fix(coding-agent): retry explicit provider retry errors

closes #6019
2026-06-24 18:12:25 +02:00
parent c29bbc0958
commit 371adcf371
7 changed files with 178 additions and 18 deletions
@@ -0,0 +1,96 @@
+import type { AssistantMessage } from "../types.ts";
+
+function buildProviderErrorPattern(patterns: readonly string[]): RegExp {
+	return new RegExp(patterns.join("|"), "i");
+}
+
+const NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([
+	// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's
+	// Zen API. These are subscription/account limits, not transient throttles.
+	"GoUsageLimitError",
+	"FreeUsageLimitError",
+
+	// OpenCode Go subscription-limit text asks users to enable available-balance
+	// usage after rolling/weekly/monthly limits are reached.
+	"Monthly usage limit reached",
+	"available balance",
+
+	// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's
+	// quota/billing error code; the other strings cover common gateway wording.
+	"insufficient_quota",
+	"out of budget",
+	"quota exceeded",
+	"billing",
+]);
+
+const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
+	// Generic provider load, HTTP status, and server-side transient failures.
+	"overloaded",
+	"rate.?limit",
+	"too many requests",
+	"429",
+	"500",
+	"502",
+	"503",
+	"504",
+	"service.?unavailable",
+	"server.?error",
+	"internal.?error",
+
+	// Wrapper/provider text for transient upstream failures, including OpenRouter
+	// "Provider returned error" responses (#2264).
+	"provider.?returned.?error",
+
+	// Network, proxy, and fetch transport failures. This includes OpenAI Codex
+	// raw-fetch failures such as "upstream connect", "connection refused", and
+	// "reset before headers" (#733), plus OpenRouter connection drops (#3317).
+	"network.?error",
+	"connection.?error",
+	"connection.?refused",
+	"connection.?lost",
+	"other side closed",
+	"fetch failed",
+	"upstream.?connect",
+	"reset before headers",
+	"socket hang up",
+	"timed? out",
+	"timeout",
+	"terminated",
+
+	// WebSocket transports can report close/error text instead of HTTP/fetch text.
+	"websocket.?closed",
+	"websocket.?error",
+
+	// Premature stream endings from SDKs and transports. Anthropic can throw
+	// "stream ended without ..." and "Anthropic stream ended before message_stop"
+	// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).
+	"ended without",
+	"stream ended before message_stop",
+	"http2 request did not get a response",
+
+	// Provider-requested retry delay cap failures should flow through the outer
+	// retry policy so callers can surface/abort the backoff (#1123).
+	"retry delay",
+
+	// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock
+	// stream exceptions (#6019).
+	"you can retry your request",
+	"try your request again",
+	"please retry your request",
+]);
+
+/**
+ * Classifies whether a failed assistant message looks like a transient provider
+ * or transport error, so callers can decide if the last assistant turn should be
+ * restarted.
+ *
+ * This does not implement retry policy. Callers should first handle context
+ * overflow separately, then apply their own retry budget, backoff, and reporting
+ * before restarting the assistant turn.
+ */
+export function isRetryableAssistantError(message: AssistantMessage): boolean {
+	if (message.stopReason !== "error" || !message.errorMessage) return false;
+	const errorMessage = message.errorMessage;
+	if (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;
+	return RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);
+}