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(providers): 60s chat timeout for zhipu, agnes, cloudflare (reasoning-model TTFB) (#521)

Tashfeene3fdd3bdbe98
diff --git a/server/src/__tests__/providers/reasoning-timeouts.test.ts b/server/src/__tests__/providers/reasoning-timeouts.test.ts
new file mode 100644
index 00000000..b2d7e8f7
--- /dev/null
+++ b/server/src/__tests__/providers/reasoning-timeouts.test.ts
@@ -0,0 +1,56 @@
+import { describe, it, expect, vi, afterEach } from 'vitest';
+import { resolveProvider } from '../../providers/index.js';
+import { CloudflareProvider } from '../../providers/cloudflare.js';
+
+// Regression guard for the 2026-07-11 live-sweep finding: platforms hosting
+// hidden-reasoning models (zhipu glm-4.7-flash 41s TTFB, agnes-2.0-flash 20s,
+// @cf/zai-org/glm-4.7-flash repeated 15s aborts) need a chat timeout above
+// the 15s fetch default or every attempt — streaming included — is aborted
+// before the first byte. The value is a private construction detail, so the
+// registry entries are asserted via the stored field and Cloudflare via the
+// setTimeout the abort rides on.
+
+describe('reasoning-model chat timeouts', () => {
+ afterEach(() => {
+ vi.restoreAllMocks();
+ });
+
+ it.each([
+ ['zhipu', 60_000],
+ ['agnes', 60_000],
+ ['ollama', 120_000], // pre-existing bump; keep it from regressing too
+ ] as const)('%s is registered with a %dms chat timeout', (platform, ms) => {
+ const provider = resolveProvider(platform);
+ expect(provider).toBeDefined();
+ expect((provider as unknown as { timeoutMs: number }).timeoutMs).toBe(ms);
+ });
+
+ it('cloudflare chat aborts on a 60s timer, not the 15s default', async () => {
+ const provider = new CloudflareProvider();
+ const delays: number[] = [];
+ const origSetTimeout = global.setTimeout;
+ vi.spyOn(global, 'setTimeout').mockImplementation(((fn: () => void, ms?: number) => {
+ delays.push(ms ?? 0);
+ return origSetTimeout(fn, ms);
+ }) as typeof setTimeout);
+ vi.spyOn(global, 'fetch').mockResolvedValue({
+ ok: true,
+ json: () => Promise.resolve({
+ id: 'chatcmpl-cf', object: 'chat.completion', created: 1,
+ model: '@cf/zai-org/glm-4.7-flash',
+ choices: [{ index: 0, message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
+ usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+ }),
+ headers: new Headers(),
+ } as unknown as Response);
+
+ await provider.chatCompletion(
+ 'acct:token',
+ [{ role: 'user', content: 'hi' }],
+ '@cf/zai-org/glm-4.7-flash',
+ );
+
+ expect(delays).toContain(60_000);
+ expect(delays).not.toContain(15_000);
+ });
+});
diff --git a/server/src/providers/cloudflare.ts b/server/src/providers/cloudflare.ts
index 7f940bf0..e5515df3 100644
--- a/server/src/providers/cloudflare.ts
+++ b/server/src/providers/cloudflare.ts
@@ -13,6 +13,12 @@ import { recordQuotaObservationsFromResponse, type QuotaObservationContext } fro
* API key format expected: "account_id:api_token"
* The account_id is extracted from the key to build the URL.
*/
+
+// Workers AI hosts reasoning models (@cf/zai-org/glm-4.7-flash) that spend
+// well over the 15s fetch default before the first byte (live sweep
+// 2026-07-11: repeated 15s aborts). Matches zhipu/agnes/ollama bumps.
+const CHAT_TIMEOUT_MS = 60_000;
+
export class CloudflareProvider extends BaseProvider {
readonly platform = 'cloudflare' as const;
readonly name = 'Cloudflare Workers AI';
@@ -59,7 +65,7 @@ export class CloudflareProvider extends BaseProvider {
parallel_tool_calls: options?.parallel_tool_calls,
...extendedBodyParams(this.platform, options),
}),
- });
+ }, CHAT_TIMEOUT_MS);
recordQuotaObservationsFromResponse(res, {
platform: this.platform,
keyId: quotaContext?.keyId,
@@ -108,7 +114,7 @@ export class CloudflareProvider extends BaseProvider {
...extendedBodyParams(this.platform, options),
stream: true,
}),
- });
+ }, CHAT_TIMEOUT_MS);
recordQuotaObservationsFromResponse(res, {
platform: this.platform,
keyId: quotaContext?.keyId,
diff --git a/server/src/providers/index.ts b/server/src/providers/index.ts
index dfd74698..80bd2a6b 100644
--- a/server/src/providers/index.ts
+++ b/server/src/providers/index.ts
@@ -83,10 +83,17 @@ register(new CohereProvider());
register(new CloudflareProvider());
// Zhipu (Z.ai / bigmodel.cn) - OpenAI-compatible
+//
+// glm-4.7-flash is a hidden-reasoning model: it burns through a long
+// reasoning_content before the first answer byte (live-probed 41s TTFB on a
+// one-word completion, 2026-07-11), and Zhipu buffers that phase even when
+// streaming — so the default 15s timeout aborted every attempt. 60s covers
+// the observed worst case with headroom.
register(new OpenAICompatProvider({
platform: 'zhipu',
name: 'Zhipu AI',
baseUrl: 'https://open.bigmodel.cn/api/paas/v4',
+ timeoutMs: 60_000,
}));
// Hugging Face Inference Providers router — re-added in V13. The V4 removal
@@ -200,10 +207,14 @@ register(new OpenAICompatProvider({
// before 429s (no documented RPM/RPD). Free key from platform.agnes-ai.com,
// no card. Catalog rows live in the catalog (premium → age into free); not
// shipped as freeapi model migrations.
+// agnes-2.0-flash reasons before answering (live-probed 20s TTFB on a
+// one-word completion, 2026-07-11), so the default 15s timeout aborted it;
+// 60s matches the other reasoning-hosting platforms.
register(new OpenAICompatProvider({
platform: 'agnes',
name: 'Agnes AI',
baseUrl: 'https://apihub.agnes-ai.com/v1',
+ timeoutMs: 60_000,
}));
// Chutes was evaluated for V11 and dropped: probe with a free-tier key