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

litellm-proxy-extras/litellm_proxy_extras/migrations/20260916000000_add_key_total_spend/migration.sql

@@ -0,0 +1,5 @@
+-- AlterTable
+ALTER TABLE "LiteLLM_VerificationToken" ADD COLUMN IF NOT EXISTS "total_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
+
+-- AlterTable
+ALTER TABLE "LiteLLM_DeletedVerificationToken" ADD COLUMN IF NOT EXISTS "total_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -426,6 +426,7 @@ model LiteLLM_VerificationToken {
     key_alias   String?
     soft_budget_cooldown Boolean @default(false) // key-level state on if budget alerts need to be cooled down
     spend      Float    @default(0.0)
+    total_spend Float   @default(0.0)
     expires    DateTime?
     models     String[]
     aliases    Json  @default("{}")
@@ -528,6 +529,7 @@ model LiteLLM_DeletedVerificationToken {
     key_alias               String?
     soft_budget_cooldown    Boolean  @default(false)
     spend                   Float    @default(0.0)
+    total_spend             Float    @default(0.0)
     expires                 DateTime?
     models                  String[]
     aliases                 Json     @default("{}")

litellm-rust/Cargo.lock

@@ -1951,6 +1951,7 @@ dependencies = [
 name = "litellm-python-bridge"
 version = "0.1.0"
 dependencies = [
+ "bytes",
  "criterion",
  "futures-util",
  "litellm-auth",

litellm-rust/crates/core/src/ocr/client.rs

@@ -54,9 +54,16 @@ impl OcrClient {
             match call.resume(result.take()).await? {
                 OcrCallStep::Host(OcrHostOperation::ProjectRequest) => {
                     result = Some(OcrHostResult::Request(Ok((
-                        Box::new(request.take().ok_or_else(|| {
-                            Error::InvalidRequest("OCR request was already projected".into())
-                        })?),
+                        Box::new(
+                            request
+                                .take()
+                                .ok_or_else(|| {
+                                    Error::InvalidRequest(
+                                        "OCR request was already projected".into(),
+                                    )
+                                })?
+                                .into(),
+                        ),
                         false,
                     ))))
                 }

litellm-rust/crates/core/src/ocr/document.rs

@@ -1,3 +1,6 @@
+use std::io::Read;
+use std::path::Path;
+
 use base64::{Engine, engine::general_purpose::STANDARD};
 use data_url::mime::Mime;
 use data_url::{DataUrl, DataUrlError, forgiving_base64::DecodeError};
@@ -5,12 +8,52 @@ use reqwest::Url;
 use serde_json::Map;
 
 use super::error::{OcrError, OcrRequestError, OcrResponseError};
-use super::types::{OcrConnection, OcrDocument};
+use super::types::{OcrConnection, OcrDocument, OcrDocumentInput};
 use crate::constants::{OCR_INLINE_MAX_BYTES, OCR_MAX_FETCH_REDIRECTS};
 use crate::media::Error as MediaError;
 use crate::media::{DownloadPolicy, MediaFetcher};
 use crate::transport::Error as TransportError;
 
+pub fn prepare_document(input: OcrDocumentInput) -> Result<OcrDocument, super::Error> {
+    match input {
+        OcrDocumentInput::Document(document) => Ok(document),
+        OcrDocumentInput::Path { path, mime_type } => {
+            read_path_document(&path, mime_type.as_deref())
+        }
+        OcrDocumentInput::Bytes {
+            bytes,
+            file_name,
+            mime_type,
+        } => Ok(encode_file_document(
+            &bytes,
+            file_name.as_deref(),
+            mime_type.as_deref(),
+        )?),
+        OcrDocumentInput::HostReader { .. } => Err(super::Error::InvalidRequest(
+            "OCR file reader was not read by the host".into(),
+        )),
+    }
+}
+
+pub fn read_path_document(
+    path: &Path,
+    mime_type: Option<&str>,
+) -> Result<OcrDocument, super::Error> {
+    let mut bytes = Vec::new();
+    std::fs::File::open(path)
+        .and_then(|file| {
+            file.take(OCR_INLINE_MAX_BYTES as u64 + 1)
+                .read_to_end(&mut bytes)
+        })
+        .map_err(|source| super::Error::FileRead {
+            path: path.to_owned(),
+            kind: source.kind(),
+            message: source.to_string(),
+        })?;
+    let name = path.file_name().map(|name| name.to_string_lossy());
+    Ok(encode_file_document(&bytes, name.as_deref(), mime_type)?)
+}
+
 pub fn encode_file_document(
     bytes: &[u8],
     file_name: Option<&str>,
@@ -75,18 +118,6 @@ pub fn mime_type_for_name(name: &str) -> &'static str {
     }
 }
 
-pub fn upload_mime_type<'a>(file_name: Option<&str>, content_type: Option<&'a str>) -> &'a str {
-    match content_type
-        .and_then(|value| value.split(';').next())
-        .map(str::trim)
-    {
-        Some(value) if !value.is_empty() && value != "application/octet-stream" => value,
-        _ => file_name
-            .map(mime_type_for_name)
-            .unwrap_or("application/octet-stream"),
-    }
-}
-
 pub(crate) struct InlineDocument<'a>(DataUrl<'a>);
 
 impl<'a> InlineDocument<'a> {
@@ -230,24 +261,65 @@ mod tests {
     }
 
     #[test]
-    fn upload_mime_mapping_matches_python() {
+    fn path_documents_are_read_and_named_by_core() {
+        let dir = std::env::temp_dir().join(format!("litellm-ocr-{}", rand::random::<u64>()));
+        std::fs::create_dir_all(&dir).unwrap();
+        let path = dir.join("scan.png");
+        std::fs::write(&path, b"abc").unwrap();
         assert_eq!(
-            upload_mime_type(Some("report.pdf"), Some("application/octet-stream")),
-            "application/pdf"
-        );
-        assert_eq!(upload_mime_type(Some("image.png"), None), "image/png");
-        assert_eq!(upload_mime_type(None, None), "application/octet-stream");
-        assert_eq!(
-            upload_mime_type(Some("doc.pdf"), Some("application/pdf; charset=utf-8")),
-            "application/pdf"
+            prepare_document(OcrDocumentInput::Path {
+                path: path.clone(),
+                mime_type: None,
+            })
+            .unwrap(),
+            OcrDocument::ImageUrl {
+                image_url: "data:image/png;base64,YWJj".into(),
+                extra_fields: Map::new(),
+            }
         );
         assert_eq!(
-            upload_mime_type(
-                Some("img.png"),
-                Some("image/png; charset=utf-8; boundary=something")
-            ),
-            "image/png"
+            prepare_document(OcrDocumentInput::Path {
+                path: path.clone(),
+                mime_type: Some("application/pdf".into()),
+            })
+            .unwrap(),
+            document("data:application/pdf;base64,YWJj")
         );
+        std::fs::write(&path, vec![b'a'; OCR_INLINE_MAX_BYTES + 1]).unwrap();
+        assert_eq!(
+            prepare_document(OcrDocumentInput::Path {
+                path: path.clone(),
+                mime_type: None,
+            }),
+            Err(OcrRequestError::InlineDocumentTooLarge.into())
+        );
+        std::fs::remove_dir_all(&dir).unwrap();
+
+        let missing = dir.join("missing.pdf");
+        let Err(super::super::Error::FileRead { path, kind, .. }) =
+            prepare_document(OcrDocumentInput::Path {
+                path: missing.clone(),
+                mime_type: None,
+            })
+        else {
+            panic!("missing paths must surface a file read error");
+        };
+        assert_eq!(path, missing);
+        assert_eq!(kind, std::io::ErrorKind::NotFound);
+    }
+
+    #[test]
+    fn byte_documents_are_encoded_and_host_readers_must_be_read_first() {
+        assert_eq!(
+            prepare_document(OcrDocumentInput::Bytes {
+                bytes: b"abc".as_slice().into(),
+                file_name: Some("scan.pdf".into()),
+                mime_type: None,
+            })
+            .unwrap(),
+            document("data:application/pdf;base64,YWJj")
+        );
+        assert!(prepare_document(OcrDocumentInput::HostReader { mime_type: None }).is_err());
     }
 
     #[test]

litellm-rust/crates/core/src/ocr/error.rs

@@ -50,6 +50,12 @@ pub enum Error {
     Connect(String),
     #[error("routing error: {0}")]
     Routing(String),
+    #[error("Failed to read OCR file {}: {message}", path.display())]
+    FileRead {
+        path: std::path::PathBuf,
+        kind: std::io::ErrorKind,
+        message: String,
+    },
     /// The request is outside the surface this route covers in Rust. Hosts that
     /// keep a reference implementation treat this as "fall back", not "fail".
     #[error("unsupported by the rust path: {0}")]

litellm-rust/crates/core/src/ocr/lifecycle.rs

@@ -9,6 +9,7 @@ use super::hooks::{
     OcrDuringCallRequest, OcrHookFuture, OcrHooks, OcrLogFuture, OcrPostCallRequest,
     OcrPreCallRequest,
 };
+use super::types::{OcrDocumentInput, OcrFileContent};
 use super::{LiteLLMOcrRequest, LiteLLMOcrResponse, OcrClient};
 use crate::call_lifecycle::host::{
     HostCall, HostCallFuture, HostCallStep, HostFailure, HostLifecycle, HostPhase,
@@ -52,6 +53,7 @@ impl OcrAdmission {
 #[derive(Clone, Debug)]
 pub enum OcrHostOperation {
     ProjectRequest,
+    ReadDocument,
     Lifecycle(HostPhase),
     ConstructResponse(Arc<LiteLLMOcrResponse>),
     MapFailure(Error),
@@ -83,7 +85,8 @@ impl OcrHostOperation {
 }
 
 pub enum OcrHostResult {
-    Request(Result<(Box<LiteLLMOcrRequest>, bool), Error>),
+    Request(Result<(Box<LiteLLMOcrRequest<OcrDocumentInput>>, bool), Error>),
+    Document(Result<OcrFileContent, Error>),
     Lifecycle(Result<(), HostFailure<Error>>),
     AzureAdToken(Result<ResolvedCredential, AuthError>),
     PreCall(Result<OcrPreCallRequest, Error>),
@@ -313,7 +316,7 @@ struct PendingOperation {
 
 struct OcrExecution {
     client: Option<OcrClient>,
-    request: Option<LiteLLMOcrRequest>,
+    request: Option<LiteLLMOcrRequest<OcrDocumentInput>>,
     operations_tx: mpsc::UnboundedSender<PendingOperation>,
     operations_rx: mpsc::UnboundedReceiver<PendingOperation>,
     pending_result: Option<oneshot::Sender<OcrHostResult>>,
@@ -397,12 +400,14 @@ impl OcrExecution {
                 },
             )));
         }
-        request.hooks = Arc::new(ProtocolHooks {
+        let hooks = Arc::new(ProtocolHooks {
             operations: self.operations_tx.clone(),
             intercepts_requests,
             terminal: self.terminal.clone(),
         });
+        request.hooks = hooks.clone();
         self.execution = Some(tokio::spawn(async move {
+            let request = prepare_request_document(request, &hooks).await?;
             perform_ocr_request(&client, request).await
         }));
     }
@@ -423,6 +428,39 @@ impl OcrExecution {
     }
 }
 
+async fn prepare_request_document(
+    request: LiteLLMOcrRequest<OcrDocumentInput>,
+    hooks: &ProtocolHooks,
+) -> Result<LiteLLMOcrRequest, Error> {
+    let request = match &request.document {
+        OcrDocumentInput::HostReader { mime_type } => {
+            let mime_type = mime_type.clone();
+            let content = match hooks.invoke(OcrHostOperation::ReadDocument).await? {
+                OcrHostResult::Document(result) => result?,
+                _ => {
+                    return Err(Error::InvalidRequest(
+                        "invalid OCR document read host result".into(),
+                    ));
+                }
+            };
+            request.with_document(OcrDocumentInput::Bytes {
+                bytes: content.bytes,
+                file_name: content.file_name,
+                mime_type,
+            })
+        }
+        _ => request,
+    };
+    if let OcrDocumentInput::Document(_) = &request.document {
+        return request.map_document(super::document::prepare_document);
+    }
+    tokio::task::spawn_blocking(move || request.map_document(super::document::prepare_document))
+        .await
+        .map_err(|error| {
+            Error::InvalidRequest(format!("OCR document preparation task failed: {error}"))
+        })?
+}
+
 impl Drop for OcrExecution {
     fn drop(&mut self) {
         if let Some(execution) = &self.execution {
@@ -567,6 +605,9 @@ impl OcrHost for NoopOcrHost {
                 OcrHostOperation::ProjectRequest => OcrHostResult::Request(Err(
                     Error::InvalidRequest("OCR host has no request projection".into()),
                 )),
+                OcrHostOperation::ReadDocument => OcrHostResult::Document(Err(
+                    Error::InvalidRequest("OCR host has no document reader".into()),
+                )),
                 OcrHostOperation::Lifecycle(_)
                 | OcrHostOperation::ConstructResponse(_)
                 | OcrHostOperation::MapFailure(_)
@@ -602,6 +643,9 @@ impl OcrHost for OcrHookHost {
                 OcrHostOperation::ProjectRequest => OcrHostResult::Request(Err(
                     Error::InvalidRequest("OCR hook host has no request projection".into()),
                 )),
+                OcrHostOperation::ReadDocument => OcrHostResult::Document(Err(
+                    Error::InvalidRequest("OCR hook host has no document reader".into()),
+                )),
                 OcrHostOperation::Success {
                     context,
                     response,

litellm-rust/crates/core/src/ocr/mod.rs

@@ -13,12 +13,15 @@ pub mod types;
 pub mod wire;
 
 pub use client::{OcrClient, ocr};
-pub use document::{encode_file_document, mime_type_for_name, upload_mime_type};
+pub use document::{encode_file_document, mime_type_for_name, read_path_document};
 pub use lifecycle::{
     NativeOutcome, NativeResult, NoopOcrHost, OcrAdmission, OcrCall, OcrCallStep, OcrDecline,
     OcrHookHost, OcrHost, OcrHostOperation, OcrHostResult,
 };
-pub use types::{LiteLLMOcrRequest, LiteLLMOcrResponse, OcrConnection, OcrDocument};
+pub use types::{
+    LiteLLMOcrRequest, LiteLLMOcrResponse, OcrConnection, OcrDocument, OcrDocumentInput,
+    OcrFileContent,
+};
 
 #[cfg(test)]
 #[path = "../../tests/azure_ai_ocr.rs"]

litellm-rust/crates/core/src/ocr/types.rs

@@ -1,7 +1,10 @@
 use std::collections::BTreeMap;
+use std::convert::Infallible;
+use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::Duration;
 
+use bytes::Bytes;
 use serde::{Deserialize, Serialize};
 use serde_json::{Map, Value};
 
@@ -50,6 +53,35 @@ impl OcrDocument {
     }
 }
 
+#[derive(Clone, Debug, PartialEq)]
+pub enum OcrDocumentInput {
+    Document(OcrDocument),
+    Path {
+        path: PathBuf,
+        mime_type: Option<String>,
+    },
+    Bytes {
+        bytes: Bytes,
+        file_name: Option<String>,
+        mime_type: Option<String>,
+    },
+    HostReader {
+        mime_type: Option<String>,
+    },
+}
+
+impl From<OcrDocument> for OcrDocumentInput {
+    fn from(document: OcrDocument) -> Self {
+        Self::Document(document)
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct OcrFileContent {
+    pub bytes: Bytes,
+    pub file_name: Option<String>,
+}
+
 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub enum OcrResponseFormat {
@@ -89,9 +121,9 @@ impl Default for OcrConnection {
     }
 }
 
-pub struct LiteLLMOcrRequest {
+pub struct LiteLLMOcrRequest<D = OcrDocument> {
     pub model: String,
-    pub document: OcrDocument,
+    pub document: D,
     pub connection: OcrConnection,
     pub hooks: Arc<dyn OcrHooks>,
     pub litellm_call_id: Option<String>,
@@ -101,10 +133,10 @@ pub struct LiteLLMOcrRequest {
     pub(crate) adapter: OcrAdapterKind,
 }
 
-impl LiteLLMOcrRequest {
+impl<D> LiteLLMOcrRequest<D> {
     pub fn new(
         model: String,
-        document: OcrDocument,
+        document: D,
         custom_llm_provider: Option<&str>,
         optional_params: Map<String, Value>,
     ) -> Result<Self, Error> {
@@ -151,6 +183,36 @@ impl LiteLLMOcrRequest {
             ..self
         }
     }
+
+    pub fn map_document<T, E>(
+        self,
+        map: impl FnOnce(D) -> Result<T, E>,
+    ) -> Result<LiteLLMOcrRequest<T>, E> {
+        Ok(LiteLLMOcrRequest {
+            model: self.model,
+            document: map(self.document)?,
+            connection: self.connection,
+            hooks: self.hooks,
+            litellm_call_id: self.litellm_call_id,
+            optional_params: self.optional_params,
+            input_sources: self.input_sources,
+            azure_ad_token_provider: self.azure_ad_token_provider,
+            adapter: self.adapter,
+        })
+    }
+
+    pub fn with_document<T>(self, document: T) -> LiteLLMOcrRequest<T> {
+        let Ok(request) = self.map_document(|_| Ok::<T, Infallible>(document));
+        request
+    }
+}
+
+impl From<LiteLLMOcrRequest> for LiteLLMOcrRequest<OcrDocumentInput> {
+    fn from(request: LiteLLMOcrRequest) -> Self {
+        let Ok(request) = request
+            .map_document(|document| Ok::<_, Infallible>(OcrDocumentInput::Document(document)));
+        request
+    }
 }
 
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]

litellm-rust/crates/core/src/ocr/wire.rs

@@ -68,9 +68,9 @@ pub struct DecodedOcrResponse<T> {
 
 #[derive(Deserialize)]
 #[serde(deny_unknown_fields)]
-pub struct OcrWireRequest {
+pub struct OcrWireRequest<D = Value> {
     pub model: String,
-    pub document: Value,
+    pub document: D,
     pub api_key: Option<String>,
     pub api_base: Option<String>,
     pub custom_llm_provider: Option<String>,
@@ -141,10 +141,34 @@ pub fn consumed_optional_params(
 }
 
 pub fn decode_request(wire: OcrWireRequest) -> Result<LiteLLMOcrRequest, Error> {
+    let OcrWireRequest {
+        model,
+        document,
+        api_key,
+        api_base,
+        custom_llm_provider,
+        extra_headers,
+        optional_params,
+        input_sources,
+        timeout_seconds,
+    } = wire;
+    decode_request_input(OcrWireRequest {
+        model,
+        document: decode_document(document)?,
+        api_key,
+        api_base,
+        custom_llm_provider,
+        extra_headers,
+        optional_params,
+        input_sources,
+        timeout_seconds,
+    })
+}
+
+pub fn decode_request_input<D>(wire: OcrWireRequest<D>) -> Result<LiteLLMOcrRequest<D>, Error> {
     let api_key_source = source_for(&wire.input_sources, "api_key");
     let api_base_source = source_for(&wire.input_sources, "api_base");
     let extra_headers_source = source_for(&wire.input_sources, "extra_headers");
-    let document = decode_document(wire.document)?;
     let headers = wire
         .extra_headers
         .unwrap_or_default()
@@ -183,7 +207,7 @@ pub fn decode_request(wire: OcrWireRequest) -> Result<LiteLLMOcrRequest, Error>
         .unwrap_or(defaults.max_response_bytes);
     let request = LiteLLMOcrRequest::new(
         wire.model,
-        document,
+        wire.document,
         wire.custom_llm_provider.as_deref(),
         wire.optional_params
             .into_iter()
@@ -209,14 +233,14 @@ pub fn decode_request(wire: OcrWireRequest) -> Result<LiteLLMOcrRequest, Error>
     })
 }
 
-fn decode_document(value: Value) -> Result<OcrDocument, OcrRequestError> {
+pub fn decode_document(value: Value) -> Result<OcrDocument, Error> {
     let kind = value.get("type").and_then(Value::as_str);
     let missing_url = matches!(kind, Some("document_url")) && value.get("document_url").is_none()
         || matches!(kind, Some("image_url")) && value.get("image_url").is_none();
     if missing_url {
-        return Err(OcrRequestError::MissingDocumentUrl);
+        return Err(OcrRequestError::MissingDocumentUrl.into());
     }
-    decode_request_value(value, "document")
+    Ok(decode_request_value(value, "document")?)
 }
 
 fn source_for(sources: &BTreeMap<String, InputSource>, name: &str) -> InputSource {
@@ -334,10 +358,7 @@ mod tests {
             serde_json::json!({"type": "document_url"}),
             serde_json::json!({"type": "image_url"}),
         ] {
-            assert_eq!(
-                decode_document(document),
-                Err(OcrRequestError::MissingDocumentUrl)
-            );
+            assert_eq!(decode_document(document), Err(Error::MissingDocumentUrl));
         }
     }
 }

litellm-rust/crates/core/tests/ocr.rs

@@ -348,13 +348,14 @@ async fn fallible_host_phases_do_not_replay_or_reach_transport() {
                     }
                     OcrHostOperation::ProjectRequest => {
                         result = Some(OcrHostResult::Request(Ok((
-                            Box::new(request.take().unwrap()),
+                            Box::new(request.take().unwrap().into()),
                             false,
                         ))))
                     }
                     OcrHostOperation::AcquireAzureAdToken => {
                         panic!("test request has no token provider")
                     }
+                    OcrHostOperation::ReadDocument => panic!("test request has no file reader"),
                     OcrHostOperation::PreCall(request) => {
                         phases.push("pre");
                         result = Some(OcrHostResult::PreCall(if failure_phase == "pre" {
@@ -405,7 +406,7 @@ async fn invalid_provider_response_runs_post_call_before_normalization_failure()
         match call.resume(result.take()).await {
             Ok(OcrCallStep::Host(OcrHostOperation::ProjectRequest)) => {
                 result = Some(OcrHostResult::Request(Ok((
-                    Box::new(request.take().unwrap()),
+                    Box::new(request.take().unwrap().into()),
                     false,
                 ))));
             }
@@ -467,9 +468,10 @@ async fn direct_native_host_drives_the_same_state_machine() {
                     _ => panic!("unexpected OCR operation"),
                 });
                 result = Some(match operation {
-                    OcrHostOperation::ProjectRequest => {
-                        OcrHostResult::Request(Ok((Box::new(request.take().unwrap()), false)))
-                    }
+                    OcrHostOperation::ProjectRequest => OcrHostResult::Request(Ok((
+                        Box::new(request.take().unwrap().into()),
+                        false,
+                    ))),
                     operation => host.invoke(operation).await,
                 });
             }
@@ -501,6 +503,137 @@ async fn direct_native_host_drives_the_same_state_machine() {
     ));
 }
 
+async fn drive_native_file_call(
+    request: super::LiteLLMOcrRequest<super::OcrDocumentInput>,
+    content: Result<super::OcrFileContent, crate::ocr::Error>,
+) -> (Result<super::LiteLLMOcrResponse, crate::ocr::Error>, usize) {
+    let NativeOutcome::Completed(mut call) =
+        OcrCall::admit(super::test_support::ocr_client(), OcrAdmission::all())
+    else {
+        panic!("supported call declined")
+    };
+    let mut request = Some(request);
+    let mut content = Some(content);
+    let mut result = None;
+    let mut reads = 0;
+    let outcome = loop {
+        match call.resume(result.take()).await {
+            Ok(OcrCallStep::Host(OcrHostOperation::ProjectRequest)) => {
+                result = Some(OcrHostResult::Request(Ok((
+                    Box::new(request.take().unwrap()),
+                    false,
+                ))));
+            }
+            Ok(OcrCallStep::Host(OcrHostOperation::ReadDocument)) => {
+                reads += 1;
+                result = Some(OcrHostResult::Document(content.take().unwrap()));
+            }
+            Ok(OcrCallStep::Host(operation)) => result = Some(NoopOcrHost.invoke(operation).await),
+            Ok(OcrCallStep::Complete(response)) => break Ok(response),
+            Err(error) => break Err(error),
+        }
+    };
+    (outcome, reads)
+}
+
+#[tokio::test]
+async fn host_reader_documents_are_read_once_at_the_core_selected_point_and_encoded() {
+    let (base, seen, server) = mock_server(vec![MockResponse::json(json!({
+        "pages":[{"index":0,"markdown":"file"}]
+    }))])
+    .await;
+    let request = wire_request("mistral/model", &base, json!({})).with_document(
+        super::OcrDocumentInput::HostReader {
+            mime_type: Some("application/pdf".into()),
+        },
+    );
+    let (response, reads) = drive_native_file_call(
+        request,
+        Ok(super::OcrFileContent {
+            bytes: b"abc".as_slice().into(),
+            file_name: Some("scan.png".into()),
+        }),
+    )
+    .await;
+    server.await.unwrap();
+    assert_eq!(response.unwrap().pages[0]["markdown"], "file");
+    assert_eq!(reads, 1);
+    assert!(seen.lock().unwrap()[0].contains("data:application/pdf;base64,YWJj"));
+}
+
+#[tokio::test]
+async fn host_reader_failures_and_empty_files_fail_before_the_provider_is_called() {
+    let (base, seen, _server) = mock_server(vec![]).await;
+    let request = wire_request("mistral/model", &base, json!({}));
+    let failure = crate::ocr::Error::InvalidRequest("reader exploded".into());
+    let (response, reads) = drive_native_file_call(
+        request.with_document(super::OcrDocumentInput::HostReader { mime_type: None }),
+        Err(failure.clone()),
+    )
+    .await;
+    assert_eq!(response.unwrap_err(), failure);
+    assert_eq!(reads, 1);
+
+    let request = wire_request("mistral/model", &base, json!({}));
+    let (response, _) = drive_native_file_call(
+        request.with_document(super::OcrDocumentInput::HostReader { mime_type: None }),
+        Ok(super::OcrFileContent {
+            bytes: Default::default(),
+            file_name: None,
+        }),
+    )
+    .await;
+    assert!(matches!(
+        response.unwrap_err(),
+        crate::ocr::Error::InvalidRequest(_)
+    ));
+    assert!(seen.lock().unwrap().is_empty());
+}
+
+#[tokio::test]
+async fn path_documents_are_read_by_core_without_a_host_operation() {
+    let (base, seen, server) = mock_server(vec![MockResponse::json(json!({
+        "pages":[{"index":0,"markdown":"path"}]
+    }))])
+    .await;
+    let dir = std::env::temp_dir().join(format!("litellm-ocr-{}", rand::random::<u64>()));
+    std::fs::create_dir_all(&dir).unwrap();
+    let path = dir.join("scan.png");
+    std::fs::write(&path, b"abc").unwrap();
+    let request = wire_request("mistral/model", &base, json!({})).with_document(
+        super::OcrDocumentInput::Path {
+            path: path.clone(),
+            mime_type: None,
+        },
+    );
+    let (response, reads) = drive_native_file_call(
+        request,
+        Err(crate::ocr::Error::InvalidRequest("unused".into())),
+    )
+    .await;
+    server.await.unwrap();
+    std::fs::remove_dir_all(&dir).unwrap();
+    assert_eq!(response.unwrap().pages[0]["markdown"], "path");
+    assert_eq!(reads, 0);
+    assert!(seen.lock().unwrap()[0].contains("data:image/png;base64,YWJj"));
+
+    let (base, seen, _server) = mock_server(vec![]).await;
+    let request = wire_request("mistral/model", &base, json!({}));
+    let (response, _) = drive_native_file_call(
+        request.with_document(super::OcrDocumentInput::Path {
+            path: path.clone(),
+            mime_type: None,
+        }),
+        Err(crate::ocr::Error::InvalidRequest("unused".into())),
+    )
+    .await;
+    assert!(matches!(
+        response.unwrap_err(),
+        crate::ocr::Error::FileRead { path: failed, kind: std::io::ErrorKind::NotFound, .. } if failed == path
+    ));
+    assert!(seen.lock().unwrap().is_empty());
+}
+
 #[tokio::test]
 async fn public_finalization_failure_never_dispatches_success_or_replays_provider() {
     use crate::call_lifecycle::host::{HostFailure, HostPhase};
@@ -543,9 +676,10 @@ async fn public_finalization_failure_never_dispatches_success_or_replays_provide
                     | OcrHostOperation::Lifecycle(HostPhase::DeploymentFailure) => {
                         panic!("finalization failure used provider/success dispatch")
                     }
-                    OcrHostOperation::ProjectRequest => {
-                        OcrHostResult::Request(Ok((Box::new(request.take().unwrap()), false)))
-                    }
+                    OcrHostOperation::ProjectRequest => OcrHostResult::Request(Ok((
+                        Box::new(request.take().unwrap().into()),
+                        false,
+                    ))),
                     operation => host.invoke(operation).await,
                 });
             }
@@ -582,7 +716,7 @@ async fn cancellation_at_provider_hook_prevents_execution_and_further_resumption
             OcrCallStep::Host(OcrHostOperation::PreCall(_)) => break,
             OcrCallStep::Host(OcrHostOperation::ProjectRequest) => {
                 result = Some(OcrHostResult::Request(Ok((
-                    Box::new(request.take().unwrap()),
+                    Box::new(request.take().unwrap().into()),
                     false,
                 ))))
             }
@@ -799,7 +933,7 @@ async fn cancellation_waits_for_provider_capture_drop_even_when_acknowledgement_
                     _ = entered.notified() => break,
                     step = call.resume(result.take()) => {
                         result = Some(match step.unwrap() {
-                            OcrCallStep::Host(OcrHostOperation::ProjectRequest) => OcrHostResult::Request(Ok((Box::new(request.take().unwrap()), false))),
+                            OcrCallStep::Host(OcrHostOperation::ProjectRequest) => OcrHostResult::Request(Ok((Box::new(request.take().unwrap().into()), false))),
                             OcrCallStep::Host(operation) => NoopOcrHost.invoke(operation).await,
                             OcrCallStep::Complete(_) => panic!("pending provider completed"),
                         });

litellm-rust/crates/python-bridge/Cargo.toml

@@ -16,6 +16,7 @@ extension-module = ["pyo3/extension-module"]
 panic-test = []
 
 [dependencies]
+bytes.workspace = true
 futures-util.workspace = true
 litellm-core.workspace = true
 litellm-auth.workspace = true

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

@@ -1,97 +1,56 @@
-use std::io::Read;
 use std::path::PathBuf;
 
-use pyo3::exceptions::{PyFileNotFoundError, PyTypeError, PyValueError};
+use bytes::Bytes;
+use pyo3::exceptions::{PyTypeError, PyValueError};
+use pyo3::gc::{PyTraverseError, PyVisit};
 use pyo3::prelude::*;
 use pyo3::pybacked::PyBackedBytes;
-#[cfg(test)]
-use pyo3::types::PyDict;
 use pyo3::types::{PyBytes, PyString};
 
-use litellm_core::constants::OCR_INLINE_MAX_BYTES;
-use litellm_core::ocr::{OcrDocument, encode_file_document, mime_type_for_name, upload_mime_type};
-use litellm_python_interop::to_py_preserving_errors;
+use litellm_core::ocr::{OcrDocumentInput, OcrFileContent};
 
-enum FileBytes {
-    Python(PyBackedBytes),
-    Native(Vec<u8>),
+#[derive(Debug)]
+pub(super) struct PythonFileReader {
+    reader: Py<PyAny>,
+    name: Option<String>,
 }
 
-impl AsRef<[u8]> for FileBytes {
-    fn as_ref(&self) -> &[u8] {
-        match self {
-            Self::Python(bytes) => bytes,
-            Self::Native(bytes) => bytes,
-        }
+impl PythonFileReader {
+    pub(super) fn read(&self, py: Python<'_>) -> PyResult<OcrFileContent> {
+        let value = self.reader.bind(py).call0()?;
+        let bytes = if value.is_instance_of::<PyString>() {
+            Bytes::from(value.extract::<String>()?)
+        } else if value.is_instance_of::<PyBytes>() {
+            extract_bytes(&value)?
+        } else {
+            return Err(PyTypeError::new_err(format!(
+                "OCR file read must return bytes or str, got {}",
+                value.get_type(),
+            )));
+        };
+        Ok(OcrFileContent {
+            bytes,
+            file_name: self.name.clone(),
+        })
+    }
+
+    pub(super) fn traverse(&self, visit: &PyVisit<'_>) -> Result<(), PyTraverseError> {
+        visit.call(&self.reader)
     }
 }
 
-fn read_file_input(
-    py: Python<'_>,
-    file: &Bound<'_, PyAny>,
-) -> PyResult<(FileBytes, Option<String>)> {
-    if file.is_instance_of::<PyString>() {
-        return Err(PyValueError::new_err(
-            "OCR file input does not accept bare str values. Pass bytes, a pathlib.Path, or a file-like object.",
-        ));
+fn extract_bytes(value: &Bound<'_, PyAny>) -> PyResult<Bytes> {
+    if value.is_exact_instance_of::<PyBytes>() {
+        return Ok(Bytes::from_owner(value.extract::<PyBackedBytes>()?));
     }
-    if file.is_instance(&py.import("os")?.getattr("PathLike")?)? {
-        let path: PathBuf = file.extract()?;
-        let name = path
-            .file_name()
-            .map(|value| value.to_string_lossy().into_owned());
-        let bytes = py
-            .detach(|| {
-                let mut bytes = Vec::new();
-                std::fs::File::open(&path)?
-                    .take(OCR_INLINE_MAX_BYTES as u64 + 1)
-                    .read_to_end(&mut bytes)?;
-                Ok::<_, std::io::Error>(bytes)
-            })
-            .map_err(|error| {
-                if error.kind() == std::io::ErrorKind::NotFound {
-                    PyFileNotFoundError::new_err(format!("File not found: {}", path.display()))
-                } else {
-                    error.into()
-                }
-            })?;
-        return Ok((FileBytes::Native(bytes), name));
-    }
-    if file.is_instance_of::<PyBytes>() {
-        return Ok((FileBytes::Python(file.extract()?), None));
-    }
-    let reader = file
-        .getattr_opt("read")?
-        .filter(|value| value.is_callable());
-    let Some(reader) = reader else {
-        return Err(PyValueError::new_err(format!(
-            "Unsupported file input type: {}. Expected pathlib.Path, bytes, or a file-like object.",
-            file.get_type(),
-        )));
-    };
-    let name = file
-        .getattr_opt("name")?
-        .filter(|value| !value.is_none())
-        .map(|value| value.extract::<String>())
-        .transpose()?;
-    let value = reader.call0()?;
-    let bytes = if value.is_instance_of::<PyString>() {
-        FileBytes::Native(value.extract::<String>()?.into_bytes())
-    } else if value.is_instance_of::<PyBytes>() {
-        FileBytes::Python(value.extract()?)
-    } else {
-        return Err(PyTypeError::new_err(format!(
-            "OCR file read must return bytes or str, got {}",
-            value.get_type(),
-        )));
-    };
-    Ok((bytes, name))
+    Ok(Bytes::copy_from_slice(
+        value.extract::<PyBackedBytes>()?.as_ref(),
+    ))
 }
 
 pub(super) struct FileDocumentInput {
-    bytes: FileBytes,
-    name: Option<String>,
-    mime_type: Option<String>,
+    pub input: OcrDocumentInput,
+    pub reader: Option<PythonFileReader>,
 }
 
 impl FromPyObject<'_, '_> for FileDocumentInput {
@@ -104,79 +63,79 @@ impl FromPyObject<'_, '_> for FileDocumentInput {
             Err(error) if error.is_instance_of::<pyo3::exceptions::PyKeyError>(py) => None,
             Err(error) => return Err(error),
         };
+        let missing = || {
+            PyValueError::new_err(
+                "document with type='file' must include a 'file' field containing a pathlib.Path, file-like object, or bytes",
+            )
+        };
         let file = document.get_item("file").map_err(|error| {
             if error.is_instance_of::<pyo3::exceptions::PyKeyError>(py) {
-                PyValueError::new_err("document with type='file' must include a 'file' field containing a pathlib.Path, file-like object, or bytes")
+                missing()
             } else {
                 error
             }
         })?;
         if file.is_none() {
+            return Err(missing());
+        }
+        if file.is_instance_of::<PyString>() {
             return Err(PyValueError::new_err(
-                "document with type='file' must include a 'file' field containing a pathlib.Path, file-like object, or bytes",
+                "OCR file input does not accept bare str values. Pass bytes, a pathlib.Path, or a file-like object.",
             ));
         }
-        let (bytes, name) = read_file_input(py, &file)?;
+        if file.is_instance(&py.import("os")?.getattr("PathLike")?)? {
+            return Ok(Self {
+                input: OcrDocumentInput::Path {
+                    path: file.extract::<PathBuf>()?,
+                    mime_type,
+                },
+                reader: None,
+            });
+        }
+        if file.is_instance_of::<PyBytes>() {
+            return Ok(Self {
+                input: OcrDocumentInput::Bytes {
+                    bytes: extract_bytes(&file)?,
+                    file_name: None,
+                    mime_type,
+                },
+                reader: None,
+            });
+        }
+        let reader = file
+            .getattr_opt("read")?
+            .filter(|value| value.is_callable());
+        let Some(reader) = reader else {
+            return Err(PyValueError::new_err(format!(
+                "Unsupported file input type: {}. Expected pathlib.Path, bytes, or a file-like object.",
+                file.get_type(),
+            )));
+        };
+        let name = file
+            .getattr_opt("name")?
+            .filter(|value| !value.is_none())
+            .map(|value| value.extract::<String>())
+            .transpose()?;
         Ok(Self {
-            bytes,
-            name,
-            mime_type,
+            input: OcrDocumentInput::HostReader { mime_type },
+            reader: Some(PythonFileReader {
+                reader: reader.unbind(),
+                name,
+            }),
         })
     }
 }
 
-pub(super) fn file_document(py: Python<'_>, document: FileDocumentInput) -> PyResult<OcrDocument> {
-    py.detach(|| {
-        encode_file_document(
-            document.bytes.as_ref(),
-            document.name.as_deref(),
-            document.mime_type.as_deref(),
-        )
-    })
-    .map_err(|error| PyValueError::new_err(error.to_string()))
-}
-
-#[pyfunction]
-fn _ocr_file_document(py: Python<'_>, document: Bound<'_, PyAny>) -> PyResult<Py<PyAny>> {
-    to_py_preserving_errors(py, &file_document(py, document.extract()?)?)
-}
-
-#[pyfunction]
-fn _ocr_mime_type(file_name: &str) -> String {
-    mime_type_for_name(file_name).into()
-}
-
-#[pyfunction]
-#[pyo3(signature = (file_content, file_name=None, content_type=None))]
-fn _ocr_upload_document(
-    py: Python<'_>,
-    file_content: &Bound<'_, PyBytes>,
-    file_name: Option<&str>,
-    content_type: Option<&str>,
-) -> PyResult<Py<PyAny>> {
-    let bytes: PyBackedBytes = file_content.extract()?;
-    let document = py
-        .detach(|| {
-            encode_file_document(
-                &bytes,
-                None,
-                Some(upload_mime_type(file_name, content_type)),
-            )
-        })
-        .map_err(|error| PyValueError::new_err(error.to_string()))?;
-    to_py_preserving_errors(py, &document)
-}
-
-pub(super) fn register(module: &Bound<'_, PyModule>) -> PyResult<()> {
-    module.add("_OCR_MAX_FILE_BYTES", OCR_INLINE_MAX_BYTES)?;
-    module.add_function(wrap_pyfunction!(_ocr_upload_document, module)?)?;
-    module.add_function(wrap_pyfunction!(_ocr_file_document, module)?)?;
-    module.add_function(wrap_pyfunction!(_ocr_mime_type, module)?)
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
+    use pyo3::types::PyDict;
+
+    fn eval<'py>(py: Python<'py>, source: &std::ffi::CStr) -> Bound<'py, PyDict> {
+        let locals = PyDict::new(py);
+        py.run(source, Some(&locals), Some(&locals)).unwrap();
+        locals
+    }
 
     #[test]
     fn extraction_validates_required_file_and_optional_mime_type() {
@@ -196,69 +155,148 @@ mod tests {
                 let error = document.extract::<FileDocumentInput>().err().unwrap();
                 assert!(error.is_instance_of::<PyTypeError>(py));
             }
-            let document = py.eval(c"{'file': b'abc'}", None, None).unwrap();
+            let error = py
+                .eval(c"{'file': 'scan.pdf'}", None, None)
+                .unwrap()
+                .extract::<FileDocumentInput>()
+                .err()
+                .unwrap();
+            assert!(error.is_instance_of::<PyValueError>(py));
+            assert!(error.to_string().contains("bare str"));
+            let document = py
+                .eval(c"{'file': b'abc', 'mime_type': 'image/png'}", None, None)
+                .unwrap();
             let input: FileDocumentInput = document.extract().unwrap();
-            assert_eq!(input.bytes.as_ref(), b"abc");
-            assert_eq!(input.name, None);
-            assert_eq!(input.mime_type, None);
+            assert!(input.reader.is_none());
+            assert_eq!(
+                input.input,
+                OcrDocumentInput::Bytes {
+                    bytes: b"abc".as_slice().into(),
+                    file_name: None,
+                    mime_type: Some("image/png".into()),
+                }
+            );
         });
     }
 
     #[test]
-    fn extraction_validates_mime_type_before_consuming_file() {
+    fn paths_and_readers_are_projected_without_io() {
         Python::initialize();
         Python::attach(|py| {
-            let locals = PyDict::new(py);
-            py.run(
-                c"class Reader:
+            let locals = eval(
+                py,
+                c"from pathlib import Path
+class Reader:
+    name = 'scan.png'
     def __init__(self):
         self.reads = 0
     def read(self):
         self.reads += 1
         return b'abc'
 reader = Reader()
-document = {'file': reader, 'mime_type': 7}",
-                Some(&locals),
-                Some(&locals),
-            )
-            .unwrap();
+document = {'file': reader, 'mime_type': 7}
+reader_document = {'file': reader}
+path_document = {'file': Path('/nonexistent/ocr-projection-test.pdf'), 'mime_type': 'image/png'}",
+            );
             let document = locals.get_item("document").unwrap().unwrap();
             let error = document.extract::<FileDocumentInput>().err().unwrap();
             assert!(error.is_instance_of::<PyTypeError>(py));
-            let reads: usize = locals
-                .get_item("reader")
-                .unwrap()
-                .unwrap()
-                .getattr("reads")
-                .unwrap()
-                .extract()
-                .unwrap();
-            assert_eq!(reads, 0);
+
+            let document = locals.get_item("reader_document").unwrap().unwrap();
+            let input: FileDocumentInput = document.extract().unwrap();
+            assert_eq!(
+                input.input,
+                OcrDocumentInput::HostReader { mime_type: None }
+            );
+            let reads = || {
+                locals
+                    .get_item("reader")
+                    .unwrap()
+                    .unwrap()
+                    .getattr("reads")
+                    .unwrap()
+                    .extract::<usize>()
+                    .unwrap()
+            };
+            assert_eq!(reads(), 0);
+            let content = input.reader.unwrap().read(py).unwrap();
+            assert_eq!(reads(), 1);
+            assert_eq!(
+                content,
+                OcrFileContent {
+                    bytes: b"abc".as_slice().into(),
+                    file_name: Some("scan.png".into()),
+                }
+            );
+
+            let document = locals.get_item("path_document").unwrap().unwrap();
+            let input: FileDocumentInput = document.extract().unwrap();
+            assert!(input.reader.is_none());
+            assert_eq!(
+                input.input,
+                OcrDocumentInput::Path {
+                    path: PathBuf::from("/nonexistent/ocr-projection-test.pdf"),
+                    mime_type: Some("image/png".into()),
+                }
+            );
         });
     }
 
     #[test]
-    fn extraction_preserves_reader_key_error_identity() {
+    fn reader_results_are_normalized_and_exceptions_keep_their_identity() {
         Python::initialize();
         Python::attach(|py| {
-            let locals = PyDict::new(py);
-            py.run(
+            let locals = eval(
+                py,
                 c"failure = KeyError('reader failed')
-class Reader:
+class Raising:
     def read(self):
         raise failure
-document = {'file': Reader()}",
-                Some(&locals),
-                Some(&locals),
-            )
-            .unwrap();
-            let document = locals.get_item("document").unwrap().unwrap();
-            let error = document.extract::<FileDocumentInput>().err().unwrap();
+class Text:
+    def read(self):
+        return 'héllo'
+class Wrong:
+    def read(self):
+        return 7
+raising = {'file': Raising()}
+text = {'file': Text()}
+wrong = {'file': Wrong()}",
+            );
+            let reader = |name: &str| {
+                locals
+                    .get_item(name)
+                    .unwrap()
+                    .unwrap()
+                    .extract::<FileDocumentInput>()
+                    .unwrap()
+                    .reader
+                    .unwrap()
+            };
+            let error = reader("raising").read(py).unwrap_err();
             assert!(
                 error
                     .value(py)
                     .is(locals.get_item("failure").unwrap().unwrap())
             );
+            assert_eq!(
+                reader("text").read(py).unwrap().bytes.as_ref(),
+                "héllo".as_bytes()
+            );
+            let error = reader("wrong").read(py).unwrap_err();
+            assert!(error.is_instance_of::<PyTypeError>(py));
+            assert!(error.to_string().contains("bytes or str"));
         });
     }
+
+    #[test]
+    fn exact_python_bytes_transfer_without_copying_and_outlive_the_input() {
+        Python::initialize();
+        let (bytes, pointer) = Python::attach(|py| {
+            let value = PyBytes::new(py, b"document bytes");
+            let pointer = value.as_bytes().as_ptr() as usize;
+            (extract_bytes(value.as_any()).unwrap(), pointer)
+        });
+        assert_eq!(bytes.as_ptr() as usize, pointer);
+        assert_eq!(bytes.as_ref(), b"document bytes");
+    }
 }

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

@@ -1,4 +1,5 @@
 use litellm_core::ocr::Error;
+use pyo3::exceptions::{PyFileNotFoundError, PyOSError};
 use pyo3::prelude::*;
 
 use crate::errors::{RustUpstreamError, core_error_to_pyerr};
@@ -7,6 +8,12 @@ pub(super) fn to_pyerr(error: Error) -> PyErr {
     let status = error.http_status_code();
     let mapped = match error {
         Error::Http { status, body } => RustUpstreamError::new_err((status, body)),
+        Error::FileRead {
+            path,
+            kind: std::io::ErrorKind::NotFound,
+            ..
+        } => PyFileNotFoundError::new_err(format!("File not found: {}", path.display())),
+        Error::FileRead { message, .. } => PyOSError::new_err(message),
         other => core_error_to_pyerr(other.into()),
     };
     attach_status(mapped, status)

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

@@ -66,13 +66,27 @@ impl PythonOcrHost {
                 retained_fields.set_item(name, value)?;
             }
         }
-        retained_fields.set_item("document", &self.projected()?.fields.document)?;
         let projected = self.projected_mut()?;
+        let document = match &projected.fields.document {
+            Some(document) => document.clone_ref(py),
+            None => to_py(py, &request.document)?,
+        };
+        retained_fields.set_item("document", &document)?;
+        projected.fields.document = Some(document);
         projected.retained_fields = Some(retained_fields.unbind());
         projected.pre_call = Some((&request).into());
         Ok(request)
     }
 
+    fn read_document(&self, py: Python<'_>) -> PyResult<litellm_core::ocr::OcrFileContent> {
+        self.projected()?
+            .fields
+            .reader
+            .as_ref()
+            .ok_or_else(missing_state)?
+            .read(py)
+    }
+
     fn acquire_azure_ad_token(&self, py: Python<'_>) -> PyResult<ResolvedCredential> {
         let provider = self
             .projected()?
@@ -193,7 +207,7 @@ impl PythonRoute for PythonOcrHost {
                 let OcrHostData::Unprojected { request } = &self.data else {
                     return Err(missing_state());
                 };
-                let projected = project_request(py, request.bind(py), self.state.kwargs.bind(py))?;
+                let projected = project_request(request.bind(py), self.state.kwargs.bind(py))?;
                 let has_token_provider = projected.fields.azure_ad_token_provider.is_some();
                 let request = projected.request;
                 self.data = OcrHostData::Projected(Box::new(ProjectedOcrHost {
@@ -205,6 +219,7 @@ impl PythonRoute for PythonOcrHost {
                 }));
                 OcrHostResult::Request(Ok((Box::new(request), has_token_provider)))
             }
+            OcrHostOperation::ReadDocument => OcrHostResult::Document(Ok(self.read_document(py)?)),
             OcrHostOperation::AcquireAzureAdToken => {
                 OcrHostResult::AzureAdToken(Ok(self.acquire_azure_ad_token(py)?))
             }
@@ -258,6 +273,9 @@ impl PythonRoute for PythonOcrHost {
             OcrHostData::Projected(projected) => {
                 visit.call(&projected.fields.boundary_request)?;
                 visit.call(&projected.fields.document)?;
+                if let Some(reader) = &projected.fields.reader {
+                    reader.traverse(visit)?;
+                }
                 visit.call(&projected.fields.api_key)?;
                 if let Some(provider) = &projected.fields.azure_ad_token_provider {
                     provider.traverse(visit)?;

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

@@ -9,6 +9,5 @@ use pyo3::prelude::*;
 
 pub(super) fn register(module: &Bound<'_, PyModule>) -> PyResult<()> {
     value::register(module)?;
-    document::register(module)?;
     lifecycle::register(module)
 }

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

@@ -1,14 +1,15 @@
 use std::sync::Arc;
 
-use litellm_core::ocr::wire::{OcrWireRequest, consumed_optional_params, decode_request};
-use litellm_core::ocr::{LiteLLMOcrRequest, NativeOutcome, OcrCall};
-use litellm_python_interop::{
-    from_py_preserving_errors as from_py, to_py_preserving_errors as to_py,
+use litellm_core::ocr::wire::{
+    OcrWireRequest, consumed_optional_params, decode_document, decode_request_input,
 };
+use litellm_core::ocr::{LiteLLMOcrRequest, NativeOutcome, OcrCall, OcrDocumentInput};
+use litellm_python_interop::from_py_preserving_errors as from_py;
 use pyo3::prelude::*;
 use pyo3::types::PyDict;
 use serde_json::{Map, Value};
 
+use super::document::{FileDocumentInput, PythonFileReader};
 use super::errors::to_pyerr as ocr_error_to_pyerr;
 use super::lifecycle::BridgeOcrHooks;
 use crate::auth::{AZURE_AD_TOKEN_PROVIDER, PythonTokenProvider};
@@ -17,7 +18,8 @@ use crate::marshal::{project_optional_fields, python_timeout_seconds, request_in
 
 pub(super) struct ProjectedOcrFields {
     pub boundary_request: Py<PyAny>,
-    pub document: Py<PyAny>,
+    pub document: Option<Py<PyAny>>,
+    pub reader: Option<PythonFileReader>,
     pub api_key: Py<PyAny>,
     pub azure_ad_token_provider: Option<PythonTokenProvider>,
     pub provider: &'static str,
@@ -25,7 +27,7 @@ pub(super) struct ProjectedOcrFields {
 }
 
 pub(super) struct ProjectedOcrCall {
-    pub request: LiteLLMOcrRequest,
+    pub request: LiteLLMOcrRequest<OcrDocumentInput>,
     pub fields: ProjectedOcrFields,
 }
 
@@ -80,12 +82,12 @@ impl<'py> OcrArguments<'_, 'py> {
 }
 
 enum ProjectedDocument {
-    File { wire: Value, retained: Py<PyAny> },
+    File(FileDocumentInput),
     Other { wire: Value, retained: Py<PyAny> },
 }
 
 impl ProjectedDocument {
-    fn project(py: Python<'_>, document: &Bound<'_, PyAny>) -> PyResult<Self> {
+    fn project(document: &Bound<'_, PyAny>) -> PyResult<Self> {
         let kind: String = document.get_item("type")?.extract()?;
         if kind != "file" {
             return Ok(Self::Other {
@@ -93,25 +95,28 @@ impl ProjectedDocument {
                 retained: document.clone().unbind(),
             });
         }
-        let input = document.extract()?;
-        let encoded = super::document::file_document(py, input)?;
-        let wire = serde_json::to_value(encoded)
-            .map_err(|error| pyo3::exceptions::PyValueError::new_err(error.to_string()))?;
-        Ok(Self::File {
-            retained: to_py(py, &wire)?,
-            wire,
-        })
+        Ok(Self::File(document.extract()?))
     }
 
-    fn into_parts(self) -> (Value, Py<PyAny>) {
+    fn into_parts(
+        self,
+    ) -> PyResult<(
+        OcrDocumentInput,
+        Option<Py<PyAny>>,
+        Option<PythonFileReader>,
+    )> {
         match self {
-            Self::File { wire, retained } | Self::Other { wire, retained } => (wire, retained),
+            Self::File(FileDocumentInput { input, reader }) => Ok((input, None, reader)),
+            Self::Other { wire, retained } => Ok((
+                decode_document(wire).map_err(ocr_error_to_pyerr)?.into(),
+                Some(retained),
+                None,
+            )),
         }
     }
 }
 
 pub(super) fn project_request(
-    py: Python<'_>,
     request: &Bound<'_, PyAny>,
     kwargs: &Bound<'_, PyDict>,
 ) -> PyResult<ProjectedOcrCall> {
@@ -119,8 +124,7 @@ pub(super) fn project_request(
     let arguments = OcrArguments { request, kwargs };
     let model = arguments.model()?;
     let custom_llm_provider = arguments.custom_llm_provider()?;
-    let (wire_document, retained_document) =
-        ProjectedDocument::project(py, &arguments.document()?)?.into_parts();
+    let document = ProjectedDocument::project(&arguments.document()?)?;
     let api_key = arguments.api_key()?;
     let specs = consumed_optional_params(&model, custom_llm_provider.as_deref())
         .map_err(ocr_error_to_pyerr)?;
@@ -136,9 +140,10 @@ pub(super) fn project_request(
     let azure_ad_token_provider = kwargs
         .get_item("azure_ad_token_provider")?
         .and_then(|provider| PythonTokenProvider::select(provider, AZURE_AD_TOKEN_PROVIDER));
+    let (document, retained_document, reader) = document.into_parts()?;
     let wire = OcrWireRequest {
         model,
-        document: wire_document,
+        document,
         api_key: api_key.extract()?,
         api_base: arguments.api_base()?,
         custom_llm_provider,
@@ -147,13 +152,14 @@ pub(super) fn project_request(
         input_sources,
         timeout_seconds: arguments.timeout_seconds()?,
     };
-    let request = decode_request(wire).map_err(ocr_error_to_pyerr)?;
+    let request = decode_request_input(wire).map_err(ocr_error_to_pyerr)?;
     let provider = request.provider_name();
     Ok(ProjectedOcrCall {
         request: request.with_host_hooks(Arc::new(BridgeOcrHooks), None),
         fields: ProjectedOcrFields {
             boundary_request,
             document: retained_document,
+            reader,
             api_key: api_key.unbind(),
             azure_ad_token_provider,
             provider,
@@ -197,10 +203,21 @@ mod tests {
     }
 
     fn project_document(
-        py: Python<'_>,
         document: &Bound<'_, PyAny>,
-    ) -> PyResult<(Value, Py<PyAny>)> {
-        ProjectedDocument::project(py, document).map(ProjectedDocument::into_parts)
+    ) -> PyResult<(
+        OcrDocumentInput,
+        Option<Py<PyAny>>,
+        Option<PythonFileReader>,
+    )> {
+        ProjectedDocument::project(document)?.into_parts()
+    }
+
+    fn url_document(url: &str) -> OcrDocumentInput {
+        litellm_core::ocr::OcrDocument::DocumentUrl {
+            document_url: url.into(),
+            extra_fields: Map::new(),
+        }
+        .into()
     }
 
     fn stub_timeout_conversion(py: Python<'_>) {
@@ -374,7 +391,7 @@ kwargs = {}
     }
 
     #[test]
-    fn document_reader_mutations_are_visible_to_later_field_reads() {
+    fn document_readers_are_not_consumed_during_projection() {
         Python::initialize();
         Python::attach(|py| {
             stub_timeout_conversion(py);
@@ -406,7 +423,12 @@ kwargs = {}
                 .unwrap();
             let arguments = arguments(&request, &kwargs);
             let document = arguments.document().unwrap();
-            project_document(py, &document).unwrap();
+            let (input, retained, reader) = project_document(&document).unwrap();
+            assert_eq!(input, OcrDocumentInput::HostReader { mime_type: None });
+            assert!(retained.is_none());
+            assert_eq!(arguments.api_base().unwrap().as_deref(), Some("original"));
+            assert_eq!(arguments.timeout_seconds().unwrap(), Some(1.0));
+            reader.unwrap().read(py).unwrap();
             assert_eq!(arguments.api_base().unwrap().as_deref(), Some("mutated"));
             assert_eq!(arguments.timeout_seconds().unwrap(), Some(9.0));
         });
@@ -444,7 +466,7 @@ kwargs = {'api_key': key}
     }
 
     #[test]
-    fn file_documents_are_encoded_and_other_documents_keep_the_python_object() {
+    fn file_documents_become_typed_inputs_and_other_documents_keep_the_python_object() {
         Python::initialize();
         Python::attach(|py| {
             let file = py
@@ -454,13 +476,17 @@ kwargs = {'api_key': key}
                     None,
                 )
                 .unwrap();
+            let (input, retained, reader) = project_document(&file).unwrap();
             assert_eq!(
-                project_document(py, &file).unwrap().0,
-                serde_json::json!({
-                    "type": "document_url",
-                    "document_url": "data:application/pdf;base64,JVBERi0xLjQ=",
-                })
+                input,
+                OcrDocumentInput::Bytes {
+                    bytes: b"%PDF-1.4".as_slice().into(),
+                    file_name: None,
+                    mime_type: Some("application/pdf".into()),
+                }
             );
+            assert!(retained.is_none());
+            assert!(reader.is_none());
 
             let original = py
                 .eval(
@@ -469,44 +495,21 @@ kwargs = {'api_key': key}
                     None,
                 )
                 .unwrap();
-            let (wire, retained) = project_document(py, &original).unwrap();
-            assert_eq!(
-                wire,
-                serde_json::json!({
-                    "type": "document_url",
-                    "document_url": "https://example.com/a.pdf",
-                })
-            );
-            assert!(retained.bind(py).is(&original));
+            let (input, retained, _) = project_document(&original).unwrap();
+            assert_eq!(input, url_document("https://example.com/a.pdf"));
+            assert!(retained.unwrap().bind(py).is(&original));
         });
     }
 
     #[test]
-    fn unknown_document_types_reach_existing_downstream_validation() {
+    fn unknown_document_types_reach_existing_core_validation() {
         Python::initialize();
         Python::attach(|py| {
             let document = py
                 .eval(c"{'type': 'mystery', 'mystery': 'x'}", None, None)
                 .unwrap();
-            let wire_document = project_document(py, &document).unwrap().0;
-            assert_eq!(
-                wire_document,
-                serde_json::json!({"type": "mystery", "mystery": "x"})
-            );
-            let error = match decode_request(OcrWireRequest {
-                model: "mistral/mistral-ocr-latest".into(),
-                document: wire_document,
-                api_key: None,
-                api_base: None,
-                custom_llm_provider: None,
-                extra_headers: None,
-                optional_params: Map::new(),
-                input_sources: Default::default(),
-                timeout_seconds: None,
-            }) {
-                Ok(_) => panic!("unknown discriminators belong to core validation"),
-                Err(error) => error,
-            };
+            let error = project_document(&document).unwrap_err();
+            assert!(error.is_instance_of::<PyValueError>(py));
             assert!(error.to_string().contains("document"));
         });
     }
@@ -517,14 +520,14 @@ kwargs = {'api_key': key}
         Python::attach(|py| {
             let missing = py.eval(c"{}", None, None).unwrap();
             assert!(
-                project_document(py, &missing)
+                project_document(&missing)
                     .unwrap_err()
                     .is_instance_of::<PyKeyError>(py)
             );
 
             let non_string = py.eval(c"{'type': 1}", None, None).unwrap();
             assert!(
-                project_document(py, &non_string)
+                project_document(&non_string)
                     .unwrap_err()
                     .is_instance_of::<PyTypeError>(py)
             );
@@ -540,7 +543,7 @@ document = Document()
 ",
             );
             let error =
-                project_document(py, &locals.get_item("document").unwrap().unwrap()).unwrap_err();
+                project_document(&locals.get_item("document").unwrap().unwrap()).unwrap_err();
             assert!(
                 error
                     .value(py)
@@ -569,9 +572,9 @@ document = Document()
 ",
             );
             let document = locals.get_item("document").unwrap().unwrap();
-            let (wire, retained) = project_document(py, &document).unwrap();
-            assert_eq!(wire["type"], "document_url");
-            assert!(!retained.bind(py).is(&document));
+            let (input, retained, _) = project_document(&document).unwrap();
+            assert!(matches!(input, OcrDocumentInput::Bytes { .. }));
+            assert!(retained.is_none());
             let reads: Vec<String> = document.getattr("reads").unwrap().extract().unwrap();
             assert_eq!(reads, ["type", "mime_type", "file"]);
         });

litellm/constants.py

@@ -54,6 +54,7 @@ S3_BOUNDED_OBJECT_KEY_HEAD_BYTES: Final = 64
 S3_PREFIX_DIGEST_CHARS: Final = 16
 # s3 allows 2048 bytes of combined metadata headers, which Content-Disposition counts against
 MAX_S3_OBJECT_DOWNLOAD_FILENAME_BYTES: Final = 1024
+S3_LOG_PROMPTS_ONLY_ENV_VAR: Final = "S3_LOG_PROMPTS_ONLY"
 MAX_FILE_LIST_LIMIT: Final = 10000
 DEFAULT_SQS_FLUSH_INTERVAL_SECONDS: Final = int(os.getenv("DEFAULT_SQS_FLUSH_INTERVAL_SECONDS", 10))
 DEFAULT_NUM_WORKERS_LITELLM_PROXY: Final = int(os.getenv("DEFAULT_NUM_WORKERS_LITELLM_PROXY", 1))

litellm/integrations/callback_configs.json

@@ -446,6 +446,12 @@
         "ui_name": "S3 Path Prefix",
         "description": "Path prefix within the bucket for organizing logs",
         "required": false
+      },
+      "s3_log_prompts_only": {
+        "type": "boolean",
+        "ui_name": "Log Prompts Only",
+        "description": "Log request messages to S3 but drop the model response from each logged object",
+        "required": false
       }
     },
     "description": "S3 Bucket (AWS) Logging Integration"

litellm/integrations/custom_logger.py

@@ -118,6 +118,7 @@ class CustomLogger:  # https://docs.litellm.ai/docs/observability/custom_callbac
 
         alias_map: Final = {
             "langfuse_otel": "langfuse",
+            "s3_v2": "s3",
         }
         lookup_name: Final = alias_map.get(normalized_name, normalized_name)
 

litellm/integrations/s3.py

@@ -2,19 +2,42 @@
 #    On success + failure, log events to Supabase
 
 import hashlib
+import os
+from collections.abc import Mapping
 from datetime import datetime
 from typing import Final, cast
 
+from pydantic import TypeAdapter, ValidationError
+
 import litellm
 from litellm._logging import print_verbose, verbose_logger
 from litellm.constants import (
     MAX_S3_OBJECT_DOWNLOAD_FILENAME_BYTES,
     MAX_S3_OBJECT_KEY_BYTES,
     S3_BOUNDED_OBJECT_KEY_HEAD_BYTES,
+    S3_LOG_PROMPTS_ONLY_ENV_VAR,
     S3_PREFIX_DIGEST_CHARS,
 )
 from litellm.types.utils import StandardLoggingPayload
 
+_S3_LOG_PROMPTS_ONLY: Final = TypeAdapter(bool)
+
+
+def resolve_s3_log_prompts_only(configured: object, environ: Mapping[str, str] | None = None) -> bool:
+    env: Final = os.environ if environ is None else environ
+    raw: Final = env.get(S3_LOG_PROMPTS_ONLY_ENV_VAR) if configured is None else configured
+    if raw is None or raw == "":
+        return False
+    try:
+        return _S3_LOG_PROMPTS_ONLY.validate_python(raw.strip() if isinstance(raw, str) else raw)
+    except ValidationError:
+        verbose_logger.warning("s3 logging: s3_log_prompts_only=%r is not a boolean, logging prompts only", raw)
+        return True
+
+
+def prompts_only_payload(payload: StandardLoggingPayload) -> StandardLoggingPayload:
+    return {**payload, "response": None}
+
 
 class S3Logger:
     # Class variables or attributes
@@ -33,6 +56,7 @@ class S3Logger:
         s3_config=None,
         s3_server_side_encryption: str | None = None,
         s3_sse_kms_key_id: str | None = None,
+        s3_log_prompts_only: bool | None = None,
         **kwargs,
     ):
         import boto3
@@ -41,29 +65,30 @@ class S3Logger:
             verbose_logger.debug("in init s3 logger - s3_callback_params %s", litellm.s3_callback_params)
 
             s3_use_team_prefix = False
+            params: Final = {
+                key: litellm.get_secret(value) if isinstance(value, str) and value.startswith("os.environ/") else value
+                for key, value in (litellm.s3_callback_params or {}).items()
+            }
 
             if litellm.s3_callback_params is not None:
-                # read in .env variables - example os.environ/AWS_BUCKET_NAME
-                for key, value in litellm.s3_callback_params.items():
-                    if isinstance(value, str) and value.startswith("os.environ/"):
-                        litellm.s3_callback_params[key] = litellm.get_secret(value)
-                # now set s3 params from litellm.s3_logger_params
-                s3_bucket_name = litellm.s3_callback_params.get("s3_bucket_name")
-                s3_region_name = litellm.s3_callback_params.get("s3_region_name")
-                s3_api_version = litellm.s3_callback_params.get("s3_api_version")
-                s3_use_ssl = litellm.s3_callback_params.get("s3_use_ssl", True)
-                s3_verify = litellm.s3_callback_params.get("s3_verify")
-                s3_endpoint_url = litellm.s3_callback_params.get("s3_endpoint_url")
-                s3_aws_access_key_id = litellm.s3_callback_params.get("s3_aws_access_key_id")
-                s3_aws_secret_access_key = litellm.s3_callback_params.get("s3_aws_secret_access_key")
-                s3_aws_session_token = litellm.s3_callback_params.get("s3_aws_session_token")
-                s3_config = litellm.s3_callback_params.get("s3_config")
-                s3_path = litellm.s3_callback_params.get("s3_path")
-                s3_server_side_encryption = litellm.s3_callback_params.get("s3_server_side_encryption")
-                s3_sse_kms_key_id = litellm.s3_callback_params.get("s3_sse_kms_key_id")
-                # done reading litellm.s3_callback_params
-                s3_use_team_prefix = bool(litellm.s3_callback_params.get("s3_use_team_prefix", False))
+                s3_bucket_name = params.get("s3_bucket_name")
+                s3_region_name = params.get("s3_region_name")
+                s3_api_version = params.get("s3_api_version")
+                s3_use_ssl = params.get("s3_use_ssl", True)
+                s3_verify = params.get("s3_verify")
+                s3_endpoint_url = params.get("s3_endpoint_url")
+                s3_aws_access_key_id = params.get("s3_aws_access_key_id")
+                s3_aws_secret_access_key = params.get("s3_aws_secret_access_key")
+                s3_aws_session_token = params.get("s3_aws_session_token")
+                s3_config = params.get("s3_config")
+                s3_path = params.get("s3_path")
+                s3_server_side_encryption = params.get("s3_server_side_encryption")
+                s3_sse_kms_key_id = params.get("s3_sse_kms_key_id")
+                s3_use_team_prefix = bool(params.get("s3_use_team_prefix", False))
             self.s3_use_team_prefix = s3_use_team_prefix
+            self.s3_log_prompts_only: object = (
+                params.get("s3_log_prompts_only") if s3_log_prompts_only is None else s3_log_prompts_only
+            )
             self.bucket_name = s3_bucket_name
             self.s3_path = s3_path
             self.s3_server_side_encryption, self.s3_sse_kms_key_id = resolve_sse_params(
@@ -144,7 +169,9 @@ class S3Logger:
 
             from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 
-            payload_str: Final = safe_dumps(payload)
+            payload_str: Final = safe_dumps(
+                prompts_only_payload(payload) if resolve_s3_log_prompts_only(self.s3_log_prompts_only) else payload
+            )
 
             print_verbose(f"\ns3 Logger - Logging payload = {payload_str}")
 

litellm/integrations/s3_v2.py

@@ -21,6 +21,8 @@ from litellm.constants import DEFAULT_S3_BATCH_SIZE, DEFAULT_S3_FLUSH_INTERVAL_S
 from litellm.integrations.s3 import (
     get_s3_object_download_filename,
     get_s3_object_key,
+    prompts_only_payload,
+    resolve_s3_log_prompts_only,
     resolve_sse_params,
 )
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
@@ -68,6 +70,7 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
         s3_use_virtual_hosted_style: bool = False,
         s3_server_side_encryption: str | None = None,
         s3_sse_kms_key_id: str | None = None,
+        s3_log_prompts_only: bool | None = None,
         s3_callback_params_override: dict | None = None,
         **kwargs,
     ):
@@ -108,6 +111,7 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
                 s3_use_virtual_hosted_style=s3_use_virtual_hosted_style,
                 s3_server_side_encryption=s3_server_side_encryption,
                 s3_sse_kms_key_id=s3_sse_kms_key_id,
+                s3_log_prompts_only=s3_log_prompts_only,
             )
             verbose_logger.debug("s3 logger using endpoint url %s", s3_endpoint_url)
 
@@ -163,6 +167,7 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
         s3_use_virtual_hosted_style: bool = False,
         s3_server_side_encryption: str | None = None,
         s3_sse_kms_key_id: str | None = None,
+        s3_log_prompts_only: bool | None = None,
         params_source: dict | None = None,
     ):
         """
@@ -212,6 +217,10 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
             bool(params.get("s3_use_virtual_hosted_style", False)) or s3_use_virtual_hosted_style
         )
 
+        self.s3_log_prompts_only: object = (
+            params.get("s3_log_prompts_only") if s3_log_prompts_only is None else s3_log_prompts_only
+        )
+
         self.s3_server_side_encryption, self.s3_sse_kms_key_id = resolve_sse_params(
             params.get("s3_server_side_encryption") or s3_server_side_encryption,
             params.get("s3_sse_kms_key_id") or s3_sse_kms_key_id,
@@ -489,8 +498,13 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
 
         s3_object_download_filename: Final = get_s3_object_download_filename(start_time, standard_logging_payload["id"])
 
+        payload: Final = (
+            prompts_only_payload(standard_logging_payload)
+            if resolve_s3_log_prompts_only(self.s3_log_prompts_only)
+            else standard_logging_payload
+        )
         return s3BatchLoggingElement(
-            payload=dict(standard_logging_payload),
+            payload=dict(payload),
             s3_object_key=s3_object_key,
             s3_object_download_filename=s3_object_download_filename,
         )

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

@@ -78,6 +78,7 @@ _PROPAGATED_METADATA_KEYS: Final = (
     "user_api_key_end_user_id",
     "user_api_end_user_max_budget",
     "user_api_key_model_max_budget",
+    "user_api_key_team_model_max_budget",
     "user_api_key_user_model_max_budget",
     "user_api_key_end_user_model_max_budget",
     "litellm_call_id",
@@ -395,9 +396,9 @@ async def _check_summary_model_budget(
     ``user_api_key_auth`` runs for the client-requested model. Returns True outside the proxy or when no
     per-model budget is configured.
 
-    All three scopes are checked because the summary's spend is charged to all
-    three: this file propagates the key, user and end-user budgets into the
-    subrequest's metadata, so enforcing only two of them would let compaction
+    Every scope is checked because the summary's spend is charged to every
+    scope: this file propagates the key, team, user and end-user budgets into the
+    subrequest's metadata, so skipping one of them would let compaction
     increment a counter it can never be refused by.
     """
     if user_api_key_auth is None:
@@ -444,6 +445,26 @@ async def _check_summary_model_budget(
             )
             return False
 
+    team_model_max_budget: Final = user_api_key_auth.team_model_max_budget
+    team_id: Final = user_api_key_auth.team_id
+    if isinstance(team_model_max_budget, dict) and team_model_max_budget and team_id is not None:
+        try:
+            await model_max_budget_limiter.is_team_within_model_budget(
+                team_id=team_id,
+                team_model_max_budget=team_model_max_budget,
+                key_model_max_budget=model_max_budget if isinstance(model_max_budget, dict) else None,
+                model=summary_model,
+            )
+        except litellm.BudgetExceededError:
+            return False
+        except Exception as e:  # noqa: BLE001  # a budget gate denies on any failure, as the other scopes do
+            verbose_logger.warning(
+                "compact_20260112: unexpected error during team model-budget check for summary_model=%s; denying: %s",
+                summary_model,
+                e,
+            )
+            return False
+
     end_user_model_max_budget: Final[dict[str, object] | None] = getattr(
         user_api_key_auth, "end_user_model_max_budget", None
     )

litellm/model_prices_and_context_window_backup.json

@@ -377,8 +377,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 1.5e-08
+        "supports_tool_choice": true
     },
     "amazon.nova-2-lite-v1:0": {
         "cache_read_input_token_cost": 7.5e-08,
@@ -561,8 +560,7 @@
         "supports_function_calling": true,
         "supports_prompt_caching": true,
         "supports_response_schema": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 8.75e-09
+        "supports_tool_choice": true
     },
     "amazon.nova-pro-v1:0": {
         "cache_read_input_token_cost": 2e-07,
@@ -578,8 +576,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 2e-07
+        "supports_tool_choice": true
     },
     "amazon.nova-sonic-v1:0": {
         "deprecation_date": "2026-09-14",
@@ -45794,8 +45791,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 1.5e-08
+        "supports_tool_choice": true
     },
     "us.amazon.nova-micro-v1:0": {
         "cache_read_input_token_cost": 8.75e-09,
@@ -45809,8 +45805,7 @@
         "supports_function_calling": true,
         "supports_prompt_caching": true,
         "supports_response_schema": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 8.75e-09
+        "supports_tool_choice": true
     },
     "us.amazon.nova-premier-v1:0": {
         "deprecation_date": "2026-09-14",
@@ -45842,8 +45837,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 2e-07
+        "supports_tool_choice": true
     },
     "us.anthropic.claude-3-5-haiku-20241022-v1:0": {
         "cache_creation_input_token_cost": 1e-06,

litellm/models/verification_token.py

@@ -18,6 +18,7 @@ class LiteLLM_VerificationToken(LiteLLMPydanticObjectBase):
     key_name: str | None = None
     key_alias: str | None = None
     spend: float = 0.0
+    total_spend: float = 0.0
     max_budget: float | None = None
     expires: str | datetime | None = None
     models: list = []
@@ -69,6 +70,7 @@ class LiteLLM_DeletedVerificationToken(LiteLLM_VerificationToken):
     """Audit record for deleted keys; mirrors the token plus deletion metadata."""
 
     id: str | None = None
+    organization_id: str | None = None
     deleted_at: datetime | None = None
     deleted_by: str | None = None
     deleted_by_api_key: str | None = None

litellm/ocr/input.py

@@ -1,112 +0,0 @@
-from collections.abc import Mapping
-from os import PathLike
-from typing import Final, Literal, Protocol, cast  # noqa: TID251  # native callables are validated when loaded
-
-from typing_extensions import NotRequired, ReadOnly, TypedDict
-
-from litellm.rust_bridge.bindings import NativeBinding
-from litellm.rust_bridge.configuration import rust_ocr_enabled
-
-
-class FileReader(Protocol):
-    def read(self) -> bytes | str: ...
-
-
-class FileDocument(TypedDict):
-    type: ReadOnly[Literal["file"]]
-    file: ReadOnly[bytes | PathLike[str] | FileReader]
-    mime_type: ReadOnly[NotRequired[str]]
-
-
-class NativeFileDocument(Protocol):
-    def __call__(self, document: Mapping[str, object]) -> dict[str, str]: ...
-
-
-class NativeUploadDocument(Protocol):
-    def __call__(self, file_content: bytes, file_name: str | None, content_type: str | None) -> dict[str, str]: ...
-
-
-class NativeMimeType(Protocol):
-    def __call__(self, file_name: str) -> str: ...
-
-
-_FILE_DOCUMENT: Final = NativeBinding(
-    "_ocr_file_document",
-    validate=lambda value: (
-        cast(  # cast-ok: native export owns the callable signature
-            NativeFileDocument, value
-        )
-        if callable(value)
-        else None
-    ),
-)
-_UPLOAD_DOCUMENT: Final = NativeBinding(
-    "_ocr_upload_document",
-    validate=lambda value: (
-        cast(  # cast-ok: native export owns the callable signature
-            NativeUploadDocument, value
-        )
-        if callable(value)
-        else None
-    ),
-)
-_MAX_FILE_BYTES: Final = NativeBinding(
-    "_OCR_MAX_FILE_BYTES", validate=lambda value: value if isinstance(value, int) and value > 0 else None
-)
-_MIME_TYPE: Final = NativeBinding(
-    "_ocr_mime_type",
-    validate=lambda value: (
-        cast(  # cast-ok: native export owns the callable signature
-            NativeMimeType, value
-        )
-        if callable(value)
-        else None
-    ),
-)
-_PYTHON_MAX_FILE_BYTES: Final = 50 * 1024 * 1024
-
-
-def get_mime_type(file_path: str) -> str:
-    native: Final = _MIME_TYPE.load() if rust_ocr_enabled() else None
-    if native is None:
-        from litellm.ocr import legacy
-
-        return legacy.get_mime_type(file_path)
-    return native(file_path)
-
-
-def get_max_file_bytes() -> int:
-    limit: Final = _MAX_FILE_BYTES.load() if rust_ocr_enabled() else None
-    if limit is None:
-        return _PYTHON_MAX_FILE_BYTES
-    return limit
-
-
-def convert_file_document_to_url_document(document: FileDocument) -> dict[str, str]:
-    native: Final = _FILE_DOCUMENT.load() if rust_ocr_enabled() else None
-    if native is None:
-        from litellm.ocr import legacy
-
-        return legacy.convert_file_document_to_url_document(document)
-    return native(document)
-
-
-def convert_upload_to_url_document(
-    file_content: bytes, filename: str | None, content_type: str | None
-) -> dict[str, str]:
-    native: Final = _UPLOAD_DOCUMENT.load() if rust_ocr_enabled() else None
-    if native is None:
-        from litellm.ocr import legacy
-
-        if len(file_content) > _PYTHON_MAX_FILE_BYTES:
-            raise ValueError("OCR file exceeds the size limit")
-        content_mime: Final = content_type.split(";")[0].strip() if content_type else None
-        mime_type: Final = (
-            legacy.get_mime_type(filename)
-            if filename and (not content_mime or content_mime == "application/octet-stream")
-            else content_mime or "application/octet-stream"
-        )
-        return legacy.convert_file_document_to_url_document(
-            {"type": "file", "file": file_content, "mime_type": mime_type}
-        )
-    return native(file_content, filename, content_type)

litellm/ocr/legacy.py

@@ -11,7 +11,7 @@ from collections.abc import Coroutine, Mapping
 from dataclasses import dataclass
 from io import IOBase
 from types import MappingProxyType
-from typing import Final, cast  # noqa: TID251  # adapters preserve the legacy untyped contracts
+from typing import Final, Protocol, cast  # noqa: TID251  # adapters preserve the legacy untyped contracts
 
 import httpx
 
@@ -26,7 +26,6 @@ from litellm.llms.base_llm.ocr.transformation import (
     parse_ocr_request_format,
 )
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
-from litellm.ocr.input import FileReader
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import CustomPricingLiteLLMParams
 from litellm.utils import ProviderConfigManager, client
@@ -34,6 +33,10 @@ from litellm.utils import ProviderConfigManager, client
 base_llm_http_handler: Final = BaseLLMHTTPHandler()
 
 
+class FileReader(Protocol):
+    def read(self) -> bytes | str: ...
+
+
 @dataclass(frozen=True, slots=True)
 class _PreparedOCRRequest:
     model: str

litellm/ocr/main.py

@@ -5,7 +5,7 @@ import httpx
 
 from litellm.llms.base_llm.ocr.transformation import OCRResponse
 from litellm.ocr import legacy
-from litellm.ocr.input import convert_file_document_to_url_document, get_mime_type
+from litellm.ocr.legacy import convert_file_document_to_url_document, get_mime_type
 from litellm.rust_bridge.bindings import native_exception_types
 from litellm.rust_bridge.configuration import rust_ocr_enabled
 from litellm.rust_bridge.ocr import LiteLLMOcrRequest

litellm/proxy/_types.py

@@ -2004,6 +2004,13 @@ RouterSettingsDict = Annotated[
 class NewTeamRequest(TeamBase):
     router_settings: RouterSettingsDict | None = None
     model_aliases: dict | None = None
+    model_max_budget: GenericBudgetConfigType | None = Field(
+        default=None,
+        description=(
+            "Max budget per model for every key on the team, overridable per key "
+            "(e.g. {'gpt-4o': {'max_budget': 10, 'budget_duration': '1d'}})"
+        ),
+    )
     tags: list | None = None
     guardrails: list[str] | None = None
     policies: list[str] | None = None
@@ -2105,6 +2112,13 @@ class UpdateTeamRequest(LiteLLMPydanticObjectBase):
     access_group_ids: list[str] | None = None
     budget_limits: list[BudgetLimitEntry] | None = None  # multiple concurrent budget windows
     default_team_member_models: list[str] | None = None  # default allowed_models seeded onto new team members
+    model_max_budget: GenericBudgetConfigType | None = Field(
+        default=None,
+        description=(
+            "Max budget per model for every key on the team, overridable per key "
+            "(e.g. {'gpt-4o': {'max_budget': 10, 'budget_duration': '1d'}})"
+        ),
+    )
 
 
 class PatchTeamRequest(UpdateTeamRequest):
@@ -3032,6 +3046,7 @@ class LiteLLM_VerificationTokenView(LiteLLM_VerificationToken):
     team_tpd_limit: int | None = None
     team_max_budget: float | None = None
     team_soft_budget: float | None = None
+    team_model_max_budget: dict[str, object] | None = None
     team_models: list = []
     team_blocked: bool = False
     soft_budget: float | None = None
@@ -3710,6 +3725,7 @@ class AllCallbacks(LiteLLMPydanticObjectBase):
             "AWS_ACCESS_KEY_ID",
             "AWS_SECRET_ACCESS_KEY",
             "AWS_REGION_NAME",
+            "S3_LOG_PROMPTS_ONLY",
         ],
     )
 
@@ -4462,6 +4478,7 @@ class TeamInfoResponseObjectTeamTable(LiteLLM_TeamTable):
     # Parent org's model ceiling, reported only to callers who can manage the team.
     # None = no org or not a manager; [] or ["all-proxy-models"] = no ceiling.
     organization_models: list[str] | None = None
+    model_max_budget_usage: Mapping[str, Mapping[str, object]] | None = None
 
 
 class TeamInfoResponseObject(TypedDict):

litellm/proxy/auth/team_grants.py

@@ -59,6 +59,7 @@ class TeamGrants(TypedDict, total=False):
     team_tpd_limit: ReadOnly[int | None]
     team_max_budget: ReadOnly[float | None]
     team_soft_budget: ReadOnly[float | None]
+    team_model_max_budget: ReadOnly[dict[str, object] | None]
     team_spend: ReadOnly[float | None]
     team_models: ReadOnly[Sequence[str]]
     team_blocked: ReadOnly[bool]
@@ -101,6 +102,7 @@ def team_grants(
         team_tpd_limit=team_object.tpd_limit,
         team_max_budget=team_object.max_budget,
         team_soft_budget=team_object.soft_budget,
+        team_model_max_budget=team_object.model_max_budget,
         team_spend=team_object.spend,
         team_models=tuple(team_object.models),
         team_blocked=team_object.blocked,

litellm/proxy/auth/user_api_key_auth.py

@@ -304,6 +304,16 @@ class _UserModelBudgetLimiter(Protocol):
     ) -> bool: ...
 
 
+class _TeamModelBudgetLimiter(Protocol):
+    async def is_team_within_model_budget(
+        self,
+        team_id: str,
+        team_model_max_budget: Mapping[str, object],
+        key_model_max_budget: Mapping[str, object] | None,
+        model: str,
+    ) -> bool: ...
+
+
 class _TokenTeamModels(Protocol):
     @property
     def team_models(self) -> list[str]: ...
@@ -374,6 +384,25 @@ async def _check_user_model_budget(
         )
 
 
+async def _check_team_model_budget(
+    valid_token: UserAPIKeyAuth,
+    model_max_budget_limiter: _TeamModelBudgetLimiter,
+    models: list[str],
+) -> None:
+    """Enforce the team's `model_max_budget` for every requested model the key does not override."""
+    team_model_max_budget: Final = valid_token.team_model_max_budget
+    if valid_token.team_id is None or not team_model_max_budget:
+        return
+    key_model_max_budget: Final[Mapping[str, object] | None] = valid_token.model_max_budget
+    for model_name in models:
+        await model_max_budget_limiter.is_team_within_model_budget(
+            team_id=valid_token.team_id,
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=key_model_max_budget,
+            model=model_name,
+        )
+
+
 async def _check_key_model_budget_with_fallback(
     valid_token: UserAPIKeyAuth,
     model_max_budget_limiter: _KeyModelBudgetLimiter,
@@ -2376,6 +2405,7 @@ async def _user_api_key_auth_builder(
                         team_id=valid_token.team_id,
                         max_budget=valid_token.team_max_budget,
                         soft_budget=valid_token.team_soft_budget,
+                        model_max_budget=valid_token.team_model_max_budget,
                         spend=valid_token.team_spend,
                         tpm_limit=valid_token.team_tpm_limit,
                         rpm_limit=valid_token.team_rpm_limit,
@@ -2530,6 +2560,7 @@ def _team_obj_from_token(valid_token: UserAPIKeyAuth) -> LiteLLM_TeamTableCached
         team_id=valid_token.team_id,
         max_budget=valid_token.team_max_budget,
         soft_budget=valid_token.team_soft_budget,
+        model_max_budget=valid_token.team_model_max_budget,
         spend=valid_token.team_spend,
         tpm_limit=valid_token.team_tpm_limit,
         rpm_limit=valid_token.team_rpm_limit,
@@ -2606,6 +2637,7 @@ async def _run_centralized_common_checks(
         litellm_proxy_admin_name,
         llm_router,
         master_key,
+        model_max_budget_limiter,
         prisma_client,
         proxy_logging_obj,
         user_api_key_cache,
@@ -2874,6 +2906,21 @@ async def _run_centralized_common_checks(
     finally:
         release_spend_counter_batch()
 
+    if not skip_budget_checks:
+        await _check_team_model_budget(
+            valid_token=user_api_key_auth_obj,
+            model_max_budget_limiter=model_max_budget_limiter,
+            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,
+                )
+            ),
+        )
+
     await _reserve_budget_after_common_checks(
         user_api_key_auth_obj=user_api_key_auth_obj,
         request=request,

litellm/proxy/db/create_views.py

@@ -78,6 +78,7 @@ async def create_missing_views(db: SupportsRawQueries) -> None:
                 v.*,
                 t.spend AS team_spend,
                 t.max_budget AS team_max_budget,
+                t.model_max_budget AS team_model_max_budget,
                 t.tpm_limit AS team_tpm_limit,
                 t.rpm_limit AS team_rpm_limit,
                 t.tpd_limit AS team_tpd_limit,

litellm/proxy/db/db_spend_update_writer.py

@@ -18,6 +18,8 @@ from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, cast, overload
 from urllib.parse import quote, unquote
 
+from typing_extensions import ReadOnly, TypedDict
+
 import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.caching import RedisCache
@@ -109,6 +111,10 @@ def _batch_cost_row_to_write(payload: SpendLogsPayload, disable_spend_logs: bool
     return MappingProxyType({field: value for field, value in payload.items() if field in _BATCH_COST_CLAIM_FIELDS})
 
 
+class _SpendIncrement(TypedDict):
+    increment: ReadOnly[float]
+
+
 class _SpendBatch(Protocol):
     litellm_usertable: BatchTable
     litellm_verificationtoken: BatchTable
@@ -1615,10 +1621,12 @@ class DBSpendUpdateWriter:
                         async with transaction.batch_() as batcher:
                             # Sort by token for consistent lock ordering across pods to prevent deadlocks.
                             for token, response_cost in sorted(key_list_transactions.items()):
+                                spend_increment: _SpendIncrement = {"increment": response_cost}
                                 batcher.litellm_verificationtoken.update_many(  # 'update_many' prevents error from being raised if no row exists
                                     where={"token": token},
                                     data={
-                                        "spend": {"increment": response_cost},
+                                        "spend": spend_increment,
+                                        "total_spend": spend_increment,
                                         "last_active": datetime.now(timezone.utc),
                                     },
                                 )

litellm/proxy/hooks/model_max_budget_limiter.py

@@ -19,12 +19,14 @@ from litellm.types.utils import BudgetConfig, StandardLoggingPayload
 VIRTUAL_KEY_SPEND_CACHE_KEY_PREFIX: Final = "virtual_key_spend"
 END_USER_SPEND_CACHE_KEY_PREFIX: Final = "end_user_model_spend"
 USER_SPEND_CACHE_KEY_PREFIX: Final = "user_model_spend"
+TEAM_SPEND_CACHE_KEY_PREFIX: Final = "team_model_spend"
 
 _SPEND_CACHE_KEY_PREFIXES: Final = MappingProxyType(
     {
         Litellm_EntityType.KEY: VIRTUAL_KEY_SPEND_CACHE_KEY_PREFIX,
         Litellm_EntityType.USER: USER_SPEND_CACHE_KEY_PREFIX,
         Litellm_EntityType.END_USER: END_USER_SPEND_CACHE_KEY_PREFIX,
+        Litellm_EntityType.TEAM: TEAM_SPEND_CACHE_KEY_PREFIX,
     }
 )
 
@@ -37,6 +39,7 @@ _BUDGET_START_TIME_KEY_PREFIXES: Final = MappingProxyType(
         Litellm_EntityType.KEY: "virtual_key_budget_start_time",
         Litellm_EntityType.USER: "user_model_budget_start_time",
         Litellm_EntityType.END_USER: "end_user_budget_start_time",
+        Litellm_EntityType.TEAM: "team_model_budget_start_time",
     }
 )
 
@@ -139,6 +142,18 @@ def resolve_model_budget(model: str, model_max_budget: Mapping[str, object]) ->
     return None
 
 
+def team_model_budget_applies(model: str, key_model_max_budget: Mapping[str, object] | None) -> bool:
+    """A key entry that spend-gates `model` overrides the team cap: it is then gated on and billed to the key alone."""
+    if not key_model_max_budget:
+        return True
+    resolved: Final = resolve_model_budget(model=model, model_max_budget=key_model_max_budget)
+    return resolved is None or not _spend_gated(resolved.budget_config)
+
+
+def _spend_gated(budget_config: BudgetConfig) -> bool:
+    return budget_config.max_budget is not None and budget_config.max_budget >= 0
+
+
 def _budget_model_candidates(model: str) -> tuple[str, ...]:
     """Names a budget may be configured under for a request on `model`, most specific first.
 
@@ -346,6 +361,30 @@ class _PROXY_VirtualKeyModelMaxBudgetLimiter(RouterBudgetLimiting):
             exceeded_message=f"LiteLLM End User: {end_user_id}, exceeded budget for model={model}",
         )
 
+    async def is_team_within_model_budget(
+        self,
+        team_id: str,
+        team_model_max_budget: Mapping[str, object],
+        key_model_max_budget: Mapping[str, object] | None,
+        model: str,
+    ) -> bool:
+        """
+        Check if the team is within the model budget, unless the key's own
+        `model_max_budget` overrides it for `model`
+
+        Raises:
+            BudgetExceededError: If the team has exceeded the model budget
+        """
+        if not team_model_budget_applies(model=model, key_model_max_budget=key_model_max_budget):
+            return True
+        return await self._is_entity_within_model_budget(
+            entity_type=Litellm_EntityType.TEAM,
+            entity_id=team_id,
+            model_max_budget=team_model_max_budget,
+            model=model,
+            exceeded_message=f"LiteLLM Team: {team_id}, exceeded budget for model={model}",
+        )
+
     async def _is_entity_within_model_budget(
         self,
         entity_type: Litellm_EntityType,
@@ -456,11 +495,26 @@ class _PROXY_VirtualKeyModelMaxBudgetLimiter(RouterBudgetLimiting):
             return
 
         response_cost: Final[float] = standard_logging_payload.get("response_cost", 0)
+        key_model_max_budget: Final = _metadata.get("user_api_key_model_max_budget")
         entity_budgets: Final = (
             (
                 Litellm_EntityType.KEY,
                 payload_metadata.get("user_api_key_hash"),
-                _metadata.get("user_api_key_model_max_budget"),
+                key_model_max_budget,
+            ),
+            (
+                Litellm_EntityType.TEAM,
+                payload_metadata.get("user_api_key_team_id"),
+                (
+                    _metadata.get("user_api_key_team_model_max_budget")
+                    if team_model_budget_applies(
+                        model=model,
+                        key_model_max_budget=(
+                            key_model_max_budget if isinstance(key_model_max_budget, Mapping) else None
+                        ),
+                    )
+                    else None
+                ),
             ),
             (
                 Litellm_EntityType.USER,
@@ -478,7 +532,7 @@ class _PROXY_VirtualKeyModelMaxBudgetLimiter(RouterBudgetLimiting):
         if not resolved_budgets:
             verbose_proxy_logger.debug(
                 "Not running _PROXY_VirtualKeyModelMaxBudgetLimiter.async_log_success_event: "
-                "no key, user or end-user model_max_budget covers model=%s",
+                "no key, team, user or end-user model_max_budget covers model=%s",
                 model,
             )
             return

litellm/proxy/hooks/parallel_request_limiter_v3.py

@@ -531,6 +531,7 @@ class RequestRateLimiterStash:
     owner_litellm_call_id: str | None = None
     rate_limit_response: RateLimitResponse | None = None
     parallel_slot: ParallelSlotAcquisition | None = None
+    parallel_slot_release_lock: asyncio.Lock = field(default_factory=asyncio.Lock, repr=False, compare=False)
     reserved_tokens: int = 0
     reserved_model: str | None = None
     reserved_scopes: frozenset[tuple[str, str]] = field(default_factory=frozenset)
@@ -1620,6 +1621,20 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             statuses.append(self._gauge_status(gauge, in_flight + 1, "OK"))
         return RateLimitResponse(overall_code="OK", statuses=statuses)
 
+    async def _release_stashed_parallel_slot(
+        self,
+        stash: RequestRateLimiterStash | None,
+        parent_otel_span: Span | None,
+    ) -> None:
+        if stash is None:
+            return
+        async with stash.parallel_slot_release_lock:
+            acquisition: Final = stash.parallel_slot
+            if acquisition is None:
+                return
+            await self._release_parallel_request_slots(acquisition, parent_otel_span)
+            stash.parallel_slot = None  # rebind-ok: marks this request's slot as released
+
     async def _release_parallel_request_slots(
         self,
         acquisition: ParallelSlotAcquisition,
@@ -3379,13 +3394,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
                     parent_otel_span=user_api_key_dict.parent_otel_span,
                 )
                 stash.reservation_released = True
-            acquisition: Final = stash.parallel_slot
-            if acquisition is not None:
-                await self._release_parallel_request_slots(
-                    acquisition=acquisition,
-                    parent_otel_span=user_api_key_dict.parent_otel_span,
-                )
-                stash.parallel_slot = None
+            await self._release_stashed_parallel_slot(stash, user_api_key_dict.parent_otel_span)
             self._handle_rate_limit_error(
                 response=io_response,
                 descriptors=descriptors,
@@ -3700,13 +3709,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
                 )
 
                 if tpm_response["overall_code"] == "OVER_LIMIT":
-                    acquisition: Final = stash.parallel_slot
-                    if acquisition is not None:
-                        await self._release_parallel_request_slots(
-                            acquisition=acquisition,
-                            parent_otel_span=user_api_key_dict.parent_otel_span,
-                        )
-                        stash.parallel_slot = None
+                    await self._release_stashed_parallel_slot(stash, user_api_key_dict.parent_otel_span)
                     self._handle_rate_limit_error(
                         response=tpm_response,
                         descriptors=descriptors,
@@ -4524,13 +4527,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             verbose_proxy_logger.debug("INSIDE parallel request limiter ASYNC SUCCESS LOGGING")
 
             stash: Final = get_request_stash_for_call(_call_id_from_callback_kwargs(kwargs))
-            acquisition: Final = stash.parallel_slot if stash is not None else None
-            if stash is not None and acquisition is not None:
-                await self._release_parallel_request_slots(
-                    acquisition=acquisition,
-                    parent_otel_span=litellm_parent_otel_span,
-                )
-                stash.parallel_slot = None
+            await self._release_stashed_parallel_slot(stash, litellm_parent_otel_span)
 
             pipeline_operations: Final = self._build_success_event_pipeline_operations(
                 kwargs=kwargs,
@@ -4650,13 +4647,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             pipeline_operations: Final[list[RedisPipelineIncrementOperation]] = []
 
             stash: Final = get_request_stash_for_call(_call_id_from_callback_kwargs(kwargs))
-            acquisition: Final = stash.parallel_slot if stash is not None else None
-            if stash is not None and acquisition is not None:
-                await self._release_parallel_request_slots(
-                    acquisition=acquisition,
-                    parent_otel_span=litellm_parent_otel_span,
-                )
-                stash.parallel_slot = None
+            await self._release_stashed_parallel_slot(stash, litellm_parent_otel_span)
 
             # Skip the reservation refund if async_post_call_failure_hook
             # already released it (proxy-level rejection that also bubbles up
@@ -4764,23 +4755,23 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
         object's current max_parallel_requests configuration, which can
         change mid-request) decides whether there is anything to release.
         """
-        stash: Final = get_request_stash()
-        if stash is None or stash.parallel_slot is None:
-            return
-
-        await self._release_parallel_request_slots(
-            acquisition=stash.parallel_slot,
-            parent_otel_span=None,
-        )
-        stash.parallel_slot = None
+        await self._release_stashed_parallel_slot(get_request_stash(), None)
 
     async def async_post_call_success_hook(self, data: dict, user_api_key_dict: UserAPIKeyAuth, response):
         """
-        Post-call hook to update rate limit headers in the response.
+        Release completed-request slots and update rate limit headers in the response.
         """
         try:
-            stash: Final = get_request_stash()
-            litellm_proxy_rate_limit_response: Final = stash.rate_limit_response if stash is not None else None
+            slot_stash: Final = get_request_stash_for_call(_call_id_from_callback_kwargs(data))
+            await self._release_stashed_parallel_slot(slot_stash, user_api_key_dict.parent_otel_span)
+        except Exception as e:
+            verbose_proxy_logger.exception("Error releasing parallel request slot in post-call hook: %s", e)
+
+        try:
+            header_stash: Final = get_request_stash()
+            litellm_proxy_rate_limit_response: Final = (
+                header_stash.rate_limit_response if header_stash is not None else None
+            )
 
             if litellm_proxy_rate_limit_response is not None and response_has_hidden_params(response):
                 additional_headers: Final = ensure_response_additional_headers(response)
@@ -4848,12 +4839,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             stash: Final = get_request_stash()
             if stash is None:
                 return
-            if stash.parallel_slot is not None:
-                await self._release_parallel_request_slots(
-                    acquisition=stash.parallel_slot,
-                    parent_otel_span=user_api_key_dict.parent_otel_span,
-                )
-                stash.parallel_slot = None
+            await self._release_stashed_parallel_slot(stash, user_api_key_dict.parent_otel_span)
 
             if stash.batch_enqueued_reservation is not None:
                 await self.batch_enqueued_token_store.refund(

litellm/proxy/litellm_pre_call_utils.py

@@ -2334,6 +2334,7 @@ async def add_litellm_data_to_request(
     # Team spend, budget - used by prometheus.py
     data[_metadata_variable_name]["user_api_key_team_max_budget"] = user_api_key_dict.team_max_budget
     data[_metadata_variable_name]["user_api_key_team_spend"] = user_api_key_dict.team_spend
+    data[_metadata_variable_name]["user_api_key_team_model_max_budget"] = user_api_key_dict.team_model_max_budget
     data[_metadata_variable_name]["user_api_key_request_route"] = user_api_key_dict.request_route
 
     # API Key spend, budget - used by prometheus.py

litellm/proxy/management_endpoints/common_utils.py

@@ -55,6 +55,7 @@ def validate_budget_duration(budget_duration: str | None, status_code: int = 400
 from litellm._logging import verbose_proxy_logger
 from litellm.caching import DualCache
 from litellm.proxy._types import (
+    CommonProxyErrors,
     KeyRequestBase,
     LiteLLM_ManagementEndpoint_MetadataFields,
     LiteLLM_ManagementEndpoint_MetadataFields_Premium,
@@ -73,12 +74,62 @@ from litellm.proxy._types import (  # noqa: F401  re-exported
 from litellm.proxy.common_utils.timezone_utils import get_budget_reset_time
 from litellm.proxy.utils import _premium_user_check
 from litellm.repositories.team_repository import TeamRepository
+from litellm.types.utils import BudgetConfig
 
 if TYPE_CHECKING:
     from litellm.proxy._types import NewProjectRequest, UpdateProjectRequest
     from litellm.proxy.utils import PrismaClient, ProxyLogging
 
 
+def validate_team_model_max_budget(
+    model_max_budget: Mapping[str, BudgetConfig] | None,
+    premium_user: bool,
+) -> None:
+    """Reject a team `model_max_budget` the limiter could not enforce (no duration, bad cap, tpm/rpm limits)."""
+    if not model_max_budget:
+        return
+    if premium_user is not True:
+        raise HTTPException(
+            status_code=403,
+            detail={
+                "error": f"Setting model_max_budget on a team is an enterprise feature. {CommonProxyErrors.not_premium_user.value}"
+            },
+        )
+    for model_name, budget_config in model_max_budget.items():
+        if not model_name.strip():
+            raise HTTPException(
+                status_code=400,
+                detail={"error": "model_max_budget keys must be non-empty model names"},
+            )
+        max_budget = budget_config.max_budget
+        if max_budget is None or not math.isfinite(max_budget) or max_budget < 0:
+            raise HTTPException(
+                status_code=400,
+                detail={
+                    "error": (
+                        f"model_max_budget[{model_name!r}].max_budget must be a non-negative finite number. "
+                        f"Received: {max_budget}"
+                    )
+                },
+            )
+        if budget_config.budget_duration is None:
+            raise HTTPException(
+                status_code=400,
+                detail={"error": f"model_max_budget[{model_name!r}] requires a budget_duration, e.g. '1d' or '30d'"},
+            )
+        validate_budget_duration(budget_config.budget_duration)
+        if budget_config.tpm_limit is not None or budget_config.rpm_limit is not None:
+            raise HTTPException(
+                status_code=400,
+                detail={
+                    "error": (
+                        f"model_max_budget[{model_name!r}] tpm_limit/rpm_limit are not enforced on a team; "
+                        "set per-model rate limits on the key instead"
+                    )
+                },
+            )
+
+
 def require_caller_user_id_for_non_admin(
     user_api_key_dict: UserAPIKeyAuth,
 ) -> str:

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -4166,7 +4166,10 @@ async def info_key_fn(
 
     Returns:
     - key: str - The key that was looked up, echoed back as it was passed in
-    - info: dict - The key's row, minus the hashed token
+    - info: dict - The key's row, minus the hashed token. Deleted keys are served from the
+      LiteLLM_DeletedVerificationToken archive and carry deleted_at / deleted_by
+        - status: "active" | "expired" | "revoked" | "deleted" - Derived from blocked, expires and
+          whether the row came from the archive
         - key_alias: str | None - User-friendly key alias
         - spend: float - Amount spent by the key. When budget_duration is set this covers only the
           current budget window, not the key's lifetime
@@ -4220,10 +4223,15 @@ async def info_key_fn(
         hashed_key: str | None = key
         if key is not None:
             hashed_key = _hash_token_if_needed(token=key)
-        key_info = await _prisma_table(VerificationTokenRepository(prisma_client)).find_unique(
+        live_key_info: Final = await _prisma_table(VerificationTokenRepository(prisma_client)).find_unique(
             where={"token": hashed_key},
             include={"litellm_budget_table": True},
         )
+        key_info: Final = (
+            live_key_info
+            if live_key_info is not None
+            else await _find_deleted_key_info(prisma_client=prisma_client, hashed_key=hashed_key)
+        )
         if key_info is None:
             raise ProxyException(
                 message="Key not found in database",
@@ -4231,7 +4239,6 @@ async def info_key_fn(
                 param="key",
                 code=status.HTTP_404_NOT_FOUND,
             )
-
         if (
             await _can_user_query_key_info(
                 user_api_key_dict=user_api_key_dict,
@@ -4245,38 +4252,46 @@ async def info_key_fn(
                 detail=f"You are not allowed to access this key's info. Your role={user_api_key_dict.user_role}",
             )
         ## REMOVE HASHED TOKEN INFO BEFORE RETURNING ##
-        try:
-            key_info = key_info.model_dump()
-        except Exception:
-            # if using pydantic v1
-            key_info = key_info.dict()  # pyright: ignore[reportDeprecated]  # deliberate pydantic v1 fallback
-        key_token_hash: Final[str | None] = key_info.pop("token")
+        key_info_dict: Final = key_info.model_dump()
+        key_token_hash: Final[str | None] = key_info_dict.pop("token")
+        key_info_dict["status"] = (
+            "deleted" if live_key_info is None else _derive_key_status(key_info_dict, now=datetime.now(timezone.utc))
+        )
 
-        model_max_budget = key_info.get("model_max_budget") or {}
-        budget_table: Final = key_info.get("litellm_budget_table") or {}
+        model_max_budget = key_info_dict.get("model_max_budget") or {}
+        budget_table: Final = key_info_dict.get("litellm_budget_table") or {}
         if not model_max_budget and isinstance(budget_table, dict):
             model_max_budget = budget_table.get("model_max_budget") or {}
         if model_max_budget and key_token_hash:
-            key_info["model_max_budget_usage"] = await _build_model_max_budget_usage(
+            key_info_dict["model_max_budget_usage"] = await _build_model_max_budget_usage(
                 api_key_hash=key_token_hash,
                 model_max_budget=model_max_budget,
                 user_api_key_cache=model_max_budget_limiter.dual_cache,
             )
         budget_limits_usage: Final = await _build_budget_limits_usage(
-            budget_limits=key_info.get("budget_limits"),
+            budget_limits=key_info_dict.get("budget_limits"),
             api_key_hash=key_token_hash,
         )
         if budget_limits_usage is not None:
-            key_info["budget_limits_usage"] = budget_limits_usage
+            key_info_dict["budget_limits_usage"] = budget_limits_usage
 
-        # Attach object_permission if object_permission_id is set
-        key_info = await attach_object_permission_to_dict(key_info, prisma_client)
-
-        return {"key": key, "info": key_info}
+        return {"key": key, "info": await attach_object_permission_to_dict(key_info_dict, prisma_client)}
     except Exception as e:
         raise handle_exception_on_proxy(e)
 
 
+async def _find_deleted_key_info(
+    prisma_client: PrismaClient, hashed_key: str | None
+) -> LiteLLM_DeletedVerificationToken | None:
+    archived_row: Final = await _deleted_verification_token_table(prisma_client).find_first(
+        where={"token": hashed_key},
+        order={"deleted_at": "desc"},
+    )
+    if archived_row is None:
+        return None
+    return LiteLLM_DeletedVerificationToken.model_validate(archived_row.model_dump())
+
+
 def _check_model_access_group(models: list[str] | None, llm_router: Router | None, premium_user: bool) -> Literal[True]:
     """
     if is_model_access_group is True + is_wildcard_route is True, check if user is a premium user
@@ -6216,6 +6231,24 @@ async def get_member_team_ids(
 
 VALID_EXPIRES_FILTER_VALUES: Final = frozenset({"active", "expired"})
 
+KeyStatus = Literal["active", "expired", "revoked", "deleted"]
+VALID_STATUS_FILTER_VALUES: Final[frozenset[KeyStatus]] = frozenset({"active", "expired", "revoked", "deleted"})
+
+
+class _KeyStatusSource(BaseModel):
+    blocked: bool | None = None
+    expires: datetime | None = None
+
+
+def _derive_key_status(row: Mapping[str, object], now: datetime) -> KeyStatus:
+    source: Final = _KeyStatusSource.model_validate(row)
+    if source.blocked is True:
+        return "revoked"
+    if source.expires is None:
+        return "active"
+    expires_utc: Final = source.expires if source.expires.tzinfo else source.expires.replace(tzinfo=timezone.utc)
+    return "expired" if expires_utc < now else "active"
+
 
 @router.get(
     "/key/list",
@@ -6252,7 +6285,10 @@ async def list_keys(
     ),
     sort_order: str = Query(default="desc", description="Sort order ('asc' or 'desc')"),
     expand: list[str] | None = Query(None, description="Expand related objects (e.g. 'user')"),
-    status: str | None = Query(None, description="Filter by status (e.g. 'deleted')"),
+    status: str | None = Query(
+        None,
+        description="Filter by status: 'active' (not blocked, not expired), 'expired' (not blocked, past expiry), 'revoked' (blocked) or 'deleted' (archived keys). Omit to return live keys regardless of status.",
+    ),
     project_id: str | None = Query(None, description="Filter keys by project ID"),
     access_group_id: str | None = Query(None, description="Filter keys by access group ID"),
     agent_id: str | None = Query(None, description="Filter keys by agent ID"),
@@ -6270,7 +6306,9 @@ async def list_keys(
 
     Parameters:
         expand: Optional[List[str]] - Expand related objects (e.g. 'user' to include user information)
-        status: Optional[str] - Filter by status. Currently supports "deleted" to query deleted keys.
+        status: Optional[str] - Filter by status: "active", "expired", "revoked" (blocked) or "deleted".
+        "deleted" reads the LiteLLM_DeletedVerificationToken archive; the other values partition the
+        live key table, so every live key matches exactly one of them.
 
     Returns:
         {
@@ -6292,11 +6330,10 @@ async def list_keys(
             verbose_proxy_logger.error("Database not connected")
             raise Exception("Database not connected")
 
-        # Validate status parameter
-        if status is not None and status != "deleted":
+        if status is not None and status not in VALID_STATUS_FILTER_VALUES:
             raise HTTPException(
                 status_code=400,
-                detail={"error": "Invalid status value. Currently only 'deleted' is supported."},
+                detail={"error": "Invalid status value. Supported: 'active', 'expired', 'revoked', 'deleted'."},
             )
 
         if isinstance(expires, str) and expires not in VALID_EXPIRES_FILTER_VALUES:
@@ -6608,6 +6645,18 @@ def _build_expires_where_clause(expires_filter: str, now: datetime) -> dict[str,
     return {"OR": [{"expires": None}, {"expires": {"gte": now}}]}
 
 
+def _not_blocked_where_clause() -> dict[str, object]:
+    return {"OR": [{"blocked": None}, {"blocked": False}]}
+
+
+def _build_status_where_clause(status_filter: str | None, now: datetime) -> dict[str, object] | None:
+    if status_filter == "revoked":
+        return {"blocked": True}
+    if status_filter in ("expired", "active"):
+        return {"AND": [_not_blocked_where_clause(), _build_expires_where_clause(status_filter, now)]}
+    return None
+
+
 def _build_key_search_where(search: str) -> KeySearchWhere:
     search_where: Final[KeySearchWhere] = {
         "OR": (
@@ -6635,6 +6684,7 @@ def _build_key_filter_conditions(
     use_key_alias_substring_matching: bool = False,
     expires_filter: str | None = None,
     search: str | None = None,
+    status_filter: str | None = None,
 ) -> Mapping[str, object]:
     """Build filter conditions for key listing.
 
@@ -6724,6 +6774,8 @@ def _build_key_filter_conditions(
 
     # Apply team_id, project_id and access_group_id as global AND filters so they
     # narrow results across all visibility conditions (own keys, team keys, etc.)
+    now: Final = datetime.now(timezone.utc)
+    status_where: Final = _build_status_where_clause(status_filter, now)
     global_filters: Final[tuple[Mapping[str, object], ...]] = (
         *(
             (
@@ -6741,10 +6793,11 @@ def _build_key_filter_conditions(
         *(({"access_group_ids": {"hasSome": [access_group_id]}},) if access_group_id else ()),
         *(({"agent_id": agent_id},) if agent_id and isinstance(agent_id, str) else ()),
         *(
-            (_build_expires_where_clause(expires_filter, datetime.now(timezone.utc)),)
+            (_build_expires_where_clause(expires_filter, now),)
             if expires_filter is not None and expires_filter in VALID_EXPIRES_FILTER_VALUES
             else ()
         ),
+        *((status_where,) if status_where is not None else ()),
     )
     combined_where: Final[Mapping[str, object]] = {"AND": [where, *global_filters]} if global_filters else where
     verbose_proxy_logger.debug("Filter conditions: %s", combined_where)
@@ -6817,6 +6870,7 @@ async def _list_key_helper(
         use_key_alias_substring_matching=use_key_alias_substring_matching,
         expires_filter=expires_filter,
         search=search,
+        status_filter=status,
     )
 
     # Calculate skip for pagination

litellm/proxy/management_endpoints/team_endpoints.py

@@ -22,7 +22,7 @@ from typing import TYPE_CHECKING, Annotated, Final, NamedTuple, NoReturn, Protoc
 
 import fastapi
 from fastapi import APIRouter, Depends, Header, HTTPException, Request, status
-from pydantic import BaseModel, JsonValue
+from pydantic import BaseModel, JsonValue, ValidationError
 from typing_extensions import ReadOnly, TypedDict
 
 import litellm
@@ -38,6 +38,7 @@ from litellm.proxy._types import (
     DeleteTeamRequest,
     LiteLLM_AuditLogs,
     LiteLLM_DeletedTeamTable,
+    Litellm_EntityType,
     LiteLLM_ManagementEndpoint_MetadataFields,
     LiteLLM_ManagementEndpoint_MetadataFields_Premium,
     LiteLLM_ModelTable,
@@ -95,6 +96,10 @@ from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 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.user_api_key_cache import UserApiKeyCache
+from litellm.proxy.hooks.model_max_budget_limiter import (
+    build_model_max_budget_usage,
+    resolve_model_budget,
+)
 from litellm.proxy.management_endpoints.common_daily_activity import (
     get_daily_activity_aggregated,
 )
@@ -108,6 +113,7 @@ from litellm.proxy.management_endpoints.common_utils import (
     _upsert_budget_and_membership,
     _user_has_admin_view,
     validate_budget_duration,
+    validate_team_model_max_budget,
 )
 from litellm.proxy.management_endpoints.organization_endpoints import (
     add_member_to_organization,
@@ -177,6 +183,7 @@ from litellm.types.proxy.management_endpoints.team_endpoints import (
     TeamUserSpendRow,
     UpdateTeamMemberPermissionsRequest,
 )
+from litellm.types.utils import BudgetConfig
 
 if TYPE_CHECKING:
     from prisma import Prisma
@@ -1170,6 +1177,62 @@ def _check_team_budget_update_authority(
         )
 
 
+def _existing_model_cap(raw_budget_config: object) -> BudgetConfig | None:
+    try:
+        return BudgetConfig.model_validate(raw_budget_config)
+    except ValidationError:
+        return None
+
+
+def _check_team_model_budget_update_authority(
+    data: UpdateTeamRequest,
+    user_api_key_dict: UserAPIKeyAuth,
+    existing_model_max_budget: Mapping[str, object] | None,
+) -> None:
+    """Like `_check_team_budget_update_authority`: only a proxy admin may raise, re-window or drop a per-model cap."""
+    if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN:
+        return
+    if "model_max_budget" not in data.model_fields_set or not existing_model_max_budget:
+        return
+    requested: Final[Mapping[str, BudgetConfig]] = data.model_max_budget or {}
+    for model_name, raw_existing in existing_model_max_budget.items():
+        existing = _existing_model_cap(raw_existing)
+        if existing is None or existing.max_budget is None or model_name in requested:
+            continue
+        raise HTTPException(
+            status_code=403,
+            detail={
+                "error": (
+                    f"Only a proxy admin can remove a team's model_max_budget for {model_name!r}. "
+                    f"Current max_budget={existing.max_budget}."
+                )
+            },
+        )
+    for model_name, proposed in requested.items():
+        governing = resolve_model_budget(model=model_name, model_max_budget=existing_model_max_budget)
+        if governing is None:
+            continue
+        cap = governing.budget_config
+        if cap.max_budget is None:
+            continue
+        if (
+            proposed.max_budget is None
+            or proposed.max_budget > cap.max_budget
+            or proposed.budget_duration != cap.budget_duration
+        ):
+            raise HTTPException(
+                status_code=403,
+                detail={
+                    "error": (
+                        f"Only a proxy admin can raise a team's model_max_budget for {model_name!r} or change its "
+                        f"budget_duration. Current max_budget={cap.max_budget} per {cap.budget_duration} "
+                        f"(entry {governing.budget_model!r}), requested={proposed.max_budget} per "
+                        f"{proposed.budget_duration}."
+                    )
+                },
+            )
+
+
 def _should_auto_add_team_creator(
     user_api_key_dict: UserAPIKeyAuth,
     general_settings: Mapping[str, object],
@@ -1230,6 +1293,7 @@ async def new_team(
     - prompts: Optional[List[str]] - List of prompts that the team is allowed to use.
     - organization_id: Optional[str] - The organization id of the team. Default is None. Create via `/organization/new`.
     - model_aliases: Optional[dict] - Model aliases for the team. [Docs](https://docs.litellm.ai/docs/proxy/team_based_routing#create-team-with-model-alias)
+    - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
     - guardrails: Optional[List[str]] - Guardrails for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails)
     - policies: Optional[List[str]] - Policies for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails/guardrail_policies)
     - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key.
@@ -1291,6 +1355,7 @@ async def new_team(
             general_settings,
             litellm_proxy_admin_name,
             llm_router,
+            premium_user,
             prisma_client,
             user_api_key_cache,
         )
@@ -1321,6 +1386,7 @@ async def new_team(
 
         validate_budget_duration(data.budget_duration)
         validate_budget_duration(data.team_member_budget_duration)
+        validate_team_model_max_budget(model_max_budget=data.model_max_budget, premium_user=premium_user)
 
         if data.soft_budget is not None:
             if data.max_budget is not None:
@@ -1980,6 +2046,7 @@ async def update_team(
     - tags: Optional[List[str]] - Tags for [tracking spend](https://litellm.vercel.app/docs/proxy/enterprise#tracking-spend-for-custom-tags) and/or doing [tag-based routing](https://litellm.vercel.app/docs/proxy/tag_routing).
     - organization_id: Optional[str] - The organization id of the team. Default is None. Create via `/organization/new`.
     - model_aliases: Optional[dict] - Model aliases for the team. [Docs](https://docs.litellm.ai/docs/proxy/team_based_routing#create-team-with-model-alias)
+    - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
     - guardrails: Optional[List[str]] - Guardrails for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails)
     - policies: Optional[List[str]] - Policies for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails/guardrail_policies)
     - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key.
@@ -2031,6 +2098,7 @@ async def update_team(
         from litellm.proxy.proxy_server import (
             litellm_proxy_admin_name,
             llm_router,
+            premium_user,
             prisma_client,
             proxy_logging_obj,
             user_api_key_cache,
@@ -2069,6 +2137,7 @@ async def update_team(
 
         validate_budget_duration(data.budget_duration)
         validate_budget_duration(data.team_member_budget_duration)
+        validate_team_model_max_budget(model_max_budget=data.model_max_budget, premium_user=premium_user)
 
         existing_team_row = await _raw_team_db(TeamRepository(prisma_client)).find_unique(
             where={"team_id": data.team_id}
@@ -2204,8 +2273,15 @@ async def update_team(
                 user_api_key_dict=user_api_key_dict,
                 existing_team_max_budget=existing_team_row.max_budget,
             )
+        _check_team_model_budget_update_authority(
+            data=data,
+            user_api_key_dict=user_api_key_dict,
+            existing_model_max_budget=existing_team_row.model_max_budget,
+        )
 
         updated_kv = data.json(exclude_unset=True)
+        if "model_max_budget" in updated_kv and updated_kv["model_max_budget"] is None:
+            updated_kv["model_max_budget"] = {}
 
         # Drop server-owned metadata keys from caller input so they can only
         # be written by the same code path that creates the underlying rows.
@@ -4473,7 +4549,7 @@ async def team_info(
     ```
     """
     from litellm.proxy._types import TeamInfoResponseObjectTeamTable
-    from litellm.proxy.proxy_server import prisma_client
+    from litellm.proxy.proxy_server import model_max_budget_limiter, prisma_client
 
     try:
         if prisma_client is None:
@@ -4573,6 +4649,12 @@ async def team_info(
             update={  # mutable-ok: pydantic update payload
                 "members_with_roles": hydrated_members,
                 "organization_models": organization_models,
+                "model_max_budget_usage": await build_model_max_budget_usage(
+                    entity_type=Litellm_EntityType.TEAM,
+                    entity_id=team_id,
+                    model_max_budget=resolved_team_info.model_max_budget,
+                    cache=model_max_budget_limiter.dual_cache,
+                ),
             }
         )
 

litellm/proxy/ocr_endpoints/endpoints.py

@@ -15,12 +15,13 @@ from litellm.llms.base_llm.ocr.transformation import (
     OCRResponse,
     parse_ocr_request_format,
 )
-from litellm.ocr.input import convert_upload_to_url_document, get_max_file_bytes
+from litellm.ocr.legacy import convert_file_document_to_url_document, get_mime_type
 from litellm.proxy._types import *
 from litellm.proxy.auth.user_api_key_auth import UserAPIKeyAuth, user_api_key_auth
 from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
 
 router: Final = APIRouter()
+_MAX_FILE_BYTES: Final = 50 * 1024 * 1024
 
 
 def _build_document_from_upload(
@@ -28,7 +29,15 @@ def _build_document_from_upload(
     filename: str | None,
     content_type: str | None,
 ) -> dict[str, str]:
-    return convert_upload_to_url_document(file_content, filename, content_type)
+    supplied_mime: Final = content_type.split(";")[0].strip() if content_type else None
+    mime_type: Final = (
+        get_mime_type(filename)
+        if filename and (not supplied_mime or supplied_mime == "application/octet-stream")
+        else supplied_mime
+    )
+    return convert_file_document_to_url_document(
+        {"type": "file", "file": file_content, "mime_type": mime_type or "application/octet-stream"}
+    )
 
 
 def _with_request_format(data: Mapping[str, Any], request: Request) -> Mapping[str, Any]:
@@ -103,9 +112,11 @@ async def _parse_multipart_form(request: Request) -> dict[str, Any]:
 
     # Seek to start in case the file was already partially read by middleware
     await uploaded_file.seek(0)
-    file_content: Final = await uploaded_file.read(get_max_file_bytes() + 1)
+    file_content: Final = await uploaded_file.read(_MAX_FILE_BYTES + 1)
     if not file_content:
         raise ValueError("Uploaded file is empty")
+    if len(file_content) > _MAX_FILE_BYTES:
+        raise ValueError("OCR file exceeds the size limit")
 
     document: Final = _build_document_from_upload(
         file_content=file_content,

litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

@@ -609,6 +609,7 @@ class HttpPassThroughEndpointHelpers(BasePassthroughUtils):
         # merely shares the name.
         if not request_dispatched_to_pass_through_endpoint(request):
             _metadata["user_api_key_model_max_budget"] = user_api_key_dict.model_max_budget
+            _metadata["user_api_key_team_model_max_budget"] = user_api_key_dict.team_model_max_budget
             _metadata["user_api_key_user_model_max_budget"] = user_api_key_dict.user_model_max_budget
             _metadata["user_api_key_end_user_model_max_budget"] = user_api_key_dict.end_user_model_max_budget
         _metadata.update(

litellm/proxy/schema.prisma

@@ -426,6 +426,7 @@ model LiteLLM_VerificationToken {
     key_alias   String?
     soft_budget_cooldown Boolean @default(false) // key-level state on if budget alerts need to be cooled down
     spend      Float    @default(0.0)
+    total_spend Float   @default(0.0)
     expires    DateTime?
     models     String[]
     aliases    Json  @default("{}")
@@ -528,6 +529,7 @@ model LiteLLM_DeletedVerificationToken {
     key_alias               String?
     soft_budget_cooldown    Boolean  @default(false)
     spend                   Float    @default(0.0)
+    total_spend             Float    @default(0.0)
     expires                 DateTime?
     models                  String[]
     aliases                 Json     @default("{}")

litellm/proxy/spend_tracking/carried_budget_state.py

@@ -25,6 +25,7 @@ def carry_team_and_user_budget_state(
             budget_reset_at=team_object.budget_reset_at,
             max_budget=team_object.max_budget,
         )
+        valid_token.team_model_max_budget = team_object.model_max_budget  # rebind-ok: caller keeps this object
     if user_object is not None:
         valid_token.user_budget_snapshot = UserBudgetSnapshot(  # rebind-ok: same object the caller keeps using
             budget_reset_at=user_object.budget_reset_at,

litellm/proxy/utils.py

@@ -4340,6 +4340,7 @@ class PrismaClient:
                             v.*,
                             t.spend AS team_spend,
                             t.max_budget AS team_max_budget,
+                            t.model_max_budget AS team_model_max_budget,
                             t.tpm_limit AS team_tpm_limit,
                             t.rpm_limit AS team_rpm_limit,
                             t.tpd_limit AS team_tpd_limit
@@ -4779,6 +4780,7 @@ class PrismaClient:
                             t.spend AS team_spend, 
                             t.max_budget AS team_max_budget,
                             t.soft_budget AS team_soft_budget,
+                            t.model_max_budget AS team_model_max_budget,
                             t.tpm_limit AS team_tpm_limit,
                             t.rpm_limit AS team_rpm_limit,
                             t.tpd_limit AS team_tpd_limit,

litellm/rust_bridge/_native.pyi

@@ -33,15 +33,6 @@ def aocr(
     timeout_seconds: float | None = None,
 ) -> Future[dict[str, object]]: ...
 
-_OCR_MAX_FILE_BYTES: int
-
-def _ocr_upload_document(
-    file_content: bytes,
-    file_name: str | None = None,
-    content_type: str | None = None,
-) -> dict[str, str]: ...
-def _ocr_file_document(document: Mapping[str, object]) -> dict[str, str]: ...
-def _ocr_mime_type(file_name: str) -> str: ...
 def _ocr_lifecycle(
     request: LiteLLMOcrRequest,
     args: tuple[object, ...],
@@ -139,15 +130,11 @@ class TokenCounter:
 def gil_stats() -> dict[str, int]: ...
 
 __all__ = [
-    "_OCR_MAX_FILE_BYTES",
     "ResponsesWebSocketConnection",
     "RustBridgeDeclined",
     "RustUpstreamError",
     "TokenCounter",
-    "_ocr_file_document",
     "_ocr_lifecycle",
-    "_ocr_mime_type",
-    "_ocr_upload_document",
     "achat_completions",
     "amessages",
     "aocr",

model_prices_and_context_window.json

@@ -377,8 +377,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 1.5e-08
+        "supports_tool_choice": true
     },
     "amazon.nova-2-lite-v1:0": {
         "cache_read_input_token_cost": 7.5e-08,
@@ -561,8 +560,7 @@
         "supports_function_calling": true,
         "supports_prompt_caching": true,
         "supports_response_schema": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 8.75e-09
+        "supports_tool_choice": true
     },
     "amazon.nova-pro-v1:0": {
         "cache_read_input_token_cost": 2e-07,
@@ -578,8 +576,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 2e-07
+        "supports_tool_choice": true
     },
     "amazon.nova-sonic-v1:0": {
         "deprecation_date": "2026-09-14",
@@ -45794,8 +45791,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 1.5e-08
+        "supports_tool_choice": true
     },
     "us.amazon.nova-micro-v1:0": {
         "cache_read_input_token_cost": 8.75e-09,
@@ -45809,8 +45805,7 @@
         "supports_function_calling": true,
         "supports_prompt_caching": true,
         "supports_response_schema": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 8.75e-09
+        "supports_tool_choice": true
     },
     "us.amazon.nova-premier-v1:0": {
         "deprecation_date": "2026-09-14",
@@ -45842,8 +45837,7 @@
         "supports_prompt_caching": true,
         "supports_response_schema": true,
         "supports_vision": true,
-        "supports_tool_choice": true,
-        "cache_read_input_token_cost": 2e-07
+        "supports_tool_choice": true
     },
     "us.anthropic.claude-3-5-haiku-20241022-v1:0": {
         "cache_creation_input_token_cost": 1e-06,

schema.prisma

@@ -426,6 +426,7 @@ model LiteLLM_VerificationToken {
     key_alias   String?
     soft_budget_cooldown Boolean @default(false) // key-level state on if budget alerts need to be cooled down
     spend      Float    @default(0.0)
+    total_spend Float   @default(0.0)
     expires    DateTime?
     models     String[]
     aliases    Json  @default("{}")
@@ -528,6 +529,7 @@ model LiteLLM_DeletedVerificationToken {
     key_alias               String?
     soft_budget_cooldown    Boolean  @default(false)
     spend                   Float    @default(0.0)
+    total_spend             Float    @default(0.0)
     expires                 DateTime?
     models                  String[]
     aliases                 Json     @default("{}")

tests/proxy_unit_tests/test_unit_test_max_model_budget_limiter.py

@@ -587,6 +587,8 @@ def _success_kwargs(
     response_cost=0.5,
     key_hash=None,
     key_model_max_budget=None,
+    team_id=None,
+    team_model_max_budget=None,
     user_id=None,
     user_model_max_budget=None,
     end_user_id=None,
@@ -600,6 +602,7 @@ def _success_kwargs(
             "end_user": end_user_id,
             "metadata": {
                 "user_api_key_hash": key_hash,
+                "user_api_key_team_id": team_id,
                 "user_api_key_user_id": user_id,
                 "user_api_key_end_user_id": end_user_id,
             },
@@ -607,6 +610,7 @@ def _success_kwargs(
         "litellm_params": {
             "metadata": {
                 "user_api_key_model_max_budget": key_model_max_budget,
+                "user_api_key_team_model_max_budget": team_model_max_budget,
                 "user_api_key_user_model_max_budget": user_model_max_budget,
                 "user_api_key_end_user_model_max_budget": end_user_model_max_budget,
             },
@@ -1417,3 +1421,266 @@ async def test_spend_logged_on_one_replica_is_enforced_and_reported_on_another()
     replica_c = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=DualCache(redis_cache=shared_redis))
     with pytest.raises(litellm.BudgetExceededError):
         await replica_c.is_key_within_model_budget(user_api_key, "gpt-4")
+
+
+def _log_success(limiter, **kwargs):
+    return limiter.async_log_success_event(
+        _success_kwargs(**kwargs), response_obj=None, start_time=None, end_time=None
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "request_model",
+    ["gpt-4", "openai/gpt-4"],
+    ids=["bare_model", "provider_prefixed_model"],
+)
+async def test_team_model_budget_is_shared_by_every_key_without_an_override(request_model):
+    """
+    Two keys on the same team, neither carrying a matching key-level entry,
+    charge one team counter and are both refused once it is spent.
+    """
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {"gpt-4": {"budget_limit": 1.0, "time_period": "1d"}}
+    check = lambda: limiter.is_team_within_model_budget(
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+        key_model_max_budget=None,
+        model=request_model,
+    )
+
+    assert await check() is True
+    await _log_success(
+        limiter,
+        model_group=request_model,
+        response_cost=0.6,
+        key_hash="vk-a",
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+    )
+    assert await check() is True
+    await _log_success(
+        limiter,
+        model_group=request_model,
+        response_cost=0.6,
+        key_hash="vk-b",
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+    )
+
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:gpt-4:1d") == pytest.approx(1.2)
+    with pytest.raises(litellm.BudgetExceededError) as exc:
+        await check()
+    assert exc.value.entity_type == Litellm_EntityType.TEAM.value
+    assert await build_model_max_budget_usage(
+        entity_type=Litellm_EntityType.TEAM,
+        entity_id="team-1",
+        model_max_budget=team_model_max_budget,
+        cache=dual_cache,
+    ) == {"gpt-4": {"current_spend": pytest.approx(1.2), "budget_limit": 1.0, "time_period": "1d"}}
+
+
+@pytest.mark.asyncio
+async def test_key_override_replaces_the_team_cap_for_that_model():
+    """
+    A key with its own entry for the model is gated on the key counter alone:
+    the exhausted team counter does not block it, and its spend never lands on
+    the team counter.
+    """
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {"gpt-4": {"budget_limit": 1.0, "time_period": "1d"}}
+    key_model_max_budget = {"gpt-4": {"budget_limit": 5.0, "time_period": "1d"}}
+    await dual_cache.async_set_cache(key="team_model_spend:team-1:gpt-4:1d", value=9.0)
+
+    assert (
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=key_model_max_budget,
+            model="openai/gpt-4",
+        )
+        is True
+    )
+
+    await _log_success(
+        limiter,
+        model_group="openai/gpt-4",
+        response_cost=2.0,
+        key_hash="vk-override",
+        key_model_max_budget=key_model_max_budget,
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+    )
+
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:gpt-4:1d") == 9.0
+    assert await dual_cache.async_get_cache(key="virtual_key_spend:vk-override:gpt-4:1d") == 2.0
+
+
+@pytest.mark.asyncio
+async def test_key_entry_for_another_model_does_not_lift_the_team_cap():
+    """A key override only covers the model it names; other models stay on the team counter."""
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {"gpt-4": {"budget_limit": 1.0, "time_period": "1d"}}
+    key_model_max_budget = {"claude-3": {"budget_limit": 5.0, "time_period": "1d"}}
+
+    await _log_success(
+        limiter,
+        model_group="gpt-4",
+        response_cost=1.5,
+        key_hash="vk-other",
+        key_model_max_budget=key_model_max_budget,
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+    )
+
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:gpt-4:1d") == 1.5
+    with pytest.raises(litellm.BudgetExceededError):
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=key_model_max_budget,
+            model="gpt-4",
+        )
+
+
+@pytest.mark.asyncio
+async def test_team_budget_leaves_unconfigured_models_alone():
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {"gpt-4": {"budget_limit": 0.0, "time_period": "1d"}}
+
+    assert (
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=None,
+            model="claude-3",
+        )
+        is True
+    )
+    with patch.object(limiter, "_increment_spend_for_key", new_callable=AsyncMock) as mock_increment:
+        await _log_success(
+            limiter,
+            model_group="claude-3",
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+        )
+    mock_increment.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_team_counters_are_isolated_by_team_model_and_window():
+    """Same model on two teams, and two models with different windows on one team, never share a counter."""
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {
+        "gpt-4": {"budget_limit": 10.0, "time_period": "1d"},
+        "claude-3": {"budget_limit": 10.0, "time_period": "30d"},
+    }
+
+    for team_id, model in (("team-1", "gpt-4"), ("team-2", "gpt-4"), ("team-1", "claude-3")):
+        await _log_success(
+            limiter,
+            model_group=model,
+            response_cost=1.0,
+            team_id=team_id,
+            team_model_max_budget=team_model_max_budget,
+        )
+
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:gpt-4:1d") == 1.0
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-2:gpt-4:1d") == 1.0
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:claude-3:30d") == 1.0
+    assert await dual_cache.async_get_cache(key="team_model_budget_start_time:team-1:claude-3:30d") is not None
+
+
+@pytest.mark.asyncio
+async def test_malformed_team_entry_is_skipped_and_its_sibling_still_enforced():
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=DualCache())
+    team_model_max_budget = {
+        "gpt-4": {"budget_limit": "not-a-number", "time_period": "1d"},
+        "claude-3": {"budget_limit": 0.0, "time_period": "1d"},
+    }
+
+    assert (
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=None,
+            model="gpt-4",
+        )
+        is True
+    )
+    with pytest.raises(litellm.BudgetExceededError):
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=None,
+            model="claude-3",
+        )
+
+
+@pytest.mark.asyncio
+async def test_malformed_key_entry_does_not_count_as_an_override():
+    """A key entry the limiter cannot enforce must not also switch the team cap off."""
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {"gpt-4": {"budget_limit": 1.0, "time_period": "1d"}}
+    key_model_max_budget = {"gpt-4": {"budget_limit": "not-a-number", "time_period": "1d"}}
+
+    await _log_success(
+        limiter,
+        model_group="gpt-4",
+        response_cost=1.5,
+        key_hash="vk-bad",
+        key_model_max_budget=key_model_max_budget,
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+    )
+
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:gpt-4:1d") == 1.5
+    with pytest.raises(litellm.BudgetExceededError):
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=key_model_max_budget,
+            model="gpt-4",
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "key_entry",
+    [
+        {"time_period": "1d", "tpm_limit": 100},
+        {"time_period": "1d", "rpm_limit": 10},
+        {"budget_limit": -1.0, "time_period": "1d"},
+    ],
+)
+async def test_key_entry_without_a_spend_cap_does_not_lift_the_team_cap(key_entry):
+    """A key row that only rate-limits the model, or has no enforceable cap, leaves the team cap in force."""
+    dual_cache = DualCache()
+    limiter = _PROXY_VirtualKeyModelMaxBudgetLimiter(dual_cache=dual_cache)
+    team_model_max_budget = {"gpt-4": {"budget_limit": 1.0, "time_period": "1d"}}
+    key_model_max_budget = {"gpt-4": key_entry}
+
+    await _log_success(
+        limiter,
+        model_group="openai/gpt-4",
+        response_cost=1.5,
+        key_hash="vk-rate-limited",
+        key_model_max_budget=key_model_max_budget,
+        team_id="team-1",
+        team_model_max_budget=team_model_max_budget,
+    )
+
+    assert await dual_cache.async_get_cache(key="team_model_spend:team-1:gpt-4:1d") == 1.5
+    with pytest.raises(litellm.BudgetExceededError):
+        await limiter.is_team_within_model_budget(
+            team_id="team-1",
+            team_model_max_budget=team_model_max_budget,
+            key_model_max_budget=key_model_max_budget,
+            model="openai/gpt-4",
+        )

tests/test_litellm/integrations/test_s3.py

@@ -1,16 +1,24 @@
+import copy
+import json
 from datetime import datetime
 from unittest.mock import MagicMock, patch
 
+import pytest
+
 import litellm
 from litellm.constants import MAX_S3_OBJECT_DOWNLOAD_FILENAME_BYTES, MAX_S3_OBJECT_KEY_BYTES
-from litellm.integrations.s3 import S3Logger
+from litellm.integrations.s3 import S3Logger, prompts_only_payload, resolve_s3_log_prompts_only
 
 TEST_KMS_KEY_ARN = "arn:aws:kms:us-east-1:111122223333:key/test-key-id"
+TEST_MESSAGES = [{"role": "user", "content": "Reply with exactly the word PINEAPPLE."}]
+TEST_RESPONSE = {"choices": [{"message": {"role": "assistant", "content": "PINEAPPLE"}}]}
 
 
 def _standard_logging_payload(response_id: str = "chatcmpl-test-id") -> dict:
     return {
         "id": response_id,
+        "messages": copy.deepcopy(TEST_MESSAGES),
+        "response": copy.deepcopy(TEST_RESPONSE),
         "metadata": {"user_api_key_team_alias": None},
     }
 
@@ -22,7 +30,9 @@ def _log_event_kwargs(response_id: str = "chatcmpl-test-id") -> dict:
     }
 
 
-def _run_log_event(callback_params: dict, response_id: str = "chatcmpl-test-id") -> MagicMock:
+def _run_log_event(
+    callback_params: dict, response_id: str = "chatcmpl-test-id", log_kwargs: dict[str, object] | None = None
+) -> MagicMock:
     original = litellm.s3_callback_params
     litellm.s3_callback_params = callback_params
     try:
@@ -31,7 +41,7 @@ def _run_log_event(callback_params: dict, response_id: str = "chatcmpl-test-id")
             mock_boto3_client.return_value = mock_s3_client
             logger = S3Logger()
             logger.log_event(
-                kwargs=_log_event_kwargs(response_id),
+                kwargs=_log_event_kwargs(response_id) if log_kwargs is None else log_kwargs,
                 response_obj={"id": response_id},
                 start_time=datetime(2026, 7, 30, 12, 0, 0),
                 end_time=datetime(2026, 7, 30, 12, 0, 1),
@@ -182,3 +192,123 @@ def test_put_object_keeps_the_configured_path_intact_when_only_the_id_has_to_shr
     key = mock_s3_client.put_object.call_args.kwargs["Key"]
     assert key.startswith(long_path + "/2026-07-30/")
     assert len(key.encode("utf-8")) == MAX_S3_OBJECT_KEY_BYTES
+
+
+def _uploaded_body(mock_s3_client: MagicMock) -> dict[str, object]:
+    return json.loads(mock_s3_client.put_object.call_args.kwargs["Body"])
+
+
+def test_log_event_prompts_only_drops_response_and_keeps_messages(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv("S3_LOG_PROMPTS_ONLY", raising=False)
+    log_kwargs = _log_event_kwargs()
+    original_payload = copy.deepcopy(log_kwargs["standard_logging_object"])
+
+    mock_s3_client = _run_log_event(
+        {"s3_bucket_name": "test-bucket", "s3_region_name": "us-east-1", "s3_log_prompts_only": True},
+        log_kwargs=log_kwargs,
+    )
+
+    body = _uploaded_body(mock_s3_client)
+    assert body["messages"] == TEST_MESSAGES
+    assert body["response"] is None
+    assert body["id"] == "chatcmpl-test-id"
+    assert log_kwargs["standard_logging_object"] == original_payload
+
+
+def test_log_event_default_keeps_response(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv("S3_LOG_PROMPTS_ONLY", raising=False)
+
+    mock_s3_client = _run_log_event({"s3_bucket_name": "test-bucket", "s3_region_name": "us-east-1"})
+
+    body = _uploaded_body(mock_s3_client)
+    assert body["response"] == TEST_RESPONSE
+    assert body["messages"] == TEST_MESSAGES
+
+
+def test_log_event_reads_prompts_only_env_var_at_log_time(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv("S3_LOG_PROMPTS_ONLY", raising=False)
+    original = litellm.s3_callback_params
+    litellm.s3_callback_params = {"s3_bucket_name": "test-bucket", "s3_region_name": "us-east-1"}
+    try:
+        with patch("boto3.client") as mock_boto3_client:
+            mock_s3_client = MagicMock()
+            mock_boto3_client.return_value = mock_s3_client
+            logger = S3Logger()
+            monkeypatch.setenv("S3_LOG_PROMPTS_ONLY", "true")
+            logger.log_event(
+                kwargs=_log_event_kwargs(),
+                response_obj={"id": "chatcmpl-test-id"},
+                start_time=datetime(2026, 7, 30, 12, 0, 0),
+                end_time=datetime(2026, 7, 30, 12, 0, 1),
+                print_verbose=lambda *args, **kwargs: None,
+            )
+    finally:
+        litellm.s3_callback_params = original
+
+    body = _uploaded_body(mock_s3_client)
+    assert body["response"] is None
+    assert body["messages"] == TEST_MESSAGES
+
+
+def test_log_event_explicit_false_param_beats_env_var(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv("S3_LOG_PROMPTS_ONLY", "true")
+
+    mock_s3_client = _run_log_event(
+        {"s3_bucket_name": "test-bucket", "s3_region_name": "us-east-1", "s3_log_prompts_only": False}
+    )
+
+    assert _uploaded_body(mock_s3_client)["response"] == TEST_RESPONSE
+
+
+def test_s3_logger_init_does_not_mutate_global_callback_params(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv("MY_S3_BUCKET", "resolved-bucket")
+    callback_params = {"s3_bucket_name": "os.environ/MY_S3_BUCKET", "s3_region_name": "us-east-1"}
+    snapshot = copy.deepcopy(callback_params)
+    original = litellm.s3_callback_params
+    litellm.s3_callback_params = callback_params
+    try:
+        with patch("boto3.client"):
+            logger = S3Logger()
+    finally:
+        litellm.s3_callback_params = original
+
+    assert logger.bucket_name == "resolved-bucket"
+    assert callback_params == snapshot
+
+
+@pytest.mark.parametrize(
+    "configured,env_value,expected",
+    [
+        (True, None, True),
+        (False, "true", False),
+        ("true", None, True),
+        ("False", "true", False),
+        ("1", None, True),
+        ("0", None, False),
+        (" yes ", None, True),
+        (None, None, False),
+        (None, "true", True),
+        (None, "false", False),
+        (None, "", False),
+        ("", "true", False),
+    ],
+)
+def test_resolve_s3_log_prompts_only(configured: object, env_value: str | None, expected: bool):
+    environ = {} if env_value is None else {"S3_LOG_PROMPTS_ONLY": env_value}
+    assert resolve_s3_log_prompts_only(configured, environ) is expected
+
+
+def test_resolve_s3_log_prompts_only_unparseable_value_fails_toward_prompts_only():
+    assert resolve_s3_log_prompts_only("enabled", {}) is True
+
+
+def test_prompts_only_payload_returns_copy_with_response_cleared():
+    payload = _standard_logging_payload()
+    snapshot = copy.deepcopy(payload)
+
+    stripped = prompts_only_payload(payload)
+
+    assert stripped["response"] is None
+    assert stripped["messages"] == TEST_MESSAGES
+    assert stripped is not payload
+    assert payload == snapshot

tests/test_litellm/integrations/test_s3_v2.py

@@ -1,8 +1,11 @@
 import asyncio
+import copy
+import json
 import re
 import sys
 import textwrap
 import uuid
+from collections.abc import Awaitable, Callable
 from contextlib import asynccontextmanager
 from datetime import datetime
 from pathlib import Path
@@ -10,6 +13,7 @@ from unittest.mock import AsyncMock, MagicMock, call, patch
 
 import httpx
 import pytest
+import respx
 
 from litellm.integrations.s3_v2 import S3Logger
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
@@ -2310,3 +2314,137 @@ def _s3_logger_for_region(region_name: str) -> S3Logger:
 )
 def test_build_object_url_uses_partition_dns_suffix(region_name: str, expected_url: str) -> None:
     assert _s3_logger_for_region(region_name)._build_object_url("2025-01-01/key.json") == expected_url
+
+
+def _prompts_only_logger(s3_log_prompts_only: bool | None = None) -> S3Logger:
+    return S3Logger(
+        s3_bucket_name="test-bucket",
+        s3_aws_access_key_id="test-key",
+        s3_aws_secret_access_key="test-secret",
+        s3_region_name="us-east-1",
+        s3_log_prompts_only=s3_log_prompts_only,
+    )
+
+
+def _chat_payload() -> StandardLoggingPayload:
+    return StandardLoggingPayload(
+        id="chatcmpl-prompts-only",
+        messages=[{"role": "user", "content": "Reply with exactly the word PINEAPPLE."}],
+        response={"choices": [{"message": {"role": "assistant", "content": "PINEAPPLE"}}]},
+        metadata={"user_api_key_team_alias": None},
+    )
+
+
+async def _queued_body_via_async_upload(
+    logger: S3Logger, log_event: Callable[..., Awaitable[None]]
+) -> dict[str, object]:
+    payload = _chat_payload()
+    original = copy.deepcopy(payload)
+    await log_event(
+        kwargs={"standard_logging_object": payload},
+        response_obj=None,
+        start_time=datetime(2026, 7, 30, 12, 0, 0),
+        end_time=datetime(2026, 7, 30, 12, 0, 1),
+    )
+    assert payload == original, "the caller's standard_logging_object must not be mutated"
+    (element,) = logger.log_queue
+
+    response = MagicMock()
+    response.status_code = 200
+    response.raise_for_status = MagicMock()
+    logger.async_httpx_client = AsyncMock()
+    logger.async_httpx_client.put.return_value = response
+    await logger.async_upload_data_to_s3(element)
+    return json.loads(logger.async_httpx_client.put.call_args.kwargs["data"])
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("event_name", ["async_log_success_event", "async_log_failure_event"])
+async def test_prompts_only_drops_response_but_keeps_messages_in_uploaded_object(
+    monkeypatch: pytest.MonkeyPatch, event_name: str
+):
+    import litellm
+
+    monkeypatch.setattr(litellm, "s3_callback_params", {"s3_log_prompts_only": True})
+    logger = _prompts_only_logger()
+
+    log_event: Callable[..., Awaitable[None]] = (
+        logger.async_log_success_event if event_name == "async_log_success_event" else logger.async_log_failure_event
+    )
+    body = await _queued_body_via_async_upload(logger, log_event)
+
+    assert body["messages"] == _chat_payload()["messages"]
+    assert body["response"] is None
+    assert body["id"] == "chatcmpl-prompts-only"
+
+
+@pytest.mark.asyncio
+async def test_prompts_only_default_off_keeps_response_in_uploaded_object(monkeypatch: pytest.MonkeyPatch):
+    import litellm
+
+    monkeypatch.setattr(litellm, "s3_callback_params", {})
+    monkeypatch.delenv("S3_LOG_PROMPTS_ONLY", raising=False)
+    logger = _prompts_only_logger()
+
+    body = await _queued_body_via_async_upload(logger, logger.async_log_success_event)
+
+    assert body["response"] == _chat_payload()["response"]
+    assert body["messages"] == _chat_payload()["messages"]
+
+
+@pytest.mark.asyncio
+async def test_prompts_only_explicit_false_in_params_beats_env_var(monkeypatch: pytest.MonkeyPatch):
+    import litellm
+
+    monkeypatch.setattr(litellm, "s3_callback_params", {"s3_log_prompts_only": False})
+    monkeypatch.setenv("S3_LOG_PROMPTS_ONLY", "true")
+    logger = _prompts_only_logger()
+
+    body = await _queued_body_via_async_upload(logger, logger.async_log_success_event)
+
+    assert body["response"] == _chat_payload()["response"]
+
+
+@pytest.mark.asyncio
+async def test_prompts_only_env_var_applies_when_param_unset(monkeypatch: pytest.MonkeyPatch):
+    import litellm
+
+    monkeypatch.setattr(litellm, "s3_callback_params", {})
+    logger = _prompts_only_logger()
+    monkeypatch.setenv("S3_LOG_PROMPTS_ONLY", "true")
+
+    body = await _queued_body_via_async_upload(logger, logger.async_log_success_event)
+
+    assert body["response"] is None
+    assert body["messages"] == _chat_payload()["messages"]
+
+
+@respx.mock
+def test_prompts_only_constructor_kwarg_applies_to_sync_upload(monkeypatch: pytest.MonkeyPatch):
+    import litellm
+
+    monkeypatch.setattr(litellm, "s3_callback_params", {})
+    monkeypatch.delenv("S3_LOG_PROMPTS_ONLY", raising=False)
+    logger = _prompts_only_logger(s3_log_prompts_only=True)
+    payload = _chat_payload()
+
+    element = logger.create_s3_batch_logging_element(
+        start_time=datetime(2026, 7, 30, 12, 0, 0),
+        standard_logging_payload=payload,
+    )
+    assert element is not None
+    assert payload["response"] == _chat_payload()["response"]
+
+    put_route = respx.put(url__regex=r"https://test-bucket\.s3\..*").mock(return_value=httpx.Response(200))
+    logger.upload_data_to_s3(element)
+
+    body = json.loads(put_route.calls.last.request.content)
+    assert body["response"] is None
+    assert body["messages"] == _chat_payload()["messages"]
+
+
+@pytest.mark.parametrize("callback_name", ["s3", "s3_v2"])
+def test_prompts_only_toggle_is_exposed_to_admin_ui_for_both_s3_callbacks(callback_name: str):
+    from litellm.integrations.custom_logger import CustomLogger
+
+    assert "S3_LOG_PROMPTS_ONLY" in CustomLogger.get_callback_env_vars(callback_name)

tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py

@@ -1200,6 +1200,7 @@ def _fake_user_api_key_auth(
     team_models=None,
     team_id=None,
     model_max_budget=None,
+    team_model_max_budget=None,
     end_user_model_max_budget=None,
     end_user_id=None,
     user_model_max_budget=None,
@@ -1220,6 +1221,7 @@ def _fake_user_api_key_auth(
     auth.team_id = team_id
     auth.team_model_aliases = None
     auth.model_max_budget = model_max_budget
+    auth.team_model_max_budget = team_model_max_budget
     auth.end_user_model_max_budget = end_user_model_max_budget
     auth.end_user_id = end_user_id
     auth.user_model_max_budget = user_model_max_budget
@@ -1860,6 +1862,78 @@ async def test_summary_model_rate_limit_skipped_for_legacy_limiter():
     assert not result.applied_edits[0].get("error")
 
 
+async def test_summary_model_denied_when_team_over_model_budget():
+    """The team per-model budget gates the summary subrequest, whose spend is
+    charged to the team counter via the propagated `user_api_key_team_model_max_budget`.
+    The key's own `model_max_budget` is handed to the limiter so a key-level
+    override keeps taking precedence over the team cap here as it does in auth."""
+    import litellm
+
+    messages = _simple_messages()
+    mock_call = AsyncMock(return_value=_make_mock_response("<summary>x</summary>"))
+    key_budget = {"claude-opus-4-8": {"budget_limit": 1}}
+    team_budget = {"claude-haiku-4-5": {"budget_limit": 5, "time_period": "1d"}}
+
+    auth = _fake_user_api_key_auth(
+        key_models=["all-proxy-models"],
+        model_max_budget=key_budget,
+        team_model_max_budget=team_budget,
+        team_id="team-over-budget",
+        token="hashed-token",
+    )
+
+    limiter = MagicMock()
+    limiter.is_key_within_model_budget = AsyncMock(return_value=True)
+    limiter.is_team_within_model_budget = AsyncMock(
+        side_effect=litellm.BudgetExceededError(
+            message="over budget", current_cost=10, max_budget=5
+        )
+    )
+
+    with (
+        patch(  # test-quality-ok: apply_compact_20260112 reads the summary model setting as a module global, no seam
+            "litellm.llms.anthropic.experimental_pass_through.context_management.editors.compact._read_summary_model_setting",
+            return_value="claude-haiku-4-5",
+        ),
+        patch("litellm.token_counter", return_value=200_000),  # test-quality-ok: forces the over-threshold branch
+        patch(  # test-quality-ok: the summary call is the observable that must NOT happen when the team is over budget
+            "litellm.llms.anthropic.experimental_pass_through.context_management.editors.compact._call_summary_model",
+            mock_call,
+        ),
+        patch(  # test-quality-ok: the limiter is a proxy_server module global the editor imports, no injection seam
+            "litellm.proxy.proxy_server.model_max_budget_limiter", limiter
+        ),
+    ):
+        result = await apply_compact_20260112(
+            model=MODEL,
+            messages=messages,
+            tools=None,
+            system=None,
+            edit_spec=_EDIT_SPEC_DEFAULT,
+            user_api_key_auth=auth,
+        )
+
+    mock_call.assert_not_awaited()
+    assert result.applied_edits[0].get("error") == "summary_model_budget_exceeded"
+    limiter.is_team_within_model_budget.assert_awaited_once_with(
+        team_id="team-over-budget",
+        team_model_max_budget=team_budget,
+        key_model_max_budget=key_budget,
+        model="claude-haiku-4-5",
+    )
+    import inspect
+
+    from litellm.proxy.hooks.model_max_budget_limiter import (
+        _PROXY_VirtualKeyModelMaxBudgetLimiter,
+    )
+
+    real_params = inspect.signature(
+        _PROXY_VirtualKeyModelMaxBudgetLimiter.is_team_within_model_budget
+    ).parameters
+    for kwarg in ("team_id", "team_model_max_budget", "key_model_max_budget", "model"):
+        assert kwarg in real_params, f"compact.py passes {kwarg}=, which the limiter does not accept"
+
+
 async def test_scoped_budget_metadata_propagated_to_summary_call():
     """The end-user/project scope identifiers and the end-user budget the post-call
     spend and rate-limit hooks key on are forwarded to the summary subrequest, and

tests/test_litellm/models/test_models.py

@@ -362,6 +362,14 @@ class TestVerificationToken:
         assert deleted.deleted_at is not None
         assert deleted.token == "t1"
 
+    def test_total_spend_is_carried_separately_from_resettable_spend(self):
+        token = LiteLLM_VerificationToken(token="t1", spend=0.0, total_spend=12.5)
+        assert token.model_dump()["total_spend"] == 12.5
+        assert token.model_dump()["spend"] == 0.0
+
+        deleted = LiteLLM_DeletedVerificationToken.model_validate({**token.model_dump(), "deleted_by": "admin"})
+        assert deleted.total_spend == 12.5
+
 
 class TestConfigTable:
     def test_config_creation(self):

tests/test_litellm/ocr/test_ocr_file_input.py

@@ -12,32 +12,16 @@ Tests that:
 import base64
 import os
 import tempfile
-from collections.abc import Generator
 from io import BytesIO
 from pathlib import Path
 from typing import Final
-from unittest.mock import AsyncMock, MagicMock, Mock
+from unittest.mock import AsyncMock, MagicMock
 
 import orjson
 import pytest
 from starlette.datastructures import FormData
 
-from litellm.ocr.input import convert_file_document_to_url_document, get_mime_type
-
-
-@pytest.fixture(autouse=True, params=["native", "disabled", "unavailable"])
-def document_runtime(request: pytest.FixtureRequest, monkeypatch: pytest.MonkeyPatch) -> Generator[None]:
-    from litellm.rust_bridge import bindings, configuration
-
-    configuration.reset_rust_configuration()
-    monkeypatch.delenv("LITELLM_RUST", raising=False)
-    if request.param == "disabled":
-        monkeypatch.setenv("LITELLM_RUST", "0")
-        monkeypatch.setattr(bindings, "get_native_bridge", Mock(side_effect=AssertionError("Rust is disabled")))
-    elif request.param == "unavailable":
-        monkeypatch.setattr(bindings, "get_native_bridge", lambda: None)
-    yield
-    configuration.reset_rust_configuration()
+from litellm.ocr.legacy import convert_file_document_to_url_document, get_mime_type
 
 
 class TestGetMimeType:
@@ -503,10 +487,9 @@ class TestProxySecurityGuard:
 async def test_proxy_upload_stops_reading_at_size_limit() -> None:
     from starlette.datastructures import UploadFile
 
-    from litellm.ocr.input import get_max_file_bytes
-    from litellm.proxy.ocr_endpoints.endpoints import _parse_multipart_form
+    from litellm.proxy.ocr_endpoints.endpoints import _MAX_FILE_BYTES, _parse_multipart_form
 
-    limit: Final = get_max_file_bytes()
+    limit: Final = _MAX_FILE_BYTES
     with tempfile.TemporaryFile() as stream:
         stream.truncate(limit * 2)
         upload: Final = UploadFile(file=stream, filename="large.pdf")

tests/test_litellm/proxy/auth/test_team_grants.py

@@ -31,6 +31,7 @@ def _full_team(model_aliases=ALIASES) -> LiteLLM_TeamTable:
         max_budget=50.0,
         soft_budget=25.0,
         spend=12.5,
+        model_max_budget={"gpt-4o": {"max_budget": 5.0, "budget_duration": "1d"}},
         models=["gpt-4o", "gpt-4o-mini"],
         blocked=True,
         metadata={"tier": "gold"},
@@ -72,6 +73,7 @@ def test_team_grants_cover_every_team_field_the_key_path_gets():
     assert token.team_max_budget == 50.0
     assert token.team_soft_budget == 25.0
     assert token.team_spend == 12.5
+    assert token.team_model_max_budget == {"gpt-4o": {"max_budget": 5.0, "budget_duration": "1d"}}
     assert token.team_models == ["gpt-4o", "gpt-4o-mini"]
     assert token.team_blocked is True
     assert token.team_metadata == {"tier": "gold"}

tests/test_litellm/proxy/auth/test_user_api_key_auth.py

@@ -4374,6 +4374,74 @@ async def test_centralized_common_checks_carries_team_and_user_budget_state_on_t
     }
 
 
+class _RecordingTeamModelBudgetLimiter:
+    def __init__(self):
+        self.calls = []
+
+    async def is_team_within_model_budget(self, team_id, team_model_max_budget, key_model_max_budget, model):
+        self.calls.append((team_id, dict(team_model_max_budget), key_model_max_budget, model))
+        return True
+
+
+@pytest.mark.asyncio
+async def test_centralized_common_checks_enforces_team_model_max_budget_from_the_resolved_team():
+    """The team's model_max_budget is enforced at the single authz gate, off the
+    team object auth resolved (not the possibly stale token copy), and the key's
+    own model_max_budget is handed to the limiter so a matching key entry can
+    override the team cap."""
+    from fastapi import Request
+    from starlette.datastructures import URL
+
+    import litellm.proxy.proxy_server as _proxy_server_mod
+
+    team_caps = {"gpt-4o": {"max_budget": 5.0, "budget_duration": "1d"}}
+    key_caps = {"claude-sonnet-4-6": {"max_budget": 1.0, "budget_duration": "1d"}}
+    token = UserAPIKeyAuth(
+        api_key="sk-test",
+        token="hashed",
+        team_id="t1",
+        team_model_max_budget={"gpt-4o": {"max_budget": 999.0, "budget_duration": "30d"}},
+        model_max_budget=key_caps,
+    )
+    request = Request(scope={"type": "http"})
+    request._url = URL(url="/chat/completions")
+
+    user_api_key_cache = DualCache()
+    await user_api_key_cache.async_set_cache(
+        key="team_id:t1",
+        value=LiteLLM_TeamTableCachedObj(team_id="t1", model_max_budget=team_caps),
+    )
+    limiter = _RecordingTeamModelBudgetLimiter()
+    attrs = {
+        **_proxy_attrs_for_centralized_checks(user_custom_auth=None),
+        "prisma_client": MagicMock(),
+        "user_api_key_cache": user_api_key_cache,
+        "model_max_budget_limiter": limiter,
+    }
+    originals = {a: getattr(_proxy_server_mod, a, None) for a in attrs}
+    try:
+        for k, v in attrs.items():
+            setattr(_proxy_server_mod, k, v)
+        with (
+            patch("litellm.proxy.auth.user_api_key_auth.common_checks", new_callable=AsyncMock),  # test-quality-ok: stubs the sibling check so only the team model-budget gate is under test
+            patch(  # test-quality-ok: stubs the budget reservation so only the team model-budget gate is under test
+                "litellm.proxy.auth.user_api_key_auth._reserve_budget_after_common_checks",
+                new_callable=AsyncMock,
+            ),
+        ):
+            await _run_centralized_common_checks(
+                user_api_key_auth_obj=token,
+                request=request,
+                request_data={"model": "gpt-4o"},
+                route="/chat/completions",
+            )
+    finally:
+        for k, v in originals.items():
+            setattr(_proxy_server_mod, k, v)
+
+    assert limiter.calls == [("t1", team_caps, key_caps, "gpt-4o")]
+
+
 @pytest.mark.asyncio
 async def test_centralized_common_checks_skipped_for_custom_auth_without_flag():
     """Existing RPS guarantee: custom-auth deployments without

tests/test_litellm/proxy/common_utils/test_reset_budget_job.py

@@ -291,6 +291,23 @@ def test_reset_budget_for_key(reset_budget_job, mock_prisma_client):
     assert set(write["data"].keys()) == {"spend", "budget_reset_at"}
 
 
+def test_reset_budget_for_key_leaves_lifetime_total_spend_alone(reset_budget_job, mock_prisma_client):
+    """A period reset zeroes spend but must neither write nor touch the lifetime total_spend."""
+    now = datetime.now(timezone.utc)
+    key = LiteLLM_VerificationToken(
+        token="tok-key-1", spend=100.0, total_spend=340.0, budget_duration="30d", budget_reset_at=now
+    )
+    mock_prisma_client.data["key"] = [key]
+
+    asyncio.run(reset_budget_job.reset_budget_for_litellm_keys())
+
+    (write,) = _batch_writes(mock_prisma_client, "key")
+    assert write["data"]["spend"] == {"decrement": 100.0}
+    assert "total_spend" not in write["data"]
+    assert key.spend == 0.0
+    assert key.total_spend == 340.0
+
+
 def test_reset_budget_for_key_honors_injected_reset_time(mock_prisma_client, mock_proxy_logging):
     """Injected BudgetResetSettings drives the written reset time end to end (DI, no globals).
 

tests/test_litellm/proxy/db/test_create_views.py

@@ -71,6 +71,7 @@ async def test_create_views_creates_view_on_does_not_exist():
     mock_db.execute_raw.assert_called_once()
     created_sql = mock_db.execute_raw.call_args[0][0]
     assert 'CREATE VIEW "LiteLLM_VerificationTokenView"' in created_sql
+    assert "t.model_max_budget AS team_model_max_budget" in created_sql
 
 
 @pytest.mark.asyncio

tests/test_litellm/proxy/db/test_db_spend_update_writer.py

@@ -1658,6 +1658,57 @@ async def test_commit_key_spend_updates_includes_last_active():
     assert before_call <= last_active <= after_call
 
 
+@pytest.mark.asyncio
+async def test_commit_spend_updates_to_db_increments_key_total_spend_alongside_spend():
+    """
+    The key table write must increment the lifetime total_spend by the same amount as the
+    resettable spend, in the same update so the two cannot drift.
+    """
+    db_writer = DBSpendUpdateWriter()
+
+    mock_batcher = MagicMock()
+    mock_batcher.litellm_verificationtoken = MagicMock()
+    mock_batcher.litellm_verificationtoken.update_many = MagicMock()
+
+    mock_transaction = AsyncMock()
+    mock_transaction.__aenter__ = AsyncMock(return_value=mock_transaction)
+    mock_transaction.__aexit__ = AsyncMock(return_value=False)
+    mock_transaction.batch_ = MagicMock(
+        return_value=AsyncMock(
+            __aenter__=AsyncMock(return_value=mock_batcher),
+            __aexit__=AsyncMock(return_value=False),
+        )
+    )
+
+    mock_prisma_client = MagicMock()
+    mock_prisma_client.db = MagicMock()
+    mock_prisma_client.db.tx = MagicMock(return_value=mock_transaction)
+
+    db_spend_update_transactions = {
+        "user_list_transactions": {},
+        "end_user_list_transactions": {},
+        "key_list_transactions": {"hashed_token_abc": 0.05, "hashed_token_def": 1.25},
+        "team_list_transactions": {},
+        "team_member_list_transactions": {},
+        "org_list_transactions": {},
+        "tag_list_transactions": {},
+        "agent_list_transactions": {},
+    }
+
+    await db_writer._commit_spend_updates_to_db(
+        prisma_client=mock_prisma_client,
+        n_retry_times=0,
+        proxy_logging_obj=MagicMock(),
+        db_spend_update_transactions=db_spend_update_transactions,
+    )
+
+    calls = mock_batcher.litellm_verificationtoken.update_many.call_args_list
+    assert [c.kwargs["where"] for c in calls] == [{"token": "hashed_token_abc"}, {"token": "hashed_token_def"}]
+    for call, expected_cost in zip(calls, (0.05, 1.25)):
+        assert call.kwargs["data"]["spend"] == {"increment": expected_cost}
+        assert call.kwargs["data"]["total_spend"] == call.kwargs["data"]["spend"]
+
+
 @pytest.mark.asyncio
 async def test_update_database_creates_single_task():
     """
@@ -2813,7 +2864,7 @@ async def test_commit_spend_updates_to_db_does_not_stamp_key_settings_updated_at
     mock_batcher.litellm_verificationtoken.update_many.assert_called_once()
     call_kwargs = mock_batcher.litellm_verificationtoken.update_many.call_args[1]
     assert call_kwargs["where"] == {"token": token}
-    assert set(call_kwargs["data"]) == {"spend", "last_active"}
+    assert set(call_kwargs["data"]) == {"spend", "total_spend", "last_active"}
     assert call_kwargs["data"]["spend"] == {"increment": response_cost}
 
 

tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py

@@ -7,6 +7,7 @@ import logging
 import os
 import sys
 import time
+from collections.abc import Sequence
 from contextlib import contextmanager
 from datetime import datetime, timedelta
 from typing import Any, Dict, List, Optional
@@ -32,6 +33,7 @@ from litellm.proxy.hooks.parallel_request_limiter_v3 import (
 )
 from litellm.proxy.utils import InternalUsageCache, ProxyLogging, hash_token
 from litellm.types.caching import RedisPipelineIncrementOperation
+from litellm.types.llms.openai import ResponsesAPIResponse
 from litellm.types.utils import (
     EmbeddingResponse,
     ModelResponse,
@@ -4054,6 +4056,125 @@ async def _seed_max_parallel_requests_slots(
     )
 
 
+@pytest.mark.asyncio
+async def test_completed_responses_post_call_releases_parallel_slot() -> None:
+    api_key = hash_token("sk-responses-post-call")
+    local_cache = DualCache()
+    handler = _PROXY_MaxParallelRequestsHandler(
+        internal_usage_cache=InternalUsageCache(local_cache)
+    )
+    user_api_key_dict = UserAPIKeyAuth(api_key=api_key, max_parallel_requests=1)
+    data = {
+        "model": "gpt-4o-mini",
+        "input": "hello",
+        "litellm_call_id": "responses-owner",
+    }
+    parallel_key = f"{{api_key:{api_key}}}:max_parallel_requests"
+
+    await handler.async_pre_call_hook(
+        user_api_key_dict=user_api_key_dict,
+        cache=local_cache,
+        data=data,
+        call_type="aresponses",
+    )
+    assert handler._gauge_in_flight_from_cache_value(
+        await local_cache.async_get_cache(key=parallel_key)
+    ) == 1
+
+    await handler.async_post_call_success_hook(
+        data=data,
+        user_api_key_dict=user_api_key_dict,
+        response=ResponsesAPIResponse(
+            id="resp_parallel_slot",
+            created_at=0,
+            model="gpt-4o-mini",
+            object="response",
+            output=[],
+            status="completed",
+        ),
+    )
+    assert handler._gauge_in_flight_from_cache_value(
+        await local_cache.async_get_cache(key=parallel_key)
+    ) == 0
+
+    await handler.async_log_success_event(
+        kwargs={"litellm_call_id": data["litellm_call_id"]},
+        response_obj=None,
+        start_time=None,
+        end_time=None,
+    )
+    assert handler._gauge_in_flight_from_cache_value(
+        await local_cache.async_get_cache(key=parallel_key)
+    ) == 0
+
+
+@pytest.mark.asyncio
+async def test_concurrent_success_callbacks_release_parallel_slot_once_when_redis_fails() -> None:
+    from unittest.mock import AsyncMock
+
+    api_key = hash_token("sk-concurrent-release")
+    local_cache = DualCache()
+    handler = _PROXY_MaxParallelRequestsHandler(
+        internal_usage_cache=InternalUsageCache(local_cache)
+    )
+    user_api_key_dict = UserAPIKeyAuth(api_key=api_key, max_parallel_requests=2)
+    call_id = "concurrent-release-owner"
+    parallel_key = f"{{api_key:{api_key}}}:max_parallel_requests"
+    release_started = asyncio.Event()
+    allow_redis_failure = asyncio.Event()
+
+    async def failing_release(
+        keys: Sequence[str], args: Sequence[object]
+    ) -> list[int]:
+        release_started.set()
+        await allow_redis_failure.wait()
+        raise ConnectionError("redis unavailable")
+
+    release_script = AsyncMock(side_effect=failing_release)
+    handler.parallel_release_script = release_script
+    await local_cache.async_set_cache(key=parallel_key, value=2, local_only=True)
+    stash = get_or_create_request_stash()
+    stash.owner_litellm_call_id = call_id
+    stash.parallel_slot = ParallelSlotAcquisition(
+        slot_id="slot-concurrent-release",
+        counter_keys=[parallel_key],
+    )
+    data = {"litellm_call_id": call_id}
+
+    post_call_task = asyncio.create_task(
+        handler.async_post_call_success_hook(
+            data=data,
+            user_api_key_dict=user_api_key_dict,
+            response=ResponsesAPIResponse(
+                id="resp_concurrent_release",
+                created_at=0,
+                model="gpt-4o-mini",
+                object="response",
+                output=[],
+                status="completed",
+            ),
+        )
+    )
+    await asyncio.wait_for(release_started.wait(), timeout=5)
+    logging_task = asyncio.create_task(
+        handler.async_log_success_event(
+            kwargs=data,
+            response_obj=None,
+            start_time=None,
+            end_time=None,
+        )
+    )
+    allow_redis_failure.set()
+    await asyncio.wait_for(
+        asyncio.gather(post_call_task, logging_task),
+        timeout=5,
+    )
+
+    assert release_script.await_count == 1
+    assert await local_cache.async_get_cache(key=parallel_key) == 1
+    assert stash.parallel_slot is None
+
+
 async def _build_seeded_limiter():
     """Build a v3 limiter whose api-key slot registry already holds the pre-call slot."""
     api_key = hash_token("sk-disconnect")

tests/test_litellm/proxy/management_endpoints/test_common_utils.py

@@ -35,6 +35,7 @@ from litellm.proxy.management_endpoints.common_utils import (
     admin_can_invite_user,
 )
 from litellm.proxy.management_endpoints.common_utils import _has_non_empty_value
+from litellm.types.utils import BudgetConfig
 
 
 class TestUpdateMetadataFieldsEmptyCollections:
@@ -1162,3 +1163,54 @@ async def test_router_weights_validate_current_deployment_scope(
         assert exc.value.detail == error
     else:
         await validation
+
+
+@pytest.mark.parametrize(
+    "model_max_budget, error",
+    [
+        ({"gpt-4o": BudgetConfig(max_budget=-1.0, budget_duration="1d")}, "non-negative finite"),
+        ({"gpt-4o": BudgetConfig(max_budget=float("inf"), budget_duration="1d")}, "non-negative finite"),
+        ({"gpt-4o": BudgetConfig(max_budget=float("nan"), budget_duration="1d")}, "non-negative finite"),
+        ({"gpt-4o": BudgetConfig(budget_duration="1d")}, "non-negative finite"),
+        ({"gpt-4o": BudgetConfig(max_budget=5.0)}, "requires a budget_duration"),
+        ({"gpt-4o": BudgetConfig(max_budget=5.0, budget_duration="fortnight")}, "budget_duration"),
+        ({"  ": BudgetConfig(max_budget=5.0, budget_duration="1d")}, "non-empty model names"),
+        ({"gpt-4o": BudgetConfig(max_budget=5.0, budget_duration="1d", tpm_limit=1000)}, "not enforced on a team"),
+        ({"gpt-4o": BudgetConfig(max_budget=5.0, budget_duration="1d", rpm_limit=10)}, "not enforced on a team"),
+    ],
+    ids=["negative", "inf", "nan", "no_cap", "no_duration", "bad_duration", "blank_model", "tpm_limit", "rpm_limit"],
+)
+def test_validate_team_model_max_budget_rejects_unenforceable_entries(model_max_budget, error) -> None:
+    from litellm.proxy.management_endpoints.common_utils import validate_team_model_max_budget
+
+    with pytest.raises(HTTPException) as exc:
+        validate_team_model_max_budget(model_max_budget=model_max_budget, premium_user=True)
+    assert exc.value.status_code == 400
+    assert error in exc.value.detail["error"]
+
+
+def test_validate_team_model_max_budget_accepts_a_zero_cap_and_prefixed_models() -> None:
+    from litellm.proxy.management_endpoints.common_utils import validate_team_model_max_budget
+
+    assert (
+        validate_team_model_max_budget(
+            model_max_budget={
+                "gpt-4o": BudgetConfig(max_budget=0.0, budget_duration="1d"),
+                "openai/gpt-4o-mini": BudgetConfig(max_budget=2.5, budget_duration="30d"),
+            },
+            premium_user=True,
+        )
+        is None
+    )
+
+
+def test_validate_team_model_max_budget_is_license_gated_only_when_set() -> None:
+    from litellm.proxy.management_endpoints.common_utils import validate_team_model_max_budget
+
+    validate_team_model_max_budget(model_max_budget=None, premium_user=False)
+    validate_team_model_max_budget(model_max_budget={}, premium_user=False)
+    with pytest.raises(HTTPException) as exc:
+        validate_team_model_max_budget(
+            model_max_budget={"gpt-4o": BudgetConfig(max_budget=1.0, budget_duration="1d")}, premium_user=False
+        )
+    assert exc.value.status_code == 403

tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py

@@ -1431,6 +1431,60 @@ async def test_key_info_returns_object_permission(monkeypatch):
     )
 
 
+def _stored_key_with_lifetime_spend(token: str, spend: float, total_spend: float) -> LiteLLM_VerificationToken:
+    return LiteLLM_VerificationToken.model_validate(
+        {"token": token, "user_id": "user123", "spend": spend, "total_spend": total_spend}
+    )
+
+
+@pytest.mark.asyncio
+async def test_key_info_returns_lifetime_total_spend_next_to_resettable_spend(monkeypatch):
+    """After a budget reset the period spend is 0 while total_spend keeps the lifetime figure."""
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(
+        return_value=_stored_key_with_lifetime_spend(token="hashed_key", spend=0.0, total_spend=3.75)
+    )
+
+    result = await info_key_fn(
+        key="sk-test-key-456",
+        user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-test-key-456"),
+    )
+
+    assert result["info"]["spend"] == 0.0
+    assert result["info"]["total_spend"] == 3.75
+
+
+@pytest.mark.asyncio
+async def test_list_keys_full_object_returns_lifetime_total_spend():
+    mock_prisma_client = AsyncMock()
+    mock_prisma_client.db.litellm_verificationtoken.find_many = AsyncMock(
+        return_value=[_stored_key_with_lifetime_spend(token="hashed_key", spend=0.0, total_spend=3.75)]
+    )
+    mock_prisma_client.db.litellm_verificationtoken.count = AsyncMock(return_value=1)
+
+    result = await _list_key_helper(
+        prisma_client=mock_prisma_client,
+        page=1,
+        size=50,
+        user_id=None,
+        team_id=None,
+        organization_id=None,
+        key_alias=None,
+        key_hash=None,
+        exclude_team_id=None,
+        return_full_object=True,
+        admin_team_ids=None,
+    )
+
+    listed_key = result["keys"][0]
+    assert isinstance(listed_key, UserAPIKeyAuth)
+    assert listed_key.spend == 0.0
+    assert listed_key.total_spend == 3.75
+
+
 @pytest.mark.asyncio
 async def test_get_new_token_with_valid_key(monkeypatch):
     """Test get_new_token function when provided with a valid key that starts with 'sk-'"""
@@ -4923,6 +4977,23 @@ def test_transform_verification_tokens_to_deleted_records():
     assert json.loads(record2["budget_fallbacks"]) == {"gpt-4": ["gpt-4o-mini"]}
 
 
+def test_transform_verification_tokens_to_deleted_records_keeps_organization_id():
+    live_row = MagicMock()
+    live_row.model_dump.return_value = {
+        "token": "hashed-token-org",
+        "user_id": "user-123",
+        "team_id": None,
+        "organization_id": "org-finops",
+    }
+
+    records = _transform_verification_tokens_to_deleted_records(
+        keys=[live_row],
+        user_api_key_dict=UserAPIKeyAuth(user_id="admin-1", api_key="sk-admin"),
+    )
+
+    assert records[0]["organization_id"] == "org-finops"
+
+
 def test_transform_verification_tokens_to_deleted_records_empty_list():
     user_api_key_dict = UserAPIKeyAuth(
         user_id="user-123",
@@ -6022,6 +6093,244 @@ async def test_list_keys_with_invalid_status():
         assert "deleted" in str(exc_info.value.message)
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("status_filter", ["active", "expired", "revoked"])
+async def test_list_keys_accepts_live_status_filters(monkeypatch, status_filter):
+    from unittest.mock import Mock
+
+    from litellm.proxy.management_endpoints.key_management_endpoints import list_keys
+
+    live_row = MagicMock()
+    live_row.model_dump.return_value = {"token": "hashed_live_token", "object_permission_id": None}
+    mock_prisma_client = AsyncMock()
+    mock_prisma_client.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[live_row])
+    mock_prisma_client.db.litellm_verificationtoken.count = AsyncMock(return_value=1)
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_many = AsyncMock(return_value=[])
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+
+    response = await list_keys(
+        request=Mock(),
+        user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN),
+        page=1,
+        size=10,
+        user_id=None,
+        team_id=None,
+        organization_id=None,
+        key_hash=None,
+        key_alias=None,
+        search=None,
+        return_full_object=False,
+        include_team_keys=False,
+        include_created_by_keys=False,
+        sort_by=None,
+        sort_order="desc",
+        expand=None,
+        status=status_filter,
+        project_id=None,
+        access_group_id=None,
+        agent_id=None,
+        substring_matching=False,
+        expires=None,
+    )
+
+    assert response["keys"] == ["hashed_live_token"]
+    assert response["total_count"] == 1
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_many.assert_not_called()
+
+
+def _status_filter_where(status_filter: str | None) -> Mapping[str, object]:
+    from litellm.proxy.management_endpoints.key_management_endpoints import _build_key_filter_conditions
+
+    return _build_key_filter_conditions(
+        user_id=None,
+        team_id=None,
+        organization_id=None,
+        key_alias=None,
+        key_hash=None,
+        exclude_team_id=None,
+        admin_team_ids=None,
+        status_filter=status_filter,
+    )
+
+
+def test_build_key_filter_conditions_status_filter_partitions_live_keys():
+    not_blocked = {"OR": [{"blocked": None}, {"blocked": False}]}
+
+    revoked_where = _status_filter_where("revoked")
+    assert {"blocked": True} in revoked_where["AND"]
+
+    expired_clause = next(clause for clause in _status_filter_where("expired")["AND"] if "AND" in clause)
+    assert expired_clause["AND"][0] == not_blocked
+    assert expired_clause["AND"][1]["AND"][0] == {"expires": {"not": None}}
+    assert "lt" in expired_clause["AND"][1]["AND"][1]["expires"]
+
+    active_clause = next(clause for clause in _status_filter_where("active")["AND"] if "AND" in clause)
+    assert active_clause["AND"][0] == not_blocked
+    assert active_clause["AND"][1]["OR"][0] == {"expires": None}
+    assert "gte" in active_clause["AND"][1]["OR"][1]["expires"]
+
+
+def test_build_key_filter_conditions_deleted_status_adds_no_live_clause():
+    assert _status_filter_where("deleted") == _status_filter_where(None)
+
+
+@pytest.mark.asyncio
+async def test_list_key_helper_revoked_status_filters_live_table_on_blocked():
+    mock_prisma_client = AsyncMock()
+    mock_find_many = AsyncMock(return_value=[])
+    mock_prisma_client.db.litellm_verificationtoken.find_many = mock_find_many
+    mock_prisma_client.db.litellm_verificationtoken.count = AsyncMock(return_value=0)
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_many = AsyncMock(return_value=[])
+
+    await _list_key_helper(
+        prisma_client=mock_prisma_client,
+        page=1,
+        size=50,
+        user_id=None,
+        team_id=None,
+        organization_id=None,
+        key_alias=None,
+        key_hash=None,
+        exclude_team_id=None,
+        return_full_object=True,
+        admin_team_ids=None,
+        include_created_by_keys=False,
+        status="revoked",
+    )
+
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_many.assert_not_called()
+    where = mock_find_many.call_args.kwargs["where"]
+    assert {"blocked": True} in where["AND"]
+
+
+def _archived_key_row(token: str, user_id: str) -> MagicMock:
+    row = MagicMock()
+    row.model_dump.return_value = {
+        "id": "archive-row-1",
+        "token": token,
+        "key_alias": "finops-2024",
+        "user_id": user_id,
+        "team_id": None,
+        "organization_id": "org-finops",
+        "blocked": None,
+        "deleted_at": datetime(2024, 11, 15, 10, 0, tzinfo=timezone.utc),
+        "deleted_by": "admin-1",
+    }
+    return row
+
+
+@pytest.mark.asyncio
+async def test_info_key_fn_serves_deleted_key_from_archive(monkeypatch):
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    hashed = "hashed_deleted_token"
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(return_value=None)
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_first = AsyncMock(
+        return_value=_archived_key_row(hashed, "user-x")
+    )
+
+    result = await info_key_fn(
+        key=hashed,
+        user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-admin"),
+    )
+
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_first.assert_awaited_once()
+    assert mock_prisma_client.db.litellm_deletedverificationtoken.find_first.await_args.kwargs["where"] == {
+        "token": hashed
+    }
+    info = result["info"]
+    assert info["status"] == "deleted"
+    assert info["key_alias"] == "finops-2024"
+    assert info["organization_id"] == "org-finops"
+    assert info["deleted_by"] == "admin-1"
+    assert info["deleted_at"] is not None
+    assert "token" not in info
+
+
+@pytest.mark.asyncio
+async def test_info_key_fn_archived_key_keeps_owner_authorization(monkeypatch):
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    hashed = "hashed_deleted_token"
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(return_value=None)
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_first = AsyncMock(
+        return_value=_archived_key_row(hashed, "owner-1")
+    )
+
+    with pytest.raises(ProxyException) as exc_info:
+        await info_key_fn(
+            key=hashed,
+            user_api_key_dict=UserAPIKeyAuth(
+                user_role=LitellmUserRoles.INTERNAL_USER, user_id="someone-else", api_key="sk-other"
+            ),
+        )
+    assert exc_info.value.code == "403"
+
+    owner_result = await info_key_fn(
+        key=hashed,
+        user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.INTERNAL_USER, user_id="owner-1", api_key="sk-own"),
+    )
+    assert owner_result["info"]["status"] == "deleted"
+
+
+@pytest.mark.asyncio
+async def test_info_key_fn_unknown_key_still_404s(monkeypatch):
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(return_value=None)
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_first = AsyncMock(return_value=None)
+
+    with pytest.raises(ProxyException) as exc_info:
+        await info_key_fn(
+            key="hashed_missing",
+            user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-admin"),
+        )
+    assert exc_info.value.code == "404"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("blocked", "expires", "expected_status"),
+    [
+        (True, None, "revoked"),
+        (True, "2020-01-01T00:00:00Z", "revoked"),
+        (False, "2020-01-01T00:00:00Z", "expired"),
+        (None, datetime(2020, 1, 1, tzinfo=timezone.utc), "expired"),
+        (False, None, "active"),
+        (None, "2999-01-01T00:00:00Z", "active"),
+    ],
+)
+async def test_info_key_fn_reports_live_key_status(monkeypatch, blocked, expires, expected_status):
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    live_row = MagicMock(spec=LiteLLM_VerificationToken)
+    live_row.model_dump.return_value = {
+        "token": "hashed_live",
+        "user_id": "user-x",
+        "team_id": None,
+        "object_permission_id": None,
+        "blocked": blocked,
+        "expires": expires,
+    }
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(return_value=live_row)
+
+    result = await info_key_fn(
+        key="hashed_live",
+        user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-admin"),
+    )
+
+    assert result["info"]["status"] == expected_status
+    mock_prisma_client.db.litellm_deletedverificationtoken.find_first.assert_not_called()
+
+
 @pytest.mark.asyncio
 async def test_list_keys_non_admin_user_id_auto_set():
     """

tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py

@@ -14651,3 +14651,246 @@ async def test_team_info_reports_parent_organization_models_only_to_team_manager
         )
 
     assert response["team_info"].organization_models == expected_models
+
+
+_EXISTING_TEAM_MODEL_CAPS: Final = {
+    "gpt-4o": {"max_budget": 10.0, "budget_duration": "1d"},
+    "claude-sonnet-4-6": {"max_budget": 5.0, "budget_duration": "7d"},
+}
+
+
+@pytest.mark.parametrize(
+    "requested",
+    [
+        {**_EXISTING_TEAM_MODEL_CAPS, "gpt-4o": {"max_budget": 20.0, "budget_duration": "1d"}},
+        {**_EXISTING_TEAM_MODEL_CAPS, "gpt-4o": {"max_budget": 10.0, "budget_duration": "30d"}},
+        {**_EXISTING_TEAM_MODEL_CAPS, "gpt-4o": {"budget_duration": "1d"}},
+        {"claude-sonnet-4-6": _EXISTING_TEAM_MODEL_CAPS["claude-sonnet-4-6"]},
+        {},
+        None,
+        {**_EXISTING_TEAM_MODEL_CAPS, "openai/gpt-4o": {"max_budget": 1000.0, "budget_duration": "1d"}},
+        {**_EXISTING_TEAM_MODEL_CAPS, "openai/gpt-4o": {"max_budget": 10.0, "budget_duration": "30d"}},
+        {**_EXISTING_TEAM_MODEL_CAPS, "anthropic/claude-sonnet-4-6": {"budget_duration": "7d"}},
+    ],
+    ids=[
+        "raise",
+        "change_duration",
+        "drop_cap_value",
+        "remove_model",
+        "clear_all",
+        "clear_with_null",
+        "raise_via_provider_alias",
+        "rewindow_via_provider_alias",
+        "uncap_via_provider_alias",
+    ],
+)
+def test_team_admin_cannot_loosen_team_model_caps(requested) -> None:
+    from litellm.proxy.management_endpoints.team_endpoints import _check_team_model_budget_update_authority
+
+    with pytest.raises(HTTPException) as exc:
+        _check_team_model_budget_update_authority(
+            data=UpdateTeamRequest(team_id="t1", model_max_budget=requested),
+            user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.INTERNAL_USER, user_id="team-admin"),
+            existing_model_max_budget=_EXISTING_TEAM_MODEL_CAPS,
+        )
+    assert exc.value.status_code == 403
+    assert "proxy admin" in exc.value.detail["error"]
+
+
+@pytest.mark.parametrize(
+    "requested",
+    [
+        {**_EXISTING_TEAM_MODEL_CAPS, "gpt-4o": {"max_budget": 2.0, "budget_duration": "1d"}},
+        {**_EXISTING_TEAM_MODEL_CAPS, "gpt-4o-mini": {"max_budget": 1.0, "budget_duration": "1d"}},
+        dict(_EXISTING_TEAM_MODEL_CAPS),
+        {**_EXISTING_TEAM_MODEL_CAPS, "openai/gpt-4o": {"max_budget": 2.0, "budget_duration": "1d"}},
+    ],
+    ids=["lower", "add_model", "unchanged", "tighten_via_provider_alias"],
+)
+def test_team_admin_can_tighten_or_keep_team_model_caps(requested) -> None:
+    from litellm.proxy.management_endpoints.team_endpoints import _check_team_model_budget_update_authority
+
+    assert (
+        _check_team_model_budget_update_authority(
+            data=UpdateTeamRequest(team_id="t1", model_max_budget=requested),
+            user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.INTERNAL_USER, user_id="team-admin"),
+            existing_model_max_budget=_EXISTING_TEAM_MODEL_CAPS,
+        )
+        is None
+    )
+
+
+def test_team_model_cap_authority_skips_omitted_field_malformed_rows_and_proxy_admins() -> None:
+    from litellm.proxy.management_endpoints.team_endpoints import _check_team_model_budget_update_authority
+
+    team_admin = UserAPIKeyAuth(user_role=LitellmUserRoles.INTERNAL_USER, user_id="team-admin")
+    outcomes = (
+        _check_team_model_budget_update_authority(
+            data=UpdateTeamRequest(team_id="t1", max_budget=1.0),
+            user_api_key_dict=team_admin,
+            existing_model_max_budget=_EXISTING_TEAM_MODEL_CAPS,
+        ),
+        _check_team_model_budget_update_authority(
+            data=UpdateTeamRequest(team_id="t1", model_max_budget={}),
+            user_api_key_dict=team_admin,
+            existing_model_max_budget={"gpt-4o": "not-a-budget", "gpt-4o-mini": {"budget_duration": "1d"}},
+        ),
+        _check_team_model_budget_update_authority(
+            data=UpdateTeamRequest(team_id="t1", model_max_budget=None),
+            user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN),
+            existing_model_max_budget=_EXISTING_TEAM_MODEL_CAPS,
+        ),
+    )
+    assert outcomes == (None, None, None)
+
+
+@pytest.mark.asyncio
+async def test_new_team_persists_model_max_budget(mock_db_client, mock_admin_auth):
+    mock_db_client.jsonify_team_object = lambda db_data: db_data
+    mock_db_client.get_data = AsyncMock(return_value=None)
+    mock_db_client.update_data = AsyncMock(return_value=MagicMock())
+    mock_db_client.db = MagicMock()
+    mock_db_client.db.litellm_modeltable = MagicMock()
+    mock_db_client.db.litellm_modeltable.create = AsyncMock(return_value=MagicMock(id="model123"))
+
+    team_create_result = MagicMock(team_id="team-model-caps")
+    team_create_result.model_dump.return_value = {"team_id": "team-model-caps"}
+    mock_team_create = AsyncMock(return_value=team_create_result)
+    mock_db_client.db.litellm_teamtable = MagicMock()
+    mock_db_client.db.litellm_teamtable.create = mock_team_create
+    _wire_team_create_tx(mock_db_client)
+    mock_db_client.db.litellm_teamtable.count = AsyncMock(return_value=0)
+    mock_db_client.db.litellm_teamtable.update = AsyncMock(return_value=team_create_result)
+    mock_db_client.db.litellm_usertable = MagicMock()
+    mock_db_client.db.litellm_usertable.update = AsyncMock(return_value=MagicMock())
+
+    from fastapi import Request
+
+    from litellm.proxy._types import NewTeamRequest
+    from litellm.proxy.management_endpoints.team_endpoints import new_team
+
+    with patch("litellm.proxy.proxy_server.premium_user", True):  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        await new_team(
+            data=NewTeamRequest(
+                team_alias="model-caps",
+                model_max_budget={"gpt-4o": {"max_budget": 10.0, "budget_duration": "1d"}},
+            ),
+            http_request=MagicMock(spec=Request),
+            user_api_key_dict=mock_admin_auth,
+        )
+
+    team_data = mock_team_create.call_args.kwargs["data"]
+    assert team_data["model_max_budget"] == {
+        "gpt-4o": {"max_budget": 10.0, "budget_duration": "1d", "tpm_limit": None, "rpm_limit": None}
+    }
+
+
+@pytest.mark.asyncio
+async def test_new_team_rejects_unenforceable_model_max_budget(mock_db_client, mock_admin_auth):
+    from fastapi import Request
+
+    from litellm.proxy._types import NewTeamRequest, ProxyException
+    from litellm.proxy.management_endpoints.team_endpoints import new_team
+
+    mock_db_client.db.litellm_teamtable.create = AsyncMock()
+
+    with patch("litellm.proxy.proxy_server.premium_user", True), pytest.raises(ProxyException) as exc:  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        await new_team(
+            data=NewTeamRequest(team_alias="model-caps", model_max_budget={"gpt-4o": {"max_budget": 10.0}}),
+            http_request=MagicMock(spec=Request),
+            user_api_key_dict=mock_admin_auth,
+        )
+
+    assert exc.value.code == "400"
+    assert "budget_duration" in str(exc.value.message)
+    mock_db_client.db.litellm_teamtable.create.assert_not_awaited()
+
+
+def _existing_team_with_model_caps(caps):
+    existing = MagicMock()
+    existing.team_id = "standalone-team-123"
+    existing.organization_id = None
+    existing.max_budget = None
+    existing.model_id = None
+    existing.model_max_budget = caps
+    existing.model_dump.return_value = {
+        "team_id": "standalone-team-123",
+        "organization_id": None,
+        "model_max_budget": caps,
+        "members_with_roles": [{"user_id": "team-admin-model-caps", "role": "admin"}],
+    }
+    return existing
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("cleared_with", [{}, None], ids=["empty_mapping", "null"])
+async def test_update_team_clearing_model_max_budget_writes_an_empty_mapping(
+    disable_audit_logging_for_mocked_team, cleared_with
+):
+    from fastapi import Request
+
+    from litellm.proxy.management_endpoints.team_endpoints import update_team
+
+    with (
+        patch("litellm.proxy.proxy_server.prisma_client") as mock_prisma,  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.user_api_key_cache") as mock_cache,  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.litellm_proxy_admin_name", "admin"),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.premium_user", True),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+    ):
+        mock_prisma.db.litellm_teamtable.find_unique = AsyncMock(
+            return_value=_existing_team_with_model_caps(_EXISTING_TEAM_MODEL_CAPS)
+        )
+        mock_prisma.jsonify_team_object = lambda db_data: db_data
+        mock_cache.async_get_cache = AsyncMock(return_value=None)
+        mock_cache.async_set_cache = AsyncMock()
+        updated = _existing_team_with_model_caps({})
+        updated.litellm_model_table = None
+        mock_prisma.db.litellm_teamtable.update = AsyncMock(return_value=updated)
+
+        await update_team(
+            data=UpdateTeamRequest(team_id="standalone-team-123", model_max_budget=cleared_with),
+            http_request=MagicMock(spec=Request),
+            user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, user_id="admin"),
+        )
+
+    assert mock_prisma.db.litellm_teamtable.update.call_args.kwargs["data"]["model_max_budget"] == {}
+
+
+@pytest.mark.asyncio
+async def test_update_team_model_max_budget_raise_blocked_for_team_admin():
+    from fastapi import Request
+
+    from litellm.proxy._types import ProxyException
+    from litellm.proxy.management_endpoints.team_endpoints import update_team
+
+    with (
+        patch("litellm.proxy.proxy_server.prisma_client") as mock_prisma,  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.user_api_key_cache") as mock_cache,  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.litellm_proxy_admin_name", "admin"),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.premium_user", True),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+        patch("litellm.proxy.proxy_server.create_audit_log_for_update", new=AsyncMock()),  # test-quality-ok: stubs the audit write so the test observes only the team update result
+    ):
+        mock_prisma.db.litellm_teamtable.find_unique = AsyncMock(
+            return_value=_existing_team_with_model_caps(_EXISTING_TEAM_MODEL_CAPS)
+        )
+        mock_cache.async_get_cache = AsyncMock(return_value=None)
+        mock_prisma.db.litellm_teamtable.update = AsyncMock()
+
+        with pytest.raises(ProxyException) as exc:
+            await update_team(
+                data=UpdateTeamRequest(
+                    team_id="standalone-team-123",
+                    model_max_budget={
+                        **_EXISTING_TEAM_MODEL_CAPS,
+                        "gpt-4o": {"max_budget": 100.0, "budget_duration": "1d"},
+                    },
+                ),
+                http_request=MagicMock(spec=Request),
+                user_api_key_dict=UserAPIKeyAuth(
+                    user_role=LitellmUserRoles.INTERNAL_USER, user_id="team-admin-model-caps", models=[]
+                ),
+            )
+
+    assert exc.value.code == "403"
+    assert "proxy admin" in str(exc.value.message).lower()
+    mock_prisma.db.litellm_teamtable.update.assert_not_awaited()

tests/test_litellm/proxy/spend_tracking/test_carried_budget_state.py

@@ -47,6 +47,19 @@ def test_team_and_user_state_round_trips_through_metadata():
     )
 
 
+def test_team_model_max_budget_rides_on_the_token():
+    """The team's per-model caps must reach the token, or the auth check and the spend hook never see them."""
+    token = UserAPIKeyAuth(token="hashed", team_id="t1")
+    team_model_max_budget = {"gpt-4o": {"max_budget": 5.0, "budget_duration": "1d"}}
+    carry_team_and_user_budget_state(
+        valid_token=token,
+        team_object=LiteLLM_TeamTable(team_id="t1", model_max_budget=team_model_max_budget),
+        user_object=None,
+    )
+
+    assert token.team_model_max_budget == team_model_max_budget
+
+
 def test_missing_objects_leave_no_metadata_and_no_snapshot():
     token = UserAPIKeyAuth(token="hashed", team_id="t1", user_id="u1")
     carry_team_and_user_budget_state(valid_token=token, team_object=None, user_object=None)

tests/test_litellm/proxy/utils/prisma_and_spend/test_prisma_client_get_data.py

@@ -401,19 +401,20 @@ async def test_check_view_exists_creates_token_view_when_missing(
     prisma_client.db.execute_raw = AsyncMock()
     prisma_client.health_check = AsyncMock(return_value=[{"?column?": 1}])
     result = await prisma_client.check_view_exists()
+    created_sql = prisma_client.db.execute_raw.await_args.args[0]
     actual = {
         "result": result,
         "create_called": prisma_client.db.execute_raw.await_count,
-        "create_sql_starts_with_create_view": prisma_client.db.execute_raw.await_args.args[
-            0
-        ]
-        .strip()
-        .startswith('CREATE VIEW "LiteLLM_VerificationTokenView"'),
+        "create_sql_starts_with_create_view": created_sql.strip().startswith(
+            'CREATE VIEW "LiteLLM_VerificationTokenView"'
+        ),
+        "projects_team_model_max_budget": "t.model_max_budget AS team_model_max_budget" in created_sql,
     }
     assert actual == {
         "result": None,
         "create_called": 1,
         "create_sql_starts_with_create_view": True,
+        "projects_team_model_max_budget": True,
     }
 
 

tests/test_litellm_rust/ocr/test_requests.py

@@ -518,32 +518,34 @@ async def test_native_ocr_inherits_named_credentials_without_overwriting_argumen
     assert ocr_server.requests[0].body["pages"] == [0, 2]
 
 
-@pytest.mark.parametrize("source", ["sdk", "proxy"])
 @pytest.mark.parametrize(
-    "filename,mime", [("scan.PNG", "image/png"), ("document.pdf", "application/pdf"), ("note.txt", "text/plain")]
+    "filename,field,mime",
+    [("scan.PNG", "image_url", "image/png"), ("document.pdf", "document_url", "application/pdf")],
 )
-def test_ocr_file_helpers_use_native_document_preparation(source: str, filename: str, mime: str) -> None:
+def test_native_ocr_infers_mime_type_from_reader_name(
+    ocr_server: RecordingServer, filename: str, field: str, mime: str
+) -> None:
     from io import BytesIO
 
-    from litellm.ocr.input import convert_file_document_to_url_document, get_mime_type
-    from litellm.proxy.ocr_endpoints.endpoints import _build_document_from_upload
-
     file: Final = BytesIO(b"abc")
     file.name = filename
-    document: Final = (
-        convert_file_document_to_url_document({"type": "file", "file": file})
-        if source == "sdk"
-        else _build_document_from_upload(b"abc", filename, "application/octet-stream; charset=utf-8")
-    )
-    field: Final = "image_url" if mime.startswith("image/") else "document_url"
-    assert get_mime_type(filename) == mime
-    assert document == {"type": field, field: f"data:{mime};base64,YWJj"}
+    call_native_ocr(ocr_server, document={"type": "file", "file": file})
+    assert ocr_server.requests[0].body["document"] == {"type": field, field: f"data:{mime};base64,YWJj"}
+
+
+def test_native_ocr_encodes_str_reader_results_as_utf8(ocr_server: RecordingServer) -> None:
+    from io import StringIO
+
+    call_native_ocr(ocr_server, document={"type": "file", "file": StringIO("abc"), "mime_type": "text/plain"})
+    assert ocr_server.requests[0].body["document"] == {
+        "type": "document_url",
+        "document_url": "data:text/plain;base64,YWJj",
+    }
 
 
 @pytest.mark.parametrize("attribute", ["read", "name"])
-def test_native_file_preparation_preserves_property_errors(attribute: str) -> None:
-    from litellm.ocr.input import convert_file_document_to_url_document
-
+def test_native_file_preparation_preserves_property_errors(ocr_server: RecordingServer, attribute: str) -> None:
+    ocr_server.expected_requests = 0
     failure: Final = LookupError("file property failed")
 
     class File:
@@ -555,16 +557,47 @@ def test_native_file_preparation_preserves_property_errors(attribute: str) -> No
         def read(self):
             return b"abc"
 
-    with pytest.raises(LookupError) as caught:
-        convert_file_document_to_url_document({"type": "file", "file": File()})
-    assert caught.value is failure
+    with pytest.raises(litellm.APIConnectionError, match="file property failed") as caught:
+        call_native_ocr(ocr_server, document={"type": "file", "file": File()})
+    assert caught.value.__context__ is failure
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("asynchronous", [False, True])
+async def test_native_file_preparation_preserves_reader_exception(
+    ocr_server: RecordingServer, asynchronous: bool
+) -> None:
+    ocr_server.expected_requests = 0
+    failure: Final = RuntimeError("reader failed")
+
+    class Reader:
+        def read(self) -> bytes:
+            raise failure
+
+    document: Final = {"type": "file", "file": Reader()}
+    with pytest.raises(litellm.APIConnectionError, match="reader failed") as caught:
+        await call_native_aocr(ocr_server, document=document) if asynchronous else call_native_ocr(
+            ocr_server, document=document
+        )
+    assert caught.value.__context__ is failure
+
+
+def test_native_file_preparation_rejects_unsupported_reader_results(ocr_server: RecordingServer) -> None:
+    ocr_server.expected_requests = 0
+
+    class Reader:
+        def read(self) -> int:
+            return 1
+
+    with pytest.raises(litellm.APIConnectionError, match="bytes or str") as caught:
+        call_native_ocr(ocr_server, document={"type": "file", "file": Reader()})
+    assert isinstance(caught.value.__context__, TypeError)
 
 
 @pytest.mark.parametrize("kind", ["bytes", "path", "reader"])
-def test_native_file_preparation_rejects_oversized_input(kind: str, tmp_path: Path) -> None:
-    from litellm.ocr.input import FileDocument, convert_file_document_to_url_document, get_max_file_bytes
-
-    limit: Final = get_max_file_bytes()
+def test_native_file_preparation_rejects_oversized_input(ocr_server: RecordingServer, kind: str, tmp_path: Path) -> None:
+    ocr_server.expected_requests = 0
+    limit: Final = 50 * 1024 * 1024
     path: Final = tmp_path / "large.pdf"
     with path.open("wb") as stream:
         stream.truncate(limit + 1)
@@ -573,53 +606,25 @@ def test_native_file_preparation_rejects_oversized_input(kind: str, tmp_path: Pa
         def read(self) -> bytes:
             return b"a" * (limit + 1)
 
-    document: Final[FileDocument] = {
+    document: Final = {
         "type": "file",
         "file": path if kind == "path" else Reader() if kind == "reader" else b"a" * (limit + 1),
     }
-    with pytest.raises(ValueError, match="exceeds the size limit"):
-        convert_file_document_to_url_document(document)
+    with pytest.raises(litellm.BadRequestError, match="exceeds the size limit"):
+        call_native_ocr(ocr_server, document=document)
 
 
-@pytest.mark.parametrize("kind", ["str", "path", "reader"])
-def test_native_upload_binding_rejects_filesystem_inputs(kind: str, tmp_path: Path) -> None:
+def test_native_file_preparation_reports_missing_paths(ocr_server: RecordingServer, tmp_path: Path) -> None:
+    ocr_server.expected_requests = 0
+    missing: Final = tmp_path / "missing.pdf"
+    with pytest.raises(litellm.APIConnectionError, match=f"File not found: {missing}") as caught:
+        call_native_ocr(ocr_server, document={"type": "file", "file": missing})
+    assert isinstance(caught.value.__context__, FileNotFoundError)
+
+
+def test_native_file_preparation_rejects_empty_readers(ocr_server: RecordingServer) -> None:
     from io import BytesIO
-    from typing import cast  # noqa: TID251  # deliberately invalid inputs exercise the native runtime boundary
 
-    from litellm.ocr.input import convert_upload_to_url_document
-
-    path: Final = tmp_path / "secret.pdf"
-    path.write_bytes(b"server secret")
-    source: Final = str(path) if kind == "str" else path if kind == "path" else BytesIO(b"abc")
-    with pytest.raises(TypeError):
-        convert_upload_to_url_document(cast(bytes, source), "document.pdf", None)
-
-
-@pytest.mark.parametrize("extra_bytes", [0, 1])
-def test_native_upload_enforces_file_size_limit(extra_bytes: int) -> None:
-    import base64
-
-    from litellm.ocr.input import convert_upload_to_url_document, get_max_file_bytes
-
-    content: Final = b"a" * (get_max_file_bytes() + extra_bytes)
-    if extra_bytes:
-        with pytest.raises(ValueError, match="exceeds the size limit"):
-            convert_upload_to_url_document(content, "scan.pdf", None)
-        return
-    document: Final = convert_upload_to_url_document(content, "scan.pdf", None)
-    assert document["type"] == "document_url"
-    assert base64.b64decode(document["document_url"].split(",", 1)[1]) == content
-
-
-def test_native_file_preparation_preserves_reader_exception() -> None:
-    from litellm.ocr.input import convert_file_document_to_url_document
-
-    failure: Final = RuntimeError("reader failed")
-
-    class Reader:
-        def read(self) -> bytes:
-            raise failure
-
-    with pytest.raises(RuntimeError) as caught:
-        convert_file_document_to_url_document({"type": "file", "file": Reader()})
-    assert caught.value is failure
+    ocr_server.expected_requests = 0
+    with pytest.raises(litellm.BadRequestError, match="File is empty"):
+        call_native_ocr(ocr_server, document={"type": "file", "file": BytesIO(b"")})

ui/litellm-dashboard/src/app/(dashboard)/skills/_components/add_plugin_form.tsx

@@ -232,7 +232,7 @@ const AddPluginForm: React.FC<AddPluginFormProps> = ({ visible, onClose, accessT
                 name="skillUrl"
                 label={labelWithHint(
                   "Source URL",
-                  "Paste an HTTPS git repository URL from GitHub, GitLab, Bitbucket, or a self-hosted host (e.g. github.com/org/repo or github.com/org/repo/tree/main/my-skill), or an HTTPS link to a .zip archive of the skill hosted on S3 or any static file server.",
+                  "Paste an HTTPS git repository URL from GitHub, GitLab, Bitbucket, or a self-hosted host (e.g. github.com/org/repo or github.com/org/repo/tree/main/my-skill), or an HTTPS link to a .zip archive of the skill hosted on S3 or any static file server. For a private repository use its SSH clone URL (git@ghe.example.com:org/repo.git) so Claude Code clones it with your own SSH key.",
                 )}
               >
                 {({ ref, onChange, ...field }) => (

ui/litellm-dashboard/src/components/DeletedKeysPage/DeletedKeysPage.test.tsx

@@ -22,6 +22,7 @@ const mockDeletedKey: DeletedKeyResponse = {
   key_name: "test-key",
   key_alias: "Test Key Alias",
   spend: 5.5,
+  total_spend: 5.5,
   max_budget: 100,
   expires: "2024-12-31T23:59:59Z",
   models: ["gpt-3.5-turbo"],

ui/litellm-dashboard/src/components/Teams.test.tsx

@@ -7,6 +7,7 @@ import { beforeEach, describe, expect, it, vi } from "vitest";
 import { useTeamMetadataSchema } from "@/app/(dashboard)/hooks/teams/useTeamMetadataSchema";
 import { toast } from "@/lib/toast";
 import { fetchAvailableModelsForTeamOrKey } from "./key_team_helpers/fetch_available_models_team_key";
+import { MODEL_MAX_BUDGET_PREMIUM_HINT } from "./key_team_helpers/ModelMaxBudgetEditor";
 import {
   fetchMCPAccessGroups,
   getDefaultTeamSettings,
@@ -1547,6 +1548,36 @@ describe("Teams - the exact bytes the create call sends", () => {
     expect(await screen.findByText("Please input a team name")).toBeInTheDocument();
     expect(teamCreateCall).not.toHaveBeenCalled();
   });
+
+  it("locks the per-model budget editor and says why when the proxy has no enterprise license", async () => {
+    await openCreateModal({ premiumUser: false });
+
+    expect(screen.getByRole("button", { name: /Add Model Budget/i })).toBeDisabled();
+    expect(screen.getByText(MODEL_MAX_BUDGET_PREMIUM_HINT)).toBeInTheDocument();
+  });
+
+  it("sends the per-model budget a licensed operator fills in, keyed by model", async () => {
+    const user = userEvent.setup({ delay: null });
+    await openCreateModal({ premiumUser: true });
+
+    await user.click(screen.getByRole("button", { name: /Add Model Budget/i }));
+    await chooseSelectOption(user, screen.getByPlaceholderText("Select model"), "gpt-4");
+    fireEvent.change(screen.getByPlaceholderText("Max spend ($)"), { target: { value: "3" } });
+
+    const payload = await submit();
+
+    expect(payload.model_max_budget).toStrictEqual({ "gpt-4": { budget_limit: 3, time_period: "30d" } });
+  });
+
+  it("leaves model_max_budget out when a started row is removed again", async () => {
+    const user = userEvent.setup({ delay: null });
+    await openCreateModal({ premiumUser: true });
+
+    await user.click(screen.getByRole("button", { name: /Add Model Budget/i }));
+    await user.click(screen.getByRole("button", { name: "Remove model budget" }));
+
+    expect(wireBody(await submit())).not.toHaveProperty("model_max_budget");
+  });
 });
 
 describe("Teams - the create form keeps the organization and models picks while it is open", () => {

ui/litellm-dashboard/src/components/Teams.tsx

@@ -48,6 +48,7 @@ import BudgetDurationDropdown, {
 } from "./common_components/budget_duration_dropdown";
 import { Organization, getDefaultTeamSettings, getGuardrailsList, getPoliciesList, teamDeleteCall } from "./networking";
 import NumericalInput from "./shared/numerical_input";
+import { ModelMaxBudget, ModelMaxBudgetField } from "./key_team_helpers/ModelMaxBudgetEditor";
 import VectorStoreSelector from "./vector_store_management/VectorStoreSelector";
 import SearchToolSelector from "./search_tools/SearchToolSelector";
 import SkillSelector from "./skills/SkillSelector";
@@ -271,6 +272,7 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
   const [policiesList, setPoliciesList] = useState<string[]>([]);
   const [loggingSettings, setLoggingSettings] = useState<any[]>([]);
   const [modelAliases, setModelAliases] = useState<{ [key: string]: string }>({});
+  const [modelMaxBudget, setModelMaxBudget] = useState<ModelMaxBudget>({});
   const [routerSettings, setRouterSettings] = useState<RouterSettingsAccordionValue | null>(null);
   const [routerSettingsKey, setRouterSettingsKey] = useState<number>(0);
 
@@ -348,6 +350,7 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
     setSearchToolSettingsOpen(false);
     setLoggingSettings([]);
     setModelAliases({});
+    setModelMaxBudget({});
     setRouterSettings(null);
     setRouterSettingsKey((prev) => prev + 1);
   };
@@ -525,6 +528,10 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
           formValues.model_aliases = modelAliases;
         }
 
+        if (Object.keys(modelMaxBudget).length > 0) {
+          formValues.model_max_budget = modelMaxBudget;
+        }
+
         // Add router_settings if any are defined
         if (routerSettings?.router_settings) {
           // Only include router_settings if it has at least one non-null value
@@ -813,6 +820,14 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
                       />
                     )}
                   </FormField>
+                  <ModelMaxBudgetField
+                    key={`model-max-budget-${routerSettingsKey}`}
+                    premiumUser={premiumUser}
+                    value={modelMaxBudget}
+                    onChange={setModelMaxBudget}
+                    availableModels={userModels}
+                    hint="Cap this team's spend on individual models, each with its own reset window. Every key on the team shares the cap unless the key sets its own budget for that model."
+                  />
                   <FormField control={form.control} name="tpm_limit" label="Tokens per minute Limit (TPM)">
                     {({ ref, value, ...field }) => (
                       <NumericalInput {...field} ref={ref} value={value ?? ""} step={1} width={400} />

ui/litellm-dashboard/src/components/VirtualKeysPage/VirtualKeysTable.test.tsx

@@ -79,6 +79,7 @@ const mockKey: KeyResponse = {
   key_name: "test-key",
   key_alias: "Test Key Alias",
   spend: 5.5,
+  total_spend: 42.25,
   max_budget: 100,
   expires: "2999-12-31T23:59:59Z",
   models: ["gpt-3.5-turbo", "gpt-4"],
@@ -236,6 +237,14 @@ it("should display key information correctly", async () => {
   });
 });
 
+it("shows lifetime spend in its own column next to the period spend meter", async () => {
+  renderWithProviders(<VirtualKeysTable />);
+
+  expect(await screen.findByText("Lifetime Spend")).toBeInTheDocument();
+  expect(screen.getByText("$42.2500")).toBeInTheDocument();
+  expect(screen.getByText("$5.5000")).toBeInTheDocument();
+});
+
 it("should display user email correctly", async () => {
   renderWithProviders(<VirtualKeysTable />);
 
@@ -638,6 +647,23 @@ describe("server-side filtering – the LIT-4080 regression guard", () => {
     });
   });
 
+  it("threads the Status drawer filter into the useKeys query and the URL", async () => {
+    const onUrlUpdate = vi.fn<OnUrlUpdateFunction>();
+    renderWithProviders(<VirtualKeysTable />, { onUrlUpdate });
+
+    openFilters();
+    const user = userEvent.setup();
+    await chooseSelectOption(user, await screen.findByRole("combobox", { name: "Status" }), "Revoked (blocked)");
+    fireEvent.click(screen.getByTestId("filter-drawer-apply"));
+
+    await waitFor(() => {
+      expect(mockUseKeys).toHaveBeenLastCalledWith(1, 50, expect.objectContaining({ status: "revoked" }));
+    });
+    await waitFor(() => {
+      expect(lastSearchParam(onUrlUpdate, "filter_status")).toBe("revoked");
+    });
+  });
+
   it("sends the search box as the combined alias-or-ID search rather than the key-alias filter", async () => {
     renderWithProviders(<VirtualKeysTable />);
 
@@ -745,6 +771,25 @@ describe("Status column reflects blocked / expiry / scim metadata", () => {
     expect(screen.queryByText(/Blocked by SCIM/i)).not.toBeInTheDocument();
   });
 
+  it("renders Deleted for an archived key, even when the archived row was also blocked", async () => {
+    mockUseKeys.mockReturnValue(
+      keysResult([
+        { ...mockKey, blocked: true, metadata: {}, deleted_at: "2024-11-15T10:00:00Z", deleted_by: "admin-1" },
+      ]),
+    );
+
+    renderWithProviders(<VirtualKeysTable />);
+
+    const tag = await screen.findByTestId(`key-status-${mockKey.token_id}`);
+    expect(tag).toHaveTextContent("Deleted");
+
+    const user = userEvent.setup();
+    await user.hover(tag);
+    await waitFor(() => {
+      expect(screen.getByText(/by admin-1/)).toBeInTheDocument();
+    });
+  });
+
   it("marks a SCIM-blocked key with the SCIM tooltip reason", async () => {
     mockUseKeys.mockReturnValue(keysResult([{ ...mockKey, blocked: true, metadata: { scim_blocked: true } }]));
 
@@ -790,6 +835,24 @@ describe("table state lives in the URL so it survives leaving and returning to t
     expect(screen.getByTestId("filter-chip-team_id")).toHaveTextContent("Test Team");
   });
 
+  it("restores the status filter from the URL and sends it to /key/list", async () => {
+    renderWithProviders(<VirtualKeysTable />, { searchParams: { filter_status: "deleted" } });
+
+    await waitFor(() => {
+      expect(mockUseKeys).toHaveBeenLastCalledWith(1, 50, expect.objectContaining({ status: "deleted" }));
+    });
+    expect(screen.getByTestId("filter-chip-status")).toHaveTextContent("Deleted");
+  });
+
+  it("ignores a hand-edited status the backend would reject instead of 400ing the page", async () => {
+    renderWithProviders(<VirtualKeysTable />, { searchParams: { filter_status: "bogus" } });
+
+    await waitFor(() => {
+      expect(mockUseKeys).toHaveBeenLastCalledWith(1, 50, expect.objectContaining({ status: undefined }));
+    });
+    expect(screen.queryByTestId("filter-chip-status")).not.toBeInTheDocument();
+  });
+
   it("writes the search term to the URL", async () => {
     const onUrlUpdate = vi.fn<OnUrlUpdateFunction>();
     renderWithProviders(<VirtualKeysTable />, { onUrlUpdate });

ui/litellm-dashboard/src/components/VirtualKeysPage/VirtualKeysTable.tsx

@@ -14,6 +14,7 @@ import {
 import { SearchSelect } from "@/components/shared/SearchSelect";
 import { PageHeader } from "@/components/shared/PageHeader";
 import { Input } from "@/components/ui/input";
+import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select";
 import { useDebouncedValue } from "@tanstack/react-pacer/debouncer";
 import { ColumnFiltersState, functionalUpdate, OnChangeFn, PaginationState, SortingState } from "@tanstack/react-table";
 import { KeyRound } from "lucide-react";
@@ -28,7 +29,7 @@ interface VirtualKeysTableProps {
   headerActions?: React.ReactNode;
 }
 
-const FILTER_COLUMNS = ["team_id", "org_id", "user_id", "key_hash"] as const;
+const FILTER_COLUMNS = ["team_id", "org_id", "user_id", "key_hash", "status"] as const;
 type FilterColumn = (typeof FILTER_COLUMNS)[number];
 
 const FILTER_LABELS: Record<FilterColumn, string> = {
@@ -36,8 +37,28 @@ const FILTER_LABELS: Record<FilterColumn, string> = {
   org_id: "Organization",
   user_id: "User ID",
   key_hash: "Key ID",
+  status: "Status",
 };
 
+const KEY_STATUS_VALUES = ["active", "expired", "revoked", "deleted"] as const;
+type KeyStatusFilter = (typeof KEY_STATUS_VALUES)[number];
+const ALL_STATUSES = "all";
+
+const KEY_STATUS_LABELS: Record<KeyStatusFilter, string> = {
+  active: "Active",
+  expired: "Expired",
+  revoked: "Revoked (blocked)",
+  deleted: "Deleted",
+};
+
+const STATUS_FILTER_ITEMS = [
+  { value: ALL_STATUSES, label: "All statuses" },
+  ...KEY_STATUS_VALUES.map((value) => ({ value, label: KEY_STATUS_LABELS[value] })),
+];
+
+const isKeyStatusFilter = (value: string): value is KeyStatusFilter =>
+  (KEY_STATUS_VALUES as readonly string[]).includes(value);
+
 const DEFAULT_SORT_BY = "created_at";
 const DEFAULT_SORT_ORDER = "desc";
 const DEFAULT_PAGE_SIZE = 50;
@@ -65,6 +86,7 @@ const TABLE_STATE = {
   filter_org: parseAsString.withDefault(""),
   filter_user: parseAsString.withDefault(""),
   filter_key_id: parseAsString.withDefault(""),
+  filter_status: parseAsString.withDefault(""),
 };
 
 const toSortOrder = (active: SortingState[number]): "asc" | "desc" => (active.desc ? "desc" : "asc");
@@ -96,15 +118,16 @@ export function VirtualKeysTable({ headerActions }: VirtualKeysTableProps) {
     () => ({ pageIndex: tableState.page - 1, pageSize: tableState.page_size }),
     [tableState.page, tableState.page_size],
   );
-  const { filter_team, filter_org, filter_user, filter_key_id } = tableState;
+  const { filter_team, filter_org, filter_user, filter_key_id, filter_status } = tableState;
   const appliedFilters = useMemo(
     () => ({
       team_id: filter_team.trim(),
       org_id: filter_org.trim(),
       user_id: filter_user.trim(),
       key_hash: filter_key_id.trim(),
+      status: isKeyStatusFilter(filter_status) ? filter_status : "",
     }),
-    [filter_team, filter_org, filter_user, filter_key_id],
+    [filter_team, filter_org, filter_user, filter_key_id, filter_status],
   );
   const columnFilters = useMemo<ColumnFiltersState>(
     () =>
@@ -121,6 +144,7 @@ export function VirtualKeysTable({ headerActions }: VirtualKeysTableProps) {
     search: searchQuery.trim() || undefined,
     userID: appliedFilters.user_id || undefined,
     keyHash: appliedFilters.key_hash || undefined,
+    status: appliedFilters.status || undefined,
     sortBy,
     sortOrder: tableState.sort_order,
     expand: "user",
@@ -164,6 +188,7 @@ export function VirtualKeysTable({ headerActions }: VirtualKeysTableProps) {
         filter_org: filterValue(next, "org_id"),
         filter_user: filterValue(next, "user_id"),
         filter_key_id: filterValue(next, "key_hash"),
+        filter_status: filterValue(next, "status"),
         page: null,
       };
       void setTableState(nextFilters);
@@ -233,6 +258,9 @@ export function VirtualKeysTable({ headerActions }: VirtualKeysTableProps) {
       if (columnId === "org_id") {
         return organizations.find((org) => org.organization_id === raw)?.organization_alias || raw;
       }
+      if (columnId === "status" && isKeyStatusFilter(raw)) {
+        return KEY_STATUS_LABELS[raw];
+      }
       return raw;
     },
     [allTeams, organizations],
@@ -340,6 +368,24 @@ export function VirtualKeysTable({ headerActions }: VirtualKeysTableProps) {
                       placeholder="Enter Key ID…"
                     />
                   </DataTableFilterField>
+                  <DataTableFilterField label="Status">
+                    <Select
+                      items={STATUS_FILTER_ITEMS}
+                      value={(get("status") as string) || ALL_STATUSES}
+                      onValueChange={(value) => set("status", value === ALL_STATUSES ? undefined : value)}
+                    >
+                      <SelectTrigger className="w-full" aria-label="Status">
+                        <SelectValue placeholder="All statuses" />
+                      </SelectTrigger>
+                      <SelectContent>
+                        {STATUS_FILTER_ITEMS.map((item) => (
+                          <SelectItem key={item.value} value={item.value}>
+                            {item.label}
+                          </SelectItem>
+                        ))}
+                      </SelectContent>
+                    </Select>
+                  </DataTableFilterField>
                 </>
               )}
             </DataTableFilterDrawer>

ui/litellm-dashboard/src/components/VirtualKeysPage/keyTableColumns.tsx

@@ -13,6 +13,7 @@ import {
   IdCell,
   IdentityCell,
   ModelsCell,
+  MoneyCell,
   SpendBudgetCell,
   StatusBadge,
   UserPopoverCell,
@@ -43,6 +44,13 @@ export const KEY_TABLE_SORT_FIELDS: readonly string[] = [
 ];
 
 const getKeyStatus = (key: KeyResponse): KeyStatus => {
+  if (key.deleted_at) {
+    return {
+      tone: "neutral",
+      label: "Deleted",
+      tooltip: `Deleted ${new Date(key.deleted_at).toLocaleString()}${key.deleted_by ? ` by ${key.deleted_by}` : ""}. Kept for audit and spend history; requests using this key are rejected.`,
+    };
+  }
   if (key.blocked === true) {
     const isScimBlocked = (key.metadata as Record<string, unknown> | null | undefined)?.scim_blocked === true;
     return {
@@ -274,6 +282,20 @@ export const getKeyTableColumns = ({
       );
     },
   },
+  {
+    id: "total_spend",
+    accessorKey: "total_spend",
+    meta: { title: "Lifetime Spend" },
+    header: () => (
+      <InfoHeader
+        label="Lifetime Spend"
+        tooltip="Cumulative spend across every budget period. Budget resets do not touch this value. Keys created before this field existed only count spend from then on."
+      />
+    ),
+    size: 130,
+    enableSorting: false,
+    cell: (info) => <MoneyCell value={info.getValue() as number | null | undefined} showZero />,
+  },
   {
     id: "budget_reset_at",
     accessorKey: "budget_reset_at",

ui/litellm-dashboard/src/components/claude_code_plugins/helpers.test.ts

@@ -156,6 +156,20 @@ describe("getSourceLink", () => {
   it("returns null when no repo or url", () => {
     expect(getSourceLink({ source: "github" })).toBeNull();
   });
+
+  it("keeps http and upper-case https urls registered through the api clickable", () => {
+    expect(getSourceLink({ source: "url", url: "http://git.internal.example/org/repo" })).toBe(
+      "http://git.internal.example/org/repo",
+    );
+    expect(getSourceLink({ source: "git-subdir", url: "HTTPS://gitlab.com/org/repo", path: "sub/dir" })).toBe(
+      "HTTPS://gitlab.com/org/repo",
+    );
+  });
+
+  it("returns null for an ssh clone url, which is not browsable", () => {
+    expect(getSourceLink({ source: "url", url: "git@ghe.example.com:org/repo.git" })).toBeNull();
+    expect(getSourceLink({ source: "url", url: "ssh://git@ghe.example.com/org/repo.git" })).toBeNull();
+  });
 });
 
 describe("getCategoryBadgeColor", () => {
@@ -466,6 +480,70 @@ describe("parseSkillSource", () => {
     expect(parseSkillSource("gitlab.com/org/repo", "a//b")).toBeNull();
   });
 
+  it("keeps an scp-style ssh clone url so private hosts authenticate with the user's key", () => {
+    expect(parseSkillSource("git@ghe.example.com:org/repo.git")?.parsed).toEqual({
+      source: "url",
+      url: "git@ghe.example.com:org/repo.git",
+    });
+    expect(parseSkillSource("git@ghe.example.com:org/repo.git")?.suggestedName).toBe("repo");
+  });
+
+  it("stores an ssh clone url exactly as typed, so a forced .git suffix cannot break azure devops or codecommit", () => {
+    for (const url of [
+      "git@ghe.example.com:org/repo",
+      "git@ssh.dev.azure.com:v3/org/project/repo",
+      "ssh://git@ghe.example.com/org/repo",
+      "ssh://apka1234@git-codecommit.us-east-1.amazonaws.com/v1/repos/my-repo",
+      "ssh://git@ghe.example.com:2222/org/nested/repo.git",
+    ]) {
+      expect(parseSkillSource(url)?.parsed).toEqual({ source: "url", url });
+    }
+    expect(parseSkillSource("git@ssh.dev.azure.com:v3/org/project/repo")?.suggestedName).toBe("repo");
+  });
+
+  it("accepts an internal host whose last label is not alphabetic, matching the https rule", () => {
+    expect(parseSkillSource("git@gitlab.internal.k8s2:org/repo.git")?.parsed).toEqual({
+      source: "url",
+      url: "git@gitlab.internal.k8s2:org/repo.git",
+    });
+    expect(parseSkillSource("https://gitlab.internal.k8s2/org/repo")?.parsed).toEqual({
+      source: "url",
+      url: "https://gitlab.internal.k8s2/org/repo",
+    });
+  });
+
+  it("combines an ssh clone url with an explicit subfolder", () => {
+    expect(parseSkillSource("git@ghe.example.com:org/repo.git", "plugins/my-skill")?.parsed).toEqual({
+      source: "git-subdir",
+      url: "git@ghe.example.com:org/repo.git",
+      path: "plugins/my-skill",
+    });
+    expect(parseSkillSource("git@ghe.example.com:org/repo.git", "../etc")).toBeNull();
+  });
+
+  it("rejects ssh-looking input without a host or repo path", () => {
+    expect(parseSkillSource("git@ghe.example.com:repo.git")).toBeNull();
+    expect(parseSkillSource("git@localhost:org/repo.git")).toBeNull();
+    expect(parseSkillSource("git@:org/repo.git")).toBeNull();
+    expect(parseSkillSource("ssh://ghe.example.com/org/repo.git")).toBeNull();
+  });
+
+  it("rejects ssh remotes with ip hosts or traversal segments", () => {
+    expect(parseSkillSource("git@10.0.0.5:org/repo.git")).toBeNull();
+    expect(parseSkillSource("ssh://git@169.254.169.254/org/repo")).toBeNull();
+    expect(parseSkillSource("git@ghe.example.com:../etc")).toBeNull();
+    expect(parseSkillSource("ssh://git@ghe.example.com/org/../repo")).toBeNull();
+    expect(parseSkillSource("git@ghe.example.com:org/../../etc/passwd")).toBeNull();
+    expect(parseSkillSource("git@ghe.example.com:org/.github")?.parsed).toEqual({
+      source: "url",
+      url: "git@ghe.example.com:org/.github",
+    });
+  });
+
+  it("rejects an ssh remote carrying a password, which would publish a secret on the feed", () => {
+    expect(parseSkillSource("ssh://git:s3cret@ghe.example.com/org/repo.git")).toBeNull();
+  });
+
   it("returns null for empty and garbage input", () => {
     expect(parseSkillSource("")).toBeNull();
     expect(parseSkillSource("   ")).toBeNull();
@@ -568,7 +646,7 @@ describe("parseSkillSource", () => {
 // Skill sources are served on the unauthenticated public feeds and cloned by clients, so the
 // parser must never publish an insecure, credentialed, internal, or malformed clone URL.
 describe("parseSkillSource — security boundary", () => {
-  it("rejects non-https schemes", () => {
+  it("rejects schemes other than https and user-qualified ssh", () => {
     for (const url of [
       "http://gitlab.com/org/repo",
       "HTTP://gitlab.com/org/repo",

ui/litellm-dashboard/src/components/claude_code_plugins/helpers.ts

@@ -29,17 +29,34 @@ export const SHA256_REGEX = /^[0-9a-fA-F]{64}$/;
 
 export const isValidSha256 = (digest: string): boolean => digest.trim() === "" || SHA256_REGEX.test(digest.trim());
 
-// WHATWG normalizes obfuscated IPv4 (e.g. 2130706433, 0x7f.0.0.1) to dotted-decimal, so this
-// catches every IPv4 form; bracketed IPv6 is rejected separately.
+// WHATWG normalizes obfuscated IPv4 (e.g. 2130706433, 0x7f.0.0.1) to dotted-decimal on https, so
+// this catches every IPv4 form there; on a non-special scheme like ssh it catches the dotted form
+// only. Bracketed IPv6 is rejected separately.
 const IPV4_HOST_REGEX = /^\d{1,3}(\.\d{1,3}){3}$/;
 
 const GITHUB_ORG_REGEX = /^[A-Za-z0-9-]+$/;
 const GITHUB_REPO_REGEX = /^[A-Za-z0-9._-]+$/;
 
+const BROWSABLE_URL_REGEX = /^https?:\/\//i;
+const SSH_SCHEME = "ssh://";
+const SSH_SCP_REGEX = /^([a-z0-9._-]+)@([^:/@]+):(?!\/)(.+)$/i;
+
 const buildRepoUrl = (url: URL): string => `${url.protocol}//${url.host}${url.pathname.replace(/\/+$/, "")}`;
 
 const pathSegments = (url: URL): string[] => url.pathname.split("/").filter((seg) => seg !== "");
 
+const toUrl = (candidate: string): URL | null => {
+  try {
+    return new URL(candidate);
+  } catch {
+    return null;
+  }
+};
+
+/** One host rule for every scheme, so an ssh remote is neither more nor less trusted than its https twin. */
+const isSafeHost = (url: URL): boolean =>
+  url.hostname.includes(".") && !url.hostname.startsWith("[") && !IPV4_HOST_REGEX.test(url.hostname);
+
 /**
  * Validate and normalize a repository URL into a parsed URL, or null. Enforces https (rejects
  * http/ssh/git/etc.), rejects embedded credentials, and requires a dotted host, so the public
@@ -52,20 +69,8 @@ const parseRepoUrl = (raw: string): URL | null => {
     return null;
   }
   const withScheme = /^[a-z][a-z0-9+.-]*:\/\//i.test(trimmed) ? trimmed : `https://${trimmed}`;
-  let url: URL;
-  try {
-    url = new URL(withScheme);
-  } catch {
-    return null;
-  }
-  if (
-    url.protocol !== "https:" ||
-    url.username !== "" ||
-    url.password !== "" ||
-    !url.hostname.includes(".") ||
-    url.hostname.startsWith("[") ||
-    IPV4_HOST_REGEX.test(url.hostname)
-  ) {
+  const url = toUrl(withScheme);
+  if (!url || url.protocol !== "https:" || url.username !== "" || url.password !== "" || !isSafeHost(url)) {
     return null;
   }
   return url;
@@ -140,13 +145,12 @@ const parseGitHubSource = (url: URL, subPath?: string): SkillSourcePreview | nul
   return repoPreview;
 };
 
-const parseRawGitSource = (url: URL, subPath?: string): SkillSourcePreview | null => {
-  if (pathSegments(url).length < 2) {
-    return null;
-  }
-
-  const repoUrl = buildRepoUrl(url);
-
+const buildGitSourcePreview = (
+  kind: "Git" | "SSH",
+  repoUrl: string,
+  repoName: string,
+  subPath?: string,
+): SkillSourcePreview | null => {
   const normalized = normalizeSubPath(subPath ?? "");
   if (normalized !== "") {
     if (!SUBDIR_PATH_REGEX.test(normalized)) {
@@ -154,18 +158,51 @@ const parseRawGitSource = (url: URL, subPath?: string): SkillSourcePreview | nul
     }
     return {
       parsed: { source: "git-subdir", url: repoUrl, path: normalized },
-      label: `Git subdir — ${repoUrl} @ ${normalized}`,
+      label: `${kind} subdir — ${repoUrl} @ ${normalized}`,
       suggestedName: toKebabCase(lastSegment(normalized)),
     };
   }
 
   return {
     parsed: { source: "url", url: repoUrl },
-    label: `Git repo — ${repoUrl}`,
-    suggestedName: toKebabCase(lastSegment(url.pathname).replace(/\.git$/, "")),
+    label: `${kind} repo — ${repoUrl}`,
+    suggestedName: toKebabCase(repoName),
   };
 };
 
+const parseRawGitSource = (url: URL, subPath?: string): SkillSourcePreview | null => {
+  if (pathSegments(url).length < 2) {
+    return null;
+  }
+  const repoName = lastSegment(url.pathname).replace(/\.git$/, "");
+  return buildGitSourcePreview("Git", buildRepoUrl(url), repoName, subPath);
+};
+
+/**
+ * Parse an scp-style `git@host:org/repo` or `ssh://git@host/org/repo` clone URL, registering it
+ * exactly as typed: git treats the `.git` suffix as optional, and forcing one on breaks hosts whose
+ * paths are not `org/repo`, like Azure DevOps `v3/...` and CodeCommit `v1/repos/...`. The scp form is
+ * rewritten to `ssh://` only to reuse the https host and credential rules, and only a URL that
+ * survives that round trip unchanged is accepted, which keeps traversal segments off the feed.
+ */
+const parseSshSource = (raw: string, subPath?: string): SkillSourcePreview | null => {
+  const trimmed = raw.trim();
+  const scp = SSH_SCP_REGEX.exec(trimmed);
+  const candidate = scp ? `${SSH_SCHEME}${scp[1]}@${scp[2]}/${scp[3]}` : trimmed;
+  if (!candidate.toLowerCase().startsWith(SSH_SCHEME)) {
+    return null;
+  }
+  const url = toUrl(candidate);
+  if (!url || url.username === "" || url.password !== "" || !isSafeHost(url)) {
+    return null;
+  }
+  const pathStart = candidate.indexOf("/", SSH_SCHEME.length);
+  if (pathStart === -1 || url.pathname !== candidate.slice(pathStart) || pathSegments(url).length < 2) {
+    return null;
+  }
+  return buildGitSourcePreview("SSH", trimmed, lastSegment(url.pathname).replace(/\.git$/i, ""), subPath);
+};
+
 const parseArchiveSource = (url: URL): SkillSourcePreview => ({
   parsed: { source: "archive", url: url.href },
   label: `Zip archive — ${url.host}${url.pathname}`,
@@ -175,10 +212,15 @@ const parseArchiveSource = (url: URL): SkillSourcePreview => ({
 /**
  * Parse any git-accessible repository URL or https zip archive URL into a registerable skill
  * source. A `.zip` path is an `archive` source (S3, Artifactory, any static host). GitHub URLs
- * keep their `github`/`git-subdir` shorthand; every other host is treated as a raw repo URL,
+ * keep their `github`/`git-subdir` shorthand; ssh clone URLs stay ssh so a private host
+ * authenticates with the user's own key; every other host is treated as a raw repo URL,
  * with an optional subfolder turning it into git-subdir.
  */
 export const parseSkillSource = (rawUrl: string, subPath?: string): SkillSourcePreview | null => {
+  const ssh = parseSshSource(rawUrl, subPath);
+  if (ssh) {
+    return ssh;
+  }
   const url = parseRepoUrl(rawUrl);
   if (!url) {
     return null;
@@ -268,14 +310,14 @@ export const getSourceDisplayText = (source: PluginSource): string => {
 };
 
 /**
- * Get clickable link for plugin source
+ * Get clickable link for plugin source. Ssh clone urls are not browsable, so they yield null.
  */
 export const getSourceLink = (source: PluginSource): string | null => {
   if (source.source === "github" && source.repo) {
     return `https://github.com/${source.repo}`;
   }
   const linksToUrl = source.source === "url" || source.source === "git-subdir" || source.source === "archive";
-  return linksToUrl && source.url ? source.url : null;
+  return linksToUrl && source.url && BROWSABLE_URL_REGEX.test(source.url) ? source.url : null;
 };
 
 /**

ui/litellm-dashboard/src/components/claude_code_plugins/skill_detail.test.tsx

@@ -0,0 +1,42 @@
+import { render, screen } from "@testing-library/react";
+import { describe, expect, it, vi } from "vitest";
+
+import { Plugin } from "./types";
+
+import SkillDetail from "./skill_detail";
+
+const buildSkill = (source: Plugin["source"]): Plugin => ({
+  id: "plugin-id",
+  name: "my-skill",
+  source,
+  enabled: true,
+});
+
+describe("SkillDetail source", () => {
+  it("links a github source to the repository", () => {
+    render(<SkillDetail skill={buildSkill({ source: "github", repo: "org/repo" })} onBack={vi.fn()} />);
+    expect(screen.getByRole("link", { name: "github.com/org/repo" })).toHaveAttribute(
+      "href",
+      "https://github.com/org/repo",
+    );
+  });
+
+  it("renders an ssh clone url as plain text instead of an unusable link", () => {
+    render(
+      <SkillDetail skill={buildSkill({ source: "url", url: "git@ghe.example.com:org/repo.git" })} onBack={vi.fn()} />,
+    );
+    expect(screen.getByText("git@ghe.example.com:org/repo.git")).toBeInTheDocument();
+    expect(screen.queryByRole("link")).not.toBeInTheDocument();
+  });
+
+  it("renders an ssh git-subdir source as plain text without a tree path", () => {
+    render(
+      <SkillDetail
+        skill={buildSkill({ source: "git-subdir", url: "git@ghe.example.com:org/repo.git", path: "plugins/x" })}
+        onBack={vi.fn()}
+      />,
+    );
+    expect(screen.getByText("git@ghe.example.com:org/repo.git @ plugins/x")).toBeInTheDocument();
+    expect(screen.queryByRole("link")).not.toBeInTheDocument();
+  });
+});

ui/litellm-dashboard/src/components/claude_code_plugins/skill_detail.tsx

@@ -1,8 +1,38 @@
 import React, { useState } from "react";
 import { ArrowLeft, Check, Copy, Link2 } from "lucide-react";
 import { cn } from "@/lib/cva.config";
-import { buildMarketplaceSettingsSnippet, formatInstallCommand } from "./helpers";
-import { Plugin } from "./types";
+import { buildMarketplaceSettingsSnippet, formatInstallCommand, getSourceDisplayText, getSourceLink } from "./helpers";
+import { Plugin, PluginSource } from "./types";
+
+const SkillSource: React.FC<{ source: PluginSource }> = ({ source }) => {
+  const link = getSourceLink(source);
+  const href = link && source.source === "git-subdir" && source.path ? `${link}/tree/main/${source.path}` : link;
+  if (href) {
+    return (
+      <div className="mb-6">
+        <div className="mb-1 text-xs text-muted-foreground">Source</div>
+        <a
+          href={href}
+          target="_blank"
+          rel="noopener noreferrer"
+          className="flex items-center gap-1 break-all text-[13px] text-info"
+        >
+          {href.replace("https://", "")}
+          <Link2 className="size-3 shrink-0" />
+        </a>
+      </div>
+    );
+  }
+  if (!source.url) {
+    return null;
+  }
+  return (
+    <div className="mb-6">
+      <div className="mb-1 text-xs text-muted-foreground">Source</div>
+      <div className="break-all text-[13px] text-foreground">{getSourceDisplayText(source)}</div>
+    </div>
+  );
+};
 
 interface SkillDetailProps {
   skill: Plugin;
@@ -22,14 +52,6 @@ const SkillDetail: React.FC<SkillDetailProps> = ({ skill, onBack }) => {
     setTimeout(() => setCopiedKey(null), 2000);
   };
 
-  const sourceUrl = (() => {
-    const src = skill.source;
-    if (src.source === "github" && src.repo) return `https://github.com/${src.repo}`;
-    if (src.source === "git-subdir" && src.url) return src.path ? `${src.url}/tree/main/${src.path}` : src.url;
-    if ((src.source === "url" || src.source === "archive") && src.url) return src.url;
-    return null;
-  })();
-
   const installCommand = formatInstallCommand(skill);
 
   const settingsSnippet = buildMarketplaceSettingsSnippet(
@@ -128,20 +150,7 @@ const SkillDetail: React.FC<SkillDetailProps> = ({ skill, onBack }) => {
               </span>
             </div>
 
-            {sourceUrl && (
-              <div className="mb-6">
-                <div className="mb-1 text-xs text-muted-foreground">Source</div>
-                <a
-                  href={sourceUrl}
-                  target="_blank"
-                  rel="noopener noreferrer"
-                  className="flex items-center gap-1 break-all text-[13px] text-info"
-                >
-                  {sourceUrl.replace("https://", "")}
-                  <Link2 className="size-3 shrink-0" />
-                </a>
-              </div>
-            )}
+            <SkillSource source={skill.source} />
 
             {skill.keywords && skill.keywords.length > 0 && (
               <div className="mb-6">

ui/litellm-dashboard/src/components/key_team_helpers/ModelMaxBudgetEditor.tsx

@@ -144,6 +144,7 @@ export function ModelMaxBudgetEditor({
               onClick={() => removeEntry(entry.id)}
               disabled={!premiumUser}
               title={hintWhenLocked}
+              aria-label="Remove model budget"
               className="absolute top-2 right-2 text-muted-foreground hover:text-destructive transition-colors p-1"
             >
               <X className="w-4 h-4" />

ui/litellm-dashboard/src/components/key_team_helpers/key_list.tsx

@@ -39,6 +39,7 @@ export interface KeyResponse {
   key_name: string;
   key_alias: string;
   spend: number;
+  total_spend: number;
   max_budget: number;
   expires: string;
   models: string[];
@@ -64,6 +65,8 @@ export interface KeyResponse {
   model_max_budget_usage?: Record<string, ModelBudgetUsage> | null;
   soft_budget_cooldown: boolean;
   blocked: boolean;
+  deleted_at?: string | null;
+  deleted_by?: string | null;
   litellm_budget_table: Record<string, unknown>;
   organization_id: string | null;
   org_id?: string | null;

ui/litellm-dashboard/src/components/settings.test.tsx

@@ -302,6 +302,113 @@ describe("Settings", () => {
     });
   });
 
+  const mockS3Callback = (variables: Record<string, string | null>, callbackName = "s3") => {
+    mockGetCallbacksCall.mockResolvedValue({
+      callbacks: [{ name: callbackName, variables }],
+      available_callbacks: {
+        s3: {
+          litellm_callback_name: "s3",
+          litellm_callback_params: [
+            "AWS_ACCESS_KEY_ID",
+            "AWS_SECRET_ACCESS_KEY",
+            "AWS_REGION_NAME",
+            "S3_LOG_PROMPTS_ONLY",
+          ],
+          ui_callback_name: "s3 Bucket (AWS)",
+        },
+      },
+      alerts: [],
+    });
+    mockGetCallbackConfigsCall.mockResolvedValue([
+      {
+        id: "s3",
+        displayName: "S3",
+        dynamic_params: {
+          s3_bucket_name: { type: "text", ui_name: "S3 Bucket Name", required: false },
+          s3_log_prompts_only: { type: "boolean", ui_name: "Log Prompts Only", required: false },
+        },
+      },
+    ]);
+  };
+
+  const openS3EditModal = async (callbackName = "s3") => {
+    const user = userEvent.setup();
+    render(<Settings {...defaultProps} />);
+    await user.click(await screen.findByTestId(`callback-actions-${callbackName}-success`));
+    await user.click(await screen.findByTestId("callback-action-edit"));
+    return user;
+  };
+
+  it("should render a saved boolean dynamic param as a checked switch and post false when toggled off", async () => {
+    mockS3Callback({ S3_LOG_PROMPTS_ONLY: "true" });
+    const user = await openS3EditModal();
+
+    const promptsOnlySwitch = await screen.findByRole("switch", { name: "Log Prompts Only" });
+    expect(promptsOnlySwitch).toBeChecked();
+
+    await user.click(promptsOnlySwitch);
+    expect(promptsOnlySwitch).not.toBeChecked();
+    await user.click(within(screen.getByRole("dialog")).getByRole("button", { name: "Save Changes" }));
+
+    await waitFor(() => {
+      expect(vi.mocked(setCallbacksCall)).toHaveBeenCalledWith(
+        "token",
+        expect.objectContaining({
+          environment_variables: expect.objectContaining({ callback: "s3", s3_log_prompts_only: "false" }),
+        }),
+      );
+    });
+  });
+
+  it("should render an unset boolean dynamic param as an unchecked switch and post true when toggled on", async () => {
+    mockS3Callback({ S3_LOG_PROMPTS_ONLY: null });
+    const user = await openS3EditModal();
+
+    const promptsOnlySwitch = await screen.findByRole("switch", { name: "Log Prompts Only" });
+    expect(promptsOnlySwitch).not.toBeChecked();
+
+    await user.click(promptsOnlySwitch);
+    await user.click(within(screen.getByRole("dialog")).getByRole("button", { name: "Save Changes" }));
+
+    await waitFor(() => {
+      expect(vi.mocked(setCallbacksCall)).toHaveBeenCalledWith(
+        "token",
+        expect.objectContaining({
+          environment_variables: expect.objectContaining({ callback: "s3", s3_log_prompts_only: "true" }),
+        }),
+      );
+    });
+  });
+
+  it.each(["True", "1"])("should render a boolean dynamic param stored as %s as a checked switch", async (stored) => {
+    mockS3Callback({ S3_LOG_PROMPTS_ONLY: stored });
+    await openS3EditModal();
+
+    expect(await screen.findByRole("switch", { name: "Log Prompts Only" })).toBeChecked();
+  });
+
+  it("should resolve the s3_v2 callback to the s3 dynamic params and post under the s3_v2 name", async () => {
+    mockS3Callback({ S3_LOG_PROMPTS_ONLY: null }, "s3_v2");
+    const user = await openS3EditModal("s3_v2");
+
+    const promptsOnlySwitch = await screen.findByRole("switch", { name: "Log Prompts Only" });
+    expect(promptsOnlySwitch).not.toBeChecked();
+    expect(within(screen.getByRole("dialog")).getByRole("combobox", { name: "Callback" })).toHaveValue("S3");
+
+    await user.click(promptsOnlySwitch);
+    await user.click(within(screen.getByRole("dialog")).getByRole("button", { name: "Save Changes" }));
+
+    await waitFor(() => {
+      expect(vi.mocked(setCallbacksCall)).toHaveBeenCalledWith(
+        "token",
+        expect.objectContaining({
+          environment_variables: expect.objectContaining({ callback: "s3_v2", s3_log_prompts_only: "true" }),
+          litellm_settings: { success_callback: ["s3_v2"] },
+        }),
+      );
+    });
+  });
+
   it("should send the typed webhook url for an alert type when the alerting tab is saved", async () => {
     const user = userEvent.setup();
     render(<Settings {...defaultProps} />);

ui/litellm-dashboard/src/components/settings.tsx

@@ -67,19 +67,20 @@ const DynamicParamsFields: React.FC<DynamicParamsFieldsProps> = ({ params, callb
     return null;
   }
 
+  const callbackConfig = findCallbackConfig(callbackConfigs, selectedCallback);
   return (
     <div className="space-y-4 mt-6 p-4 bg-muted rounded-lg border">
       {params.map((param) => {
-        const callbackConfig = callbackConfigs.find((config) => config.id === selectedCallback);
         const paramConfig = callbackConfig?.dynamic_params?.[param] || {};
         const paramType = paramConfig.type || "text";
         const fieldLabel = paramConfig.ui_name || param.replace(/_/g, " ").replace(/\b\w/g, (l) => l.toUpperCase());
         const isRequired = paramConfig.required || false;
         const selectOptions: string[] = Array.isArray(paramConfig.options) ? paramConfig.options : [];
         const isSelect = paramType === "select" && selectOptions.length > 0;
+        const isBoolean = paramType === "boolean";
         const fieldId = `${fieldIdPrefix}-${param}`;
         const validationRules = isRequired ? { required: `Please enter the ${fieldLabel.toLowerCase()}` } : undefined;
-        const registration = isSelect ? undefined : register(param, validationRules);
+        const registration = isSelect || isBoolean ? undefined : register(param, validationRules);
 
         return (
           <Field key={param} className="mb-4">
@@ -111,7 +112,22 @@ const DynamicParamsFields: React.FC<DynamicParamsFieldsProps> = ({ params, callb
                 )}
               />
             )}
+            {isBoolean && (
+              <Controller
+                control={control}
+                name={param}
+                render={({ field }) => (
+                  <Switch
+                    id={fieldId}
+                    checked={/^(true|1)$/i.test(String(field.value ?? ""))}
+                    onCheckedChange={(checked: boolean) => field.onChange(checked ? "true" : "false")}
+                    onBlur={field.onBlur}
+                  />
+                )}
+              />
+            )}
             {!isSelect &&
+              !isBoolean &&
               (paramType === "password" ? (
                 <Input
                   id={fieldId}
@@ -162,7 +178,7 @@ export const CallbackSelector: React.FC<CallbackSelectorProps> = ({
 }) => {
   const { control } = useFormContext<CallbackFormValues>();
   const inputId = React.useId();
-  const selectedConfig = callbackConfigs.find((config) => config.id === selectedCallback) ?? null;
+  const selectedConfig = findCallbackConfig(callbackConfigs, selectedCallback) ?? null;
 
   return (
     <Controller
@@ -221,6 +237,31 @@ export const CallbackSelector: React.FC<CallbackSelectorProps> = ({
   );
 };
 
+const CALLBACK_CONFIG_ALIASES: Record<string, string> = { s3_v2: "s3" };
+
+interface DynamicParamConfig {
+  type?: string;
+  ui_name?: string;
+  required?: boolean;
+  options?: string[];
+}
+
+interface CallbackConfigWithParams {
+  id: string;
+  dynamic_params?: Record<string, DynamicParamConfig>;
+}
+
+const findCallbackConfig = <T extends { id: string }>(
+  callbackConfigs: readonly T[],
+  callbackName: string | null,
+): T | undefined => {
+  if (!callbackName) {
+    return undefined;
+  }
+  const configId = CALLBACK_CONFIG_ALIASES[callbackName] ?? callbackName;
+  return callbackConfigs.find((config) => config.id === configId);
+};
+
 // Shared helper function to get dynamic params for a callback
 const getDynamicParamsForCallback = (
   callbackName: string | null,
@@ -231,7 +272,7 @@ const getDynamicParamsForCallback = (
     return fallbackVariables ? Object.keys(fallbackVariables) : [];
   }
 
-  const callbackConfig = callbackConfigs.find((config) => config.id === callbackName);
+  const callbackConfig = findCallbackConfig(callbackConfigs, callbackName);
   if (callbackConfig?.dynamic_params) {
     return Object.keys(callbackConfig.dynamic_params);
   }

ui/litellm-dashboard/src/components/team/TeamInfo.test.tsx

@@ -1609,6 +1609,99 @@ describe("TeamInfoView", () => {
     });
   });
 
+  describe("per-model budgets", () => {
+    const teamWithModelBudget = () =>
+      createMockTeamData({
+        models: ["gpt-4"],
+        model_max_budget: { "gpt-4": { max_budget: 5, budget_duration: "1d" } },
+        model_max_budget_usage: { "gpt-4": { current_spend: 1.25, budget_limit: 5, time_period: "1d" } },
+      });
+
+    const openSettingsEditor = async (user: ReturnType<typeof userEvent.setup>) => {
+      await waitFor(() => {
+        expect(screen.queryAllByText("Test Team").length).toBeGreaterThan(0);
+      });
+      await user.click(screen.getByRole("tab", { name: "Settings" }));
+      await user.click(await screen.findByRole("button", { name: /edit settings/i }));
+      await screen.findByLabelText("Team Name");
+    };
+
+    const savedPayload = async () => {
+      await waitFor(() => {
+        expect(networking.teamUpdateCall).toHaveBeenCalled();
+      });
+      return vi.mocked(networking.teamUpdateCall).mock.calls[0][1] as Record<string, unknown>;
+    };
+
+    it("shows the stored per-model budget and its current spend in the read-only settings view", async () => {
+      const user = userEvent.setup({ delay: null });
+      vi.mocked(networking.teamInfoCall).mockResolvedValue(teamWithModelBudget());
+
+      renderWithProviders(<TeamInfoView {...defaultProps} />);
+
+      await waitFor(() => {
+        expect(screen.queryAllByText("Test Team").length).toBeGreaterThan(0);
+      });
+      await user.click(screen.getByRole("tab", { name: "Settings" }));
+
+      expect(await screen.findByText("Per-Model Budget (gpt-4): $5 per 1d, spent $1.25")).toBeInTheDocument();
+    });
+
+    it("seeds the editor from the stored budget and keeps it read-only without an enterprise license", async () => {
+      const user = userEvent.setup({ delay: null });
+      vi.mocked(networking.teamInfoCall).mockResolvedValue(teamWithModelBudget());
+
+      renderWithProviders(<TeamInfoView {...defaultProps} premiumUser={false} />);
+
+      await openSettingsEditor(user);
+
+      expect(screen.getByPlaceholderText("Max spend ($)")).toHaveValue(5);
+      expect(screen.getByPlaceholderText("Max spend ($)")).toBeDisabled();
+      expect(screen.getByRole("button", { name: /Add Model Budget/i })).toBeDisabled();
+    });
+
+    it("leaves model_max_budget out of a save that did not touch it", async () => {
+      const user = userEvent.setup({ delay: null });
+      vi.mocked(networking.teamInfoCall).mockResolvedValue(teamWithModelBudget());
+      vi.mocked(networking.teamUpdateCall).mockResolvedValue({ data: {}, team_id: "123" } as any);
+
+      renderWithProviders(<TeamInfoView {...defaultProps} premiumUser={true} />);
+
+      await openSettingsEditor(user);
+      await user.click(screen.getByRole("button", { name: /save changes/i }));
+
+      expect(await savedPayload()).not.toHaveProperty("model_max_budget");
+    });
+
+    it("sends the edited cap for the model", async () => {
+      const user = userEvent.setup({ delay: null });
+      vi.mocked(networking.teamInfoCall).mockResolvedValue(teamWithModelBudget());
+      vi.mocked(networking.teamUpdateCall).mockResolvedValue({ data: {}, team_id: "123" } as any);
+
+      renderWithProviders(<TeamInfoView {...defaultProps} premiumUser={true} />);
+
+      await openSettingsEditor(user);
+      fireEvent.change(screen.getByPlaceholderText("Max spend ($)"), { target: { value: "2.5" } });
+      await user.click(screen.getByRole("button", { name: /save changes/i }));
+
+      expect((await savedPayload()).model_max_budget).toEqual({ "gpt-4": { budget_limit: 2.5, time_period: "1d" } });
+    });
+
+    it("sends an empty model_max_budget when the last row is removed, so the stored cap is cleared", async () => {
+      const user = userEvent.setup({ delay: null });
+      vi.mocked(networking.teamInfoCall).mockResolvedValue(teamWithModelBudget());
+      vi.mocked(networking.teamUpdateCall).mockResolvedValue({ data: {}, team_id: "123" } as any);
+
+      renderWithProviders(<TeamInfoView {...defaultProps} premiumUser={true} />);
+
+      await openSettingsEditor(user);
+      await user.click(screen.getByRole("button", { name: "Remove model budget" }));
+      await user.click(screen.getByRole("button", { name: /save changes/i }));
+
+      expect((await savedPayload()).model_max_budget).toEqual({});
+    });
+  });
+
   describe("team member settings", () => {
     it("should populate Default Key Duration from the team's stored metadata", async () => {
       const user = userEvent.setup({ delay: null });

ui/litellm-dashboard/src/components/team/TeamInfo.tsx

@@ -51,6 +51,13 @@ import GuardrailsSelect from "./GuardrailsSelect";
 import { copyToClipboard as utilCopyToClipboard } from "../../utils/dataUtils";
 import AccessGroupSelector from "../common_components/AccessGroupSelector";
 import BudgetDurationDropdown, { NEVER_RESETS_BUDGET_DURATION } from "../common_components/budget_duration_dropdown";
+import {
+  ModelBudgetUsage,
+  ModelMaxBudget,
+  ModelMaxBudgetField,
+  modelMaxBudgetToEntries,
+} from "../key_team_helpers/ModelMaxBudgetEditor";
+import { modelMaxBudgetUpdate, StoredModelMaxBudget } from "../key_team_helpers/modelMaxBudgetPayload";
 import {
   computeTeamModelBadges,
   normalizeTeamModelSelection,
@@ -268,6 +275,8 @@ export interface TeamData {
     max_budget: number | null;
     soft_budget?: number | null;
     budget_duration: string | null;
+    model_max_budget?: StoredModelMaxBudget | null;
+    model_max_budget_usage?: Record<string, ModelBudgetUsage> | null;
     models: string[];
     blocked: boolean;
     spend: number;
@@ -563,6 +572,7 @@ const TeamInfoView: React.FC<TeamInfoProps> = ({
   const [isDeleting, setIsDeleting] = useState(false);
   const [isTeamSaving, setIsTeamSaving] = useState(false);
   const [teamModelAliases, setTeamModelAliases] = useState<Record<string, string>>({});
+  const [teamModelMaxBudget, setTeamModelMaxBudget] = useState<ModelMaxBudget>({});
   const routerSettingsRef = React.useRef<RouterSettingsAccordionRef>(null);
   const [organization, setOrganization] = useState<Organization | null>(null);
   const { userRole, userId } = useAuthorized();
@@ -628,6 +638,7 @@ const TeamInfoView: React.FC<TeamInfoProps> = ({
 
   const startEditing = () => {
     form.reset(teamFormValues());
+    setTeamModelMaxBudget((teamData?.team_info?.model_max_budget ?? {}) as ModelMaxBudget);
     setTeamMemberSettingsOpen(false);
     setSearchToolSettingsOpen(false);
     setIsEditing(true);
@@ -1078,6 +1089,11 @@ const TeamInfoView: React.FC<TeamInfoProps> = ({
         updateData.model_aliases = teamModelAliases;
       }
 
+      const modelBudgets = modelMaxBudgetUpdate(teamModelMaxBudget, info.model_max_budget);
+      if (modelBudgets !== undefined) {
+        updateData.model_max_budget = modelBudgets;
+      }
+
       // Handle router_settings - read fresh values from DOM at save time.
       const currentRouterSettings = routerSettingsRef.current?.getValue();
       if (currentRouterSettings?.router_settings) {
@@ -1536,6 +1552,15 @@ const TeamInfoView: React.FC<TeamInfoProps> = ({
                     )}
                   </FormField>
 
+                  <ModelMaxBudgetField
+                    premiumUser={premiumUser}
+                    value={teamModelMaxBudget}
+                    onChange={setTeamModelMaxBudget}
+                    availableModels={availableRateLimitModels}
+                    usage={info.model_max_budget_usage}
+                    hint="Cap this team's spend on individual models, each with its own reset window. Every key on the team shares the cap unless the key sets its own budget for that model."
+                  />
+
                   <FormField control={form.control} name="tpm_limit" label="Tokens per minute Limit (TPM)">
                     {({ ref, value, ...field }) => <NumericalInput {...field} ref={ref} value={value ?? ""} step={1} />}
                   </FormField>
@@ -2051,6 +2076,17 @@ const TeamInfoView: React.FC<TeamInfoProps> = ({
                     : "No Limit"}
                 </div>
                 <div>Budget Reset: {info.budget_duration || "Never"}</div>
+                {modelMaxBudgetToEntries(info.model_max_budget as ModelMaxBudget | null | undefined).map(
+                  ({ model, budgetLimit, timePeriod }) => {
+                    const spent = model === null ? undefined : info.model_max_budget_usage?.[model]?.current_spend;
+                    return (
+                      <div key={model}>
+                        Per-Model Budget ({model}): ${budgetLimit ?? "?"} per {timePeriod}
+                        {spent !== undefined && `, spent $${spent}`}
+                      </div>
+                    );
+                  },
+                )}
                 {info.metadata?.soft_budget_alerting_emails &&
                   Array.isArray(info.metadata.soft_budget_alerting_emails) &&
                   info.metadata.soft_budget_alerting_emails.length > 0 && (

ui/litellm-dashboard/src/components/templates/key_edit_view.integration.test.tsx

@@ -174,6 +174,7 @@ describe("KeyEditView", () => {
     key_name: "sk-...TUuw",
     key_alias: "asdasdas",
     spend: 0,
+    total_spend: 0,
     max_budget: 0,
     expires: "null",
     models: [],

ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx

@@ -119,6 +119,7 @@ describe("KeyInfoView", () => {
     key_name: "sk-...TUuw",
     key_alias: "asdasdas",
     spend: 0,
+    total_spend: 0,
     max_budget: 0,
     expires: "null",
     models: [],
@@ -272,6 +273,23 @@ describe("KeyInfoView", () => {
     });
   });
 
+  it("shows lifetime spend separately from the resettable period spend", async () => {
+    vi.mocked(useAuthorized).mockReturnValue(baseUseAuthorizedMock);
+
+    renderWithProviders(
+      <KeyInfoView
+        keyData={{ ...MOCK_KEY_DATA, spend: 0.25, total_spend: 340.5 }}
+        onClose={() => {}}
+        keyId={"test-key-id"}
+        onKeyDataUpdate={() => {}}
+        teams={[]}
+      />,
+    );
+
+    expect(await screen.findByText("$0.2500")).toBeInTheDocument();
+    expect(screen.getByTestId("key-lifetime-spend")).toHaveTextContent("Lifetime spend: $340.5000");
+  });
+
   it("should render the key's saved router fallbacks", async () => {
     vi.mocked(useAuthorized).mockReturnValue(baseUseAuthorizedMock);
 

ui/litellm-dashboard/src/components/templates/key_info_view.tsx

@@ -677,6 +677,9 @@ export default function KeyInfoView({
                   {currentKeyData.budget_reset_at && (
                     <p className="text-sm">Resets {formatTimestamp(currentKeyData.budget_reset_at)}</p>
                   )}
+                  <p className="text-sm mt-2" data-testid="key-lifetime-spend">
+                    Lifetime spend: ${formatNumberWithCommas(currentKeyData.total_spend ?? 0, 4)}
+                  </p>
                 </div>
               </Card>
 
@@ -935,6 +938,11 @@ export default function KeyInfoView({
                     <p className="text-sm">${formatNumberWithCommas(currentKeyData.spend, 4)} USD</p>
                   </div>
 
+                  <div>
+                    <p className="text-sm font-medium">Lifetime Spend</p>
+                    <p className="text-sm">${formatNumberWithCommas(currentKeyData.total_spend ?? 0, 4)} USD</p>
+                  </div>
+
                   <div>
                     <p className="text-sm font-medium">Budget</p>
                     <p className="text-sm">

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -7859,7 +7859,10 @@ export interface paths {
          *
          *     Returns:
          *     - key: str - The key that was looked up, echoed back as it was passed in
-         *     - info: dict - The key's row, minus the hashed token
+         *     - info: dict - The key's row, minus the hashed token. Deleted keys are served from the
+         *       LiteLLM_DeletedVerificationToken archive and carry deleted_at / deleted_by
+         *         - status: "active" | "expired" | "revoked" | "deleted" - Derived from blocked, expires and
+         *           whether the row came from the archive
          *         - key_alias: str | None - User-friendly key alias
          *         - spend: float - Amount spent by the key. When budget_duration is set this covers only the
          *           current budget window, not the key's lifetime
@@ -7917,7 +7920,9 @@ export interface paths {
          *
          *     Parameters:
          *         expand: Optional[List[str]] - Expand related objects (e.g. 'user' to include user information)
-         *         status: Optional[str] - Filter by status. Currently supports "deleted" to query deleted keys.
+         *         status: Optional[str] - Filter by status: "active", "expired", "revoked" (blocked) or "deleted".
+         *         "deleted" reads the LiteLLM_DeletedVerificationToken archive; the other values partition the
+         *         live key table, so every live key matches exactly one of them.
          *
          *     Returns:
          *         {
@@ -15682,6 +15687,7 @@ export interface paths {
          *     - prompts: Optional[List[str]] - List of prompts that the team is allowed to use.
          *     - organization_id: Optional[str] - The organization id of the team. Default is None. Create via `/organization/new`.
          *     - model_aliases: Optional[dict] - Model aliases for the team. [Docs](https://docs.litellm.ai/docs/proxy/team_based_routing#create-team-with-model-alias)
+         *     - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
          *     - guardrails: Optional[List[str]] - Guardrails for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails)
          *     - policies: Optional[List[str]] - Policies for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails/guardrail_policies)
          *     - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key.
@@ -15908,6 +15914,7 @@ export interface paths {
          *     - tags: Optional[List[str]] - Tags for [tracking spend](https://litellm.vercel.app/docs/proxy/enterprise#tracking-spend-for-custom-tags) and/or doing [tag-based routing](https://litellm.vercel.app/docs/proxy/tag_routing).
          *     - organization_id: Optional[str] - The organization id of the team. Default is None. Create via `/organization/new`.
          *     - model_aliases: Optional[dict] - Model aliases for the team. [Docs](https://docs.litellm.ai/docs/proxy/team_based_routing#create-team-with-model-alias)
+         *     - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
          *     - guardrails: Optional[List[str]] - Guardrails for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails)
          *     - policies: Optional[List[str]] - Policies for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails/guardrail_policies)
          *     - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key.
@@ -29648,6 +29655,8 @@ export interface components {
             object_permission_id?: string | null;
             /** Org Id */
             org_id?: string | null;
+            /** Organization Id */
+            organization_id?: string | null;
             /**
              * Permissions
              * @default {}
@@ -29686,6 +29695,11 @@ export interface components {
             team_id?: string | null;
             /** Token */
             token?: string | null;
+            /**
+             * Total Spend
+             * @default 0
+             */
+            total_spend: number;
             /** Tpd Limit */
             tpd_limit?: number | null;
             /** Tpm Limit */
@@ -31259,6 +31273,11 @@ export interface components {
             team_id?: string | null;
             /** Token */
             token?: string | null;
+            /**
+             * Total Spend
+             * @default 0
+             */
+            total_spend: number;
             /** Tpd Limit */
             tpd_limit?: number | null;
             /** Tpm Limit */
@@ -33419,6 +33438,13 @@ export interface components {
             model_aliases?: {
                 [key: string]: unknown;
             } | null;
+            /**
+             * Model Max Budget
+             * @description Max budget per model for every key on the team, overridable per key (e.g. {'gpt-4o': {'max_budget': 10, 'budget_duration': '1d'}})
+             */
+            model_max_budget?: {
+                [key: string]: components["schemas"]["BudgetConfig"];
+            } | null;
             /** Model Rpm Limit */
             model_rpm_limit?: {
                 [key: string]: number;
@@ -34177,6 +34203,13 @@ export interface components {
             model_aliases?: {
                 [key: string]: unknown;
             } | null;
+            /**
+             * Model Max Budget
+             * @description Max budget per model for every key on the team, overridable per key (e.g. {'gpt-4o': {'max_budget': 10, 'budget_duration': '1d'}})
+             */
+            model_max_budget?: {
+                [key: string]: components["schemas"]["BudgetConfig"];
+            } | null;
             /** Model Rpm Limit */
             model_rpm_limit?: {
                 [key: string]: number;
@@ -39326,6 +39359,13 @@ export interface components {
             model_aliases?: {
                 [key: string]: unknown;
             } | null;
+            /**
+             * Model Max Budget
+             * @description Max budget per model for every key on the team, overridable per key (e.g. {'gpt-4o': {'max_budget': 10, 'budget_duration': '1d'}})
+             */
+            model_max_budget?: {
+                [key: string]: components["schemas"]["BudgetConfig"];
+            } | null;
             /** Model Rpm Limit */
             model_rpm_limit?: {
                 [key: string]: number;
@@ -40037,6 +40077,10 @@ export interface components {
             team_model_aliases?: {
                 [key: string]: unknown;
             } | null;
+            /** Team Model Max Budget */
+            team_model_max_budget?: {
+                [key: string]: unknown;
+            } | null;
             /**
              * Team Models
              * @default []
@@ -40057,6 +40101,11 @@ export interface components {
             team_tpm_limit?: number | null;
             /** Token */
             token?: string | null;
+            /**
+             * Total Spend
+             * @default 0
+             */
+            total_spend: number;
             /** Tpd Limit */
             tpd_limit?: number | null;
             /** Tpm Limit */
@@ -51366,7 +51415,7 @@ export interface operations {
                 sort_order?: string;
                 /** @description Expand related objects (e.g. 'user') */
                 expand?: string[] | null;
-                /** @description Filter by status (e.g. 'deleted') */
+                /** @description Filter by status: 'active' (not blocked, not expired), 'expired' (not blocked, past expiry), 'revoked' (blocked) or 'deleted' (archived keys). Omit to return live keys regardless of status. */
                 status?: string | null;
                 /** @description Filter keys by project ID */
                 project_id?: string | null;