agent: Handle context window exceeded errors from Anthropic (#28688)

![CleanShot 2025-04-14 at 11 15
38@2x](https://github.com/user-attachments/assets/9e803ffb-74fd-486b-bebc-2155a407a9fa)

Release Notes:

- agent: Handle context window exceeded errors from Anthropic

Cargo.lock

@@ -7708,6 +7708,7 @@ dependencies = [
  "smol",
  "strum",
  "theme",
+ "thiserror 2.0.12",
  "tiktoken-rs",
  "tokio",
  "ui",

crates/agent/src/message_editor.rs

@@ -761,13 +761,29 @@ impl MessageEditor {
             })
     }
 
-    fn render_reaching_token_limit(&self, line_height: Pixels, cx: &mut Context<Self>) -> Div {
+    fn render_token_limit_callout(
+        &self,
+        line_height: Pixels,
+        token_usage_ratio: TokenUsageRatio,
+        cx: &mut Context<Self>,
+    ) -> Div {
+        let heading = if token_usage_ratio == TokenUsageRatio::Exceeded {
+            "Thread reached the token limit"
+        } else {
+            "Thread reaching the token limit soon"
+        };
+
         h_flex()
             .p_2()
             .gap_2()
             .flex_wrap()
             .justify_between()
-            .bg(cx.theme().status().warning_background.opacity(0.1))
+            .bg(
+                if token_usage_ratio == TokenUsageRatio::Exceeded {
+                    cx.theme().status().error_background.opacity(0.1)
+                } else {
+                    cx.theme().status().warning_background.opacity(0.1)
+                })
             .border_t_1()
             .border_color(cx.theme().colors().border)
             .child(
@@ -779,15 +795,21 @@ impl MessageEditor {
                             .h(line_height)
                             .justify_center()
                             .child(
-                                Icon::new(IconName::Warning)
-                                    .color(Color::Warning)
-                                    .size(IconSize::XSmall),
+                                if token_usage_ratio == TokenUsageRatio::Exceeded {
+                                    Icon::new(IconName::X)
+                                        .color(Color::Error)
+                                        .size(IconSize::XSmall)
+                                } else {
+                                    Icon::new(IconName::Warning)
+                                        .color(Color::Warning)
+                                        .size(IconSize::XSmall)
+                                }
                             ),
                     )
                     .child(
                         v_flex()
                             .mr_auto()
-                            .child(Label::new("Thread reaching the token limit soon").size(LabelSize::Small))
+                            .child(Label::new(heading).size(LabelSize::Small))
                             .child(
                                 Label::new(
                                     "Start a new thread from a summary to continue the conversation.",
@@ -875,7 +897,13 @@ impl Render for MessageEditor {
             .child(self.render_editor(font_size, line_height, window, cx))
             .when(
                 total_token_usage.ratio != TokenUsageRatio::Normal,
-                |parent| parent.child(self.render_reaching_token_limit(line_height, cx)),
+                |parent| {
+                    parent.child(self.render_token_limit_callout(
+                        line_height,
+                        total_token_usage.ratio,
+                        cx,
+                    ))
+                },
             )
     }
 }

crates/agent/src/thread.rs

@@ -15,10 +15,11 @@ use futures::{FutureExt, StreamExt as _};
 use git::repository::DiffType;
 use gpui::{App, AppContext, Context, Entity, EventEmitter, SharedString, Task, WeakEntity};
 use language_model::{
-    ConfiguredModel, LanguageModel, LanguageModelCompletionEvent, LanguageModelRegistry,
-    LanguageModelRequest, LanguageModelRequestMessage, LanguageModelRequestTool,
-    LanguageModelToolResult, LanguageModelToolUseId, MaxMonthlySpendReachedError, MessageContent,
-    PaymentRequiredError, Role, StopReason, TokenUsage,
+    ConfiguredModel, LanguageModel, LanguageModelCompletionEvent, LanguageModelId,
+    LanguageModelKnownError, LanguageModelRegistry, LanguageModelRequest,
+    LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
+    LanguageModelToolUseId, MaxMonthlySpendReachedError, MessageContent, PaymentRequiredError,
+    Role, StopReason, TokenUsage,
 };
 use project::Project;
 use project::git_store::{GitStore, GitStoreCheckpoint, RepositoryState};
@@ -228,7 +229,7 @@ pub struct TotalTokenUsage {
     pub ratio: TokenUsageRatio,
 }
 
-#[derive(Default, PartialEq, Eq)]
+#[derive(Debug, Default, PartialEq, Eq)]
 pub enum TokenUsageRatio {
     #[default]
     Normal,
@@ -260,11 +261,20 @@ pub struct Thread {
     pending_checkpoint: Option<ThreadCheckpoint>,
     initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
     cumulative_token_usage: TokenUsage,
+    exceeded_window_error: Option<ExceededWindowError>,
     feedback: Option<ThreadFeedback>,
     message_feedback: HashMap<MessageId, ThreadFeedback>,
     last_auto_capture_at: Option<Instant>,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ExceededWindowError {
+    /// Model used when last message exceeded context window
+    model_id: LanguageModelId,
+    /// Token count including last message
+    token_count: usize,
+}
+
 impl Thread {
     pub fn new(
         project: Entity<Project>,
@@ -301,6 +311,7 @@ impl Thread {
                     .shared()
             },
             cumulative_token_usage: TokenUsage::default(),
+            exceeded_window_error: None,
             feedback: None,
             message_feedback: HashMap::default(),
             last_auto_capture_at: None,
@@ -367,6 +378,7 @@ impl Thread {
             action_log: cx.new(|_| ActionLog::new(project)),
             initial_project_snapshot: Task::ready(serialized.initial_project_snapshot).shared(),
             cumulative_token_usage: serialized.cumulative_token_usage,
+            exceeded_window_error: None,
             feedback: None,
             message_feedback: HashMap::default(),
             last_auto_capture_at: None,
@@ -817,6 +829,7 @@ impl Thread {
                 initial_project_snapshot,
                 cumulative_token_usage: this.cumulative_token_usage.clone(),
                 detailed_summary_state: this.detailed_summary_state.clone(),
+                exceeded_window_error: this.exceeded_window_error.clone(),
             })
         })
     }
@@ -1129,6 +1142,20 @@ impl Thread {
                                 cx.emit(ThreadEvent::ShowError(
                                     ThreadError::MaxMonthlySpendReached,
                                 ));
+                            } else if let Some(known_error) =
+                                error.downcast_ref::<LanguageModelKnownError>()
+                            {
+                                match known_error {
+                                    LanguageModelKnownError::ContextWindowLimitExceeded {
+                                        tokens,
+                                    } => {
+                                        thread.exceeded_window_error = Some(ExceededWindowError {
+                                            model_id: model.id(),
+                                            token_count: *tokens,
+                                        });
+                                        cx.notify();
+                                    }
+                                }
                             } else {
                                 let error_message = error
                                     .chain()
@@ -1784,10 +1811,6 @@ impl Thread {
         &self.project
     }
 
-    pub fn cumulative_token_usage(&self) -> TokenUsage {
-        self.cumulative_token_usage.clone()
-    }
-
     pub fn auto_capture_telemetry(&mut self, cx: &mut Context<Self>) {
         if !cx.has_flag::<feature_flags::ThreadAutoCapture>() {
             return;
@@ -1840,6 +1863,16 @@ impl Thread {
 
         let max = model.model.max_token_count();
 
+        if let Some(exceeded_error) = &self.exceeded_window_error {
+            if model.model.id() == exceeded_error.model_id {
+                return TotalTokenUsage {
+                    total: exceeded_error.token_count,
+                    max,
+                    ratio: TokenUsageRatio::Exceeded,
+                };
+            }
+        }
+
         #[cfg(debug_assertions)]
         let warning_threshold: f32 = std::env::var("ZED_THREAD_WARNING_THRESHOLD")
             .unwrap_or("0.8".to_string())

crates/agent/src/thread_store.rs

@@ -27,7 +27,9 @@ use serde::{Deserialize, Serialize};
 use settings::{Settings as _, SettingsStore};
 use util::ResultExt as _;
 
-use crate::thread::{DetailedSummaryState, MessageId, ProjectSnapshot, Thread, ThreadId};
+use crate::thread::{
+    DetailedSummaryState, ExceededWindowError, MessageId, ProjectSnapshot, Thread, ThreadId,
+};
 
 const RULES_FILE_NAMES: [&'static str; 6] = [
     ".rules",
@@ -491,6 +493,8 @@ pub struct SerializedThread {
     pub cumulative_token_usage: TokenUsage,
     #[serde(default)]
     pub detailed_summary_state: DetailedSummaryState,
+    #[serde(default)]
+    pub exceeded_window_error: Option<ExceededWindowError>,
 }
 
 impl SerializedThread {
@@ -577,6 +581,7 @@ impl LegacySerializedThread {
             initial_project_snapshot: self.initial_project_snapshot,
             cumulative_token_usage: TokenUsage::default(),
             detailed_summary_state: DetailedSummaryState::default(),
+            exceeded_window_error: None,
         }
     }
 }

crates/anthropic/src/anthropic.rs

@@ -724,4 +724,54 @@ impl ApiError {
     pub fn is_rate_limit_error(&self) -> bool {
         matches!(self.error_type.as_str(), "rate_limit_error")
     }
+
+    pub fn match_window_exceeded(&self) -> Option<usize> {
+        let Some(ApiErrorCode::InvalidRequestError) = self.code() else {
+            return None;
+        };
+
+        parse_prompt_too_long(&self.message)
+    }
+}
+
+pub fn parse_prompt_too_long(message: &str) -> Option<usize> {
+    message
+        .strip_prefix("prompt is too long: ")?
+        .split_once(" tokens")?
+        .0
+        .parse::<usize>()
+        .ok()
+}
+
+#[test]
+fn test_match_window_exceeded() {
+    let error = ApiError {
+        error_type: "invalid_request_error".to_string(),
+        message: "prompt is too long: 220000 tokens > 200000".to_string(),
+    };
+    assert_eq!(error.match_window_exceeded(), Some(220_000));
+
+    let error = ApiError {
+        error_type: "invalid_request_error".to_string(),
+        message: "prompt is too long: 1234953 tokens".to_string(),
+    };
+    assert_eq!(error.match_window_exceeded(), Some(1234953));
+
+    let error = ApiError {
+        error_type: "invalid_request_error".to_string(),
+        message: "not a prompt length error".to_string(),
+    };
+    assert_eq!(error.match_window_exceeded(), None);
+
+    let error = ApiError {
+        error_type: "rate_limit_error".to_string(),
+        message: "prompt is too long: 12345 tokens".to_string(),
+    };
+    assert_eq!(error.match_window_exceeded(), None);
+
+    let error = ApiError {
+        error_type: "invalid_request_error".to_string(),
+        message: "prompt is too long: invalid tokens".to_string(),
+    };
+    assert_eq!(error.match_window_exceeded(), None);
 }

crates/language_model/src/language_model.rs

@@ -278,6 +278,12 @@ pub trait LanguageModel: Send + Sync {
     }
 }
 
+#[derive(Debug, Error)]
+pub enum LanguageModelKnownError {
+    #[error("Context window limit exceeded ({tokens})")]
+    ContextWindowLimitExceeded { tokens: usize },
+}
+
 pub trait LanguageModelTool: 'static + DeserializeOwned + JsonSchema {
     fn name() -> String;
     fn description() -> String;
@@ -347,7 +353,7 @@ pub trait LanguageModelProviderState: 'static {
     }
 }
 
-#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
+#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd, Serialize, Deserialize)]
 pub struct LanguageModelId(pub SharedString);
 
 #[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]

crates/language_models/Cargo.toml

@@ -47,6 +47,7 @@ settings.workspace = true
 smol.workspace = true
 strum.workspace = true
 theme.workspace = true
+thiserror.workspace = true
 tiktoken-rs.workspace = true
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 ui.workspace = true

crates/language_models/src/provider/anthropic.rs

@@ -13,8 +13,9 @@ use gpui::{
 use http_client::HttpClient;
 use language_model::{
     AuthenticateError, LanguageModel, LanguageModelCacheConfiguration, LanguageModelId,
-    LanguageModelName, LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
-    LanguageModelProviderState, LanguageModelRequest, MessageContent, RateLimiter, Role,
+    LanguageModelKnownError, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
+    LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest, MessageContent,
+    RateLimiter, Role,
 };
 use language_model::{LanguageModelCompletionEvent, LanguageModelToolUse, StopReason};
 use schemars::JsonSchema;
@@ -454,7 +455,12 @@ impl LanguageModel for AnthropicModel {
         );
         let request = self.stream_completion(request, cx);
         let future = self.request_limiter.stream(async move {
-            let response = request.await.map_err(|err| anyhow!(err))?;
+            let response = request
+                .await
+                .map_err(|err| match err.downcast::<AnthropicError>() {
+                    Ok(anthropic_err) => anthropic_err_to_anyhow(anthropic_err),
+                    Err(err) => anyhow!(err),
+                })?;
             Ok(map_to_language_model_completion_events(response))
         });
         async move { Ok(future.await?.boxed()) }.boxed()
@@ -746,7 +752,7 @@ pub fn map_to_language_model_completion_events(
                         _ => {}
                     },
                     Err(err) => {
-                        return Some((vec![Err(anyhow!(err))], state));
+                        return Some((vec![Err(anthropic_err_to_anyhow(err))], state));
                     }
                 }
             }
@@ -757,6 +763,16 @@ pub fn map_to_language_model_completion_events(
     .flat_map(futures::stream::iter)
 }
 
+pub fn anthropic_err_to_anyhow(err: AnthropicError) -> anyhow::Error {
+    if let AnthropicError::ApiError(api_err) = &err {
+        if let Some(tokens) = api_err.match_window_exceeded() {
+            return anyhow!(LanguageModelKnownError::ContextWindowLimitExceeded { tokens });
+        }
+    }
+
+    anyhow!(err)
+}
+
 /// Updates usage data by preferring counts from `new`.
 fn update_usage(usage: &mut Usage, new: &Usage) {
     if let Some(input_tokens) = new.input_tokens {

crates/language_models/src/provider/cloud.rs

@@ -1,4 +1,4 @@
-use anthropic::{AnthropicError, AnthropicModelMode};
+use anthropic::{AnthropicError, AnthropicModelMode, parse_prompt_too_long};
 use anyhow::{Result, anyhow};
 use client::{
     Client, EXPIRED_LLM_TOKEN_HEADER_NAME, MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
@@ -14,7 +14,7 @@ use gpui::{AnyElement, AnyView, App, AsyncApp, Context, Entity, Subscription, Ta
 use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
 use language_model::{
     AuthenticateError, CloudModel, LanguageModel, LanguageModelCacheConfiguration, LanguageModelId,
-    LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
+    LanguageModelKnownError, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
     LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest,
     LanguageModelToolSchemaFormat, RateLimiter, ZED_CLOUD_PROVIDER_ID,
 };
@@ -33,6 +33,7 @@ use std::{
     time::Duration,
 };
 use strum::IntoEnumIterator;
+use thiserror::Error;
 use ui::{TintColor, prelude::*};
 
 use crate::AllLanguageModelSettings;
@@ -575,14 +576,19 @@ impl CloudLanguageModel {
             } else {
                 let mut body = String::new();
                 response.body_mut().read_to_string(&mut body).await?;
-                return Err(anyhow!(
-                    "cloud language model completion failed with status {status}: {body}",
-                ));
+                return Err(anyhow!(ApiError { status, body }));
             }
         }
     }
 }
 
+#[derive(Debug, Error)]
+#[error("cloud language model completion failed with status {status}: {body}")]
+struct ApiError {
+    status: StatusCode,
+    body: String,
+}
+
 impl LanguageModel for CloudLanguageModel {
     fn id(&self) -> LanguageModelId {
         self.id.clone()
@@ -696,7 +702,23 @@ impl LanguageModel for CloudLanguageModel {
                             )?)?,
                         },
                     )
-                    .await?;
+                    .await
+                    .map_err(|err| match err.downcast::<ApiError>() {
+                        Ok(api_err) => {
+                            if api_err.status == StatusCode::BAD_REQUEST {
+                                if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
+                                    return anyhow!(
+                                        LanguageModelKnownError::ContextWindowLimitExceeded {
+                                            tokens
+                                        }
+                                    );
+                                }
+                            }
+                            anyhow!(api_err)
+                        }
+                        Err(err) => anyhow!(err),
+                    })?;
+
                     Ok(
                         crate::provider::anthropic::map_to_language_model_completion_events(
                             Box::pin(response_lines(response).map_err(AnthropicError::Other)),