fix(llm, openrouter): Report mid-stream provider errors

crates/jp_llm/src/provider/openrouter.rs

@@ -498,12 +498,27 @@ fn map_completion(
         "Received event from OpenRouter API."
     );
 
+    // A failure after the response is committed arrives as a top-level `error`
+    // on an otherwise ordinary chunk. The accompanying `choices` array may be
+    // empty, in which case this is the only place the failure is reported.
+    if let Some(error) = v.error {
+        return vec![Err(map_error(error))];
+    }
+
     v.choices
         .into_iter()
         .flat_map(|v| map_event(v, state))
         .collect()
 }
 
+/// Classify an OpenRouter error payload by its status code.
+fn map_error(error: response::ErrorResponse) -> StreamError {
+    StreamError::from(jp_openrouter::Error::Api {
+        code: error.code,
+        message: error.message,
+    })
+}
+
 #[expect(clippy::too_many_lines)]
 fn map_event(
     choice: types::response::Choice,
@@ -539,17 +554,7 @@ fn map_event(
     let reasoning_details = MultiProviderMetadata::from_details(reasoning_details);
 
     if let Some(error) = error {
-        if looks_like_quota_error(&error.message) {
-            return vec![Err(StreamError::new(
-                StreamErrorKind::InsufficientQuota,
-                format!(
-                    "Insufficient API quota. Check your credits \
-                     at https://openrouter.ai/settings/credits. ({})",
-                    error.message
-                ),
-            ))];
-        }
-        return vec![Err(StreamError::other(error.message))];
+        return vec![Err(map_error(error))];
     }
 
     let mut events = vec![];

crates/jp_llm/src/provider/openrouter_tests.rs

@@ -196,6 +196,108 @@ fn tool_call_finish_is_a_clean_completion() -> Result {
     Ok(())
 }
 
+/// A prompt rejected after the response is committed arrives as a 200 stream
+/// whose only chunk carries a top-level `error` and an empty `choices` array.
+/// The rejection must reach the caller instead of being dropped, which would
+/// leave the stream looking merely truncated.
+#[test]
+fn top_level_error_with_empty_choices_is_reported() -> Result {
+    let chunk: OpenRouterChunk = serde_json::from_value(serde_json::json!({
+        "id": "gen-1786537769-1XuUJY00pSj6s5GLfq9d",
+        "provider": "Azure",
+        "choices": [],
+        "created": 1_786_537_769,
+        "model": "unknown",
+        "object": "chat.completion.chunk",
+        "system_fingerprint": null,
+        "usage": null,
+        "error": {
+            "code": 400,
+            "message": "prompt is too long: 1284668 tokens > 1000000 maximum",
+            "metadata": null
+        }
+    }))?;
+    let mut state = AggregationState {
+        tool_call_indices: vec![],
+        aggregating_reasoning: false,
+        aggregating_message: false,
+        is_structured: false,
+    };
+
+    let events = map_completion(chunk, &mut state);
+
+    let [Err(error)] = events.as_slice() else {
+        panic!("expected exactly one error event, got {events:?}");
+    };
+    assert_eq!(
+        error.message(),
+        "API error (status 400): prompt is too long: 1284668 tokens > 1000000 maximum"
+    );
+    // A 400 is the provider rejecting this request as-is. Retrying resends the
+    // identical body, so it must not be classified as transient.
+    assert_eq!(error.kind, StreamErrorKind::Other);
+    assert!(!error.is_retryable());
+    Ok(())
+}
+
+/// A quota failure delivered as a top-level error is classified by its status
+/// code, so the caller can surface the credits hint rather than a bare message.
+#[test]
+fn top_level_payment_error_is_classified_as_quota() -> Result {
+    let chunk: OpenRouterChunk = serde_json::from_value(serde_json::json!({
+        "id": "gen-1",
+        "provider": "Azure",
+        "choices": [],
+        "created": 1_786_537_769,
+        "model": "unknown",
+        "object": "chat.completion.chunk",
+        "system_fingerprint": null,
+        "usage": null,
+        "error": { "code": 402, "message": "Payment required", "metadata": null }
+    }))?;
+    let mut state = AggregationState {
+        tool_call_indices: vec![],
+        aggregating_reasoning: false,
+        aggregating_message: false,
+        is_structured: false,
+    };
+
+    let events = map_completion(chunk, &mut state);
+
+    let [Err(error)] = events.as_slice() else {
+        panic!("expected exactly one error event, got {events:?}");
+    };
+    assert_eq!(error.kind, StreamErrorKind::InsufficientQuota);
+    Ok(())
+}
+
+/// A per-choice error is classified by status code on the same path as a
+/// top-level one: a 503 from the upstream provider is worth retrying.
+#[test]
+fn per_choice_error_is_classified_by_status_code() -> Result {
+    let choice: response::Choice = serde_json::from_value(serde_json::json!({
+        "finish_reason": "error",
+        "native_finish_reason": "error",
+        "delta": { "role": null, "content": "", "reasoning": null, "tool_calls": [] },
+        "error": { "code": 503, "message": "Provider disconnected unexpectedly", "metadata": null }
+    }))?;
+    let mut state = AggregationState {
+        tool_call_indices: vec![],
+        aggregating_reasoning: false,
+        aggregating_message: false,
+        is_structured: false,
+    };
+
+    let events = map_event(choice, &mut state);
+
+    let [Err(error)] = events.as_slice() else {
+        panic!("expected exactly one error event, got {events:?}");
+    };
+    assert_eq!(error.kind, StreamErrorKind::Transient);
+    assert!(error.is_retryable());
+    Ok(())
+}
+
 fn forced_tool_request(
     reasoning: ReasoningDetails,
     enable_reasoning: bool,

crates/jp_openrouter/src/types/response.rs

@@ -65,6 +65,15 @@ pub struct ChatCompletion {
 
     /// Usage statistics for the completion request.
     pub usage: Option<Usage>,
+
+    /// The error that terminated the response, if any.
+    ///
+    /// A request that fails once the response has been committed cannot change
+    /// its HTTP status, so the failure arrives as a regular chunk carrying this
+    /// field alongside a `choices` entry with `finish_reason: "error"` (or an
+    /// empty `choices` array).
+    /// Present only on such terminating chunks.
+    pub error: Option<ErrorResponse>,
 }
 
 #[derive(Debug, Clone, Deserialize)]

crates/jp_openrouter/tests/fixtures/test_chat_completion.snap

@@ -49,5 +49,6 @@ Ok(
                 total_tokens: 51,
             },
         ),
+        error: None,
     },
 )

crates/jp_openrouter/tests/fixtures/test_chat_completion_stream.snap

@@ -31,6 +31,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -60,6 +61,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -89,6 +91,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -118,6 +121,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -147,6 +151,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -176,6 +181,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -205,6 +211,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -234,6 +241,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -263,6 +271,7 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
     ChatCompletion {
         id: "gen-1771242469-Y6Pl7tMhHtBIFcYGon8E",
@@ -292,5 +301,6 @@ expression: v
         object: ChatCompletionChunk,
         system_fingerprint: None,
         usage: None,
+        error: None,
     },
 ]