Fix disconnected stream retention

web/src/app/api/v1/chat/completions/_post.ts

@@ -874,6 +874,9 @@ export async function postChatCompletions(params: {
 
       // Log detailed error information for debugging
       const errorDetails = openrouterError?.toJSON()
+      const shouldRecordMessages = freebuffAccessTier !== 'limited'
+      const { messages: _messages, ...bodyWithoutMessages } = body
+      const telemetryBody = shouldRecordMessages ? body : bodyWithoutMessages
       const providerLabel = siliconflowError
         ? 'SiliconFlow'
         : opencodeZenError
@@ -901,7 +904,9 @@ export async function postChatCompletions(params: {
           messageCount: Array.isArray(typedBody.messages)
             ? typedBody.messages.length
             : 0,
-          messages: typedBody.messages,
+          ...(shouldRecordMessages
+            ? { messages: typedBody.messages }
+            : { messagesOmitted: true, accessTier: freebuffAccessTier }),
           providerStatusCode: (
             openrouterError ??
             fireworksError ??
@@ -935,7 +940,7 @@ export async function postChatCompletions(params: {
         userId,
         properties: {
           error: error instanceof Error ? error.message : 'Unknown error',
-          body,
+          body: telemetryBody,
           agentId,
           streaming: bodyStream,
         },

web/src/llm-api/deepseek.ts

@@ -74,6 +74,13 @@ export function isDeepSeekModel(model: string): boolean {
   return DEEPSEEK_ROUTED_MODELS.has(model)
 }
 
+function isDeepSeekV4FlashModel(model: string): boolean {
+  return (
+    model === deepseekModels.deepseekV4Flash ||
+    model === deepseekModels.deepseekV4FlashDirect
+  )
+}
+
 function getDeepSeekPricing(model: string): DeepSeekPricing {
   const entry = DEEPSEEK_MODELS[model]
   if (!entry) {
@@ -279,6 +286,7 @@ export async function handleDeepSeekStream({
     body,
     logger,
   })
+  const skipDisconnectedBilling = isDeepSeekV4FlashModel(body.model)
 
   const response = await createDeepSeekRequest({ body, originalModel, fetch })
 
@@ -392,13 +400,26 @@ export async function handleDeepSeekStream({
     cancel() {
       clearInterval(heartbeatInterval)
       clientDisconnected = true
+      if (skipDisconnectedBilling) {
+        reader
+          .cancel('client disconnected from DeepSeek V4 Flash stream')
+          .catch((error) => {
+            logger.warn(
+              { error },
+              'Failed to cancel disconnected DeepSeek V4 Flash stream',
+            )
+          })
+      }
       logger.warn(
         {
           clientDisconnected,
           responseTextLength: state.responseText.length,
           reasoningTextLength: state.reasoningText.length,
+          skippedBilling: skipDisconnectedBilling,
         },
-        'Client cancelled stream, continuing DeepSeek consumption for billing',
+        skipDisconnectedBilling
+          ? 'Client cancelled DeepSeek V4 Flash stream, ending without billing'
+          : 'Client cancelled stream, continuing DeepSeek consumption for billing',
       )
     },
   })

web/src/llm-api/openrouter.ts

@@ -40,6 +40,7 @@ type StreamState = {
 // endpoint. OR finalizes generation records asynchronously; 500ms is enough
 // in practice and keeps the delay off the client response path.
 const GENERATION_LOOKUP_DELAY_MS = 500
+const DISCONNECTED_STREAM_DRAIN_TIMEOUT_MS = 2 * 60 * 1000
 
 // Extended timeout for deep-thinking models (e.g., gpt-5) that can take
 // a long time to start streaming.
@@ -363,6 +364,7 @@ export async function handleOpenRouterStream({
     billed: false,
   }
   let clientDisconnected = false
+  let disconnectedStreamDrainTimeout: NodeJS.Timeout | null = null
 
   // Runs once on any stream-exit path. If we didn't bill through the normal
   // path (stream ended without a usage chunk, got a provider error chunk,
@@ -488,12 +490,41 @@ export async function handleOpenRouterStream({
         }
         await ensureBilled()
       } finally {
+        if (disconnectedStreamDrainTimeout) {
+          clearTimeout(disconnectedStreamDrainTimeout)
+        }
         clearInterval(heartbeatInterval)
       }
     },
     cancel() {
       clearInterval(heartbeatInterval)
       clientDisconnected = true
+      disconnectedStreamDrainTimeout = setTimeout(() => {
+        const stateSummary = {
+          clientDisconnected,
+          responseTextLength: state.responseText.length,
+          reasoningTextLength: state.reasoningText.length,
+          generationId: state.generationId,
+          billed: state.billed,
+        }
+        if (!state.billed && !state.generationId) {
+          logger.warn(
+            stateSummary,
+            'Disconnected OpenRouter stream exceeded drain timeout before fallback billing was possible; continuing to drain',
+          )
+          return
+        }
+        logger.warn(
+          stateSummary,
+          'Cancelling disconnected OpenRouter stream after drain timeout',
+        )
+        reader.cancel('client disconnected drain timeout').catch((error) => {
+          logger.warn(
+            { error },
+            'Failed to cancel disconnected OpenRouter stream',
+          )
+        })
+      }, DISCONNECTED_STREAM_DRAIN_TIMEOUT_MS)
       // Log truncated state to prevent OOM during logging (state can be up to 2MB)
       logger.warn(
         {