Skip to content

7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.

TypeScript

fix: fail over on 410 Gone and un-enumerated upstream statuses (#337, #339) (#346)

Tashfeen5cc48f6d1a56
diff --git a/README.md b/README.md
index b78923d6..a410dcc3 100644
--- a/README.md
+++ b/README.md
@@ -564,6 +564,8 @@ PRs should include a test, keep the existing test suite green, and match the `.e
<a href="https://github.com/dashitongzhi"><img src="https://images.weserv.nl/?url=github.com/dashitongzhi.png&w=60&h=60&fit=cover&mask=circle" width="60" alt="@dashitongzhi" /></a>
<a href="https://github.com/QingJ01"><img src="https://images.weserv.nl/?url=github.com/QingJ01.png&w=60&h=60&fit=cover&mask=circle" width="60" alt="@QingJ01" /></a>
<a href="https://github.com/3215"><img src="https://images.weserv.nl/?url=github.com/3215.png&w=60&h=60&fit=cover&mask=circle" width="60" alt="@3215" /></a>
+<a href="https://github.com/saifulaiub123"><img src="https://images.weserv.nl/?url=github.com/saifulaiub123.png&w=60&h=60&fit=cover&mask=circle" width="60" alt="@saifulaiub123" /></a>
+<a href="https://github.com/PietFourie"><img src="https://images.weserv.nl/?url=github.com/PietFourie.png&w=60&h=60&fit=cover&mask=circle" width="60" alt="@PietFourie" /></a>
## Terms of Service review
diff --git a/server/src/__tests__/routes/proxy-retry.test.ts b/server/src/__tests__/routes/proxy-retry.test.ts
index cbda67ea..a1825363 100644
--- a/server/src/__tests__/routes/proxy-retry.test.ts
+++ b/server/src/__tests__/routes/proxy-retry.test.ts
@@ -26,6 +26,11 @@ describe('isModelNotFoundError (drives whole-model skip within a request)', () =
expect(isModelNotFoundError(new Error('No endpoints found for openrouter/minimax/minimax-m2.5:free'))).toBe(true);
});
+ it('flags 410 Gone (model pulled upstream) by message or attached status — #339', () => {
+ expect(isModelNotFoundError(new Error('Ollama Cloud API error 410: Gone'))).toBe(true);
+ expect(isModelNotFoundError(Object.assign(new Error('Gone'), { status: 410 }))).toBe(true);
+ });
+
it('does not flag rate limits, 5xx, or payment errors', () => {
expect(isModelNotFoundError(new Error('429 Too Many Requests'))).toBe(false);
expect(isModelNotFoundError(new Error('503 Service Unavailable'))).toBe(false);
@@ -120,6 +125,37 @@ describe('isRetryableError', () => {
});
});
+ describe('410 Gone & un-enumerated upstream statuses fail over instead of 502 (#337/#339)', () => {
+ // The headline bug: a provider error whose HTTP status the substring allowlist
+ // never enumerated (410 Gone, 502, 504, 408 …) used to abort the whole chain
+ // with a 502 — stranding the healthy paid routes still queued later in the
+ // fallback order. It must rotate to the next route instead.
+ it('treats an Ollama "410: Gone" as retryable, by message and by attached status', () => {
+ // openai-compat throws via providerHttpError, so the real error carries both.
+ expect(isRetryableError(new Error('Ollama Cloud API error 410: Gone'))).toBe(true);
+ expect(isRetryableError(Object.assign(new Error('Ollama Cloud API error 410: Gone'), { status: 410 }))).toBe(true);
+ });
+
+ it('fails over on any 5xx the substring rules never listed, via the structured status', () => {
+ // No '502'/'504'/'507' substring rule exists; the err.status catch-all covers them.
+ expect(isRetryableError(Object.assign(new Error('Bad Gateway'), { status: 502 }))).toBe(true);
+ expect(isRetryableError(Object.assign(new Error('Gateway Timeout'), { status: 504 }))).toBe(true);
+ expect(isRetryableError(Object.assign(new Error('Insufficient Storage'), { status: 507 }))).toBe(true);
+ });
+
+ it('fails over on 408 request-timeout / 409 conflict by status', () => {
+ expect(isRetryableError(Object.assign(new Error('Request Timeout'), { status: 408 }))).toBe(true);
+ expect(isRetryableError(Object.assign(new Error('Conflict'), { status: 409 }))).toBe(true);
+ });
+
+ it('still treats genuinely-fatal 400/401 as NON-retryable even with an attached status', () => {
+ // The structured catch-all must not swallow client-fatal errors — they fail on
+ // every provider identically, so aborting the request is the correct behavior.
+ expect(isRetryableError(Object.assign(new Error('Bad Request'), { status: 400 }))).toBe(false);
+ expect(isRetryableError(Object.assign(new Error('Unauthorized'), { status: 401 }))).toBe(false);
+ });
+ });
+
describe('existing categories still classify correctly', () => {
it('429 / rate limits are retryable', () => {
expect(isRetryableError(new Error('429 Too Many Requests'))).toBe(true);
diff --git a/server/src/lib/error-classify.ts b/server/src/lib/error-classify.ts
index fa9c9eb9..e1664851 100644
--- a/server/src/lib/error-classify.ts
+++ b/server/src/lib/error-classify.ts
@@ -5,6 +5,20 @@
export function isRetryableError(err: any): boolean {
const msg = (err.message ?? '').toLowerCase();
+ // Trust the upstream HTTP status the provider attached to the error first
+ // (providerHttpError in providers/base.ts sets err.status on every adapter).
+ // This structured check is the robust primary signal; the message-substring
+ // rules below are the fallback for errors that carry a code in their text but
+ // no numeric status. It's the fix for #337/#339: an Ollama "410 Gone", or any
+ // upstream 5xx the substring allowlist never enumerated (502/504/507…), used to
+ // fall through to a 502 and STRAND the healthy paid routes still queued later in
+ // the chain — because the old code matched specific substrings and ignored
+ // err.status for every code except 403. 408 (request timeout), 409 (conflict),
+ // 410 (model pulled upstream), 429 (rate limit) and all 5xx are transient or
+ // fail-over-able; 400/401 stay fatal (status 0 here, handled by the absence of a
+ // matching rule) and 403 is handled by isModelAccessForbiddenError below.
+ const status = typeof err?.status === 'number' ? err.status : 0;
+ if (status === 408 || status === 409 || status === 410 || status === 429 || status >= 500) return true;
return msg.includes('429') || msg.includes('rate limit') || msg.includes('too many requests')
|| msg.includes('quota') || msg.includes('resource_exhausted')
|| msg.includes('aborted') || msg.includes('timeout') || msg.includes('etimedout')
@@ -20,6 +34,12 @@ export function isRetryableError(err: any): boolean {
// for a model that's been pulled). Rotate to the next model in the chain —
// setCooldown + the health checker will avoid this model on subsequent requests.
|| msg.includes('404') || msg.includes('not found') || msg.includes('no endpoints found')
+ // 410: the model/endpoint was permanently removed upstream (e.g. Ollama Cloud
+ // "API error 410: Gone", #339). Like a 404 it won't return on this provider, so
+ // rotate to the next route; isModelNotFoundError benches the whole model. The
+ // structured status check above already catches the 410 when the provider
+ // attaches err.status — this is the text fallback for errors that don't.
+ || msg.includes('410') || msg.includes('gone')
// 403: the key is valid (it passed validateKey, and the health checker
// disables truly-forbidden keys) but this specific model is off-limits to
// the key's tier — e.g. gpt-4o on GitHub Models' free tier, subscription-only
@@ -64,8 +84,14 @@ export function isPaymentRequiredError(err: any): boolean {
// burning one fallback attempt per key on the same dead route.
// (PR #111, credits @barbotkonv.)
export function isModelNotFoundError(err: any): boolean {
- const msg = (err.message ?? '').toLowerCase();
- return msg.includes('404') || msg.includes('not found') || msg.includes('no endpoints found');
+ // 404 (removed/deprecated) and 410 (permanently Gone) are both MODEL-level: every
+ // key for the platform fails the same way, so skip the whole model for the rest
+ // of the request instead of burning one fallback attempt per sibling key. 410
+ // added for #339 (Ollama Cloud "Gone"); prefer the structured status when present.
+ if (err?.status === 404 || err?.status === 410) return true;
+ const msg = (err?.message ?? '').toLowerCase();
+ return msg.includes('404') || msg.includes('not found') || msg.includes('no endpoints found')
+ || msg.includes('410') || msg.includes('gone');
}
// A 403 Forbidden returned for a specific model behind an otherwise-valid key.