From 97aeb5631dbf8d897628ff50562ff682bb57e6a8 Mon Sep 17 00:00:00 2001
From: Bryan Helmkamp <bryan@brynary.com>
Date: Sat, 1 Aug 2026 09:26:47 -0400
Subject: [PATCH] refactor(llm): tighten provider error classification

---
 .../src/codec/anthropic_messages/stream.rs    |  4 +--
 .../src/codec/openai_responses/stream.rs      |  4 +--
 lib/components/fabro-llm/src/error.rs         | 13 +------
 .../fabro-llm/src/providers/openai.rs         | 34 +++++++++++++++++++
 4 files changed, 39 insertions(+), 16 deletions(-)

diff --git a/lib/components/fabro-llm/src/codec/anthropic_messages/stream.rs b/lib/components/fabro-llm/src/codec/anthropic_messages/stream.rs
index 15e1e71cd..bb8d79027 100644
--- a/lib/components/fabro-llm/src/codec/anthropic_messages/stream.rs
+++ b/lib/components/fabro-llm/src/codec/anthropic_messages/stream.rs
@@ -8,7 +8,7 @@
 use super::SYNTHETIC_TOOL_NAME;
 use super::decode::{convert_synthetic_tool_to_text, map_finish_reason, refusal_error};
 use crate::codec::{RawEvent, StreamDecoder, parse_tool_arguments_or_empty};
-use crate::error::{Error, ProviderErrorDetail, ProviderErrorKind, kind_from_error_code};
+use crate::error::{self, Error, ProviderErrorDetail, ProviderErrorKind};
 use crate::types::{
     ContentPart, FinishReason, Message, RateLimitInfo, Response, Role, StreamEvent, ThinkingData,
     TokenCounts, ToolCall,
@@ -358,7 +358,7 @@ fn stream_error_event_to_provider_error(data: &serde_json::Value, provider_name:
     // overloaded_error, api_error, and unknown stream errors are transient.
     let kind = error_code
         .as_deref()
-        .and_then(kind_from_error_code)
+        .and_then(error::kind_from_error_code)
         .unwrap_or(ProviderErrorKind::Server);
 
     Error::Provider {
diff --git a/lib/components/fabro-llm/src/codec/openai_responses/stream.rs b/lib/components/fabro-llm/src/codec/openai_responses/stream.rs
index 643369688..9bca719ec 100644
--- a/lib/components/fabro-llm/src/codec/openai_responses/stream.rs
+++ b/lib/components/fabro-llm/src/codec/openai_responses/stream.rs
@@ -11,7 +11,7 @@ use serde::Deserialize;
 use super::decode::{map_finish_reason, token_counts_from_api_usage, tool_call_from_item};
 use super::wire::ApiUsage;
 use crate::codec::{CodecCtx, RawEvent, StreamDecoder};
-use crate::error::{Error, ProviderErrorDetail, ProviderErrorKind, kind_from_error_code};
+use crate::error::{self, Error, ProviderErrorDetail, ProviderErrorKind};
 use crate::types::{
     ContentPart, FinishReason, Message, RateLimitInfo, Response, Role, StreamEvent, TokenCounts,
     ToolCall,
@@ -38,7 +38,7 @@ fn provider_error_from_openai_error_json(error: &serde_json::Value, provider: &s
 
     // Unrecognized and absent codes are treated as transient.
     let kind = classifier
-        .and_then(kind_from_error_code)
+        .and_then(error::kind_from_error_code)
         .unwrap_or(ProviderErrorKind::Server);
 
     Error::Provider {
diff --git a/lib/components/fabro-llm/src/error.rs b/lib/components/fabro-llm/src/error.rs
index 7897eebe7..6ec6d7886 100644
--- a/lib/components/fabro-llm/src/error.rs
+++ b/lib/components/fabro-llm/src/error.rs
@@ -295,7 +295,7 @@ impl Error {
 /// `insufficient_quota` means the same thing whether it arrives in an HTTP
 /// error body or in a mid-stream error event.
 #[must_use]
-pub fn kind_from_error_code(code: &str) -> Option<ProviderErrorKind> {
+pub(crate) fn kind_from_error_code(code: &str) -> Option<ProviderErrorKind> {
     Some(match code {
         // Out of credit, or over a billing cap. Distinct from RateLimit:
         // backoff never clears it, but another provider has its own quota.
@@ -726,17 +726,6 @@ mod tests {
             None,
         );
         assert_eq!(err.provider_kind(), Some(ProviderErrorKind::QuotaExceeded));
-
-        // No code, so the message fallback still runs.
-        let err = error_from_status_code(
-            400,
-            "This model's maximum context length is 4096 tokens".into(),
-            "openai".into(),
-            None,
-            None,
-            None,
-        );
-        assert_eq!(err.provider_kind(), Some(ProviderErrorKind::ContextLength));
     }
 
     #[test]
diff --git a/lib/components/fabro-llm/src/providers/openai.rs b/lib/components/fabro-llm/src/providers/openai.rs
index 56eb76c3d..b56159c2c 100644
--- a/lib/components/fabro-llm/src/providers/openai.rs
+++ b/lib/components/fabro-llm/src/providers/openai.rs
@@ -520,6 +520,40 @@ mod tests {
         assert!(matches!(err, Error::Configuration { .. }));
     }
 
+    #[tokio::test]
+    async fn complete_classifies_insufficient_quota_as_quota_exceeded() {
+        let server = MockServer::start();
+        let mock = server.mock(|when, then| {
+            when.method(POST).path("/responses");
+            then.status(429)
+                .header("content-type", "application/json")
+                .json_body(serde_json::json!({
+                    "error": {
+                        "message": "You exceeded your current quota.",
+                        "type": "insufficient_quota"
+                    }
+                }));
+        });
+        let adapter = Adapter::new("sk-test").with_base_url(server.base_url());
+
+        let err = adapter
+            .complete(&minimal_request())
+            .await
+            .expect_err("spent quota should fail the completion");
+
+        mock.assert();
+        assert_eq!(err.provider_kind(), Some(ProviderErrorKind::QuotaExceeded));
+        assert_eq!(err.status_code(), Some(429));
+        assert!(!err.retryable());
+        assert!(err.failover_eligible());
+        match err {
+            Error::Provider { detail, .. } => {
+                assert_eq!(detail.error_code.as_deref(), Some("insufficient_quota"));
+            }
+            other => panic!("expected provider error, got {other:?}"),
+        }
+    }
+
     #[tokio::test]
     async fn codex_complete_via_stream_propagates_stream_errors() {
         let server = MockServer::start();