -> Mapping[str, object] | None: ...
+
+
+def model_max_budget_of(holder: ModelMaxBudgetHolder) -> Mapping[str, object] | None:
+    return holder.model_max_budget
+
+
 def route_skips_budget_checks(route: str) -> bool:
     return route not in BUDGET_ENFORCED_SIDE_EFFECT_ROUTES and (
         route in MODEL_DISCOVERY_ROUTES or not RouteChecks.is_llm_api_route(route=route)
@@ -8075,7 +8084,6 @@ async def vector_store_access_check(
     rag_ids: Final = (rag_vector_store_id,) if rag_vector_store_id is not None else ()
     vector_store_ids_to_run: Final = tuple(dict.fromkeys((*registry_ids, *rag_ids)))
     if not vector_store_ids_to_run:
-        verbose_proxy_logger.debug("Vector store to run not found, skipping vector store access check")
         return True
 
     #########################################################

litellm/proxy/auth/user_api_key_auth.py

@@ -40,6 +40,7 @@ from litellm.integrations.otel.model.config import is_otel_v2_enabled
 from litellm.integrations.otel.runtime import phase_event, phase_span, seed_request_identity
 from litellm.litellm_core_utils.dd_tracing import tracer
 from litellm.litellm_core_utils.dot_notation_indexing import get_nested_value
+from litellm.models.end_user import LiteLLM_EndUserTable
 from litellm.proxy._types import *
 from litellm.proxy.agent_endpoints.auth.agent_caller import agent_caller_from_headers
 from litellm.proxy.auth.auth_checks import (  # noqa: F401  # legacy module exports
@@ -85,6 +86,7 @@ from litellm.proxy.auth.auth_checks import (  # noqa: F401  # legacy module expo
     jwt_key_mapping_cache_key,
     key_access_group_grants_model,
     key_model_aliases_for_auth_check,
+    model_max_budget_of,
     resolve_and_validate_end_user_id,
     resolve_default_end_user_budget,
     route_skips_budget_checks,
@@ -384,6 +386,16 @@ class _TeamModelBudgetLimiter(Protocol):
     ) -> bool: ...
 
 
+class _EndUserModelBudgetLimiter(Protocol):
+    async def is_end_user_within_model_budget(
+        self,
+        end_user_id: str,
+        end_user_model_max_budget: Mapping[str, object],
+        model: str,
+        team_id: str | None = None,
+    ) -> bool: ...
+
+
 class _EveryModelBudgetLimiter(Protocol):
     async def is_key_within_model_budget(self, user_api_key_dict: UserAPIKeyAuth, model: str) -> bool: ...
 
@@ -530,6 +542,49 @@ async def _check_team_model_budget(
         )
 
 
+async def _end_user_model_max_budget(
+    end_user_id: str | None,
+    end_user_object: LiteLLM_EndUserTable | None,
+    key_end_user_budget_id: str | None,
+    prisma_client: PrismaClient | None,
+    user_api_key_cache: UserApiKeyCache,
+    parent_otel_span: Span | None,
+) -> Mapping[str, object] | None:
+    budget_table: Final = (
+        end_user_object.litellm_budget_table
+        if end_user_object is not None
+        else (
+            await resolve_default_end_user_budget(
+                prisma_client=prisma_client,
+                user_api_key_cache=user_api_key_cache,
+                key_end_user_budget_id=key_end_user_budget_id,
+                parent_otel_span=parent_otel_span,
+            )
+            if end_user_id and prisma_client is not None
+            else None
+        )
+    )
+    return model_max_budget_of(budget_table) if budget_table is not None else None
+
+
+async def _check_end_user_model_budget(
+    end_user_id: str | None,
+    end_user_model_max_budget: Mapping[str, object] | None,
+    model_max_budget_limiter: _EndUserModelBudgetLimiter,
+    models: Sequence[str],
+    team_id: str | None,
+) -> None:
+    if not end_user_id or not end_user_model_max_budget:
+        return
+    for model_name in models:
+        await model_max_budget_limiter.is_end_user_within_model_budget(
+            end_user_id=end_user_id,
+            end_user_model_max_budget=end_user_model_max_budget,
+            model=model_name,
+            team_id=team_id,
+        )
+
+
 async def _is_within_every_model_budget(
     valid_token: UserAPIKeyAuth,
     model_max_budget_limiter: _EveryModelBudgetLimiter,
@@ -2094,6 +2149,10 @@ async def _resolve_object_permission_for_unresolvable_team(
     )
 
 
+def _custom_auth_token_for_this_request(response: object) -> UserAPIKeyAuth:
+    return UserAPIKeyAuth.model_validate(response).model_copy()
+
+
 async def user_api_key_auth_builder(
     request: Request,
     api_key: str,
@@ -2172,7 +2231,7 @@ async def user_api_key_auth_builder(
                     request=request, api_key=api_key, user_custom_auth=user_custom_auth
                 )
             if response is not None and isinstance(response, UserAPIKeyAuth):
-                validated = UserAPIKeyAuth.model_validate(response)
+                validated = _custom_auth_token_for_this_request(response)
                 if getattr(litellm, "enable_post_custom_auth_checks", False):
                     validated = await _run_post_custom_auth_checks(
                         valid_token=validated,
@@ -2188,7 +2247,7 @@ async def user_api_key_auth_builder(
                 custom_auth_api_key = True
         elif user_custom_auth is not None:
             response = await user_custom_auth(request=request, api_key=api_key)
-            validated = UserAPIKeyAuth.model_validate(response)
+            validated = _custom_auth_token_for_this_request(response)
             if getattr(litellm, "enable_post_custom_auth_checks", False):
                 validated = await _run_post_custom_auth_checks(
                     valid_token=validated,
@@ -3129,7 +3188,7 @@ async def validate_resolved_virtual_key(  # noqa: C901  # Preserve ordering of e
                         )
 
                     # Recompute after a potential budget-fallback rewrite so
-                    # the end-user check below validates the final model
+                    # the user check below validates the final model
                     current_model = _get_model_from_request_context(
                         request_data=request_data,
                         route=route,
@@ -3147,23 +3206,6 @@ async def validate_resolved_virtual_key(  # noqa: C901  # Preserve ordering of e
                         models=current_models,
                     )
 
-                # Check 5b. End-user model max budget
-                end_user_mmb: Final = valid_token.end_user_model_max_budget
-                if (
-                    end_user_mmb is not None
-                    and isinstance(end_user_mmb, dict)
-                    and len(end_user_mmb) > 0
-                    and current_models
-                    and valid_token.end_user_id is not None
-                ):
-                    for model_name in current_models:
-                        await model_max_budget_limiter.is_end_user_within_model_budget(
-                            end_user_id=valid_token.end_user_id,
-                            end_user_model_max_budget=end_user_mmb,
-                            model=model_name,
-                            team_id=valid_token.team_id,
-                        )
-
         # Check 6: Additional Common Checks across jwt + key auth
         if valid_token.team_id is not None:
             try:
@@ -3733,9 +3775,7 @@ async def run_centralized_common_checks(
     project_object: Final[LiteLLM_ProjectTableCachedObj | None] = (
         None if isinstance(project_result, BaseException) else project_result
     )
-    end_user_object: Final[LiteLLM_EndUserTable | None] = (
-        None if isinstance(end_user_result, BaseException) else end_user_result
-    )
+    end_user_object: Final = end_user_result if isinstance(end_user_result, LiteLLM_EndUserTable) else None
     global_proxy_spend: float | None = None if isinstance(global_spend_result, BaseException) else global_spend_result
     carry_team_and_user_budget_state(
         valid_token=user_api_key_auth_obj,
@@ -3795,6 +3835,21 @@ async def run_centralized_common_checks(
             parent_otel_span=parent_otel_span,
             keep_token_limits=user_custom_auth is not None,
         )
+    end_user_model_max_budget: Final = await _end_user_model_max_budget(
+        end_user_id=end_user_id,
+        end_user_object=end_user_object,
+        key_end_user_budget_id=key_end_user_budget_id,
+        prisma_client=prisma_client,
+        user_api_key_cache=user_api_key_cache,
+        parent_otel_span=parent_otel_span,
+    )
+    if end_user_id and (
+        not user_api_key_auth_obj.authenticated_by_custom_auth
+        or user_api_key_auth_obj.end_user_model_max_budget is None
+    ):
+        user_api_key_auth_obj.end_user_model_max_budget = (  # rebind-ok: the spend hook charges from this token
+            dict(end_user_model_max_budget) if end_user_model_max_budget else None
+        )
 
     skip_budget_checks: Final = _should_skip_budget_checks(
         request_data=request_data,
@@ -3887,6 +3942,21 @@ async def run_centralized_common_checks(
                     user_api_key_cache=user_api_key_cache,
                     proxy_logging_obj=proxy_logging_obj,
                 )
+                await _check_end_user_model_budget(
+                    end_user_id=end_user_id,
+                    end_user_model_max_budget=user_api_key_auth_obj.end_user_model_max_budget,
+                    model_max_budget_limiter=model_max_budget_limiter,
+                    team_id=user_api_key_auth_obj.team_id,
+                    models=_get_model_names_for_budget_checks(
+                        model=_get_model_from_request_context(
+                            request_data=request_data,
+                            route=route,
+                            request=request,
+                            llm_router=llm_router,
+                            team_id=user_api_key_auth_obj.team_id,
+                        )
+                    ),
+                )
             if sub_call and RouteChecks.is_llm_api_route(route=route):
                 await virtual_key_max_budget_check(
                     valid_token=user_api_key_auth_obj,

litellm/proxy/client/cli/commands/agents.py

@@ -5,16 +5,18 @@ import shutil
 import subprocess
 import sys
 import tempfile
-from collections.abc import Callable, Mapping, Sequence
+from collections import Counter
+from collections.abc import Callable, Hashable, Iterable, Mapping, Sequence
 from dataclasses import dataclass
 from itertools import chain
 from pathlib import Path
 from types import MappingProxyType
 from typing import Final, TypeAlias, TypeVar
+from urllib.parse import urlsplit
 
 import click
 import requests
-from pydantic import JsonValue, TypeAdapter, ValidationError
+from pydantic import ConfigDict, JsonValue, TypeAdapter, ValidationError
 
 from litellm.proxy.common_utils.codex_model_catalog import (
     CODEX_BASE_INSTRUCTIONS_PATH,
@@ -497,6 +499,113 @@ def _fetch_model_listing(
     return listing.data
 
 
+class _DeploymentParams(LiteLLMBaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    model: str
+    custom_llm_provider: str | None = None
+    api_base: str | None = None
+    litellm_credential_name: str | None = None
+    reasoning_effort: JsonValue = None
+
+
+class _DeploymentInfo(LiteLLMBaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    team_public_model_name: str | None = None
+    blocked: bool | None = None
+
+
+class _Deployment(LiteLLMBaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    model_name: str
+    litellm_params: _DeploymentParams
+    model_info: _DeploymentInfo | None = None
+
+
+class _DeploymentListing(LiteLLMBaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    data: tuple[_Deployment, ...]
+
+
+@dataclass(frozen=True, slots=True)
+class ModelUpstream:
+    """The OpenAI model every deployment behind a public name routes to, and the reasoning effort they all pin."""
+
+    model: str
+    reasoning_effort: str | None = None
+
+
+_NO_UPSTREAMS: Final[Mapping[str, ModelUpstream]] = MappingProxyType({})
+_Pinned: Final = TypeVar("_Pinned", bound=Hashable)
+
+
+def _public_name(deployment: _Deployment) -> str:
+    team_name: Final = deployment.model_info.team_public_model_name if deployment.model_info is not None else None
+    return team_name or deployment.model_name
+
+
+def _is_paused(deployment: _Deployment) -> bool:
+    return deployment.model_info is not None and deployment.model_info.blocked is True
+
+
+def _is_openai_host(params: _DeploymentParams) -> bool:
+    if params.api_base is None:
+        return not params.litellm_credential_name
+    host: Final = urlsplit(params.api_base).hostname or ""
+    return host == "api.openai.com" or host.endswith(".api.openai.com")
+
+
+def _openai_model(params: _DeploymentParams) -> str | None:
+    if params.custom_llm_provider not in (None, "openai") or not _is_openai_host(params):
+        return None
+    provider, slash, name = params.model.partition("/")
+    if not slash:
+        return params.model
+    return name if provider == "openai" else None
+
+
+def _pinned_effort(params: _DeploymentParams) -> str | None:
+    return params.reasoning_effort if isinstance(params.reasoning_effort, str) else None
+
+
+def _shared_by_every_deployment(values: Iterable[tuple[str, _Pinned]]) -> Mapping[str, _Pinned]:
+    distinct: Final = frozenset(values)
+    values_per_name: Final = Counter(name for name, _ in distinct)
+    return MappingProxyType({name: value for name, value in distinct if values_per_name[name] == 1})
+
+
+def _upstreams_by_public_name(deployments: Sequence[_Deployment]) -> Mapping[str, ModelUpstream]:
+    live: Final = tuple(d for d in deployments if not _is_paused(d))
+    models: Final = _shared_by_every_deployment((_public_name(d), _openai_model(d.litellm_params)) for d in live)
+    efforts: Final = _shared_by_every_deployment((_public_name(d), _pinned_effort(d.litellm_params)) for d in live)
+    return MappingProxyType(
+        {name: ModelUpstream(model, efforts.get(name)) for name, model in models.items() if model is not None}
+    )
+
+
+def _fetch_model_upstreams(
+    base_url: str,
+    api_key: str,
+    *,
+    get: Callable[..., requests.Response],
+) -> Mapping[str, ModelUpstream]:
+    url: Final = base_url.rstrip("/") + "/v1/model/info"
+    try:
+        resp: Final = get(url, headers=MappingProxyType({"Authorization": f"Bearer {api_key}"}), timeout=10)
+    except requests.RequestException:
+        return _NO_UPSTREAMS
+    if resp.status_code != 200:
+        return _NO_UPSTREAMS
+    try:
+        listing: Final = _DeploymentListing.model_validate_json(resp.content)
+    except ValidationError:
+        return _NO_UPSTREAMS
+    return _upstreams_by_public_name(listing.data)
+
+
 class _OpenCodeLimit(LiteLLMBaseModel):
     context: int
     output: int
@@ -576,8 +685,21 @@ def opencode_model_sync_env(
     return MappingProxyType({OPENCODE_CONFIG_CONTENT_ENV: opencode_provider_config(base_url, listing)})
 
 
+def _codex_row(model: ListedModel, upstream: ModelUpstream | None) -> CodexCatalogRow:
+    return CodexCatalogRow(
+        id=model.id,
+        mode=model.mode,
+        max_input_tokens=model.max_input_tokens,
+        upstream_model=upstream.model if upstream is not None else None,
+        reasoning_effort=upstream.reasoning_effort if upstream is not None else None,
+    )
+
+
 def codex_model_catalog(
-    models: Sequence[ListedModel], stock: Sequence[CodexStockModel], instructions: str
+    models: Sequence[ListedModel],
+    stock: Sequence[CodexStockModel],
+    instructions: str,
+    upstreams: Mapping[str, ModelUpstream],
 ) -> str | None:
     """The `model_catalog_json` body listing the proxy's chat models, or None if there are none.
 
@@ -586,14 +708,23 @@ def codex_model_catalog(
     installed Codex knows keeps that Codex's own entry and the proxy only
     decides its place in the picker: the listing orders it, lists it even when
     Codex hides it or keeps it off the API, and keeps Codex's upgrade nudge only
-    when the model it points at is served too. A model Codex does not know gets the fallback
+    when the model it points at is served too. An alias of such a model, per
+    `upstreams`, takes that entry too, minus the defaults Codex picks for the
+    model: no service tier or upgrade nudge, the reasoning effort its
+    deployments pin, and a window no larger than its listed max_input_tokens.
+    A model Codex does not know gets the fallback
     entry, with the same base instructions Codex itself uses so the agent never
-    runs without a system prompt. The catalog is the one `/v1/models?client_version=...`
-    serves, built from the installed Codex's own entries instead of the proxy's vendored ones.
+    runs without a system prompt. Apart from aliases, the catalog is the one
+    `/v1/models?client_version=...` serves, built from the installed Codex's own
+    entries instead of the proxy's vendored ones.
     """
-    rows: Final = tuple(CodexCatalogRow(id=m.id, mode=m.mode, max_input_tokens=m.max_input_tokens) for m in models)
+    rows: Final = tuple(_codex_row(m, upstreams.get(m.id)) for m in models)
     body: Final = codex_models_response_json(
-        rows, stock=MappingProxyType({m.slug: m for m in stock}), instructions=instructions, byte_limit=None
+        rows,
+        stock=MappingProxyType({m.slug: m for m in stock}),
+        instructions=instructions,
+        byte_limit=None,
+        alias_defaults="deployment",
     )
     return body.json if body.listed else None
 
@@ -678,12 +809,14 @@ def codex_model_sync_args(
     Codex has no env or inline equivalent of OPENCODE_CONFIG_CONTENT: the catalog
     must be a file, so it is written under $CODEX_HOME (default ~/.codex) and
     atomically replaced on every launch. The Codex at `binary` first lists its
-    own models, so the ones the proxy serves keep that Codex's entries, and then
+    own models, so the ones the proxy serves keep that Codex's entries, an alias
+    included when every unpaused deployment /v1/model/info lists for it routes
+    to the same OpenAI model and Codex knows that model, and then
     reads the file back once before it is handed over. The key never lands in
     the file. A failed fetch, read, listing, write or read-back is reported
     rather than raised: Codex still launches with its built-in catalog and takes
     a proxy model by name via -m, and a rejected file stays on disk to be looked
-    at.
+    at. An unavailable /v1/model/info only leaves every alias on the fallback entry.
     """
     listing: Final = _fetch_model_listing(base_url, api_key, get=get)
     if isinstance(listing, ModelSyncSkipped):
@@ -700,7 +833,8 @@ def codex_model_sync_args(
     stock: Final = _stock_codex_models(binary, base_env, run=run)
     if isinstance(stock, ModelSyncSkipped):
         return stock
-    catalog: Final = codex_model_catalog(listing, stock, instructions)
+    upstreams: Final = _fetch_model_upstreams(base_url, api_key, get=get)
+    catalog: Final = codex_model_catalog(listing, stock, instructions, upstreams)
     if catalog is None:
         return ModelSyncSkipped(f"{base_url.rstrip('/')}/v1/models lists no chat models")
     try:

litellm/proxy/common_request_processing.py

@@ -2603,7 +2603,6 @@ class ProxyBaseLLMRequestProcessing:
         # have mutated `self.data` in place, and the audit-trail snapshot taken in
         # add_litellm_data_to_request predates that mutation.
         refresh_proxy_server_request_body_snapshot(self.data, guardrails_applied=True)
-        verbose_proxy_logger.debug("receiving data: %s", self.data)
 
         if "messages" in self.data and self.data["messages"]:
             logging_obj.update_messages(self.data["messages"])

litellm/proxy/common_utils/codex_model_catalog.py

@@ -24,7 +24,7 @@ from functools import cache
 from itertools import accumulate, pairwise
 from pathlib import Path
 from types import MappingProxyType, NoneType
-from typing import TYPE_CHECKING, Annotated, Final, Literal
+from typing import TYPE_CHECKING, Annotated, Final, Literal, TypeAlias
 
 from pydantic import ConfigDict, StringConstraints, TypeAdapter, ValidationError
 
@@ -110,6 +110,12 @@ class CodexStockUpgrade(LiteLLMBaseModel):
     model: str
 
 
+class CodexReasoningLevel(LiteLLMBaseModel):
+    model_config = ConfigDict(frozen=True, extra="allow")
+
+    effort: str
+
+
 class CodexStockModel(LiteLLMBaseModel):
     """One `ModelInfo` entry as Codex ships it or prints it from `codex debug models`.
 
@@ -128,6 +134,8 @@ class CodexStockModel(LiteLLMBaseModel):
     upgrade: CodexStockUpgrade | None = None
     service_tiers: tuple[CodexServiceTier, ...] = ()
     default_service_tier: str | None = None
+    supported_reasoning_levels: tuple[CodexReasoningLevel, ...] = ()
+    context_window: int | None = None
 
 
 class CodexStockCatalog(LiteLLMBaseModel):
@@ -135,19 +143,21 @@ class CodexStockCatalog(LiteLLMBaseModel):
 
 
 CodexCatalogEntry = CodexFallbackModel | CodexStockModel
+CodexAliasDefaults: TypeAlias = Literal["codex", "deployment"]
 
 
 @dataclass(frozen=True, slots=True)
 class CodexCatalogRow:
     """What the catalog needs to know about one listed model: its public id, the listing's mode and
-    input limit, the upstream model behind it, and the `model_info` values the operator set;
-    `service_tiers` holds one raw `model_info.service_tiers` value per deployment behind the id, None
-    where a deployment sets none."""
+    input limit, the upstream model behind it, the reasoning effort its deployments pin, and the
+    `model_info` values the operator set; `service_tiers` holds one raw `model_info.service_tiers`
+    value per deployment behind the id, None where a deployment sets none."""
 
     id: str
     mode: str | None = None
     max_input_tokens: int | None = None
     upstream_model: str | None = None
+    reasoning_effort: str | None = None
     display_name: str | None = None
     service_tiers: tuple[object, ...] = ()
 
@@ -254,12 +264,42 @@ def _stock_for(row: CodexCatalogRow, stock: Mapping[str, CodexStockModel]) -> Co
     return upstream if upstream is not None else stock.get(row.upstream_model.partition("/")[2])
 
 
+def _alias_context_window(row: CodexCatalogRow, stock_model: CodexStockModel) -> Mapping[str, int]:
+    if row.max_input_tokens is None or stock_model.context_window is None:
+        return MappingProxyType({})
+    return MappingProxyType({"context_window": min(stock_model.context_window, row.max_input_tokens)})
+
+
+def _alias_reasoning_level(row: CodexCatalogRow, stock_model: CodexStockModel) -> str | None:
+    efforts: Final = frozenset(level.effort for level in stock_model.supported_reasoning_levels)
+    return row.reasoning_effort if row.reasoning_effort in efforts else None
+
+
+def _deployment_defaults(row: CodexCatalogRow, stock_model: CodexStockModel) -> Mapping[str, object]:
+    """What an alias's deployments decide in place of the defaults Codex picks for the model they route to:
+    no service tier, verbosity, upgrade nudge or startup promo of Codex's own, so what the deployment
+    pins (or OpenAI's API default) stays in force, the pinned reasoning effort when Codex offers it and
+    none otherwise, and a window no larger than the alias's own limit."""
+    return MappingProxyType(
+        {
+            "service_tiers": (),
+            "default_service_tier": None,
+            "default_verbosity": None,
+            "upgrade": None,
+            "availability_nux": None,
+            "default_reasoning_level": _alias_reasoning_level(row, stock_model),
+            **_alias_context_window(row, stock_model),
+        }
+    )
+
+
 def _entry(
     index: int,
     row: CodexCatalogRow,
     stock: Mapping[str, CodexStockModel],
     served: frozenset[str],
     instructions: str,
+    alias_defaults: CodexAliasDefaults,
 ) -> CodexCatalogEntry:
     stock_model: Final = _stock_for(row, stock)
     stock_tiers: Final = MappingProxyType(
@@ -278,7 +318,8 @@ def _entry(
     upgrade: Final = (
         stock_model.upgrade if stock_model.upgrade is not None and stock_model.upgrade.model in served else None
     )
-    stock_name: Final = stock_model.display_name if row.id == stock_model.slug else row.id
+    is_alias: Final = row.id != stock_model.slug
+    stock_name: Final = row.id if is_alias else stock_model.display_name
     listing: Final = {
         "slug": row.id,
         "display_name": row.display_name if row.display_name is not None else stock_name,
@@ -286,6 +327,7 @@ def _entry(
         "visibility": "list",
         "supported_in_api": True,
         "upgrade": upgrade,
+        **(_deployment_defaults(row, stock_model) if is_alias and alias_defaults == "deployment" else {}),
     }
     if configured is None:
         return stock_model.model_copy(update=listing)
@@ -351,6 +393,7 @@ def codex_models_response_json(
     stock: Mapping[str, CodexStockModel] | None = None,
     instructions: str | None = None,
     byte_limit: int | None = CODEX_CATALOG_BYTE_LIMIT,
+    alias_defaults: CodexAliasDefaults = "codex",
 ) -> CodexCatalogBody:
     """Codex's `ModelsResponse` body for `rows`, in listing order, cut to Codex's byte limit.
 
@@ -359,13 +402,18 @@ def codex_models_response_json(
     byte limit silently and keeps its bundled one, so the body holds only the entries that fit,
     the ones offering a service tier taken first and the rest in listing order, an entry too large
     for the bytes left passed over, every kept entry at its listing position; `byte_limit=None`
-    keeps every entry.
+    keeps every entry. A row listed under another name than the Codex model its `upstream_model`
+    names (an alias) takes that model's entry; `alias_defaults="deployment"` drops the defaults
+    Codex picks for that model in favor of what the alias's deployments pin, while `"codex"` keeps
+    them, as Codex would use them for the model's own name.
     """
     known: Final = bundled_codex_models() if stock is None else stock
     prompt: Final = codex_base_instructions() if instructions is None else instructions
     drivable: Final = tuple(row for row in rows if _codex_can_drive(row))
     served: Final = frozenset(row.id for row in drivable)
-    entries: Final = tuple(_entry(index, row, known, served, prompt) for index, row in enumerate(drivable))
+    entries: Final = tuple(
+        _entry(index, row, known, served, prompt, alias_defaults) for index, row in enumerate(drivable)
+    )
     serialized: Final = tuple(_serialized(entry) for entry in entries)
     kept: Final = _kept_positions(entries, serialized, byte_limit)
     return CodexCatalogBody(

litellm/proxy/common_utils/http_parsing_utils.py

@@ -845,5 +845,3 @@ def _add_vector_store_id_from_path(request_data: dict, request: Request) -> None
             "populate_request_with_path_params: Updated request_data with vector_store_ids=%s",
             request_data.get("vector_store_ids"),
         )
-    else:
-        verbose_proxy_logger.debug("populate_request_with_path_params: No vector_store_id present in path=%s", path)

litellm/proxy/common_utils/openai_error_payload.py

@@ -2,14 +2,17 @@
 ``type`` is a required string and ``param`` is nullable, neither of which the literal
 string ``"None"`` satisfies."""
 
+import sys
 from collections.abc import Mapping
 from types import MappingProxyType
 from typing import Final
 
-from fastapi import status
+from fastapi import HTTPException, status
 
 from litellm.constants import STRINGIFIED_NONE
-from litellm.proxy._types import ProxyException
+from litellm.proxy._types import ProxyErrorTypes, ProxyException, proxy_exception_status_code
+from litellm.proxy.db.db_lookup_gate import DBLookupDeadlineExceeded
+from litellm.proxy.db.exception_handler import PrismaDBExceptionHandler, exception_chain
 
 LITELLM_CALL_ID_HEADER: Final = "x-litellm-call-id"
 
@@ -21,6 +24,19 @@ _OPENAI_ERROR_TYPE_BY_STATUS: Final[Mapping[int, str]] = MappingProxyType(
     }
 )
 
+_AUTH_FAILURE_STATUS_CODES: Final = frozenset({status.HTTP_401_UNAUTHORIZED, status.HTTP_403_FORBIDDEN})
+
+_FAILOVER_SQLSTATES: Final = frozenset({"25006", "57P01", "57P02", "57P03"})
+
+_FAILOVER_REFUSALS: Final = (
+    "is not currently accepting connections",
+    "the database system is starting up",
+    "the database system is shutting down",
+    "the database system is in recovery mode",
+    "the database system is not yet accepting connections",
+    "terminating connection due to administrator command",
+)
+
 
 def attribute_of(value: object, name: str, default: object = None) -> object:
     return getattr(value, name, default)
@@ -61,6 +77,82 @@ def openai_error_param(exc: object) -> str | None:
     return carried if isinstance(carried, str) and carried != STRINGIFIED_NONE else None
 
 
+def proxy_exception_for(exc: Exception, default_status_code: int) -> ProxyException:
+    match exc:
+        case ProxyException():
+            return exc
+        case _ if _is_database_outage(exc):
+            return ProxyException(
+                message=PrismaDBExceptionHandler.database_unavailable_message(exc),
+                type=ProxyErrorTypes.no_db_connection.value,
+                param=None,
+                code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            )
+        case HTTPException(status_code=status_code, detail=detail):
+            return _openai_shaped_proxy_exception(exc, str(detail), status_code)
+        case _:
+            return _openai_shaped_proxy_exception(
+                exc, str(exc) or type(exc).__name__, error_status_code(exc, default_status_code)
+            )
+
+
+def _is_database_outage(exc: Exception) -> bool:
+    return any(map(_is_database_outage_link, exception_chain(exc)))
+
+
+def _is_database_outage_link(exc: BaseException) -> bool:
+    match exc:
+        case DBLookupDeadlineExceeded():
+            return True
+        case Exception() if _prisma_is_loaded():
+            return _is_prisma_outage(exc)
+        case _:
+            return False
+
+
+def _prisma_is_loaded() -> bool:
+    return sys.modules.get("prisma") is not None
+
+
+def _is_prisma_outage(exc: Exception) -> bool:
+    if PrismaDBExceptionHandler.is_prisma_engine_internal_error(exc):
+        return True
+    if type(exc).__module__.partition(".")[0] != "prisma":
+        return False
+    return PrismaDBExceptionHandler.is_database_service_unavailable_error(exc) or _is_failover_refusal(exc)
+
+
+def _is_failover_refusal(exc: Exception) -> bool:
+    if PrismaDBExceptionHandler.is_read_only_transaction_error(exc):
+        return True
+    if PrismaDBExceptionHandler.postgres_sqlstate(exc) in _FAILOVER_SQLSTATES:
+        return True
+    message: Final = str(exc).lower()
+    return any(refusal in message for refusal in _FAILOVER_REFUSALS)
+
+
+def _openai_shaped_proxy_exception(exc: Exception, message: str, status_code: int) -> ProxyException:
+    answered_status: Final = int(proxy_exception_status_code(message, str(status_code)))
+    return ProxyException(
+        message=message,
+        type=_catch_all_error_type(exc, answered_status),
+        param=openai_error_param(exc),
+        code=answered_status,
+    )
+
+
+def _catch_all_error_type(exc: Exception, status_code: int) -> str:
+    if _rejects_the_caller(exc, status_code):
+        return ProxyErrorTypes.auth_error.value
+    return openai_error_type(exc, status_code)
+
+
+def _rejects_the_caller(exc: Exception, status_code: int) -> bool:
+    raised_as_http_answer: Final = isinstance(exc, HTTPException)
+    status_from_upstream: Final = error_status_code(exc, 0) == status_code
+    return status_code in _AUTH_FAILURE_STATUS_CODES and (raised_as_http_answer or not status_from_upstream)
+
+
 def litellm_call_id_headers(litellm_call_id: str | None) -> dict[str, str] | None:  # mutable-ok: ProxyException.headers
     if litellm_call_id is None:
         return None

litellm/proxy/db/exception_handler.py

@@ -36,7 +36,7 @@ _LOCK_TIMEOUT_SQLSTATE: Final = "55P03"
 _LOCK_TIMEOUT_PHRASE: Final = "canceling statement due to lock timeout"
 
 
-def _exception_chain(e: BaseException) -> Iterator[BaseException]:
+def exception_chain(e: BaseException) -> Iterator[BaseException]:
     current = e  # rebind-ok: advances one link per iteration of the bounded walk
     for _ in range(_MAX_EXCEPTION_CHAIN_DEPTH):
         yield current
@@ -49,7 +49,7 @@ def _exception_chain(e: BaseException) -> Iterator[BaseException]:
 def _database_service_unavailable_errors(e: BaseException) -> tuple[Exception, ...]:
     return tuple(
         link
-        for link in _exception_chain(e)
+        for link in exception_chain(e)
         if isinstance(link, Exception) and PrismaDBExceptionHandler.is_database_service_unavailable_error(link)
     )
 

litellm/proxy/google_endpoints/endpoints.py

@@ -1,11 +1,11 @@
 from collections.abc import Mapping
 from itertools import chain
 from types import MappingProxyType
-from typing import Final
+from typing import Annotated, Final
 
 from fastapi import APIRouter, Depends, HTTPException, Request, Response
 from fastapi.responses import JSONResponse, ORJSONResponse
-from pydantic import BaseModel, ConfigDict, Field, JsonValue, TypeAdapter, ValidationError
+from pydantic import BaseModel, BeforeValidator, ConfigDict, Field, JsonValue, TypeAdapter, ValidationError
 
 from litellm.litellm_core_utils.json_validation_rule import normalize_json_schema_types
 from litellm.proxy._types import *
@@ -27,14 +27,67 @@ from litellm.proxy.common_utils.token_count_settings import count_source_fields
 from litellm.types.llms.vertex_ai import TokenCountDetailsResponse
 
 
+def _google_repeated_field(value: JsonValue) -> JsonValue:
+    match value:
+        case None:
+            return []
+        case dict():
+            return [value]
+        case list():
+            return [item for item in value if item is not None]
+        case _:
+            return value
+
+
+def _google_content(value: JsonValue) -> JsonValue:
+    match value:
+        case {"parts": parts}:
+            return {**value, "parts": _google_repeated_field(parts)}
+        case _:
+            return value
+
+
+def _google_contents(value: JsonValue) -> JsonValue:
+    match _google_repeated_field(value):
+        case list() as contents:
+            return [_google_content(content) for content in contents]
+        case other:
+            return other
+
+
+_GOOGLE_PART_TEXT_AND_METADATA_FIELDS: Final = frozenset(
+    {"text", "thought", "thoughtSignature", "thought_signature", "partMetadata", "part_metadata"}
+)
+
+
+def _google_part_has_non_text_content(part: Mapping[str, JsonValue]) -> bool:
+    return any(field is not None for key, field in part.items() if key not in _GOOGLE_PART_TEXT_AND_METADATA_FIELDS)
+
+
+def _google_system_part(value: JsonValue) -> JsonValue:
+    match value:
+        case dict() if value.get("text") is None and _google_part_has_non_text_content(value):
+            return {key: field for key, field in value.items() if key != "text"}
+        case dict():
+            return {"text": None, **value}
+        case _:
+            return value
+
+
+_GOOGLE_REPEATED_FIELD: Final = BeforeValidator(_google_repeated_field)
+
+
 class _GoogleCountTokensTextPart(BaseModel):
     model_config = ConfigDict(frozen=True)
-    text: str
+    text: str | None
 
 
 class _GoogleCountTokensSystemInstruction(BaseModel):
     model_config = ConfigDict(frozen=True)
-    parts: tuple[_GoogleCountTokensTextPart, ...]
+    parts: Annotated[
+        tuple[Annotated[_GoogleCountTokensTextPart, BeforeValidator(_google_system_part)], ...],
+        _GOOGLE_REPEATED_FIELD,
+    ] = ()
 
 
 class _GoogleCountTokensFunctionDeclaration(BaseModel):
@@ -47,16 +100,16 @@ class _GoogleCountTokensFunctionDeclaration(BaseModel):
 
 class _GoogleCountTokensTool(BaseModel):
     model_config = ConfigDict(frozen=True, populate_by_name=True)
-    function_declarations: tuple[_GoogleCountTokensFunctionDeclaration, ...] = Field(
+    function_declarations: Annotated[tuple[_GoogleCountTokensFunctionDeclaration, ...], _GOOGLE_REPEATED_FIELD] = Field(
         default=(), alias="functionDeclarations"
     )
 
 
 class _GoogleCountTokensRequest(BaseModel):
     model_config = ConfigDict(frozen=True, extra="ignore", populate_by_name=True)
-    contents: tuple[Mapping[str, JsonValue], ...] = ()
+    contents: Annotated[tuple[Mapping[str, JsonValue], ...], BeforeValidator(_google_contents)] = ()
     system_instruction: _GoogleCountTokensSystemInstruction | None = Field(default=None, alias="systemInstruction")
-    tools: tuple[_GoogleCountTokensTool, ...] = ()
+    tools: Annotated[tuple[_GoogleCountTokensTool, ...], _GOOGLE_REPEATED_FIELD] = ()
 
 
 router: Final = APIRouter(
@@ -112,7 +165,11 @@ def _google_count_tokens_request(model_name: str, data: object) -> TokenCountReq
         GoogleGenAIAdapter().translate_generate_content_to_completion(model_name, contents).get("messages", [])
     )
     system: Final = (
-        [{"type": "text", "text": part.text} for part in validated_data.system_instruction.parts]
+        [
+            {"type": "text", "text": part.text}
+            for part in validated_data.system_instruction.parts
+            if part.text is not None
+        ]
         if validated_data.system_instruction is not None
         else None
     )

litellm/proxy/guardrails/_content_utils.py

@@ -62,6 +62,20 @@ def is_non_conversational_call_type(call_type: str) -> bool:
     return call_type in NON_CONVERSATIONAL_CALL_TYPES
 
 
+_PASSTHROUGH_EMBEDDINGS_ENDPOINT: Final = re.compile(r"/?(?:[\w-]+/)*embeddings/?")
+
+
+def is_non_conversational_request(call_type: str, data: Mapping[str, object]) -> bool:
+    if is_non_conversational_call_type(call_type):
+        return True
+    endpoint: Final = data.get("endpoint")
+    return (
+        call_type == "allm_passthrough_route"
+        and isinstance(endpoint, str)
+        and _PASSTHROUGH_EMBEDDINGS_ENDPOINT.fullmatch(endpoint) is not None
+    )
+
+
 TEXT_PART_TYPES: Final[frozenset[str]] = frozenset(
     {"text", "input_text", "output_text", "summary_text", "reasoning_text"}
 )

litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py

@@ -49,6 +49,7 @@ from litellm.llms.base_llm.guardrail_translation.utils import (
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, bedrock_bearer_token, run_aws_signing
 from litellm.llms.bedrock.guardrail_attachments import find_request_attachments
 from litellm.llms.custom_httpx.http_handler import (
+    AsyncHTTPHandler,
     get_async_httpx_client,
     httpxSpecialProvider,
 )
@@ -286,9 +287,12 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
         streaming_sampling_rate: int | None = None,
         streaming_end_of_stream_only: bool | None = None,
         streaming_buffer_release_on_scan: bool | None = None,
+        async_handler: AsyncHTTPHandler | None = None,
         **kwargs,
     ):
-        self.async_handler = get_async_httpx_client(llm_provider=httpxSpecialProvider.GuardrailCallback)
+        self.async_handler = async_handler or get_async_httpx_client(
+            llm_provider=httpxSpecialProvider.GuardrailCallback
+        )
         self._set_streaming_params(
             BedrockGuardrailStreamingParams.from_extras(
                 MappingProxyType(

litellm/proxy/guardrails/guardrail_hooks/generic_guardrail_api/example_config.yaml

@@ -6,20 +6,19 @@ model_list:
       model: gpt-4
       api_key: os.environ/OPENAI_API_KEY
 
-litellm_settings:
-  guardrails:
-    - guardrail_name: "my-generic-guardrail"
-      litellm_params:
-        guardrail: generic_guardrail_api
-        mode: pre_call  # Options: pre_call, post_call, during_call, [pre_call, post_call]
-        api_key: os.environ/GENERIC_GUARDRAIL_API_KEY  # Optional if using Bearer auth
-        api_base: http://localhost:8080  # Required. Endpoint /beta/litellm_basic_guardrail_api is automatically appended
-        unreachable_fallback: fail_closed  # Options: fail_closed (default, raise), fail_open (proceed if endpoint unreachable or upstream returns 502/503/504)
-        default_on: false  # Set to true to apply to all requests by default
-        additional_provider_specific_params:
-          # Any additional parameters your guardrail API needs
-          api_version: "v1"
-          custom_param: "value"
+guardrails:
+  - guardrail_name: "my-generic-guardrail"
+    litellm_params:
+      guardrail: generic_guardrail_api
+      mode: pre_call  # Options: pre_call, post_call, during_call, [pre_call, post_call]
+      api_key: os.environ/GENERIC_GUARDRAIL_API_KEY  # Optional if using Bearer auth
+      api_base: http://localhost:8080  # Required. Endpoint /beta/litellm_basic_guardrail_api is automatically appended
+      unreachable_fallback: fail_closed  # Options: fail_closed (default, raise), fail_open (proceed if endpoint unreachable or upstream returns 502/503/504)
+      default_on: false  # Set to true to apply to all requests by default
+      additional_provider_specific_params:
+        # Any additional parameters your guardrail API needs
+        api_version: "v1"
+        custom_param: "value"
 
 # Usage examples:
 

litellm/proxy/guardrails/guardrail_hooks/panw_prisma_airs/panw_prisma_airs.py

@@ -30,6 +30,7 @@ from litellm.integrations.custom_guardrail import (
 from litellm.llms.base_llm.guardrail_translation.utils import (
     effective_scan_only_tool_results_for_guardrail,
     effective_skip_system_message_for_guardrail,
+    effective_skip_tool_message_for_guardrail,
     role_out_of_guardrail_scope,
 )
 from litellm.llms.custom_httpx.http_handler import (
@@ -37,12 +38,16 @@ from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
 )
-from litellm.llms.openai.responses.guardrail_translation.handler import scannable_instructions
+from litellm.llms.openai.responses.guardrail_translation.handler import (
+    scannable_instructions,
+    tool_output_texts,
+)
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.common_utils.callback_utils import (
     add_guardrail_scan_id,
     add_guardrail_to_applied_guardrails_header,
 )
+from litellm.responses.litellm_completion_transformation.transformation import TOOL_CALL_OUTPUT_ITEM_TYPES
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.llms.openai import AllMessageValues
@@ -112,12 +117,16 @@ class _ResponsesInputItem(LiteLLMBaseModel):
     type: str | None = None
     role: str | None = None
     content: str | tuple[_ResponsesContentPart, ...] | None = None
+    output: object = None
+    call_id: object = None
 
-    def text_count(self, *, skip_system: bool) -> int:
-        if role_out_of_guardrail_scope(
-            (self.role or "").lower(), skip_system_message=skip_system, skip_tool_message=False
-        ):
+    def text_count(self, *, skip_system: bool, skip_tool: bool) -> int:
+        is_tool_output: Final = self.type in TOOL_CALL_OUTPUT_ITEM_TYPES
+        role: Final = "tool" if is_tool_output else (self.role or "").lower()
+        if role_out_of_guardrail_scope(role, skip_system_message=skip_system, skip_tool_message=skip_tool):
             return 0
+        if is_tool_output:
+            return len(tool_output_texts(self))
         if isinstance(self.content, str):
             return 1
         if self.content is None:
@@ -1692,13 +1701,14 @@ class PanwPrismaAirsHandler(CustomGuardrail):
         request_data: Mapping[str, object],
         *,
         skip_system: bool,
+        skip_tool: bool,
     ) -> frozenset[int] | None:
         """Return the ``texts`` indices flattened from Responses ``reasoning`` input items.
 
         The Responses translation handler gives those model-authored items the default
         ``user`` role, so the latest-turn selection must not mistake one for a human turn.
         Empty for requests without a Responses ``input`` item list; None when the raw items
-        (after the leading ``instructions`` text, both minus whatever ``skip_system`` drops)
+        (after the leading ``instructions`` text, both minus whatever ``skip_system`` and ``skip_tool`` drop)
         do not account for every entry of ``texts``.
         """
         try:
@@ -1708,7 +1718,7 @@ class PanwPrismaAirsHandler(CustomGuardrail):
         if not isinstance(raw_input, tuple):
             return frozenset()
         offset: Final = 0 if scannable_instructions(request_data, skip_system=skip_system) is None else 1
-        counts: Final = tuple(item.text_count(skip_system=skip_system) for item in raw_input)
+        counts: Final = tuple(item.text_count(skip_system=skip_system, skip_tool=skip_tool) for item in raw_input)
         if offset + sum(counts) != len(texts):
             return None
         starts: Final = itertools.accumulate(counts, initial=offset)
@@ -1738,7 +1748,10 @@ class PanwPrismaAirsHandler(CustomGuardrail):
         if sources is None:
             return None
         reasoning: Final = self._reasoning_item_text_indices(
-            texts, request_data, skip_system=effective_skip_system_message_for_guardrail(self)
+            texts,
+            request_data,
+            skip_system=effective_skip_system_message_for_guardrail(self),
+            skip_tool=effective_skip_tool_message_for_guardrail(self),
         )
         if reasoning is None:
             return None

litellm/proxy/health_endpoints/_health_endpoints.py

@@ -32,7 +32,6 @@ from litellm.proxy._types import (
     EnterpriseLicenseData,
     Litellm_EntityType,
     LitellmUserRoles,
-    ProxyErrorTypes,
     ProxyException,
     SpecialModelNames,
     UserAPIKeyAuth,
@@ -47,6 +46,7 @@ from litellm.proxy.auth.auth_utils import (
 from litellm.proxy.auth.model_checks import get_key_models
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_utils import http_parsing_utils
+from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for
 from litellm.proxy.db.db_lookup_gate import db_lookup_stall_tracker
 from litellm.proxy.db.exception_handler import PrismaDBExceptionHandler
 from litellm.proxy.db.health_check_latest import (
@@ -683,21 +683,9 @@ async def health_services_endpoint(
     except Exception as e:
         verbose_proxy_logger.error("litellm.proxy.proxy_server.health_services_endpoint(): Exception occured - %s", e)
         verbose_proxy_logger.debug(traceback.format_exc())
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_500_INTERNAL_SERVER_ERROR,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_500_INTERNAL_SERVER_ERROR) from e
 
 
 def convert_health_check_to_dict(check) -> dict:

litellm/proxy/litellm_pre_call_utils.py

@@ -3024,13 +3024,7 @@ async def add_litellm_data_to_request(
         getattr(request.state, "litellm_roi_estimator", False) is True
     )
 
-    verbose_proxy_logger.debug(
-        "[PROXY] returned data from litellm_pre_call_utils: %s",
-        data,
-    )
-
     # Team/Project credential overrides from model_config
-    # Placed after the debug log to avoid leaking credential secrets in logs
     _apply_credential_overrides_from_model_config(
         data=data,
         user_api_key_dict=user_api_key_dict,

litellm/proxy/management_endpoints/internal_user_endpoints.py

@@ -46,6 +46,7 @@ from litellm.proxy.auth.password_policy import (
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import evict_and_broadcast
+from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for
 from litellm.proxy.common_utils.user_api_key_cache import (
     AUTH_OBJECTS_TARGET,
     object_permission_cache_key,
@@ -1729,21 +1730,9 @@ async def user_update(
     except Exception as e:
         verbose_proxy_logger.exception("litellm.proxy.proxy_server.user_update(): Exception occured - %s", e)
         verbose_proxy_logger.debug(traceback.format_exc())
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 async def bulk_update_processed_users(

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -75,6 +75,7 @@ from litellm.proxy.common_utils.config_sync_pubsub import (
     coordination_redis_cache,
     publish_config_change,
 )
+from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for
 from litellm.proxy.common_utils.rbac_utils import check_org_admin_can_generate_keys
 from litellm.proxy.common_utils.timezone_utils import get_budget_reset_time
 from litellm.proxy.common_utils.user_api_key_cache import AUTH_OBJECTS_TARGET, UserApiKeyCache
@@ -4087,21 +4088,9 @@ async def update_key_fn(
         # update based on remaining passed in values
     except Exception as e:
         verbose_proxy_logger.exception("litellm.proxy.proxy_server.update_key_fn(): Exception occured - %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 @router.post(
@@ -7244,21 +7233,9 @@ async def list_keys(
 
     except Exception as e:
         verbose_proxy_logger.exception("Error in list_keys: %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"error({e})"),
-                type=ProxyErrorTypes.internal_server_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", fastapi.status.HTTP_500_INTERNAL_SERVER_ERROR),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.internal_server_error,
-            param=getattr(e, "param", "None"),
-            code=fastapi.status.HTTP_500_INTERNAL_SERVER_ERROR,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=fastapi.status.HTTP_500_INTERNAL_SERVER_ERROR) from e
 
 
 async def _apply_non_admin_alias_scope(
@@ -7402,21 +7379,9 @@ async def key_aliases(
 
     except Exception as e:
         verbose_proxy_logger.exception("Error in key_aliases: %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"error({e})"),
-                type=ProxyErrorTypes.internal_server_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.internal_server_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_500_INTERNAL_SERVER_ERROR,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_500_INTERNAL_SERVER_ERROR) from e
 
 
 def _validate_sort_params(

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -77,6 +77,7 @@ from litellm.proxy.common_utils.encrypt_decrypt_utils import (
     decrypt_value_helper,
     encrypt_value_helper,
 )
+from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for
 from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
 from litellm.proxy.db.routing_prisma_wrapper import WriterPinnedClient
 from litellm.proxy.management.teams.authz import TEAM_ADMIN_ONLY, is_team_admin
@@ -2845,21 +2846,9 @@ async def delete_model(
 
     except Exception as e:
         verbose_proxy_logger.exception("Failed to delete model. Due to error - %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 async def delete_team_model_alias(
@@ -2899,6 +2888,13 @@ async def delete_team_model_alias(
     return removed_model_aliases
 
 
+def _model_not_added_to_db() -> HTTPException:
+    return HTTPException(
+        status_code=500,
+        detail={"error": "Failed to add model to db. Check your server logs for more details."},
+    )
+
+
 @router.post(
     "/model/new",
     description="Allows adding new models to the model list in the config.yaml",
@@ -3080,6 +3076,8 @@ async def add_new_model(
                 if isinstance(e, HTTPException):
                     raise
                 verbose_proxy_logger.exception("Exception in add_new_model: %s", e)
+                if model_response is None:
+                    raise _model_not_added_to_db() from e
 
         else:
             raise HTTPException(
@@ -3088,10 +3086,7 @@ async def add_new_model(
             )
 
         if model_response is None:
-            raise HTTPException(
-                status_code=500,
-                detail={"error": "Failed to add model to db. Check your server logs for more details."},
-            )
+            raise _model_not_added_to_db()
 
         ## CREATE AUDIT LOG ##
         asyncio.create_task(
@@ -3121,21 +3116,9 @@ async def add_new_model(
 
     except Exception as e:
         verbose_proxy_logger.exception("litellm.proxy.proxy_server.add_new_model(): Exception occured - %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 #### MODEL MANAGEMENT ####
@@ -3364,21 +3347,9 @@ async def update_model(
             return None if model_response is None else _model_write_response(model_response, member_write)
     except Exception as e:
         verbose_proxy_logger.exception("litellm.proxy.proxy_server.update_model(): Exception occured - %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 @router.post(

litellm/proxy/management_endpoints/organization_endpoints.py

@@ -41,6 +41,7 @@ from litellm.proxy.auth.auth_checks import (
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import evict_and_broadcast
+from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for
 from litellm.proxy.common_utils.timezone_utils import get_budget_reset_time
 from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
 from litellm.proxy.management_endpoints.budget_management_endpoints import (
@@ -1396,21 +1397,9 @@ async def organization_member_add(
         )
     except Exception as e:
         verbose_proxy_logger.exception("Error adding member to organization: %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_500_INTERNAL_SERVER_ERROR,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_500_INTERNAL_SERVER_ERROR) from e
 
 
 async def find_member_if_email(user_email: str, prisma_client: PrismaClient) -> LiteLLM_UserTable:

litellm/proxy/management_endpoints/team_endpoints.py

@@ -132,6 +132,7 @@ from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import (
 )
 from litellm.proxy.common_utils.callback_utils import encrypt_callback_vars
 from litellm.proxy.common_utils.json_merge_patch import apply_json_merge_patch
+from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for
 from litellm.proxy.common_utils.user_api_key_cache import (
     UserApiKeyCache,
     team_membership_auth_cache_key,
@@ -5526,21 +5527,9 @@ async def team_info(
             e,
             traceback.format_exc(),
         )
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 @router.get(

litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py

@@ -50,6 +50,7 @@ from litellm.constants import (
 from litellm.exceptions import BlockedPiiEntityError, GuardrailRaisedException
 from litellm.integrations.custom_guardrail import guardrail_request_data_with_streaming, is_guardrail_intervention
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
+from litellm.litellm_core_utils.fallback_generalizations import match_routing_generalization
 from litellm.llms.anthropic.common_utils import AnthropicModelInfo, merge_anthropic_beta_headers
 from litellm.llms.azure.passthrough.transformation import (
     foreign_azure_deployment,
@@ -698,6 +699,20 @@ async def typesafe_proxy_route(
     return await endpoint_func(request, fastapi_response, user_api_key_dict)
 
 
+_COST_MAP_ENTRIES: Final = TypeAdapter(Mapping[str, object])
+
+
+def typesafe_allowlist_model_names() -> tuple[str, ...]:
+    if not passthrough_endpoint_router.has_credentials(custom_llm_provider="typesafe"):
+        return ()
+    cost_map: Final = _COST_MAP_ENTRIES.validate_python(litellm.model_cost)  # pyright: ignore[reportUnknownMemberType]  # legacy registry is validated at this boundary
+    priced_names: Final = (name.removeprefix("typesafe/") for name in cost_map if name.startswith("typesafe/"))
+    jev_models_typesafe_serves: Final = sorted(
+        name for name in priced_names if match_routing_generalization(name) == "typesafe"
+    )
+    return ("typesafe/*", *jev_models_typesafe_serves)
+
+
 @router.post(
     "/laya/v1/systemone",
     tags=["Laya Pass-through", "pass-through"],

litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py

@@ -1,4 +1,5 @@
 import json
+import os
 from collections.abc import Callable
 from typing import TYPE_CHECKING, Final
 
@@ -88,6 +89,14 @@ class PassthroughEndpointRouter:
         )
         return _resolve_oidc_reference(get_secret_str(_env_variable_name))
 
+    def has_credentials(self, custom_llm_provider: str) -> bool:
+        if self._get_deployment_api_key(custom_llm_provider=custom_llm_provider, region_name=None) is not None:
+            return True
+        env_variable_name: Final = self._get_default_env_variable_name_passthrough_endpoint(
+            custom_llm_provider=custom_llm_provider,
+        )
+        return bool(os.environ.get(env_variable_name))
+
     def _get_deployment_api_key(
         self,
         custom_llm_provider: str,

litellm/proxy/proxy_server.py

@@ -524,6 +524,7 @@ from litellm.proxy.common_utils.openai_error_payload import (
     LITELLM_CALL_ID_HEADER,
     headers_with_litellm_call_id,
     litellm_call_id_headers,
+    proxy_exception_for,
     with_litellm_call_id,
 )
 from litellm.proxy.common_utils.periodic_reload_schedule import (
@@ -12390,6 +12391,12 @@ async def model_list(
 
     hidden_names: Final = blocked_names | unhealthy_names
 
+    from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import typesafe_allowlist_model_names
+
+    asks_for_allowlist_names: Final = (
+        include_model_access_groups is True and return_wildcard_routes is True and only_model_access_groups is not True
+    )
+
     # If scope=expand and user has admin privileges, return all proxy models
     if should_expand_scope:
         # Get all proxy models as if user is a proxy admin
@@ -12410,7 +12417,10 @@ async def model_list(
         all_models = get_complete_model_list(
             key_models=[],
             team_models=[],
-            proxy_model_list=proxy_model_list,
+            proxy_model_list=[
+                *proxy_model_list,
+                *(typesafe_allowlist_model_names() if asks_for_allowlist_names else ()),
+            ],
             user_model=None,
             infer_model_from_keys=False,
             return_wildcard_routes=include_wildcard_routes,
@@ -12471,7 +12481,7 @@ async def model_list(
 
     # Otherwise, use the normal behavior (current implementation)
     # Get available models for the user
-    all_models = await get_available_models_for_user(
+    available_models: Final = await get_available_models_for_user(
         user_api_key_dict=user_api_key_dict,
         llm_router=llm_router,
         general_settings=general_settings,
@@ -12484,6 +12494,11 @@ async def model_list(
         return_wildcard_routes=include_wildcard_routes,
         user_api_key_cache=user_api_key_cache,
     )
+    offers_typesafe_allowlist_names: Final = (
+        asks_for_allowlist_names and team_id is None and user_api_key_has_admin_view(user_api_key_dict)
+    )
+    typesafe_names_to_offer: Final = typesafe_allowlist_model_names() if offers_typesafe_allowlist_names else ()
+    all_models = [*available_models, *(name for name in typesafe_names_to_offer if name not in available_models)]
 
     undiscoverable_names: Final = undiscoverable_model_names(
         all_models, llm_router, user_api_key_dict, team_id or user_api_key_dict.team_id
@@ -18437,7 +18452,6 @@ async def async_queue_request(
             "body": copy.copy(data),  # use copy instead of deepcopy
         }
 
-        verbose_proxy_logger.debug("receiving data: %s", data)
         data["model"] = (
             general_settings.get("completion_model", None)  # server default
             or user_model  # model name passed via cli args
@@ -18523,25 +18537,13 @@ async def async_queue_request(
         await proxy_logging_obj.post_call_failure_hook(
             user_api_key_dict=user_api_key_dict, original_exception=e, request_data=data
         )
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        elif isinstance(e, litellm.BadRequestError):
+        if isinstance(e, litellm.BadRequestError):
             raise ProxyException(
                 message=e.message, type="invalid_request_error", param=getattr(e, "param", None), code=e.status_code
-            )
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+            ) from e
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 @app.get("/fallback/login", tags=["experimental"], include_in_schema=False)
@@ -19820,21 +19822,9 @@ async def update_config(
     except Exception as e:
         verbose_proxy_logger.error("litellm.proxy.proxy_server.update_config(): Exception occured - %s", e)
         verbose_proxy_logger.debug(traceback.format_exc())
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_400_BAD_REQUEST) from e
 
 
 ### CONFIG GENERAL SETTINGS
@@ -21017,21 +21007,9 @@ async def get_config(
         }
     except Exception as e:
         verbose_proxy_logger.exception("litellm.proxy.proxy_server.get_config(): Exception occured - %s", e)
-        if isinstance(e, HTTPException):
-            raise ProxyException(
-                message=getattr(e, "detail", f"Authentication Error({e})"),
-                type=ProxyErrorTypes.auth_error,
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
-            )
-        elif isinstance(e, ProxyException):
-            raise e
-        raise ProxyException(
-            message="Authentication Error, " + str(e),
-            type=ProxyErrorTypes.auth_error,
-            param=getattr(e, "param", "None"),
-            code=status.HTTP_400_BAD_REQUEST,
-        )
+        if isinstance(e, ProxyException):
+            raise
+        raise proxy_exception_for(e, default_status_code=status.HTTP_500_INTERNAL_SERVER_ERROR) from e
 
 
 @router.get(

litellm/proxy/public_endpoints/agent_create_fields.json

@@ -127,8 +127,8 @@
       {
         "key": "agent_id",
         "label": "Agent ID",
-        "placeholder": "asst_abc123",
-        "tooltip": "The agent/assistant ID from your Azure AI Foundry project (e.g., asst_abc123)",
+        "placeholder": "my-agent",
+        "tooltip": "The agent name from your Azure AI Foundry project, optionally with a version (e.g., my-agent or my-agent:2), or a classic assistant ID (e.g., asst_abc123)",
         "required": true,
         "field_type": "text",
         "default_value": null,

litellm/proxy/public_endpoints/provider_create_fields.json

@@ -3519,6 +3519,34 @@
     ],
     "default_model_placeholder": "inferenceapis/deepseek-ai/DeepSeek-V4-Flash"
   },
+  {
+    "provider": "StreamLake",
+    "provider_display_name": "StreamLake",
+    "litellm_provider": "streamlake",
+    "credential_fields": [
+      {
+        "key": "api_base",
+        "label": "API Base",
+        "placeholder": "https://vanchin.streamlake.ai/api/gateway/v1/endpoints",
+        "tooltip": null,
+        "required": false,
+        "field_type": "text",
+        "options": null,
+        "default_value": null
+      },
+      {
+        "key": "api_key",
+        "label": "API Key",
+        "placeholder": null,
+        "tooltip": null,
+        "required": true,
+        "field_type": "password",
+        "options": null,
+        "default_value": null
+      }
+    ],
+    "default_model_placeholder": "streamlake/<model-or-endpoint>"
+  },
   {
     "provider": "Snowflake",
     "provider_display_name": "Snowflake",

litellm/responses/bridge_check.py

@@ -9,6 +9,7 @@ from litellm.llms.azure_ai.chat.transformation import AzureAIGPT5Config
 from litellm.llms.azure_ai.common_utils import (
     azure_ai_supports_native_responses,
     foundry_chat_rejects_function_tools_while_reasoning,
+    is_foundry_project_agent,
 )
 from litellm.llms.openai.chat.gpt_reasoning_transformation import OpenAIGPT5Config
 from litellm.secret_managers.main import get_secret_str
@@ -181,6 +182,9 @@ def responses_api_bridge_check(
     elif pinned_mode == "chat" and backend_serves_chat_completions(model_info):
         model_info["mode"] = "chat"
 
+    if custom_llm_provider == LlmProviders.AZURE_AI.value and is_foundry_project_agent(model):
+        model_info["mode"] = "responses"
+
     if web_search_options is not None and custom_llm_provider == "xai":
         model_info["mode"] = "responses"
         model = model.replace("responses/", "")

litellm/responses/litellm_completion_transformation/transformation.py

@@ -321,7 +321,7 @@ def _responses_file_part(item: Mapping[str, object]) -> _ResponsesFilePart:
     )
 
 
-def _decode_responses_media(item: Mapping[str, object]) -> _ResponsesMediaPart | None:
+def decode_responses_media(item: Mapping[str, object]) -> _ResponsesMediaPart | None:
     kind: Final = item.get("type")
     if kind in ("input_image", "image_url", "computer_screenshot") and any(key in item for key in ("image_url", "url")):
         image: Final = _responses_image_part(item)
@@ -383,6 +383,52 @@ def _attribute_fields(value: object) -> dict[str, object]:
     return dict(cast("Iterable[tuple[str, object]]", value))  # cast-ok: dict() raises on non-pair values, as before
 
 
+def tool_output_to_chat_content(
+    output: object,
+) -> str | list[ToolMessageContentPart]:
+    if output is None:
+        return ""
+    if isinstance(output, str):
+        return output
+
+    single_media: Final = decode_responses_media(output) if isinstance(output, Mapping) else None
+    if single_media is not None:
+        return [single_media]
+    if isinstance(output, list):
+        normalized_blocks: Final[list[ToolMessageContentPart]] = []
+        text_acc: Final[list[str]] = []
+        for part in output:
+            if not isinstance(part, dict):
+                continue
+            part_type = part.get("type")
+            if part_type in ("input_text", "output_text", "text"):
+                txt = part.get("text")
+                if isinstance(txt, str) and txt:
+                    text_acc.append(txt)
+                    normalized_blocks.append({"type": "text", "text": txt})
+            elif (media := decode_responses_media(part)) is not None:
+                normalized_blocks.append(media)
+
+        if any(b.get("type") != "text" for b in normalized_blocks):
+            return normalized_blocks
+        if text_acc:
+            return "".join(text_acc)
+        try:
+            return json.dumps(output)
+        except Exception:
+            return str(output)
+
+    try:
+        return json.dumps(output)
+    except Exception:
+        return str(output)
+
+
+TOOL_CALL_OUTPUT_ITEM_TYPES: Final = frozenset(
+    {"function_call_output", "custom_tool_call_output", "computer_call_output", "tool_result"}
+)
+
+
 def _input_item_role(input_item: Mapping[str, object]) -> str:
     return cast(str, input_item.get("role") or "user")  # cast-ok: client-supplied role forwarded verbatim, unvalidated
 
@@ -2371,12 +2417,7 @@ class LiteLLMCompletionResponsesConfig:
         """
         Check if the input item is a tool call output
         """
-        return input_item.get("type") in [
-            "function_call_output",
-            "custom_tool_call_output",
-            "computer_call_output",
-            "tool_result",  # Anthropic/MCP format
-        ]
+        return input_item.get("type") in TOOL_CALL_OUTPUT_ITEM_TYPES
 
     @staticmethod
     def _is_input_item_function_call(input_item: Mapping[str, object]) -> bool:
@@ -2400,58 +2441,7 @@ class LiteLLMCompletionResponsesConfig:
         if not call_id:
             return []
 
-        def _normalize_function_call_output_to_tool_content(
-            output: object,
-        ) -> str | list[ToolMessageContentPart]:
-            """
-            Preserve structured media for downstream chat adapters and token estimation.
-            Text-only outputs retain their existing string representation.
-            """
-            if output is None:
-                return ""
-            if isinstance(output, str):
-                return output
-
-            single_media: Final = _decode_responses_media(output) if isinstance(output, Mapping) else None
-            if single_media is not None:
-                return [single_media]
-            # Some adapters represent tool output as a list of "input_*" parts
-            if isinstance(output, list):
-                normalized_blocks: Final[list[ToolMessageContentPart]] = []
-                text_acc: Final[list[str]] = []
-                for part in output:
-                    if not isinstance(part, dict):
-                        continue
-                    part_type = part.get("type")
-                    if part_type in ("input_text", "output_text", "text"):
-                        txt = part.get("text")
-                        if isinstance(txt, str) and txt:
-                            text_acc.append(txt)
-                            normalized_blocks.append({"type": "text", "text": txt})
-                    elif (media := _decode_responses_media(part)) is not None:
-                        normalized_blocks.append(media)
-
-                if any(b.get("type") != "text" for b in normalized_blocks):
-                    return normalized_blocks
-                if text_acc:
-                    return "".join(text_acc)
-                try:
-                    # last resort: keep something meaningful for providers that require a string
-                    import json as _json
-
-                    return _json.dumps(output)
-                except Exception:
-                    return str(output)
-
-            # Fallback for dict/number/etc.
-            try:
-                import json as _json
-
-                return _json.dumps(output)
-            except Exception:
-                return str(output)
-
-        tool_content: Final = _normalize_function_call_output_to_tool_content(tool_call_output.get("output"))
+        tool_content: Final = tool_output_to_chat_content(tool_call_output.get("output"))
         output_cache_control: Final = _input_item_cache_control(tool_call_output)
         tool_output_message: Final = (
             ChatCompletionToolMessage(role="tool", content=tool_content, tool_call_id=str(call_id))
@@ -2606,7 +2596,7 @@ class LiteLLMCompletionResponsesConfig:
                     content_list.append(item)
                 elif isinstance(item, dict):
                     if (
-                        media := _decode_responses_media(
+                        media := decode_responses_media(
                             fields := _validate_responses_media(_STR_KEY_DICT_ADAPTER, item)
                         )
                     ) is not None:

litellm/responses/streaming_iterator.py

@@ -39,7 +39,11 @@ from litellm.constants import (
 )
 from litellm.exceptions import MidStreamFallbackError, RateLimitError
 from litellm.litellm_core_utils.asyncify import run_async_function
-from litellm.litellm_core_utils.core_helpers import process_response_headers
+from litellm.litellm_core_utils.core_helpers import (
+    process_response_headers,
+    provider_reported_usage_cost,
+    set_response_cost_in_hidden_params,
+)
 from litellm.litellm_core_utils.exception_mapping_utils import with_deployment_api_base
 from litellm.litellm_core_utils.internal_call_metadata import (
     is_background_response,
@@ -67,6 +71,7 @@ from litellm.types.llms.openai import (
     ResponsesAPIResponse,
     ResponsesAPIStreamEvents,
     ResponsesAPIStreamingResponse,
+    replayable_responses_input,
 )
 from litellm.types.responses.main import GenericResponseOutputItem, OutputText
 from litellm.types.utils import CallTypes, ModelResponse
@@ -1825,20 +1830,32 @@ def _estimate_usage_safely(
 def stamp_responses_usage_cost(
     response_obj: ResponsesAPIResponse | None, logging_obj: LiteLLMLoggingObj | None
 ) -> None:
-    if response_obj is None or logging_obj is None:
+    """The terminal response's ``usage.cost`` is the bill litellm writes to the spend row. It is priced on a copy
+    shaped the way the success handler shapes the spend row's result (the reported ``usage.cost`` in the
+    response-cost header, the usage in chat shape), so the client and the row read the same number."""
+    if response_obj is None:
         return
     usage_obj: Final[ResponseAPIUsage | None] = _usage_as_model(getattr(response_obj, "usage", None))
     if usage_obj is None:
         return
-    response_obj.usage = usage_obj  # rebind-ok: the stamped cost has to ride on the response the client receives
-    if isinstance(getattr(usage_obj, "cost", None), (int, float)):
+    response_obj.usage = usage_obj  # rebind-ok: the typed usage has to ride on the response the client receives
+    if logging_obj is None:
         return
+    priced: Final = response_obj.model_copy(
+        update={"usage": ResponseAPILoggingUtils.transform_response_api_usage_to_chat_usage(usage_obj)}
+    )
+    priced.hidden_params = {}
+    set_response_cost_in_hidden_params(
+        priced, provider_reported_usage_cost(usage_obj, logging_obj.model_call_details.get("custom_llm_provider"))
+    )
     try:
-        cost: Final[float | None] = logging_obj.response_cost_calculator(result=response_obj)
-    except Exception:
+        bill: Final = logging_obj.response_cost_calculator(result=priced)
+    except Exception:  # noqa: BLE001  # a pricing failure leaves the client usage as the upstream sent it
         return
-    if isinstance(cost, (int, float)) and cost > 0:
-        setattr(usage_obj, "cost", cost)
+    if isinstance(bill, bool) or not isinstance(bill, (int, float)):
+        return
+    if bill > 0 or isinstance(getattr(usage_obj, "cost", None), (int, float)):
+        setattr(usage_obj, "cost", float(bill))
 
 
 def opening_item_payload(payload: Mapping[str, object]) -> Mapping[str, object]:
@@ -2349,15 +2366,17 @@ def _ws_event_error(event: Mapping[str, object]) -> object:
     return response.get("error") if _is_json_object(response) else None
 
 
-def _restore_input_item_ids(items: Sequence[object]) -> Sequence[object]:
-    return ResponsesAPIRequestUtils.restore_encrypted_content_item_ids_in_input(copy.deepcopy(list(items)))
+def _replayable_input_items(items: Sequence[object]) -> object:
+    return replayable_responses_input(
+        ResponsesAPIRequestUtils.restore_encrypted_content_item_ids_in_input(copy.deepcopy(list(items)))
+    )
 
 
 def _restored_container_fields(container: Mapping[str, object]) -> Mapping[str, object]:
     input_items: Final = container.get("input")
     previous_response_id: Final = container.get("previous_response_id")
     restored: Final = {
-        "input": _restore_input_item_ids(input_items) if _is_json_array(input_items) else input_items,
+        "input": _replayable_input_items(input_items) if _is_json_array(input_items) else input_items,
         "previous_response_id": (
             ResponsesAPIRequestUtils.decode_previous_response_id_to_original_previous_response_id(previous_response_id)
             if isinstance(previous_response_id, str)

litellm/responses/utils.py

@@ -766,6 +766,12 @@ class ResponsesAPIRequestUtils:
                 model_id=model_id,
                 response_id=decoded_response_id,
             )
+        except ValueError:
+            return DecodedResponseId(
+                custom_llm_provider=None,
+                model_id=None,
+                response_id=response_id,
+            )
         except Exception as e:
             verbose_logger.debug("Error decoding response_id '%s': %s", response_id, e)
             return DecodedResponseId(

litellm/router.py

@@ -4921,6 +4921,7 @@ class Router:
             deployment=deployment, forwarded_keys=kwargs.pop(_ALIAS_MARKER_FORWARDED_PARAMS_KWARG, ())
         ):
             kwargs.pop(key, None)
+        kwargs.pop("daily_reset_time", None)  # pyright: ignore[reportUnknownMemberType]  # kwargs is an untyped dict
         defaults: Final = _LITELLM_PARAMS_ADAPTER.validate_python(self.default_litellm_params)
         self._merge_tools_from_deployment(deployment=deployment, kwargs=kwargs, defaults=defaults)
 
@@ -5668,7 +5669,7 @@ class Router:
     def _image_generation(self, prompt: str, model: str, **kwargs):
         model_name: Final = ""
         try:
-            verbose_router_logger.debug("Inside _image_generation()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _image_generation()- model: %s", model)
             deployment: Final = self.get_available_deployment(
                 model=model,
                 messages=[{"role": "user", "content": "prompt"}],
@@ -5730,7 +5731,7 @@ class Router:
     async def _aimage_generation(self, prompt: str, model: str, **kwargs):
         model_name = model
         try:
-            verbose_router_logger.debug("Inside _image_generation()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _image_generation()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,
@@ -5815,7 +5816,7 @@ class Router:
     async def _atranscription(self, file: FileTypes, model: str, **kwargs):
         model_name: Final = model
         try:
-            verbose_router_logger.debug("Inside _atranscription()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _atranscription()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,
@@ -5908,7 +5909,7 @@ class Router:
     async def _aspeech(self, model: str, input: str, voice: str | None = None, **kwargs):
         model_name: Final = model
         try:
-            verbose_router_logger.debug("Inside _aspeech()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _aspeech()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,
@@ -5970,7 +5971,7 @@ class Router:
     async def _arerank(self, model: str, **kwargs):
         model_name = None
         try:
-            verbose_router_logger.debug("Inside _rerank()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _rerank()- model: %s", model)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,
                 specific_deployment=kwargs.pop("specific_deployment", None),
@@ -6079,7 +6080,7 @@ class Router:
 
     async def _atext_completion(self, model: str, prompt: str, **kwargs):
         try:
-            verbose_router_logger.debug("Inside _atext_completion()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _atext_completion()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,
@@ -6149,7 +6150,7 @@ class Router:
 
     async def _aadapter_completion(self, adapter_id: str, model: str, **kwargs):
         try:
-            verbose_router_logger.debug("Inside _aadapter_completion()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _aadapter_completion()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,
@@ -6243,7 +6244,7 @@ class Router:
             kwargs=kwargs,
             metadata_variable_name="litellm_metadata",
         )
-        verbose_router_logger.debug("Inside aguardrail() - guardrail_name: %s; kwargs: %s", guardrail_name, kwargs)
+        verbose_router_logger.debug("Inside aguardrail() - guardrail_name: %s", guardrail_name)
         response: Final = await self.async_function_with_fallbacks(**kwargs)
         return response
 
@@ -6324,9 +6325,7 @@ class Router:
                 controls: Final = per_request_fallback_controls(kwargs)
                 kwargs[MID_STREAM_FALLBACK_CONTROLS_KEY] = controls  # rebind-ok: forwarded to every hop
             self._update_kwargs_before_fallbacks(model=model, kwargs=kwargs, metadata_variable_name="litellm_metadata")
-            verbose_router_logger.debug(
-                "Inside ageneric_api_call_with_fallbacks() - model: %s; kwargs: %s", model, kwargs
-            )
+            verbose_router_logger.debug("Inside ageneric_api_call_with_fallbacks() - model: %s", model)
             response: Final = await self.async_function_with_fallbacks(**kwargs)
             return response
         except Exception as e:
@@ -7146,9 +7145,7 @@ class Router:
         handler_name: Final = original_function.__name__
         metadata_variable_name: Final = get_router_metadata_variable_name(function_name="generic_api_call")
         try:
-            verbose_router_logger.debug(
-                "Inside _generic_api_call() - handler: %s, model: %s; kwargs: %s", handler_name, model, kwargs
-            )
+            verbose_router_logger.debug("Inside _generic_api_call() - handler: %s, model: %s", handler_name, model)
             self._update_kwargs_before_fallbacks(
                 model=model,
                 kwargs=kwargs,
@@ -7223,7 +7220,7 @@ class Router:
     def _embedding(self, input: str | list[object], model: str, **kwargs):
         model_name = None
         try:
-            verbose_router_logger.debug("Inside embedding()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside embedding()- model: %s", model)
             with embedding_deployment_selection():
                 deployment: Final = self.get_available_deployment(
                     model=model,
@@ -7312,7 +7309,7 @@ class Router:
     async def _aembedding(self, input: str | list[object], model: str, **kwargs):
         model_name = None
         try:
-            verbose_router_logger.debug("Inside _aembedding()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _aembedding()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             with embedding_deployment_selection():
                 deployment: Final = await self.async_get_available_deployment(
@@ -7396,7 +7393,7 @@ class Router:
         try:
             from litellm.router_utils.common_utils import add_model_file_id_mappings
 
-            verbose_router_logger.debug("Inside _atext_completion()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _atext_completion()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             healthy_deployments: Final = await self.async_get_healthy_deployments(
                 model=model,
@@ -7634,7 +7631,7 @@ class Router:
         **kwargs,
     ) -> LiteLLMBatch:
         try:
-            verbose_router_logger.debug("Inside _acreate_batch()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _acreate_batch()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             with batch_create_routing():
                 deployment: Final = await self.async_get_available_deployment(
@@ -7833,7 +7830,7 @@ class Router:
         **kwargs,
     ) -> LiteLLMBatch:
         try:
-            verbose_router_logger.debug("Inside _acancel_batch()- model: %s; kwargs: %s", model, kwargs)
+            verbose_router_logger.debug("Inside _acancel_batch()- model: %s", model)
             parent_otel_span: Final = get_parent_otel_span_from_kwargs(kwargs)
             deployment: Final = await self.async_get_available_deployment(
                 model=model,

litellm/router_utils/pre_call_checks/model_rate_limit_check.py

@@ -12,7 +12,7 @@ is logged the first time such a deployment is seen.
 """
 
 import contextlib
-from collections.abc import Callable
+from collections.abc import Callable, Mapping
 from datetime import datetime, timedelta
 from typing import TYPE_CHECKING, Any, Final, Literal
 
@@ -38,7 +38,7 @@ from litellm.router_utils.pre_call_checks.io_token_rate_limit_check import (
 )
 from litellm.router_utils.routing_read_batch import ROUTER_USAGE_TARGET
 from litellm.types.caching import RedisPipelineIncrementOperation
-from litellm.types.router import RouterErrors
+from litellm.types.router import RouterErrors, parse_daily_reset_time
 from litellm.types.utils import StandardLoggingPayload
 from litellm.utils import get_utc_datetime
 
@@ -58,42 +58,49 @@ class RoutingArgs:
 DAY_FORMAT: Final = "%Y-%m-%d"
 
 
-def seconds_until_utc_midnight(now: datetime) -> int:
-    next_midnight: Final = (now + timedelta(days=1)).replace(hour=0, minute=0, second=0, microsecond=0)
-    return max(1, int((next_midnight - now).total_seconds()))
+def daily_window_day(now: datetime, reset_offset: timedelta) -> str:
+    return (now - reset_offset).strftime(DAY_FORMAT)
 
 
-def _first_set(*values: int | None) -> int | None:
-    return next((value for value in values if value is not None), None)
+def seconds_until_daily_reset(now: datetime, reset_offset: timedelta = timedelta(0)) -> int:
+    window_start: Final = (now - reset_offset).replace(hour=0, minute=0, second=0, microsecond=0)
+    next_reset: Final = window_start + timedelta(days=1) + reset_offset
+    return max(1, int((next_reset - now).total_seconds()))
 
 
-class _DailyLimitFields(BaseModel):
+class _DailyLimitParams(BaseModel):
+    """rpd, tpd and daily_reset_time live in litellm_params only."""
+
     model_config = ConfigDict(frozen=True, extra="ignore", from_attributes=True)
+    model: str | None = None
     tpd: int | None = None
     rpd: int | None = None
+    daily_reset_time: str | None = None
+
+    def reset_offset(self) -> timedelta:
+        return parse_daily_reset_time(self.daily_reset_time)
 
 
-class _DailyLimitParams(_DailyLimitFields):
-    model: str | None = None
-
-
-class _DailyLimitModelInfo(_DailyLimitFields):
+class _DailyLimitModelInfo(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore", from_attributes=True)
     id: str | None = None
 
 
-class _DailyLimitDeployment(_DailyLimitFields):
+class _DailyLimitDeployment(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore", from_attributes=True)
     model_name: str = ""
     litellm_params: _DailyLimitParams = _DailyLimitParams()
     model_info: _DailyLimitModelInfo = _DailyLimitModelInfo()
 
     def limits(self) -> tuple[int | None, int | None]:
-        return (
-            _first_set(self.tpd, self.litellm_params.tpd, self.model_info.tpd),
-            _first_set(self.rpd, self.litellm_params.rpd, self.model_info.rpd),
-        )
+        return self.litellm_params.tpd, self.litellm_params.rpd
 
-    def cache_keys(self, current_day: str) -> tuple[str, str]:
+    def reset_offset(self) -> timedelta:
+        return self.litellm_params.reset_offset()
+
+    def cache_keys(self, now: datetime) -> tuple[str, str]:
         prefix: Final = f"{self.model_info.id}:{self.litellm_params.model}"
+        current_day: Final = daily_window_day(now, self.reset_offset())
         return f"{prefix}:tpd:{current_day}", f"{prefix}:rpd:{current_day}"
 
     def limit_error(self, kind: Literal["tpd", "rpd"], limit: int, usage: int, now: datetime) -> litellm.RateLimitError:
@@ -104,12 +111,19 @@ class _DailyLimitDeployment(_DailyLimitFields):
             response=httpx.Response(
                 status_code=429,
                 content=f"{RouterErrors.user_defined_ratelimit_error.value} {kind} limit={limit}. current usage={usage}. id={self.model_info.id}, model_group={self.model_name}",
-                headers={"retry-after": str(seconds_until_utc_midnight(now))},
+                headers={"retry-after": str(seconds_until_daily_reset(now, self.reset_offset()))},
                 request=httpx.Request(method="model_rate_limit_check", url="https://github.com/BerriAI/litellm"),
             ),
         )
 
 
+def _reset_offset_from_kwargs(kwargs: Mapping[str, object]) -> timedelta:
+    litellm_params: Final = kwargs.get("litellm_params")
+    if litellm_params is None:
+        return timedelta(0)
+    return _DailyLimitParams.model_validate(litellm_params).reset_offset()
+
+
 class ModelRateLimitingCheck(CustomLogger):
     """
     Pre-call check that enforces TPM/RPM or ITPM/OTPM limits on model deployments.
@@ -242,7 +256,7 @@ class ModelRateLimitingCheck(CustomLogger):
         tpd_limit, rpd_limit = deployment.limits()
         if tpd_limit is None and rpd_limit is None:
             return
-        tpd_key, rpd_key = deployment.cache_keys(now.strftime(DAY_FORMAT))
+        tpd_key, rpd_key = deployment.cache_keys(now)
         if tpd_limit is not None:
             current_tpd: Final = self._get_current_tpm(tpd_key, tpd_limit)
             if current_tpd is not None and current_tpd >= tpd_limit:
@@ -394,7 +408,7 @@ class ModelRateLimitingCheck(CustomLogger):
 
             current_minute: Final = dt.strftime("%H-%M")
             tpm_key, rpm_key = self._get_cache_keys(deployment, current_minute)
-            tpd_key, rpd_key = daily.cache_keys(dt.strftime(DAY_FORMAT))
+            tpd_key, rpd_key = daily.cache_keys(dt)
 
             model_id: Final = deployment.get("model_info", {}).get("id")
             model_name: Final = deployment.get("litellm_params", {}).get("model")
@@ -510,7 +524,8 @@ class ModelRateLimitingCheck(CustomLogger):
             dt: Final = self.clock()
             current_minute: Final = dt.strftime("%H-%M")
             tpm_key: Final = f"{model_id}:{model}:tpm:{current_minute}"
-            tpd_key: Final = f"{model_id}:{model}:tpd:{dt.strftime(DAY_FORMAT)}"
+            reset_offset: Final = _reset_offset_from_kwargs(kwargs)  # pyright: ignore[reportUnknownArgumentType]  # CustomLogger passes an untyped kwargs dict
+            tpd_key: Final = f"{model_id}:{model}:tpd:{daily_window_day(dt, reset_offset)}"
 
             verbose_router_logger.debug("[TPM TRACKING] Incrementing %s by %s", tpm_key, total_tokens)
 
@@ -570,7 +585,8 @@ class ModelRateLimitingCheck(CustomLogger):
             dt: Final = self.clock()
             current_minute: Final = dt.strftime("%H-%M")
             tpm_key: Final = f"{model_id}:{model}:tpm:{current_minute}"
-            tpd_key: Final = f"{model_id}:{model}:tpd:{dt.strftime(DAY_FORMAT)}"
+            reset_offset: Final = _reset_offset_from_kwargs(kwargs)  # pyright: ignore[reportUnknownArgumentType]  # CustomLogger passes an untyped kwargs dict
+            tpd_key: Final = f"{model_id}:{model}:tpd:{daily_window_day(dt, reset_offset)}"
 
             tokens: Final = int(total_tokens)
             self._increment(tpm_key, tokens, RoutingArgs.ttl)

litellm/router_utils/search_api_router.py

@@ -155,10 +155,9 @@ class SearchAPIRouter:
 
             available_search_tool_names: Final = [tool.get("search_tool_name") for tool in router_instance.search_tools]
             verbose_router_logger.debug(
-                "Inside SearchAPIRouter.async_search_with_fallbacks() - search_tool_name: %s, Available Search Tools: %s, kwargs: %s",
+                "Inside SearchAPIRouter.async_search_with_fallbacks() - search_tool_name: %s, Available Search Tools: %s",
                 search_tool_name,
                 available_search_tool_names,
-                kwargs,
             )
 
             # Use the existing retry/fallback infrastructure

litellm/rust_bridge/host/routes/ocr.py

@@ -7,12 +7,13 @@ from typing import Final
 from pydantic import TypeAdapter
 
 import litellm
+from litellm.litellm_core_utils.core_helpers import drops_unsupported_params
 from litellm.llms.base_llm.ocr.transformation import PROVIDER_NATIVE_RESPONSE_KEY, OCRResponse
 from litellm.rust_bridge import failures
 from litellm.rust_bridge.failures import UpstreamFailure
 from litellm.rust_bridge.public_call import optional_str
 
-__all__ = ("UpstreamFailure", "map_failure", "response")
+__all__ = ("UpstreamFailure", "drops_unsupported_params", "map_failure", "response")
 
 _RESPONSE_ADAPTER: Final = TypeAdapter(dict[str, object])
 

litellm/types/litellm_params.py

@@ -191,6 +191,7 @@ class DeploymentOptions:
     otpm: int | None = None
     rpd: int | None = None
     tpd: int | None = None
+    daily_reset_time: str | None = None
     kubernetes_pod_discovery: bool | None = None
     default_api_key_rpm_limit: int | None = None
     default_api_key_tpm_limit: int | None = None

litellm/types/llms/azure_ai.py

@@ -1,6 +1,6 @@
 from typing import Literal
 
-from typing_extensions import Required, TypedDict
+from typing_extensions import NotRequired, ReadOnly, Required, TypedDict
 
 
 class ImageEmbeddingInput(TypedDict, total=False):
@@ -15,3 +15,9 @@ class ImageEmbeddingRequest(TypedDict, total=False):
     input: Required[list[ImageEmbeddingInput]]
     dimensions: int
     encoding_format: EncodingFormat
+
+
+class FoundryAgentReference(TypedDict):
+    type: ReadOnly[Literal["agent_reference"]]
+    name: ReadOnly[str]
+    version: ReadOnly[NotRequired[str]]

litellm/types/llms/openai.py

@@ -1,5 +1,5 @@
 import builtins
-from collections.abc import Iterable, Mapping
+from collections.abc import Iterable, Mapping, Sequence
 from enum import Enum
 from os import PathLike
 from typing import IO, TYPE_CHECKING, Any, Final, Generic, Literal, Optional, TypeAlias, Union
@@ -66,6 +66,7 @@ from pydantic import (
     ConfigDict,
     Discriminator,
     Field,
+    FieldSerializationInfo,
     NonNegativeInt,
     PrivateAttr,
     SerializerFunctionWrapHandler,
@@ -1487,15 +1488,66 @@ One of: completed, failed, in_progress, cancelled, queued, or incomplete.
 """
 
 
-_NULL_DROPPED_OUTPUT_ITEM_KEYS: Final = frozenset(
-    (
-        ("reasoning", "status"),
-        ("reasoning", "content"),
-        ("reasoning", "encrypted_content"),
-        ("tool_search_call", "created_by"),
-        ("tool_search_output", "created_by"),
-    )
-)
+_NULL_DROPPED_REASONING_ITEM_KEYS: Final = frozenset(("status", "content", "encrypted_content"))
+
+
+def _as_json_object(value: object) -> Mapping[str, object] | None:
+    return value if isinstance(value, dict) else None
+
+
+def _as_json_array(value: object) -> Sequence[object] | None:
+    return value if isinstance(value, list) else None
+
+
+def _drops_created_by(value: object, *, drops_set_created_by: bool) -> bool:
+    return value is None or drops_set_created_by
+
+
+def _drops(item_type: object, key: object, value: object, *, drops_set_created_by: bool) -> bool:
+    if key == "created_by":
+        return _drops_created_by(value, drops_set_created_by=drops_set_created_by)
+    return value is None and item_type == "reasoning" and key in _NULL_DROPPED_REASONING_ITEM_KEYS
+
+
+def _without_created_by(entry: object, *, drops_set_created_by: bool) -> object:
+    entry_object: Final = _as_json_object(entry)
+    if entry_object is None:
+        return entry
+    return {
+        k: v
+        for k, v in entry_object.items()
+        if k != "created_by" or not _drops_created_by(v, drops_set_created_by=drops_set_created_by)
+    }
+
+
+def _replayable_output_value(item_type: object, key: object, value: object, *, drops_set_created_by: bool) -> object:
+    entries: Final = _as_json_array(value)
+    if item_type != "shell_call_output" or key != "output" or entries is None:
+        return value
+    return [_without_created_by(entry, drops_set_created_by=drops_set_created_by) for entry in entries]
+
+
+def replayable_responses_item(
+    item: Mapping[str, object], *, drops_set_created_by: bool = False
+) -> Mapping[str, object]:
+    item_type: Final = item.get("type")
+    return {
+        k: _replayable_output_value(item_type, k, v, drops_set_created_by=drops_set_created_by)
+        for k, v in item.items()
+        if not _drops(item_type, k, v, drops_set_created_by=drops_set_created_by)
+    }
+
+
+def replayable_responses_input(value: object, *, drops_set_created_by: bool = False) -> object:
+    items: Final = _as_json_array(value)
+    if items is None:
+        return value
+    return [
+        replayable_responses_item(obj, drops_set_created_by=drops_set_created_by)
+        if (obj := _as_json_object(item)) is not None
+        else item
+        for item in items
+    ]
 
 
 class ResponsesAPIResponse(BaseLiteLLMOpenAIResponseObject):
@@ -1570,34 +1622,16 @@ class ResponsesAPIResponse(BaseLiteLLMOpenAIResponseObject):
 
     @field_serializer("output", mode="wrap")
     @classmethod
-    def _serialize_output_filter_reasoning_nulls(cls, value, handler, _info):
+    def _serialize_output_as_replayable(
+        cls, value: builtins.object, handler: SerializerFunctionWrapHandler, _info: FieldSerializationInfo
+    ):
         """
-        Filter null status/content/encrypted_content from reasoning output items.
-
-        Mirrors the request-side filtering in
-        OpenAIResponsesAPIConfig._handle_reasoning_item() which filters these
-        same fields before sending requests to providers.
-
-        Without this, reasoning items include null fields that cause SDK errors
-        (e.g., the OpenAI C# SDK crashes on status=null).
-
-        Issue: https://github.com/BerriAI/litellm/issues/16824
+        Drop the nulls the SDK types default in but providers never send: created_by on every output item
+        (OpenAI rejects it when `output` is replayed as `input`), and status/content/encrypted_content on
+        reasoning items (the OpenAI C# SDK crashes on status=null, issue #16824).
         """
-        serialized: Final = handler(value)
-        if not isinstance(serialized, list):
-            return serialized
-        return [
-            (
-                {
-                    k: v
-                    for k, v in item.items()
-                    if v is not None or (item.get("type"), k) not in _NULL_DROPPED_OUTPUT_ITEM_KEYS
-                }
-                if isinstance(item, dict)
-                else item
-            )
-            for item in serialized
-        ]
+        serialized: Final[object] = handler(value)
+        return replayable_responses_input(serialized)
 
     @property
     def output_text(self) -> str:

litellm/types/router.py

@@ -4,6 +4,7 @@ litellm.Router Types - includes RouterConfig, UpdateRouterConfig, ModelInfo etc
 
 import datetime
 import enum
+import re
 from collections.abc import Container, Mapping, Sequence
 from dataclasses import dataclass, field
 from typing import (
@@ -518,6 +519,19 @@ def holds_secret_pointer(param_name: str) -> bool:
 _RESERVED_INIT_KEYS: Final = frozenset({"self", "params", "__class__"})
 
 
+_DAILY_RESET_TIME_PATTERN: Final = re.compile(r"\d{2}:\d{2}")
+
+
+def parse_daily_reset_time(value: str | None) -> datetime.timedelta:
+    """`daily_reset_time` is "HH:MM" in UTC, when the rpd/tpd window starts over; unset means midnight UTC."""
+    if value is None:
+        return datetime.timedelta(0)
+    if _DAILY_RESET_TIME_PATTERN.fullmatch(value) is None:
+        raise ValueError(f"daily_reset_time must be HH:MM in UTC, got {value!r}")
+    parsed: Final = datetime.time.fromisoformat(value)
+    return datetime.timedelta(hours=parsed.hour, minutes=parsed.minute)
+
+
 class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
     """
     LiteLLM Params without 'model' arg (used across completion / assistants api)
@@ -535,6 +549,7 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
     otpm: int | None = None
     tpd: int | None = None
     rpd: int | None = None
+    daily_reset_time: str | None = None
     timeout: float | str | httpx.Timeout | SDKTimeout | None = None  # if str, pass in as os.environ/
     # read timeout of a live stream, a faked stream takes the longer of this and timeout. If str, pass in as os.environ/
     stream_timeout: float | str | None = None
@@ -626,6 +641,13 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
     valkey_text_field: str | None = None
     valkey_embedding_field: str | None = None
 
+    @field_validator("daily_reset_time")
+    @classmethod
+    def _daily_reset_time_is_utc_hh_mm(cls, value: str | None) -> str | None:
+        if value is not None:
+            parse_daily_reset_time(value)
+        return value
+
     @field_validator("timeout")
     @classmethod
     def normalize_timeout(
@@ -742,6 +764,7 @@ class LiteLLMParamsTypedDict(TypedDict, total=False):
     otpm: int | None
     tpd: ReadOnly[int | None]
     rpd: ReadOnly[int | None]
+    daily_reset_time: ReadOnly[str | None]
     order: int | None
     weight: int | None
     max_parallel_requests: int | None

litellm/types/utils.py

@@ -2878,7 +2878,7 @@ class TranscriptionUsageTokensObject(LiteLLMBaseModel):
     input_tokens: int
     output_tokens: int
     total_tokens: int
-    input_token_details: TranscriptionUsageInputTokenDetailsObject
+    input_token_details: TranscriptionUsageInputTokenDetailsObject | None = None
 
 
 class TranscriptionResponse(OpenAIObject):
@@ -4593,6 +4593,7 @@ class LlmProviders(str, Enum):
     META = "meta"
     SAIL = "sail"
     REKA = "reka"
+    STREAMLAKE = "streamlake"
     LITELLM_AGENT = "litellm_agent"
     CURSOR = "cursor"
     BEDROCK_MANTLE = "bedrock_mantle"

litellm/utils.py

@@ -304,7 +304,7 @@ except (ImportError, AttributeError, TypeError):
 # Convert to str (if necessary)
 claude_json_str = json.dumps(json_data)
 import importlib.metadata
-from collections.abc import AsyncIterator, Callable, Collection, Iterable, Iterator, Mapping, Sequence
+from collections.abc import AsyncIterator, Callable, Collection, Iterable, Iterator, Mapping, Sequence, Sized
 from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, TypeAlias, TypeVar, cast, runtime_checkable
 
 from typing_extensions import assert_never
@@ -606,9 +606,23 @@ def custom_llm_setup():
             litellm.provider_list.append(custom_llm["provider"])
 
         if custom_llm["provider"] not in litellm._custom_providers:
+            _warn_if_built_in_provider_shadows_custom_handler(custom_llm["provider"])
             litellm._custom_providers.append(custom_llm["provider"])
 
 
+def _warn_if_built_in_provider_shadows_custom_handler(provider: str) -> None:
+    from litellm.llms.openai_like.json_loader import JSONProviderRegistry
+
+    if provider not in LlmProvidersSet and not JSONProviderRegistry.exists(provider):
+        return
+    verbose_logger.warning(
+        "custom_provider_map provider '%s' has the same name as a built-in LiteLLM provider, which handles "
+        "'%s/' models instead of your custom_handler. Rename the provider and its models to use your handler",
+        provider,
+        provider,
+    )
+
+
 def add_custom_logger_callback_to_specific_event(callback: str, logging_event: Literal["success", "failure"]) -> None:
     """
     Add a custom logger callback to the specific event
@@ -4067,7 +4081,9 @@ def get_optional_params_transcription(
         "timestamp_granularities": None,
     }
 
-    non_default_params = {k: v for k, v in passed_params.items() if (k in default_params and v != default_params[k])}
+    non_default_params: Final = {
+        k: v for k, v in passed_params.items() if (k in default_params and v != default_params[k])
+    }
     optional_params = {}
 
     ## raise exception if non-default value passed for non-openai/azure embedding calls
@@ -4104,6 +4120,19 @@ def get_optional_params_transcription(
             model=model,
             drop_params=drop_params if drop_params is not None else False,
         )
+    elif custom_llm_provider == "bedrock":
+        from litellm.llms.bedrock.audio_transcription import (
+            UnsupportedTranscriptionParams,
+            map_openai_transcription_params,
+        )
+
+        match map_openai_transcription_params(
+            non_default_params, drop_params=drop_params is True or litellm.drop_params is True
+        ):
+            case UnsupportedTranscriptionParams(message=message):
+                raise UnsupportedParamsError(message=message, llm_provider="bedrock")
+            case mapped_params:
+                optional_params = dict(mapped_params)
     elif provider_config is not None:  # custom audio transcription config
         supported_params = provider_config.get_supported_openai_params(model=model)
         _check_valid_arg(supported_params=supported_params)
@@ -4118,7 +4147,7 @@ def get_optional_params_transcription(
         optional_params=optional_params,
         passed_params=passed_params,
         custom_llm_provider=custom_llm_provider,
-        openai_params=OPENAI_TRANSCRIPTION_PARAMS,
+        openai_params=list(default_params) if custom_llm_provider == "bedrock" else OPENAI_TRANSCRIPTION_PARAMS,
         additional_drop_params=kwargs.get("additional_drop_params", None),
     )
 
@@ -5059,8 +5088,10 @@ def get_optional_params(
             supported_params: List[str] - supported params from the litellm config
         """
         verbose_logger.info("\nLiteLLM completion() model= %s; provider = %s", model, custom_llm_provider)
-        verbose_logger.debug("\nLiteLLM: Params passed to completion() %s", passed_params)
-        verbose_logger.debug("\nLiteLLM: Non-Default params passed to completion() %s", non_default_params)
+        if verbose_logger.isEnabledFor(logging.DEBUG):
+            verbose_logger.debug(
+                "\nLiteLLM: Non-Default params passed to completion() %s", _debug_params_repr(non_default_params)
+            )
         unsupported_params: Final = {}
         for k in non_default_params:
             if k not in supported_params:
@@ -5601,7 +5632,9 @@ def get_optional_params(
         additional_drop_params=additional_drop_params,
     )
     if _print_verbose_is_active():
-        print_verbose(f"Final returned optional params: {redact_credentials_in_payload(optional_params)}")
+        print_verbose(
+            f"Final returned optional params: {_debug_params_repr(redact_credentials_in_payload(optional_params))}"
+        )
     optional_params = apply_openai_param_overrides(
         optional_params=optional_params,
         non_default_params=non_default_params,
@@ -6610,8 +6643,6 @@ def get_model_info_helper(
             model=model, custom_llm_provider=custom_llm_provider or declared_authenticating_provider(model)
         )
 
-        verbose_logger.debug("checking potential_model_names in litellm.model_cost: %s", potential_model_names)
-
         combined_model_name: Final = potential_model_names["combined_model_name"]
         stripped_model_name: Final = potential_model_names["stripped_model_name"]
         combined_stripped_model_name: Final = potential_model_names["combined_stripped_model_name"]
@@ -6714,6 +6745,7 @@ def get_model_info_helper(
                     key, _model_info = generalization
 
             if _model_info is None or key is None:
+                verbose_logger.debug("model not found in litellm.model_cost, tried: %s", potential_model_names)
                 raise ModelNotMappedError(_model_not_mapped_message(model, custom_llm_provider))
             _input_cost_per_token: float | None = _model_info.get("input_cost_per_token")
             if _input_cost_per_token is None:
@@ -8533,6 +8565,23 @@ def get_valid_models(
         return []  # NON-Blocking
 
 
+_MAX_INLINE_DEBUG_ARG_CHARS: Final = 500
+
+
+def _debug_arg_repr(value: object) -> str:
+    """Short form of a call argument, since the outgoing provider request log already prints the request body."""
+    rendered: Final = repr(value)
+    if len(rendered) <= _MAX_INLINE_DEBUG_ARG_CHARS:
+        return rendered
+    if isinstance(value, Sized):
+        return f"<{type(value).__name__} len={len(value)}>"
+    return f"<{type(value).__name__}>"
+
+
+def _debug_params_repr(params: Mapping[str, object]) -> str:
+    return "{" + ", ".join(f"{key!r}: {_debug_arg_repr(value)}" for key, value in params.items()) + "}"
+
+
 def print_args_passed_to_litellm(original_function, args, kwargs):
     if not is_debugging_on():
         return
@@ -8553,9 +8602,9 @@ def print_args_passed_to_litellm(original_function, args, kwargs):
         ):
             return
 
-        args_str: Final = ", ".join(map(repr, args))
+        args_str: Final = ", ".join(map(_debug_arg_repr, args))
         redacted_kwargs: Final = redact_credentials_in_payload(kwargs)
-        kwargs_str: Final = ", ".join(f"{key}={value!r}" for key, value in redacted_kwargs.items())
+        kwargs_str: Final = ", ".join(f"{key}={_debug_arg_repr(value)}" for key, value in redacted_kwargs.items())
         print_verbose(
             "\n",
         )  # new line before
@@ -9899,8 +9948,14 @@ class ProviderConfigManager:
         elif litellm.LlmProviders.AZURE_AI == provider:
             from litellm.llms.azure_ai.common_utils import (
                 azure_ai_supports_native_responses,
+                is_foundry_project_agent,
+            )
+            from litellm.llms.azure_ai.responses.transformation import (
+                AzureAIAgentResponsesAPIConfig,
             )
 
+            if is_foundry_project_agent(model):
+                return AzureAIAgentResponsesAPIConfig()
             if azure_ai_supports_native_responses(model, api_base):
                 return litellm.AzureAIResponsesAPIConfig()
             return None

model_prices_and_context_window.json

@@ -41939,6 +41939,120 @@
         "supports_system_messages": true,
         "supports_native_structured_output": true
     },
+    "bedrock/ap-southeast-2/mistral.voxtral-mini-3b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 4.12e-08,
+        "input_cost_per_token_batches": 2.06e-08,
+        "input_cost_per_token_flex": 2.06e-08,
+        "input_cost_per_token_priority": 7.21e-08,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 4.12e-08,
+        "output_cost_per_token_batches": 2.06e-08,
+        "output_cost_per_token_flex": 2.06e-08,
+        "output_cost_per_token_priority": 7.21e-08,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-southeast-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/ap-southeast-2/mistral.voxtral-small-24b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 1.03e-07,
+        "input_cost_per_token_batches": 5.15e-08,
+        "input_cost_per_token_flex": 5.15e-08,
+        "input_cost_per_token_priority": 1.8025e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 3.09e-07,
+        "output_cost_per_token_batches": 1.545e-07,
+        "output_cost_per_token_flex": 1.545e-07,
+        "output_cost_per_token_priority": 5.4075e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-southeast-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/ap-northeast-1/mistral.voxtral-mini-3b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 5e-08,
+        "input_cost_per_token_batches": 3e-08,
+        "input_cost_per_token_flex": 3e-08,
+        "input_cost_per_token_priority": 9e-08,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 5e-08,
+        "output_cost_per_token_batches": 3e-08,
+        "output_cost_per_token_flex": 3e-08,
+        "output_cost_per_token_priority": 9e-08,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-northeast-1/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/ap-northeast-1/mistral.voxtral-small-24b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 1.2e-07,
+        "input_cost_per_token_batches": 6e-08,
+        "input_cost_per_token_flex": 6e-08,
+        "input_cost_per_token_priority": 2.1e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 3.6e-07,
+        "output_cost_per_token_batches": 1.8e-07,
+        "output_cost_per_token_flex": 1.8e-07,
+        "output_cost_per_token_priority": 6.3e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/ap-northeast-1/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/eu-west-2/mistral.voxtral-mini-3b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 6e-08,
+        "input_cost_per_token_batches": 3e-08,
+        "input_cost_per_token_flex": 3e-08,
+        "input_cost_per_token_priority": 1.1e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 6e-08,
+        "output_cost_per_token_batches": 3e-08,
+        "output_cost_per_token_flex": 3e-08,
+        "output_cost_per_token_priority": 1.1e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/eu-west-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
+    "bedrock/eu-west-2/mistral.voxtral-small-24b-2507": {
+        "deprecation_date": "2027-03-30",
+        "input_cost_per_token": 1.6e-07,
+        "input_cost_per_token_batches": 8e-08,
+        "input_cost_per_token_flex": 8e-08,
+        "input_cost_per_token_priority": 2.8e-07,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 32768,
+        "max_output_tokens": 8192,
+        "mode": "chat",
+        "output_cost_per_token": 4.7e-07,
+        "output_cost_per_token_batches": 2.3e-07,
+        "output_cost_per_token_flex": 2.3e-07,
+        "output_cost_per_token_priority": 8.2e-07,
+        "source": "https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/eu-west-2/index.json",
+        "supports_audio_input": true,
+        "supports_system_messages": true,
+        "supports_native_structured_output": true
+    },
     "mistral/codestral-2508": {
         "cache_read_input_token_cost": 3e-08,
         "input_cost_per_token": 3e-07,
@@ -61275,6 +61389,359 @@
         "supports_tool_choice": true,
         "supports_vision": true
     },
+    "streamlake/deepseek-v4-flash-0731": {
+        "cache_read_input_token_cost": 1.4e-08,
+        "input_cost_per_token": 4.4e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 1.32e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/DeepSeek-V4-Pro-0813": {
+        "cache_read_input_token_cost": 4.4e-08,
+        "input_cost_per_token": 1.32e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048566,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 3.96e-06,
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/DeepSeek-V3": {
+        "cache_read_input_token_cost": 1.14e-07,
+        "input_cost_per_token": 2.86e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 56000,
+        "max_output_tokens": 8000,
+        "mode": "chat",
+        "output_cost_per_token": 1.143e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": false,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/GLM-5.3": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_assistant_prefill": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/GLM-5.3-Flash": {
+        "cache_read_input_token_cost": 3e-08,
+        "input_cost_per_token": 1.5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 5e-07,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_vision": true,
+        "supports_assistant_prefill": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/Kimi-K2-Instruct-0905": {
+        "cache_read_input_token_cost": 1.43e-07,
+        "input_cost_per_token": 5.71e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 224000,
+        "max_output_tokens": 32000,
+        "mode": "chat",
+        "output_cost_per_token": 2.286e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": false,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-r1": {
+        "cache_read_input_token_cost": 2.29e-07,
+        "input_cost_per_token": 5.71e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 98304,
+        "max_output_tokens": 16000,
+        "mode": "chat",
+        "output_cost_per_token": 2.286e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v3.1": {
+        "cache_read_input_token_cost": 2.29e-07,
+        "input_cost_per_token": 5.71e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 98304,
+        "max_output_tokens": 32000,
+        "mode": "chat",
+        "output_cost_per_token": 1.714e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4-flash": {
+        "cache_read_input_token_cost": 2.8e-08,
+        "input_cost_per_token": 1.4e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 2.8e-07,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4-pro": {
+        "cache_read_input_token_cost": 1.45e-07,
+        "input_cost_per_token": 1.74e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 3.48e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4.1-flash-1017": {
+        "cache_read_input_token_cost": 6e-09,
+        "input_cost_per_token": 3e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 1.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/deepseek-v4.1-flash-260910": {
+        "cache_read_input_token_cost": 6e-09,
+        "input_cost_per_token": 3e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048568,
+        "max_output_tokens": 384000,
+        "mode": "chat",
+        "output_cost_per_token": 1.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/glm-5": {
+        "cache_read_input_token_cost": 2e-07,
+        "input_cost_per_token": 1e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 202752,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/glm-5.1": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 202752,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/glm-5.2": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1048566,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/kimi-k2.6": {
+        "cache_read_input_token_cost": 1.6e-07,
+        "input_cost_per_token": 9.5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 262144,
+        "max_output_tokens": 256000,
+        "mode": "chat",
+        "output_cost_per_token": 4e-06,
+        "supports_function_calling": true,
+        "supports_vision": true,
+        "supports_reasoning": false,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/kimi-k2.7-code": {
+        "cache_read_input_token_cost": 1.9e-07,
+        "input_cost_per_token": 9.5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 262144,
+        "max_output_tokens": 32000,
+        "mode": "chat",
+        "output_cost_per_token": 4e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/mimo-v2.6-flash": {
+        "cache_read_input_token_cost": 2.8e-09,
+        "input_cost_per_token": 1.4e-07,
+        "litellm_provider": "streamlake",
+        "mode": "chat",
+        "output_cost_per_token": 2.8e-07,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/mimo-v2.6-pro": {
+        "cache_read_input_token_cost": 3.6e-09,
+        "input_cost_per_token": 4.35e-07,
+        "litellm_provider": "streamlake",
+        "mode": "chat",
+        "output_cost_per_token": 8.7e-07,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/minimax-m2.5": {
+        "cache_creation_input_token_cost": 3.75e-07,
+        "cache_read_input_token_cost": 3e-08,
+        "input_cost_per_token": 3e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 196608,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 1.2e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/minimax-m3": {
+        "cache_read_input_token_cost": 1.2e-07,
+        "cache_read_input_token_cost_above_512k_tokens": 2.4e-07,
+        "input_cost_per_token": 6e-07,
+        "input_cost_per_token_above_512k_tokens": 1.2e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 512000,
+        "mode": "chat",
+        "output_cost_per_token": 2.4e-06,
+        "output_cost_per_token_above_512k_tokens": 4.8e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.5-397b-a17b": {
+        "cache_read_input_token_cost": 1.2e-07,
+        "input_cost_per_token": 6e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 258048,
+        "max_output_tokens": 64000,
+        "mode": "chat",
+        "output_cost_per_token": 3.6e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.6-27b": {
+        "input_cost_per_token": 6e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 229376,
+        "max_output_tokens": 64000,
+        "mode": "chat",
+        "output_cost_per_token": 3.6e-06,
+        "supports_function_calling": true,
+        "supports_vision": true,
+        "supports_reasoning": false,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.6-35b-a3b": {
+        "input_cost_per_token": 3.75e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 229376,
+        "max_output_tokens": 64000,
+        "mode": "chat",
+        "output_cost_per_token": 2.25e-06,
+        "supports_function_calling": true,
+        "supports_vision": true,
+        "supports_reasoning": false,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.8-2.4t-a95b": {
+        "cache_read_input_token_cost": 2.5e-07,
+        "input_cost_per_token": 2e-06,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 983616,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 6e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "cache_creation_input_token_cost": 2.5e-06,
+        "explicit_cache_read_input_token_cost": 1.7e-07,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
+    "streamlake/qwen3.8-27b": {
+        "cache_read_input_token_cost": 1e-07,
+        "input_cost_per_token": 5e-07,
+        "litellm_provider": "streamlake",
+        "max_input_tokens": 983616,
+        "max_output_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3e-06,
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_vision": true,
+        "cache_creation_input_token_cost": 6.25e-07,
+        "explicit_cache_read_input_token_cost": 5e-08,
+        "supports_prompt_caching": true,
+        "source": "https://www.streamlake.ai/document/DOC/mgrnm4xm362hvp5wyce"
+    },
     "deepseek-v4-flash": {
         "cache_creation_input_token_cost": 0.0,
         "cache_read_input_token_cost": 6e-09,
@@ -62590,6 +63057,14 @@
                     "litellm_provider": "anthropic"
                 }
             },
+            {
+                "name": "typesafe-jev-ids",
+                "pattern": "^jev-(?:latest|preview|1\\.13\\.0|1\\.14\\.0)$",
+                "description": "A bare Jev id that has a typesafe/ price entry (jev-latest, jev-preview, jev-1.13.0, jev-1.14.0), the name the /typesafe pass-through forwards to TypeSafe. Listed by name, matched case-insensitively like every rule here, so foo/jev-latest or a model group named like a Jev id TypeSafe does not serve, such as jev-1.0 or jev-internal-router, never resolves to typesafe. A new Jev version goes here together with its price entry. Provider inference feeds the proxy's provider/* wildcard access checks, so this rule is what lets typesafe/* on a key, team or project allowlist cover a jev-latest request.",
+                "model_info": {
+                    "litellm_provider": "typesafe"
+                }
+            },
             {
                 "name": "claude-family-baseline",
                 "pattern": "claude-[a-z]+-\\d+(?:[-.]\\d+)?(?:-\\d{8})?",

provider_endpoints_support.json

@@ -3785,6 +3785,32 @@
         "generateContent": true
       }
     },
+    "streamlake": {
+      "display_name": "StreamLake (`streamlake`)",
+      "url": "https://docs.litellm.ai/docs/providers/openai_compatible",
+      "endpoints": {
+        "chat_completions": true,
+        "responses": true,
+        "text_completion": true,
+        "embeddings": false,
+        "image_generations": false,
+        "image_edits": false,
+        "audio_transcriptions": false,
+        "audio_speech": false,
+        "rerank": false,
+        "ocr": false,
+        "search": false,
+        "moderations": false,
+        "realtime": false,
+        "video_generations": false,
+        "batches": false,
+        "messages": true,
+        "interactions": true,
+        "systemone": false,
+        "a2a": true,
+        "generateContent": true
+      }
+    },
     "tencent": {
       "display_name": "Tencent TokenHub (`tencent`)",
       "url": "https://docs.litellm.ai/docs/providers/tencent",

tests/benchmarks/test_json_fragment_accumulator_benchmarks.py

@@ -0,0 +1,26 @@
+from collections.abc import Iterator
+from typing import Final
+
+import pytest
+
+from litellm.litellm_core_utils.json_fragment_accumulator import JSONFragmentAccumulator
+
+CONCATENATED_VALUES: Final = 20_000
+
+
+def _popped_values(accumulator: JSONFragmentAccumulator) -> Iterator[object]:
+    while True:
+        found, value = accumulator.pop_next_value()
+        if not found:
+            return
+        yield value
+
+
+@pytest.mark.benchmark
+def test_draining_many_concatenated_values() -> None:
+    accumulator: Final = JSONFragmentAccumulator()
+    accumulator.append('{"a": 1}' * CONCATENATED_VALUES)
+
+    drained: Final = sum(1 for _ in _popped_values(accumulator))
+
+    assert drained == CONCATENATED_VALUES

tests/integration/authorization/test_team_admin_gate.py

@@ -746,7 +746,7 @@ def test_internal_user_model_creation_prohibition_refuses_team_admin_but_not_pro
                     "error": {
                         "message": "Model creation is disabled for internal users by disable_model_add_for_internal_users.",
                         "type": "auth_error",
-                        "param": "None",
+                        "param": None,
                         "code": "403",
                     }
                 },

tests/integration/authorization/test_user_self_update_guard.py

@@ -38,7 +38,7 @@ def _escalation_error(field: str) -> dict[str, JsonValue]:
         "error": {
             "message": f"{{'error': \"Non-admin users cannot modify '{field}' on their own record. Contact your proxy admin.\"}}",
             "type": "auth_error",
-            "param": "None",
+            "param": None,
             "code": "403",
         }
     }
@@ -49,7 +49,7 @@ def _role_change_error() -> dict[str, JsonValue]:
         "error": {
             "message": "Only proxy admins can modify user roles.",
             "type": "auth_error",
-            "param": "None",
+            "param": None,
             "code": "403",
         }
     }

tests/integration/management/test_team_delete_large_membership.py

@@ -255,8 +255,8 @@ def _team_not_found_body(team_id: str) -> dict[str, JsonValue]:
     return {
         "error": {
             "message": f"{{'message': 'Team not found, passed team id: {team_id}.'}}",
-            "type": "auth_error",
-            "param": "None",
+            "type": "invalid_request_error",
+            "param": None,
             "code": "404",
         }
     }

tests/integration/providers/test_responses_compact_wire.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import json
 import os
 import uuid
+from types import MappingProxyType
 from typing import Final
 
 import openai
@@ -65,6 +66,11 @@ def test_responses_compact_routes_forward_complete_request_and_typed_items(gatew
         "call_id": "call_previous",
         "output": "The record is ready.",
     }
+    saved_compaction_item: Final = {
+        "id": f"cmp_{uuid.uuid4().hex}",
+        "type": "compaction",
+        "encrypted_content": "synthetic-saved-state",
+    }
     requests: Final = (
         {
             "input": "Compact this conversation.",
@@ -75,6 +81,20 @@ def test_responses_compact_routes_forward_complete_request_and_typed_items(gatew
             "tools": [_FUNCTION_TOOL],
         },
     )
+    saved_compaction_request: Final = {
+        "input": [{**saved_compaction_item, "created_by": None}, {**function_output_item, "created_by": None}],
+        "instructions": "Compact the saved conversation again.",
+    }
+    forwarded_bodies: Final = MappingProxyType(
+        {
+            None: requests[0],
+            requests[1]["instructions"]: requests[1],
+            saved_compaction_request["instructions"]: {
+                **saved_compaction_request,
+                "input": [saved_compaction_item, function_output_item],
+            },
+        }
+    )
     output_items: Final = [
         {
             "id": f"rs_compact_{uuid.uuid4().hex}",
@@ -98,18 +118,7 @@ def test_responses_compact_routes_forward_complete_request_and_typed_items(gatew
         assert request.method == "POST"
         assert request.headers["authorization"] == f"Bearer {_API_KEY}", dict(request.headers)
         body: Final = _JSON_OBJECT.validate_json(request.body)
-        compact_params: Final = (
-            {
-                "model": _MODEL,
-                **requests[0],
-            }
-            if body.get("input") == requests[0]["input"]
-            else {
-                "model": _MODEL,
-                **requests[1],
-            }
-        )
-        assert body == compact_params, request.body
+        assert body == {"model": _MODEL, **forwarded_bodies[body.get("instructions")]}, request.body
         return Reply(
             body=json.dumps(
                 {
@@ -128,16 +137,20 @@ def test_responses_compact_routes_forward_complete_request_and_typed_items(gatew
             ).encode()
         )
 
+    sent_bodies: Final = (*requests, saved_compaction_request)
     with wire_server(respond) as wire, gateway.scenario() as scenario:
         model: Final = _deployment(scenario, wire)
         for alias in _ALIASES:
-            for request_body in requests:
+            for request_body in sent_bodies:
                 response: Final = gateway.request(
                     "POST",
                     alias,
                     {"model": model, **request_body},
                 )
                 assert response.status_code == 200, response.text
+                wire_output: Final = _JSON_OBJECT.validate_json(response.content)["output"]
+                assert isinstance(wire_output, list), response.text
+                assert [item for item in wire_output if not isinstance(item, dict) or "created_by" in item] == []
                 compacted: Final = CompactedResponse.model_validate_json(response.content)
                 assert compacted.id != completed_id, compacted.model_dump_json()
                 assert same_response(compacted.id, completed_id), compacted.model_dump_json()
@@ -201,7 +214,7 @@ def test_responses_compact_routes_forward_complete_request_and_typed_items(gatew
                     "user": None,
                     "store": None,
                 }, response.text
-        assert [request.target for request in drain_contract_requests(wire)] == ["/v1/responses/compact"] * 6
+        assert [request.target for request in drain_contract_requests(wire)] == ["/v1/responses/compact"] * (len(_ALIASES) * len(sent_bodies))
 
 
 @pytest.mark.parametrize("prefix", ("/v1", "", "/openai/v1"))

tests/test_litellm_rust/ocr/test_failures.py

@@ -108,7 +108,10 @@ async def test_ocr_contract_malformed_document_is_actionable(
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("asynchronous", [False, True], ids=["sync", "async"])
-@pytest.mark.parametrize("option,value,field", [("pages", [-1], "pages"), ("features", [1], "features")])
+@pytest.mark.parametrize(
+    "option,value,field",
+    [("pages", [-1], "pages"), ("features", [1], "features"), ("outputContentFormat", "html", "drop_params")],
+)
 async def test_ocr_contract_azure_invalid_options_are_bad_requests(
     server: RecordingServer,
     asynchronous: bool,

tests/test_litellm_rust/ocr/test_requests.py

@@ -3,6 +3,7 @@ from typing import Final
 
 import pytest
 
+import litellm
 from tests.test_litellm_rust.ocr.data import (
     OCR_DOCUMENT,
     OCR_RESPONSE,
@@ -102,3 +103,32 @@ def test_native_ocr_encodes_python_file_input_and_drops_unknown_arguments(server
         "model": "mistral-ocr-latest",
         "document": {"type": "image_url", "image_url": "data:image/png;base64,YWJj"},
     }
+
+
+@pytest.mark.parametrize(
+    "drop_params,global_drop_params",
+    [(True, False), (" TRUE ", False), (None, True), (False, True)],
+    ids=["per_call", "per_call_spelled", "global", "global_wins"],
+)
+def test_azure_document_intelligence_drop_params_drops_an_unsupported_output_content_format(
+    server: RecordingServer,
+    monkeypatch: pytest.MonkeyPatch,
+    drop_params: object,
+    global_drop_params: bool,
+) -> None:
+    monkeypatch.setattr(litellm, "drop_params", global_drop_params)
+    server.default_response = ResponseSpec(
+        body={"status": "succeeded", "analyzeResult": {"pages": [{"pageNumber": 1, "lines": [{"content": "hello"}]}]}}
+    )
+
+    response: Final = call_ocr(
+        server,
+        model="azure_ai/doc-intelligence/prebuilt-layout",
+        outputContentFormat="html",
+        drop_params=drop_params,
+    )
+
+    assert_served_natively(server, response)
+    assert response.pages[0].markdown == "hello"
+    assert "outputContentFormat" not in server.requests[0].path
+    assert "drop_params" not in server.requests[0].path

tests/unit/AGENTS.md

@@ -8,6 +8,10 @@ the same tests against Google's live spec
 Unit tests must not depend on the current time or execution speed. Use fixed timestamps for fixtures
 and controlled clocks for time-dependent behavior. Incidental clock reads inside production code are allowed
 
+A unix socket goes under the `socket_dir` or `socket_path` fixture, never `tmp_path`, whose paths run past macOS's
+104-byte AF_UNIX limit and fail only on a Mac. Code that opens a socket in its own `tempfile` directory runs under
+`socket_tempdir`
+
 ## What good looks like
 
 ```python

tests/unit/conftest.py

@@ -2,6 +2,8 @@ import asyncio
 import base64
 import importlib
 import os
+import shutil
+import tempfile
 from collections.abc import Callable, Coroutine, Iterator
 from dataclasses import dataclass, field
 from pathlib import Path
@@ -217,6 +219,38 @@ def _flush_client_caches() -> None:
     _reset_aws_auth_caches()
 
 
+_MACOS_LONGEST_SOCKET_PATH: Final = 103
+_LONGEST_SOCKET_SUFFIX: Final = len("/sock12345678/d/litellm-pgbouncer-12345678/.s.PGSQL.65535")
+
+
+def _socket_root() -> str:
+    default: Final = tempfile.gettempdir()
+    fits: Final = len(os.fsencode(default)) + _LONGEST_SOCKET_SUFFIX <= _MACOS_LONGEST_SOCKET_PATH
+    return default if fits else "/tmp"
+
+
+@pytest.fixture
+def socket_dir(tmp_path: Path) -> Iterator[Path]:
+    """``tmp_path`` behind a short link, so unix socket paths fit the sun_path limit that it exceeds on macOS."""
+    link_dir: Final = Path(tempfile.mkdtemp(prefix="sock", dir=_socket_root()))
+    link: Final = link_dir / "d"
+    link.symlink_to(tmp_path, target_is_directory=True)
+    yield link
+    shutil.rmtree(link_dir, ignore_errors=True)
+
+
+@pytest.fixture
+def socket_path(socket_dir: Path) -> Path:
+    return socket_dir / "spend.sock"
+
+
+@pytest.fixture
+def socket_tempdir(socket_dir: Path, monkeypatch: pytest.MonkeyPatch) -> Path:
+    """Point ``tempfile``'s default directory at ``socket_dir``, for code that picks its own socket directory."""
+    monkeypatch.setattr(tempfile, "tempdir", str(socket_dir))
+    return socket_dir
+
+
 @pytest.fixture(autouse=True, scope="session")
 def bundled_tiktoken_cache() -> None:
     importlib.import_module("litellm.litellm_core_utils.default_encoding")

tests/unit/enterprise/enterprise_callbacks/test_callback_controls.py

@@ -1,5 +1,5 @@
 import unittest.mock as mock
-from typing import cast
+from typing import Final, cast
 from unittest.mock import MagicMock, patch
 
 import pytest
@@ -197,6 +197,17 @@ class TestEnterpriseCallbackControls:
         )
         assert result is False
 
+    def test_no_disable_request_logs_nothing(self, caplog: pytest.LogCaptureFixture) -> None:
+        litellm_params: Final = {"proxy_server_request": {"url": "test", "headers": {"user-agent": "codex"}}}
+
+        with caplog.at_level("DEBUG", logger="LiteLLM"):
+            result: Final = EnterpriseCallbackControls.is_callback_disabled_dynamically(
+                DataDogLogger.__new__(DataDogLogger), litellm_params, StandardCallbackDynamicParams()
+            )
+
+        assert result is False
+        assert caplog.records == []
+
     def test_callback_not_disabled_when_header_none(
         self, mock_premium_user, mock_request_headers
     ):

tests/unit/gateway/test_launch.py

@@ -140,7 +140,7 @@ class TestPoolDatabaseUrl:
         assert "DATABASE_URL" in outcome.reason
 
     def test_token_auth_hands_the_workers_the_pool_user_not_the_token(
-        self, password_env: dict[str, str], monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+        self, password_env: dict[str, str], monkeypatch: pytest.MonkeyPatch, socket_tempdir: Path
     ):
         monkeypatch.setenv("IAM_TOKEN_DB_AUTH", "true")
         monkeypatch.setenv("AWS_REGION_NAME", "us-east-1")
@@ -149,7 +149,7 @@ class TestPoolDatabaseUrl:
         with _minted_iam_token("MINTED_TOKEN"):
             outcome: Final = pool_database_url(
                 DatabaseURLSettings.from_env(),
-                PgBouncerSettings(enabled=True, port=port, binary=str(_fake_pooler(tmp_path))),
+                PgBouncerSettings(enabled=True, port=port, binary=str(_fake_pooler(socket_tempdir))),
                 environ,
             )
         assert isinstance(outcome, str), outcome
@@ -160,12 +160,12 @@ class TestPoolDatabaseUrl:
 
 class TestMain:
     def test_workers_inherit_the_loopback_url_the_supervisor_installed(
-        self, password_env: dict[str, str], monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+        self, password_env: dict[str, str], monkeypatch: pytest.MonkeyPatch, socket_tempdir: Path
     ):
         port: Final = _free_port()
         monkeypatch.setenv("LITELLM_PGBOUNCER_ENABLED", "true")
         monkeypatch.setenv("LITELLM_PGBOUNCER_PORT", str(port))
-        monkeypatch.setenv("LITELLM_PGBOUNCER_BINARY", str(_fake_pooler(tmp_path)))
+        monkeypatch.setenv("LITELLM_PGBOUNCER_BINARY", str(_fake_pooler(socket_tempdir)))
         monkeypatch.setenv("KEEPALIVE_TIMEOUT", "75")
         served: Final[list[tuple[str, ...]]] = []
         main(("--workers", "4"), serve=lambda argv: served.append(tuple(argv)))
@@ -183,7 +183,7 @@ class TestMain:
         assert _query(os.environ["DATABASE_URL"])["pgbouncer"] == "true"
 
     def test_iam_workers_keep_the_loopback_url_instead_of_minting_their_own(
-        self, password_env: dict[str, str], monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+        self, password_env: dict[str, str], monkeypatch: pytest.MonkeyPatch, socket_tempdir: Path
     ):
         port: Final = _free_port()
         monkeypatch.delenv("DATABASE_PASSWORD")
@@ -191,7 +191,7 @@ class TestMain:
         monkeypatch.setenv("AWS_REGION_NAME", "us-east-1")
         monkeypatch.setenv("LITELLM_PGBOUNCER_ENABLED", "true")
         monkeypatch.setenv("LITELLM_PGBOUNCER_PORT", str(port))
-        monkeypatch.setenv("LITELLM_PGBOUNCER_BINARY", str(_fake_pooler(tmp_path)))
+        monkeypatch.setenv("LITELLM_PGBOUNCER_BINARY", str(_fake_pooler(socket_tempdir)))
         served: Final[list[tuple[str, ...]]] = []
         with _minted_iam_token("SUPERVISOR_TOKEN"):
             main(("--workers", "4"), serve=lambda argv: served.append(tuple(argv)))

tests/unit/google_genai/test_google_genai_adapter.py

@@ -158,6 +158,18 @@ def test_native_function_response_preserves_json_when_media_is_malformed(invalid
     assert messages == [{"role": "tool", "tool_call_id": "call_lookup", "content": '{"ok": true}'}]
 
 
+@pytest.mark.parametrize("role", ["user", "model"])
+def test_one_object_content_parts_translate_like_their_list_form(role: str) -> None:
+    part: Final = {"text": "Reply with OK"}
+
+    def messages(parts: object) -> object:
+        return GoogleGenAIAdapter().translate_generate_content_to_completion(
+            model="anthropic/claude-haiku-5-5", contents=[{"role": role, "parts": parts}]
+        )["messages"]
+
+    assert messages(part) == messages([part]) != []
+
+
 @pytest.mark.parametrize(
     ("args", "expected_arguments"),
     [(None, "null"), ('{"query":"x"}', '"{\\"query\\":\\"x\\"}"')],

tests/unit/integrations/datadog/test_datadog.py

@@ -905,10 +905,10 @@ async def test_a_failed_request_is_delivered_as_an_error_log_that_keeps_the_erro
 
 @pytest.mark.asyncio
 async def test_a_failing_redis_cache_is_delivered_to_datadog_as_redis_warnings(
-    delivery: tuple[DataDogLogger, respx.Route], monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+    delivery: tuple[DataDogLogger, respx.Route], monkeypatch: pytest.MonkeyPatch, socket_dir: Path
 ) -> None:
     datadog_logger, intake = delivery
-    absent_socket: Final = str(tmp_path / "absent.sock")
+    absent_socket: Final = str(socket_dir / "absent.sock")
     redis_cache: Final = Cache(type="redis", url=f"unix://{absent_socket}")
     monkeypatch.setattr(redis_cache.cache.service_logger_obj, "dd_logger", datadog_logger, raising=False)
     monkeypatch.setattr(litellm, "service_callback", ["datadog"])

tests/unit/litellm_core_utils/test_get_llm_provider_logic.py

@@ -714,6 +714,48 @@ class TestClaudeModelPatternMatching:
             litellm.get_llm_provider(model="claude-opus-4-9")
 
 
+@pytest.mark.usefixtures("_vcr_outcome_gate", "isolate_litellm_state", "setup_and_teardown")
+class TestTypesafeJevPatternMatching:
+    def test_every_priced_jev_id_routes_to_typesafe(
+        self, shipped_generalizations: None
+    ) -> None:
+        from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap
+
+        priced_ids: Final = tuple(
+            key.removeprefix("typesafe/")
+            for key in GetModelCostMap.load_local_model_cost_map()
+            if key.startswith("typesafe/")
+        )
+        assert "jev-latest" in priced_ids
+        assert {
+            model: litellm.get_llm_provider(model=model)[1] for model in priced_ids
+        } == dict.fromkeys(priced_ids, "typesafe")
+
+    @pytest.mark.parametrize(
+        "model",
+        [
+            "foo/jev-latest",
+            "openjev-latest",
+            "jev-",
+            "jev-latest model",
+            "jev-internal-router",
+            "jev-2.0-internal",
+            "jev-1.0-acme.router",
+            "jev-1.0",
+            "jev-2.0",
+            "jev-1.13",
+        ],
+    )
+    def test_namespaced_or_partial_jev_name_does_not_match_rule(
+        self, model: str, shipped_generalizations: None
+    ) -> None:
+        from litellm.litellm_core_utils.fallback_generalizations import (
+            match_routing_generalization,
+        )
+
+        assert match_routing_generalization(model) is None
+
+
 @pytest.mark.parametrize("prefix", ["openai", "anthropic", "custom_openai", "openai_like"])
 def test_opencode_affinity_resolution_preserves_existing_transport(prefix: str) -> None:
     api_base: Final = "https://opencode.ai/zen/go/v1"

tests/unit/litellm_core_utils/test_json_fragment_accumulator.py

@@ -1,5 +1,8 @@
 import json
 import time
+import tracemalloc
+from collections.abc import Iterator
+from typing import Final
 from unittest.mock import patch
 
 from litellm.litellm_core_utils.json_fragment_accumulator import JSONFragmentAccumulator
@@ -175,40 +178,48 @@ def test_accumulation_of_many_fragments_is_not_quadratic():
     assert elapsed_ms < 50, f"1000-fragment append took {elapsed_ms:.1f} ms (expected < 50 ms); O(n^2) regression?"
 
 
-def test_draining_many_concatenated_values_is_not_quadratic():
-    """
-    Regression guard: peeling N JSON values already sitting in one buffer,
-    one pop_next_value() call per value with no new fragments in between,
-    must be O(n) total. Re-copying the shrinking remainder on every pop
-    (slicing a new string instead of advancing a cursor) makes total drain
-    time scale with the square of the buffer size.
+def _popped_values(accumulator: JSONFragmentAccumulator) -> Iterator[object]:
+    while True:
+        found, value = accumulator.pop_next_value()
+        if not found:
+            return
+        yield value
 
-    Uses a doubling ratio rather than an absolute ms budget so it isn't
-    flaky on a slower or busier CI runner: doubling the input should
-    roughly double an O(n) drain's time but roughly quadruple an O(n^2)
-    drain's time, and that ratio holds regardless of machine speed.
-    """
 
-    def drain_time_ms(n: int) -> float:
-        accumulator = JSONFragmentAccumulator()
-        accumulator.append('{"a": 1}' * n)
-        start = time.perf_counter()
-        drained = 0
-        while True:
-            found, _ = accumulator.pop_next_value()
-            if not found:
-                break
-            drained += 1
-        assert drained == n
-        return (time.perf_counter() - start) * 1000
+def _streamed(fragments: tuple[str, ...]) -> JSONFragmentAccumulator:
+    accumulator: Final = JSONFragmentAccumulator()
+    for fragment in fragments:
+        accumulator.append(fragment)
+    return accumulator
 
-    small_ms = drain_time_ms(40_000)
-    large_ms = drain_time_ms(80_000)
 
-    ratio = large_ms / max(small_ms, 0.001)
-    assert ratio < 3.0, (
-        f"doubling drained values scaled time by {ratio:.2f}x ({small_ms:.1f} ms -> {large_ms:.1f} ms); "
-        "expected roughly 2x for O(n); O(n^2) regression?"
+def _drain_peak_bytes(fragments: tuple[str, ...]) -> int:
+    accumulator: Final = _streamed(fragments)
+    first_found, _ = accumulator.pop_next_value()
+    assert first_found is True
+    already_tracing: Final = tracemalloc.is_tracing()
+    if not already_tracing:
+        tracemalloc.start()
+    tracemalloc.reset_peak()
+    baseline, _ = tracemalloc.get_traced_memory()
+    try:
+        drained: Final = sum(1 for _ in _popped_values(accumulator))
+        _, peak = tracemalloc.get_traced_memory()
+    finally:
+        if not already_tracing:
+            tracemalloc.stop()
+    assert drained == len(fragments) - 1
+    return peak - baseline
+
+
+def test_draining_many_concatenated_values_never_copies_the_remainder():
+    value: Final = '{"text": "' + "x" * 20 + '"}'
+    fragments: Final = tuple(value if index % 2 else f" {value}" for index in range(40_000))
+    payload_bytes: Final = sum(len(fragment) for fragment in fragments)
+    peak_bytes: Final = _drain_peak_bytes(fragments)
+    assert peak_bytes < payload_bytes // 4, (
+        f"draining {len(fragments)} values allocated a peak of {peak_bytes} bytes over a {payload_bytes} byte payload, "
+        "copying the remainder on every pop?"
     )
 
 

tests/unit/litellm_core_utils/test_streaming_handler.py

@@ -6316,6 +6316,22 @@ def test_chunk_creator_preserves_hidden_provider_specific_fields_from_parsed_chu
     assert assembled._hidden_params["provider_specific_fields"] == {"traffic_type": "ON_DEMAND_FLEX"}
 
 
+def test_the_usage_chunk_carries_the_bill_the_builder_priced_for_the_spend_row() -> None:
+    from litellm.litellm_core_utils.streaming_handler import (
+        _stamp_usage_cost,  # pyright: ignore[reportPrivateUsage]  # the helper under test
+    )
+
+    complete: Final = ModelResponse(model="omni-model", usage=Usage(prompt_tokens=100, completion_tokens=20))
+    complete.hidden_params = {"response_cost": 0.00032}
+    chunk_usage: Final = Usage(prompt_tokens=100, completion_tokens=20, total_tokens=120)
+    setattr(chunk_usage, "cost", 0.037)
+    usage_chunk: Final = ModelResponseStream(id="chunk-1", model="omni-model", choices=[], usage=chunk_usage)
+
+    _stamp_usage_cost(usage_chunk, complete)
+
+    assert usage_chunk.usage is not None and usage_chunk.usage.cost == 0.00032, "the client sees what the row bills"
+
+
 def test_chunk_creator_keeps_provider_model_private_across_stream():
     from litellm.router_utils.add_retry_fallback_headers import (
         get_hidden_params_dict,

tests/unit/llms/anthropic/messages/context_management/test_clear_tool_uses.py

@@ -10,7 +10,7 @@ from unittest.mock import AsyncMock, MagicMock
 import pytest
 
 import litellm
-from litellm.llms.anthropic.messages.adapters.handler import _run_polyfill_if_enabled
+from litellm.llms.anthropic.messages.adapters.handler import _run_polyfill_if_enabled, prepare_context_managed_request
 from litellm.llms.anthropic.messages.context_management import (
     AnthropicContextManagementError,
     PolyfillResult,
@@ -690,6 +690,115 @@ def test_tool_uses_trigger_does_not_count_excluded_tools(trigger_value: int, tri
     assert (applied is not None) == triggers
 
 
+@pytest.mark.parametrize(
+    "system",
+    ["Reference notes. " * 400, [{"type": "text", "text": "Reference notes. " * 400}]],
+    ids=["string", "blocks"],
+)
+def test_input_tokens_trigger_counts_the_system_prompt(system: str | list[dict[str, object]]) -> None:
+    """Like the API, the trigger compares the whole prompt: a system prompt can push it over."""
+    messages = _make_history(n_pairs=3)
+    history_tokens: Final = litellm.token_counter(model=MODEL, messages=messages)
+    edit_spec: Final = {
+        "type": "clear_tool_uses_20250919",
+        "trigger": {"type": "input_tokens", "value": history_tokens + 100},
+        "keep": {"type": "tool_uses", "value": 1},
+    }
+    _, without_system = apply_clear_tool_uses_20250919(
+        model=MODEL, messages=deepcopy(messages), tools=None, system=None, edit_spec=edit_spec
+    )
+    _, with_system = apply_clear_tool_uses_20250919(
+        model=MODEL, messages=deepcopy(messages), tools=None, system=system, edit_spec=edit_spec
+    )
+    _, met_without_system = apply_clear_tool_uses_20250919(
+        model=MODEL,
+        messages=deepcopy(messages),
+        tools=None,
+        system=None,
+        edit_spec={**edit_spec, "trigger": {"type": "input_tokens", "value": 0}},
+    )
+    assert without_system is None
+    assert with_system is not None
+    assert met_without_system is not None
+    assert with_system["cleared_tool_uses"] == 2
+    assert with_system["cleared_input_tokens"] == met_without_system["cleared_input_tokens"] > 0
+
+
+@pytest.mark.parametrize("compact_position", [None, "after", "before"], ids=["alone", "before_compact", "after_compact"])
+async def test_input_tokens_trigger_counts_a_client_compaction_summary(compact_position: str | None) -> None:
+    """A client compaction summary counts toward the trigger and calls beside it stay clearable, in any edit order."""
+    tail: Final = _make_history(n_pairs=3)
+    edit_spec: Final = {
+        "type": "clear_tool_uses_20250919",
+        "trigger": {"type": "input_tokens", "value": litellm.token_counter(model=MODEL, messages=tail) + 100},
+        "keep": {"type": "tool_uses", "value": 1},
+    }
+    compaction: Final = {"type": "compaction", "content": "Earlier work. " * 400}
+    compacted_history: Final = [tail[0], {**tail[1], "content": [compaction, *tail[1]["content"]]}, *tail[2:]]
+    compact_spec: Final = {"type": "compact_20260112", "trigger": {"type": "input_tokens", "value": 50_000}}
+    edits: Final = {None: [edit_spec], "after": [edit_spec, compact_spec], "before": [compact_spec, edit_spec]}[
+        compact_position
+    ]
+
+    async def prepare(messages: list[dict[str, object]]) -> PolyfillResult | None:
+        return await prepare_context_managed_request(
+            model=MODEL,
+            messages=messages,
+            tools=None,
+            system=None,
+            context_management_spec={"edits": edits},
+            litellm_metadata=None,
+            additional_drop_params=None,
+            llm_router=None,
+        )
+
+    def cleared(result: PolyfillResult | None) -> list[object]:
+        applied: Final = result.applied_edits if result is not None else []
+        return [edit.get("cleared_tool_uses") for edit in applied if edit.get("type") == "clear_tool_uses_20250919"]
+
+    without_summary: Final = await prepare(deepcopy(tail))
+    with_summary: Final = await prepare(deepcopy(compacted_history))
+
+    assert cleared(without_summary) == []
+    assert cleared(with_summary) == [2]
+
+
+async def test_compact_after_clearing_replaces_the_client_compaction_summary(monkeypatch: pytest.MonkeyPatch) -> None:
+    from litellm.proxy.proxy_server import general_settings
+
+    monkeypatch.setitem(general_settings, COMPACT_SUMMARY_MODEL_SETTING_KEY, "summary-model")
+    summary: Final = MagicMock()
+    summary.choices = [MagicMock(message=MagicMock(content="<summary>Fresh summary</summary>"))]
+    summary.usage = MagicMock(prompt_tokens=50, completion_tokens=100)
+    router: Final = MagicMock()
+    router.acompletion = AsyncMock(return_value=summary)
+    tail: Final = _make_history(n_pairs=3)
+    result: Final = await prepare_context_managed_request(
+        model=MODEL,
+        messages=[{"role": "assistant", "content": [{"type": "compaction", "content": "Old summary. " * 20_000}]}, *tail],
+        tools=None,
+        system="You are a helpful assistant.",
+        context_management_spec={
+            "edits": [
+                {
+                    "type": "clear_tool_uses_20250919",
+                    "trigger": {"type": "input_tokens", "value": 1_000},
+                    "keep": {"type": "tool_uses", "value": 1},
+                },
+                {"type": "compact_20260112", "trigger": {"type": "input_tokens", "value": 50_000}},
+            ]
+        },
+        litellm_metadata=None,
+        additional_drop_params=None,
+        llm_router=router,
+    )
+    assert result is not None
+    assert isinstance(result.system, str)
+    assert "Fresh summary" in result.system
+    assert "Old summary" not in result.system
+    assert result.system.endswith("You are a helpful assistant.")
+
+
 @pytest.mark.parametrize(("clear_tool_inputs", "triggers"), [(False, False), (True, True)])
 def test_tool_uses_trigger_counts_tiny_results_only_when_their_inputs_are_cleared(
     clear_tool_inputs: bool, triggers: bool

tests/unit/llms/anthropic/messages/context_management/test_dispatcher.py

@@ -151,6 +151,7 @@ async def test_sync_editor_counts_tokens_off_the_event_loop():
                     {
                         "type": "clear_tool_uses_20250919",
                         "trigger": {"type": "input_tokens", "value": 10_000_000},
+                        "keep": {"type": "tool_uses", "value": 0},
                     }
                 ]
             },