fix(providers): pin transport policy and lift the 60s cap on Groq and Cerebras

waleedlatif1 · waleedlatif1 · commit bbc5853dc13b · 2026-08-05T19:58:25.000-07:00
diff --git a/apps/sim/providers/baseten/index.ts b/apps/sim/providers/baseten/index.ts
@@ -19,6 +19,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   FunctionCallResponse,
   Message,
@@ -86,6 +87,7 @@ export const basetenProvider: ProviderConfig = {
     }
 
     const client = new OpenAI({
+      ...openAICompatTransport(),
       apiKey: request.apiKey,
       baseURL: 'https://inference.baseten.co/v1',
     })
diff --git a/apps/sim/providers/cerebras/index.ts b/apps/sim/providers/cerebras/index.ts
@@ -15,6 +15,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -54,6 +55,7 @@ export const cerebrasProvider: ProviderConfig = {
     try {
       const client = new Cerebras({
         apiKey: request.apiKey,
+        ...openAICompatTransport(),
       })
 
       const allMessages = []
diff --git a/apps/sim/providers/deepseek/index.ts b/apps/sim/providers/deepseek/index.ts
@@ -13,6 +13,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -50,6 +51,7 @@ export const deepseekProvider: ProviderConfig = {
 
     try {
       const deepseek = new OpenAI({
+        ...openAICompatTransport(),
         apiKey: request.apiKey,
         baseURL: 'https://api.deepseek.com',
       })
diff --git a/apps/sim/providers/fireworks/index.ts b/apps/sim/providers/fireworks/index.ts
@@ -20,6 +20,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   FunctionCallResponse,
   Message,
@@ -87,6 +88,7 @@ export const fireworksProvider: ProviderConfig = {
     }
 
     const client = new OpenAI({
+      ...openAICompatTransport(),
       apiKey: request.apiKey,
       baseURL: 'https://api.fireworks.ai/inference/v1',
     })
diff --git a/apps/sim/providers/groq/index.ts b/apps/sim/providers/groq/index.ts
@@ -18,6 +18,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -50,7 +51,7 @@ export const groqProvider: ProviderConfig = {
       throw new Error('API key is required for Groq')
     }
 
-    const groq = new Groq({ apiKey: request.apiKey })
+    const groq = new Groq({ apiKey: request.apiKey, ...openAICompatTransport() })
 
     const allMessages = []
 
diff --git a/apps/sim/providers/kimi/index.ts b/apps/sim/providers/kimi/index.ts
@@ -19,6 +19,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -100,6 +101,7 @@ export const kimiProvider: ProviderConfig = {
 
     try {
       const kimi = new OpenAI({
+        ...openAICompatTransport(),
         apiKey: request.apiKey,
         baseURL: KIMI_BASE_URL,
       })
diff --git a/apps/sim/providers/litellm/index.ts b/apps/sim/providers/litellm/index.ts
@@ -16,6 +16,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   Message,
   ProviderConfig,
@@ -110,6 +111,7 @@ export const litellmProvider: ProviderConfig = {
 
     const apiKey = request.apiKey || env.LITELLM_API_KEY || 'empty'
     const litellm = new OpenAI({
+      ...openAICompatTransport(),
       apiKey,
       baseURL: `${baseUrl}/v1`,
     })
diff --git a/apps/sim/providers/meta/index.ts b/apps/sim/providers/meta/index.ts
@@ -14,6 +14,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -53,6 +54,7 @@ export const metaProvider: ProviderConfig = {
 
     try {
       const meta = new OpenAI({
+        ...openAICompatTransport(),
         apiKey: request.apiKey,
         baseURL: META_BASE_URL,
       })
diff --git a/apps/sim/providers/mistral/index.ts b/apps/sim/providers/mistral/index.ts
@@ -14,6 +14,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -61,6 +62,7 @@ export const mistralProvider: ProviderConfig = {
     }
 
     const mistral = new OpenAI({
+      ...openAICompatTransport(),
       apiKey: request.apiKey,
       baseURL: 'https://api.mistral.ai/v1',
     })
diff --git a/apps/sim/providers/nvidia/index.ts b/apps/sim/providers/nvidia/index.ts
@@ -15,6 +15,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -60,6 +61,7 @@ export const nvidiaProvider: ProviderConfig = {
 
     try {
       const nvidia = new OpenAI({
+        ...openAICompatTransport(),
         apiKey: request.apiKey,
         baseURL: NVIDIA_BASE_URL,
       })
diff --git a/apps/sim/providers/ollama-cloud/index.ts b/apps/sim/providers/ollama-cloud/index.ts
@@ -4,6 +4,7 @@ import type { StreamingExecution } from '@/executor/types'
 import { getProviderDefaultModel, getProviderModels } from '@/providers/models'
 import { executeOllamaProviderRequest } from '@/providers/ollama/core'
 import { createReadableStreamFromOllamaCloudStream } from '@/providers/ollama-cloud/utils'
+import { openAICompatTransport } from '@/providers/transport'
 import type { ProviderConfig, ProviderRequest, ProviderResponse } from '@/providers/types'
 
 const logger = createLogger('OllamaCloudProvider')
@@ -36,6 +37,7 @@ export const ollamaCloudProvider: ProviderConfig = {
         providerLabel: 'Ollama Cloud',
         createClient: () =>
           new OpenAI({
+            ...openAICompatTransport(),
             apiKey,
             baseURL: OLLAMA_CLOUD_BASE_URL,
           }),
diff --git a/apps/sim/providers/ollama/index.ts b/apps/sim/providers/ollama/index.ts
@@ -6,6 +6,7 @@ import type { StreamingExecution } from '@/executor/types'
 import { executeOllamaProviderRequest } from '@/providers/ollama/core'
 import type { ModelsObject } from '@/providers/ollama/types'
 import { createReadableStreamFromOllamaStream } from '@/providers/ollama/utils'
+import { openAICompatTransport } from '@/providers/transport'
 import type { ProviderConfig, ProviderRequest, ProviderResponse } from '@/providers/types'
 import { useProvidersStore } from '@/stores/providers'
 
@@ -52,6 +53,7 @@ export const ollamaProvider: ProviderConfig = {
       providerLabel: 'Ollama',
       createClient: () =>
         new OpenAI({
+          ...openAICompatTransport(),
           apiKey: 'empty',
           baseURL: `${OLLAMA_HOST}/v1`,
         }),
diff --git a/apps/sim/providers/openrouter/index.ts b/apps/sim/providers/openrouter/index.ts
@@ -26,6 +26,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   FunctionCallResponse,
   Message,
@@ -101,6 +102,7 @@ export const openRouterProvider: ProviderConfig = {
     }
 
     const client = new OpenAI({
+      ...openAICompatTransport(),
       apiKey: request.apiKey,
       baseURL: 'https://openrouter.ai/api/v1',
     })
diff --git a/apps/sim/providers/sakana/index.ts b/apps/sim/providers/sakana/index.ts
@@ -14,6 +14,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -54,6 +55,7 @@ export const sakanaProvider: ProviderConfig = {
 
     try {
       const sakana = new OpenAI({
+        ...openAICompatTransport(),
         apiKey: request.apiKey,
         baseURL: SAKANA_BASE_URL,
       })
diff --git a/apps/sim/providers/together/index.ts b/apps/sim/providers/together/index.ts
@@ -19,6 +19,7 @@ import {
 } from '@/providers/together/utils'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   FunctionCallResponse,
   Message,
@@ -86,6 +87,7 @@ export const togetherProvider: ProviderConfig = {
     }
 
     const client = new OpenAI({
+      ...openAICompatTransport(),
       apiKey: request.apiKey,
       baseURL: 'https://api.together.ai/v1',
     })
diff --git a/apps/sim/providers/transport.test.ts b/apps/sim/providers/transport.test.ts
@@ -0,0 +1,42 @@
+/**
+ * @vitest-environment node
+ *
+ * These pin transport policy against the vendored SDK rather than against a number
+ * typed from memory: if an `openai` bump moves `DEFAULT_TIMEOUT` or the retry default,
+ * this fails and the divergence is a decision rather than a surprise.
+ */
+import { describe, expect, it } from 'vitest'
+import {
+  openAICompatTransport,
+  PROVIDER_DISCOVERY_TIMEOUT_MS,
+  PROVIDER_HEADERS_TIMEOUT_MS,
+  PROVIDER_MAX_RETRIES,
+} from '@/providers/transport'
+
+describe('provider transport policy', () => {
+  it('pins the headers budget to the vendored OpenAI client default', async () => {
+    const { default: OpenAI } = await import('openai')
+    expect(PROVIDER_HEADERS_TIMEOUT_MS).toBe(
+      (OpenAI as unknown as { DEFAULT_TIMEOUT: number }).DEFAULT_TIMEOUT
+    )
+    expect(PROVIDER_HEADERS_TIMEOUT_MS).toBe(600_000)
+  })
+
+  /**
+   * Not lowered to 0 on purpose. A chat completion is non-idempotent and carries no
+   * idempotency key, and on the non-streaming path the response exists only once the
+   * generation is already billed — so a replay re-bills completed work.
+   */
+  it('keeps the vendor retry default rather than hand-rolling one', () => {
+    expect(PROVIDER_MAX_RETRIES).toBe(2)
+  })
+
+  it('stamps only constructor-safe options, since clients are memoised', () => {
+    expect(openAICompatTransport()).toEqual({ timeout: 600_000, maxRetries: 2 })
+  })
+
+  /** A catalog probe must not inherit a generation-sized budget. */
+  it('bounds discovery far below a generation', () => {
+    expect(PROVIDER_DISCOVERY_TIMEOUT_MS).toBeLessThan(PROVIDER_HEADERS_TIMEOUT_MS / 10)
+  })
+})
diff --git a/apps/sim/providers/transport.ts b/apps/sim/providers/transport.ts
@@ -0,0 +1,59 @@
+/**
+ * Transport policy for provider requests, in one place.
+ *
+ * The values here mostly *pin* what the vendor SDKs already default to, so an SDK
+ * bump cannot silently move production behaviour. Two are deliberate divergences,
+ * marked below.
+ *
+ * What these deliberately do NOT do: bound a stalled stream. Under Bun a `fetch`
+ * carries a socket-scoped idle wall (~300s, reduced by however long a pooled socket
+ * sat idle before reuse) that no SDK, undici `Agent`, or `RequestInit` option can
+ * reach. Raising a timeout number therefore cannot rescue a long silent generation —
+ * only keeping bytes on the socket can. That work belongs in the stream pump.
+ */
+
+/**
+ * Time-to-headers budget for a single attempt, matching `openai@7`'s own
+ * `DEFAULT_TIMEOUT` so this is behaviour-preserving for every provider already on
+ * that client.
+ *
+ * It must stay generous: `deepseek-reasoner`, `kimi-k3`, `grok-4.5-reasoning` and
+ * every dynamic-catalog provider can legitimately generate for minutes with zero
+ * bytes on the wire, and on a non-streaming call headers do not arrive until the
+ * generation completes. A tighter value converts today's successes into failures.
+ */
+export const PROVIDER_HEADERS_TIMEOUT_MS = 600_000
+
+/**
+ * Vendor default, pinned rather than changed.
+ *
+ * Deliberately not lowered to 0 in favour of a hand-rolled loop: a chat completion
+ * is non-idempotent and carries no idempotency key, and on the non-streaming path
+ * the response only exists once the generation has already been billed — so a
+ * replay re-bills completed work, multiplied by every turn of the tool loop.
+ */
+export const PROVIDER_MAX_RETRIES = 2
+
+/** Liveness probe for model catalogs, never a generation. Failure degrades gracefully. */
+export const PROVIDER_DISCOVERY_TIMEOUT_MS = 15_000
+
+export interface OpenAICompatTransport {
+  timeout: number
+  maxRetries: number
+}
+
+/**
+ * Transport options for the OpenAI-compatible clients.
+ *
+ * Stamps `timeout` and `maxRetries` only. Both are process-wide constants, which is
+ * what makes them safe to set in a constructor that {@link getCachedProviderClient}
+ * memoises — anything varying per request must be passed per call instead.
+ *
+ * Deliberately stamps no `fetch`: a wrapper would turn the SDK's header-only timer
+ * into a total deadline that truncates a live stream, and would drop the caller's
+ * abort signal.
+ */
+export const openAICompatTransport = (): OpenAICompatTransport => ({
+  timeout: PROVIDER_HEADERS_TIMEOUT_MS,
+  maxRetries: PROVIDER_MAX_RETRIES,
+})
diff --git a/apps/sim/providers/vllm/index.ts b/apps/sim/providers/vllm/index.ts
@@ -17,6 +17,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   Message,
   ProviderConfig,
@@ -144,6 +145,7 @@ export const vllmProvider: ProviderConfig = {
       `vllm::${apiKey}::${baseUrl}::${pinnedIP ?? 'no-pin'}`,
       () =>
         new OpenAI({
+          ...openAICompatTransport(),
           apiKey,
           baseURL: `${baseUrl}/v1`,
           ...(pinnedFetch ? { fetch: pinnedFetch } : {}),
diff --git a/apps/sim/providers/xai/index.ts b/apps/sim/providers/xai/index.ts
@@ -14,6 +14,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   Message,
   ProviderConfig,
@@ -53,6 +54,7 @@ export const xAIProvider: ProviderConfig = {
     }
 
     const xai = new OpenAI({
+      ...openAICompatTransport(),
       apiKey: request.apiKey,
       baseURL: 'https://api.x.ai/v1',
     })
diff --git a/apps/sim/providers/zai/index.ts b/apps/sim/providers/zai/index.ts
@@ -14,6 +14,7 @@ import { createStreamingExecution } from '@/providers/streaming-execution'
 import { isAbortError, parseToolArguments } from '@/providers/streaming-tool-loop-shared'
 import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
 import { enrichLastModelSegmentFromChatCompletions } from '@/providers/trace-enrichment'
+import { openAICompatTransport } from '@/providers/transport'
 import type {
   ProviderConfig,
   ProviderRequest,
@@ -83,6 +84,7 @@ export const zaiProvider: ProviderConfig = {
 
     try {
       const zai = new OpenAI({
+        ...openAICompatTransport(),
         apiKey: request.apiKey,
         baseURL: ZAI_BASE_URL,
       })
