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

apps/sim/app/api/guardrails/validate/route.test.ts

@@ -435,4 +435,31 @@ describe('POST /api/guardrails/validate', () => {
  await expect(res.json()).resolves.toEqual({ error: 'Failed to resolve billing attribution' })
  expect(mockValidateHallucination).not.toHaveBeenCalled()
  })
+
+ /**
+ * The signal now reaches the scoring model, so cancellation is reachable here.
+ * `passed: false` would read to a consumer as the guardrail rejecting the content,
+ * blocking a run that was abandoned rather than judged.
+ */
+ it('reports a cancelled run as cancellation rather than a failed guardrail', async () => {
+ mockAuthorizeCredentialUse.mockResolvedValue({ ok: true })
+ mockValidateHallucination.mockRejectedValueOnce(
+ Object.assign(new Error('The operation was aborted.'), { name: 'AbortError' })
+ )
+
+ const res = await POST(
+ createMockRequest('POST', {
+ validationType: 'hallucination',
+ input: 'test input',
+ knowledgeBaseId: 'kb-1',
+ model: 'openai/gpt-4o',
+ workflowId: 'wf-1',
+ })
+ )
+
+ expect(res.status).toBe(499)
+ const json = await res.json()
+ expect(json.success).toBe(false)
+ expect(json.output?.passed).toBeUndefined()
+ })
 })

apps/sim/app/api/guardrails/validate/route.ts

@@ -28,6 +28,7 @@ import {
  ProviderNotAllowedError,
 } from '@/ee/access-control/utils/permission-check'
 import type { ResolvedSecretTraceRegistry } from '@/executor/utils/resolved-secret-trace-registry'
+import { isAbortError } from '@/providers/streaming-tool-loop-shared'
 import { getProviderFromModel } from '@/providers/utils'
 
 const logger = createLogger('GuardrailsValidateAPI')
@@ -316,7 +317,8 @@ export const POST = withRouteHandler(async (request: NextRequest) => {
  auth.userId,
  billingAttribution,
  requestId,
- resolvedSecretTraceRegistry
+ resolvedSecretTraceRegistry,
+ request.signal
  )
 
  /**
@@ -371,6 +373,19 @@ export const POST = withRouteHandler(async (request: NextRequest) => {
  },
  })
  } catch (error: any) {
+ /**
+ * A cancelled run must not be reshaped into a verdict. `passed: false` here reads
+ * to a consumer as the guardrail rejecting the content, so an abandoned run would
+ * block content that was never actually judged. 499 matches the convention the
+ * workflow execute route already uses for a client-cancelled request.
+ */
+ if (isAbortError(error)) {
+ logger.info(`[${requestId}] Guardrails validation cancelled by client`)
+ return NextResponse.json(
+ { success: false, error: 'Client cancelled request' },
+ { status: 499 }
+ )
+ }
  logger.error(`[${requestId}] Guardrails validation failed`, { error })
  return NextResponse.json({
  success: true,
@@ -431,7 +446,8 @@ async function executeValidation(
  actorUserId: string,
  billingAttribution: BillingAttributionSnapshot | undefined,
  requestId: string,
- resolvedSecretTraceRegistry: ResolvedSecretTraceRegistry | undefined
+ resolvedSecretTraceRegistry: ResolvedSecretTraceRegistry | undefined,
+ abortSignal: AbortSignal | undefined
 ): Promise<{
  passed: boolean
  error?: string
@@ -488,6 +504,7 @@ async function executeValidation(
  billingAttribution,
  requestId,
  resolvedSecretTraceRegistry,
+ abortSignal,
  })
  }
  if (validationType === 'pii') {

apps/sim/app/workspace/[workspaceId]/w/[workflowId]/components/panel/components/editor/components/sub-block/components/text/text.tsx

@@ -33,7 +33,7 @@ export function Text({ blockId, subBlockId, content, className }: TextProps) {
  className={`rounded-md border bg-[var(--surface-2)] p-4 shadow-sm ${className || ''}`}
  >
  <div
- className='max-w-none break-words text-[var(--text-secondary)] text-sm [&_a]:text-[var(--brand-secondary)] [&_a]:underline [&_a]:underline-offset-2 [&_a]:hover-hover:brightness-110 [&_code]:rounded [&_code]:bg-[var(--surface-5)] [&_code]:px-1 [&_code]:py-0.5 [&_code]:text-[var(--text-tertiary)] [&_code]:text-xs [&_ul]:ml-5 [&_ul]:list-disc [&_ul]:marker:text-[var(--text-muted)] [&_strong]:font-medium [&_strong]:text-[var(--text-primary)]'
+ className='max-w-none break-words text-[var(--text-secondary)] text-sm [&_a]:text-[var(--brand-secondary)] [&_a]:underline [&_a]:underline-offset-2 [&_a]:hover-hover:brightness-110 [&_code]:rounded [&_code]:bg-[var(--surface-5)] [&_code]:px-1 [&_code]:py-0.5 [&_code]:text-[var(--text-tertiary)] [&_code]:text-xs [&_strong]:font-medium [&_strong]:text-[var(--text-primary)] [&_ul]:ml-5 [&_ul]:list-disc [&_ul]:marker:text-[var(--text-muted)]'
  dangerouslySetInnerHTML={{ __html: content }}
  />
  </div>

apps/sim/lib/guardrails/validate_hallucination.test.ts

@@ -182,4 +182,25 @@ describe('validateHallucination', () => {
  expect(mockExecuteProviderRequest).not.toHaveBeenCalled()
  expect(registry.isComplete()).toBe(true)
  })
+
+ /**
+ * Forwarding the caller's signal means the scoring model can now be aborted. A
+ * cancelled run must not be reported as a guardrail verdict — `passed: false` would
+ * block content on a run the caller abandoned, which is indistinguishable to a
+ * consumer from the model actually hallucinating.
+ */
+ it('surfaces a cancelled run as cancellation, not as a failed guardrail', async () => {
+ const registry = new ResolvedSecretTraceRegistry()
+ const fetchMock = vi.fn(async () =>
+ createPrivateKnowledgeResponse({ data: { results: [{ content: 'reference' }] } })
+ )
+ vi.stubGlobal('fetch', fetchMock)
+
+ const abort = Object.assign(new Error('The operation was aborted.'), { name: 'AbortError' })
+ mockExecuteProviderRequest.mockRejectedValueOnce(abort)
+
+ await expect(validateHallucination(createInput(registry))).rejects.toMatchObject({
+ name: 'AbortError',
+ })
+ })
 })

apps/sim/lib/guardrails/validate_hallucination.ts

@@ -24,6 +24,7 @@ import { refreshTokenIfNeeded } from '@/app/api/auth/oauth/utils'
 import { projectResolvedSecretModelContent } from '@/executor/utils/resolved-secret-content-projection'
 import type { ResolvedSecretTraceRegistry } from '@/executor/utils/resolved-secret-trace-registry'
 import { executeProviderRequest } from '@/providers'
+import { isAbortError } from '@/providers/streaming-tool-loop-shared'
 import { getProviderFromModel } from '@/providers/utils'
 
 const logger = createLogger('HallucinationValidator')
@@ -68,6 +69,12 @@ export interface HallucinationValidationInput {
  billingAttribution: BillingAttributionSnapshot
  requestId: string
  resolvedSecretTraceRegistry: ResolvedSecretTraceRegistry
+ /**
+ * The caller's cancellation signal, forwarded to the scoring model exactly as the
+ * agent handler forwards `ctx.abortSignal`. Without it the scoring request outlives
+ * a cancelled request and keeps burning a provider slot until the transport gives up.
+ */
+ abortSignal?: AbortSignal
 }
 
 /**
@@ -176,7 +183,8 @@ async function scoreHallucinationWithLLM(
  providerCredentials: HallucinationValidationInput['providerCredentials'],
  workspaceId: string | undefined,
  requestId: string,
- resolvedSecretTraceRegistry: ResolvedSecretTraceRegistry
+ resolvedSecretTraceRegistry: ResolvedSecretTraceRegistry,
+ abortSignal: AbortSignal | undefined
 ): Promise<{ score: number; reasoning: string; cost: number }> {
  try {
  const contextText = ragContext.join('\n\n---\n\n')
@@ -251,6 +259,7 @@ Evaluate the consistency and provide your score and reasoning in JSON format.`
  bedrockSecretKey: providerCredentials?.bedrockSecretKey,
  bedrockRegion: providerCredentials?.bedrockRegion,
  workspaceId,
+ abortSignal,
  },
  { resolvedSecretTraceRegistry }
  )
@@ -290,6 +299,11 @@ Evaluate the consistency and provide your score and reasoning in JSON format.`
  cost,
  }
  } catch (error: any) {
+ /**
+ * A cancelled run is not a scoring failure. Rewrapping it would erase the
+ * `AbortError` name the outer handler classifies on, so it propagates as-is.
+ */
+ if (isAbortError(error)) throw error
  logger.error(`[${requestId}] Error scoring with LLM`, {
  error: error.message,
  })
@@ -317,6 +331,7 @@ export async function validateHallucination(
  billingAttribution,
  requestId,
  resolvedSecretTraceRegistry,
+ abortSignal,
  } = input
 
  try {
@@ -371,7 +386,8 @@ export async function validateHallucination(
  providerCredentials,
  workspaceId,
  requestId,
- providerRegistry
+ providerRegistry,
+ abortSignal
  )
 
  logger.info(`[${requestId}] Confidence score: ${score}`, {
@@ -392,6 +408,12 @@ export async function validateHallucination(
  : `Low confidence: score ${score}/10 is below threshold ${threshold}`,
  }
  } catch (error: any) {
+ /**
+ * Cancellation is surfaced as cancellation, not as a guardrail verdict. Returning
+ * `passed: false` here would fail content on a run the caller abandoned, which is
+ * indistinguishable to a consumer from the model actually hallucinating.
+ */
+ if (isAbortError(error)) throw error
  logger.error(`[${requestId}] Hallucination validation error`, {
  error: error.message,
  })

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',
  })

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 = []

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',
  })

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',
  })

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 = []
 

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,
  })

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`,
  })

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,
  })

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',
  })

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,
  })

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,
  }),

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`,
  }),