[codex] fix empty stream success handling and unknown usage logs (#343)

* fix: preserve repeated short responses deltas

* fix empty stream success handling and unknown usage logs

* fix proxy usage metadata typing

* fix stream bookkeeping and explicit zero usage source

* fix sticky fast path and usage placeholder detection

src/server/proxy-core/surfaces/chatSurface.ts

@@ -3,7 +3,7 @@ import type { FastifyRequest, FastifyReply } from 'fastify';
 import { config } from '../../config.js';
 import { tokenRouter } from '../../services/tokenRouter.js';
 import { reportProxyAllFailed } from '../../services/alertService.js';
-import { mergeProxyUsage, parseProxyUsage } from '../../services/proxyUsageParser.js';
+import { hasProxyUsagePayload, mergeProxyUsage, parseProxyUsage } from '../../services/proxyUsageParser.js';
 import { type DownstreamFormat } from '../../transformers/shared/normalized.js';
 import {
   buildClaudeCountTokensUpstreamRequest,
@@ -489,7 +489,10 @@ export async function handleChatSurfaceRequest(
 
       if (isStream) {
         const upstreamContentType = (upstream.headers.get('content-type') || '').toLowerCase();
+        let streamStarted = false;
         const startSseResponse = () => {
+          if (streamStarted) return;
+          streamStarted = true;
           reply.hijack();
           reply.raw.statusCode = 200;
           reply.raw.setHeader('Content-Type', 'text/event-stream; charset=utf-8');
@@ -506,23 +509,54 @@ export async function handleChatSurfaceRequest(
           cacheCreationTokens: 0,
           promptTokensIncludeCache: null,
         };
+        let upstreamUsagePresent = false;
+        const recordStreamSuccess = async (latencyMs: number) => {
+          await recordSurfaceSuccess({
+            selected,
+            requestedModel,
+            modelName,
+            parsedUsage,
+            upstreamUsagePresent,
+            requestStartedAtMs: startTime,
+            latencyMs,
+            retryCount,
+            upstreamPath: successfulUpstreamPath,
+            logSuccess: failureToolkit.log,
+            recordDownstreamCost: (estimatedCost) => {
+              recordDownstreamCostUsage(request, estimatedCost);
+            },
+            bestEffortMetrics: {
+              errorLabel: '[proxy/chat] failed to record success metrics',
+            },
+          });
+        };
 
         const writeLines = (lines: string[]) => {
+          startSseResponse();
           for (const line of lines) {
             reply.raw.write(line);
           }
         };
+        const streamResponse = {
+          end() {
+            if (streamStarted) {
+              reply.raw.end();
+            }
+          },
+        };
         const streamSession = openAiChatTransformer.proxyStream.createSession({
           downstreamFormat,
           modelName,
           successfulUpstreamPath,
           onParsedPayload: (payload) => {
             if (payload && typeof payload === 'object') {
+              upstreamUsagePresent = upstreamUsagePresent || hasProxyUsagePayload(payload);
               parsedUsage = mergeProxyUsage(parsedUsage, parseProxyUsage(payload));
             }
           },
           writeLines,
           writeRaw: (chunk) => {
+            startSseResponse();
             reply.raw.write(chunk);
           },
         });
@@ -531,10 +565,9 @@ export async function handleChatSurfaceRequest(
           const fallbackText = await readRuntimeResponseText(upstream);
           rawText = fallbackText;
           if (looksLikeResponsesSseText(fallbackText)) {
-            startSseResponse();
             const streamResult = await streamSession.run(
               createSingleChunkStreamReader(fallbackText),
-              reply.raw,
+              streamResponse,
             );
             const latency = Date.now() - startTime;
             if (streamResult.status === 'failed') {
@@ -560,8 +593,17 @@ export async function handleChatSurfaceRequest(
                   type: 'stream_error',
                 },
               }, successfulUpstreamPath);
+              if (!streamStarted) {
+                return reply.code(502).send({
+                  error: {
+                    message: streamResult.errorMessage,
+                    type: 'upstream_error',
+                  },
+                });
+              }
               return;
             }
+            await recordStreamSuccess(latency);
             await finalizeDebugSuccess(
               200,
               successfulUpstreamPath,
@@ -588,6 +630,7 @@ export async function handleChatSurfaceRequest(
           if (String(selected.site.platform || '').trim().toLowerCase() === 'gemini-cli') {
             fallbackData = unwrapGeminiCliPayload(fallbackData);
           }
+          upstreamUsagePresent = upstreamUsagePresent || hasProxyUsagePayload(fallbackData);
           parsedUsage = mergeProxyUsage(parsedUsage, parseProxyUsage(fallbackData));
           const latency = Date.now() - startTime;
           const failure = detectProxyFailure({ rawText, usage: parsedUsage });
@@ -620,8 +663,7 @@ export async function handleChatSurfaceRequest(
             return reply.code(failureOutcome.status).send(failureOutcome.payload);
           }
 
-          startSseResponse();
-          const streamResult = streamSession.consumeUpstreamFinalPayload(fallbackData, fallbackText, reply.raw);
+          const streamResult = streamSession.consumeUpstreamFinalPayload(fallbackData, fallbackText, streamResponse);
           if (streamResult.status === 'failed') {
             clearSurfaceStickyChannel({
               stickySessionKey,
@@ -646,8 +688,17 @@ export async function handleChatSurfaceRequest(
                 type: 'stream_error',
               },
             }, successfulUpstreamPath);
+            if (!streamStarted) {
+              return reply.code(502).send({
+                error: {
+                  message: streamResult.errorMessage,
+                  type: 'upstream_error',
+                },
+              });
+            }
             return;
           }
+          await recordStreamSuccess(latency);
           await finalizeDebugSuccess(
             200,
             successfulUpstreamPath,
@@ -663,8 +714,8 @@ export async function handleChatSurfaceRequest(
             stickySessionKey,
             selected,
           });
+          return;
         } else {
-          startSseResponse();
           const upstreamReader = getRuntimeResponseReader(upstream);
           const baseReader = String(selected.site.platform || '').trim().toLowerCase() === 'gemini-cli' && upstreamReader
             ? createGeminiCliStreamReader(upstreamReader)
@@ -687,7 +738,7 @@ export async function handleChatSurfaceRequest(
               },
             }
             : baseReader;
-          const streamResult = await streamSession.run(reader, reply.raw);
+          const streamResult = await streamSession.run(reader, streamResponse);
           rawText += decoder.decode();
 
           const latency = Date.now() - startTime;
@@ -715,6 +766,14 @@ export async function handleChatSurfaceRequest(
                 type: 'stream_error',
               },
             }, successfulUpstreamPath);
+            if (!streamStarted) {
+              return reply.code(502).send({
+                error: {
+                  message: streamResult.errorMessage,
+                  type: 'upstream_error',
+                },
+              });
+            }
             return;
           }
 
@@ -725,23 +784,7 @@ export async function handleChatSurfaceRequest(
         }
 
         const latency = Date.now() - startTime;
-        await recordSurfaceSuccess({
-          selected,
-          requestedModel,
-          modelName,
-          parsedUsage,
-          requestStartedAtMs: startTime,
-          latencyMs: latency,
-          retryCount,
-          upstreamPath: successfulUpstreamPath,
-          logSuccess: failureToolkit.log,
-          recordDownstreamCost: (estimatedCost) => {
-            recordDownstreamCostUsage(request, estimatedCost);
-          },
-          bestEffortMetrics: {
-            errorLabel: '[proxy/chat] failed to record success metrics',
-          },
-        });
+        await recordStreamSuccess(latency);
         await finalizeDebugSuccess(
           200,
           successfulUpstreamPath,
@@ -786,6 +829,7 @@ export async function handleChatSurfaceRequest(
 
       const latency = Date.now() - startTime;
       const parsedUsage = parseProxyUsage(upstreamData);
+      const upstreamUsagePresent = hasProxyUsagePayload(upstreamData);
       const failure = detectProxyFailure({ rawText, usage: parsedUsage });
       if (failure) {
         clearSurfaceStickyChannel({
@@ -823,6 +867,7 @@ export async function handleChatSurfaceRequest(
         requestedModel,
         modelName,
         parsedUsage,
+        upstreamUsagePresent,
         requestStartedAtMs: startTime,
         latencyMs: latency,
         retryCount,

src/server/proxy-core/surfaces/openAiResponsesSurface.ts

@@ -2,7 +2,7 @@ import { TextDecoder } from 'node:util';
 import type { FastifyReply, FastifyRequest } from 'fastify';
 import { config } from '../../config.js';
 import { reportProxyAllFailed } from '../../services/alertService.js';
-import { mergeProxyUsage, parseProxyUsage } from '../../services/proxyUsageParser.js';
+import { hasProxyUsagePayload, mergeProxyUsage, parseProxyUsage } from '../../services/proxyUsageParser.js';
 import { openAiResponsesTransformer } from '../../transformers/openai/responses/index.js';
 import {
   buildUpstreamEndpointRequest,
@@ -604,13 +604,19 @@ export async function handleOpenAiResponsesSurfaceRequest(
 
       const upstream = endpointResult.upstream;
       const successfulUpstreamPath = endpointResult.upstreamPath;
-      const finalizeStreamSuccess = async (parsedUsage: UsageSummary, latency: number, streamDebugBody: unknown) => {
+      const finalizeStreamSuccess = async (
+        parsedUsage: UsageSummary,
+        latency: number,
+        streamDebugBody: unknown,
+        upstreamUsagePresent: boolean,
+      ) => {
         try {
           await recordSurfaceSuccess({
             selected,
             requestedModel,
             modelName,
             parsedUsage,
+            upstreamUsagePresent,
             requestStartedAtMs: startTime,
             latencyMs: latency,
             retryCount,
@@ -653,6 +659,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             cacheCreationTokens: 0,
             promptTokensIncludeCache: null,
           };
+          let upstreamUsagePresent = false;
           const writeLines = (lines: string[]) => {
             for (const line of lines) reply.raw.write(line);
           };
@@ -663,6 +670,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             getUsage: () => parsedUsage,
             onParsedPayload: (payload) => {
               if (payload && typeof payload === 'object') {
+                upstreamUsagePresent = upstreamUsagePresent || hasProxyUsagePayload(payload);
                 parsedUsage = mergeProxyUsage(parsedUsage, parseProxyUsage(payload));
               }
             },
@@ -710,6 +718,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
                   parsedUsage,
                   latency,
                   debugTrace?.options.captureStreamChunks ? rawText : { stream: true, usage: parsedUsage },
+                  upstreamUsagePresent,
                 );
 	              bindSurfaceStickyChannel({
 	                stickySessionKey,
@@ -728,6 +737,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             }
 
             parsedUsage = parseProxyUsage(upstreamData);
+            upstreamUsagePresent = upstreamUsagePresent || hasProxyUsagePayload(upstreamData);
             const latency = Date.now() - startTime;
             const failure = detectProxyFailure({ rawText, usage: parsedUsage });
 	            if (failure) {
@@ -792,6 +802,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
                 parsedUsage,
                 latency,
                 debugTrace?.options.captureStreamChunks ? rawText : upstreamData,
+                upstreamUsagePresent,
               );
 	            bindSurfaceStickyChannel({
 	              stickySessionKey,
@@ -808,6 +819,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             if (looksLikeResponsesSseText(rawText)) {
               try {
                 const collectedPayload = collectResponsesFinalPayloadFromSseText(rawText, modelName).payload;
+                upstreamUsagePresent = upstreamUsagePresent || hasProxyUsagePayload(collectedPayload);
                 parsedUsage = mergeProxyUsage(parsedUsage, parseProxyUsage(collectedPayload));
                 const createdPayload = {
                   ...collectedPayload,
@@ -829,7 +841,12 @@ export async function handleOpenAiResponsesSurfaceRequest(
                   parsedUsage,
                   latency,
                   debugTrace?.options.captureStreamChunks ? rawText : collectedPayload,
+                  upstreamUsagePresent,
                 );
+                bindSurfaceStickyChannel({
+                  stickySessionKey,
+                  selected,
+                });
                 return;
               } catch {
                 // Fall through to the generic stream session for response.failed/error terminals.
@@ -867,6 +884,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
                 parsedUsage,
                 latency,
                 debugTrace?.options.captureStreamChunks ? rawText : { stream: true, usage: parsedUsage },
+                upstreamUsagePresent,
               );
               return;
             }
@@ -937,6 +955,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
               parsedUsage,
               latency,
               debugTrace?.options.captureStreamChunks ? rawText : { stream: true, usage: parsedUsage },
+              upstreamUsagePresent,
             );
 	          bindSurfaceStickyChannel({
 	            stickySessionKey,
@@ -976,6 +995,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
         }
         const latency = Date.now() - startTime;
         const parsedUsage = parseProxyUsage(upstreamData);
+        const upstreamUsagePresent = hasProxyUsagePayload(upstreamData);
         const failure = detectProxyFailure({ rawText, usage: parsedUsage });
 	        if (failure) {
 	          clearSurfaceStickyChannel({
@@ -1022,6 +1042,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             requestedModel,
             modelName,
             parsedUsage,
+            upstreamUsagePresent,
             requestStartedAtMs: startTime,
             latencyMs: latency,
             retryCount,

src/server/proxy-core/surfaces/sharedSurface.test.ts

@@ -275,6 +275,7 @@ describe('selectSurfaceChannelForAttempt', () => {
       estimatedCost: 0.42,
       billingDetails: { source: 'test' },
       upstreamPath: '/v1/responses',
+      usageSource: 'self-log',
       clientContext: {
         clientKind: 'codex',
         clientAppId: 'app-id',
@@ -292,6 +293,7 @@ describe('selectSurfaceChannelForAttempt', () => {
       traceHint: 'trace-1',
       downstreamPath: '/v1/chat/completions',
       upstreamPath: '/v1/responses',
+      usageSource: 'self-log',
       errorMessage: 'upstream failed',
     });
     expect(insertProxyLogMock).toHaveBeenCalledWith({
@@ -828,6 +830,10 @@ describe('selectSurfaceChannelForAttempt', () => {
       promptTokens: 20,
       completionTokens: 8,
       totalTokens: 28,
+      recoveredFromSelfLog: true,
+      estimatedCostFromQuota: 0.42,
+      selfLogBillingMeta: null,
+      usageSource: 'self-log',
     });
     resolveProxyLogBillingMock.mockResolvedValue({
       estimatedCost: 0.42,
@@ -870,6 +876,7 @@ describe('selectSurfaceChannelForAttempt', () => {
       requestStartedAtMs: 1000,
       requestEndedAtMs: 1250,
       localLatencyMs: 250,
+      upstreamUsagePresent: true,
       usage: {
         promptTokens: 10,
         completionTokens: 5,
@@ -889,6 +896,10 @@ describe('selectSurfaceChannelForAttempt', () => {
         promptTokens: 20,
         completionTokens: 8,
         totalTokens: 28,
+        recoveredFromSelfLog: true,
+        estimatedCostFromQuota: 0.42,
+        selfLogBillingMeta: null,
+        usageSource: 'self-log',
       },
     });
     expect(recordSuccessMock).toHaveBeenCalledWith(11, 250, 0.42, 'upstream-model');
@@ -911,6 +922,7 @@ describe('selectSurfaceChannelForAttempt', () => {
       promptTokens: 20,
       completionTokens: 8,
       totalTokens: 28,
+      usageSource: 'self-log',
       estimatedCost: 0.42,
       billingDetails: { source: 'pricing-test' },
       upstreamPath: '/v1/responses',
@@ -920,12 +932,67 @@ describe('selectSurfaceChannelForAttempt', () => {
         promptTokens: 20,
         completionTokens: 8,
         totalTokens: 28,
+        recoveredFromSelfLog: true,
+        estimatedCostFromQuota: 0.42,
+        selfLogBillingMeta: null,
+        usageSource: 'self-log',
       },
       estimatedCost: 0.42,
       billingDetails: { source: 'pricing-test' },
     });
   });
 
+  it('logs unknown usage as null tokens while preserving success bookkeeping', async () => {
+    resolveProxyUsageWithSelfLogFallbackMock.mockResolvedValue({
+      promptTokens: 0,
+      completionTokens: 0,
+      totalTokens: 0,
+      recoveredFromSelfLog: false,
+      estimatedCostFromQuota: 0,
+      selfLogBillingMeta: null,
+      usageSource: 'unknown',
+    });
+    resolveProxyLogBillingMock.mockResolvedValue({
+      estimatedCost: 0,
+      billingDetails: null,
+    });
+    const logSuccess = vi.fn().mockResolvedValue(undefined);
+
+    const { recordSurfaceSuccess } = await import('./sharedSurface.js');
+    await recordSurfaceSuccess({
+      selected: {
+        channel: { id: 11, routeId: 22 },
+        account: { id: 33, username: 'oauth-user' },
+        site: { id: 44, url: 'https://upstream.example.com', platform: 'new-api', name: 'Upstream' },
+        tokenValue: 'live-token',
+        tokenName: 'default',
+        actualModel: 'upstream-model',
+      },
+      requestedModel: 'gpt-5.2',
+      modelName: 'upstream-model',
+      parsedUsage: {
+        promptTokens: 0,
+        completionTokens: 0,
+        totalTokens: 0,
+      },
+      requestStartedAtMs: 1000,
+      latencyMs: 250,
+      retryCount: 0,
+      upstreamPath: '/v1/chat/completions',
+      logSuccess,
+    });
+
+    expect(resolveProxyUsageWithSelfLogFallbackMock).toHaveBeenCalledWith(expect.objectContaining({
+      upstreamUsagePresent: false,
+    }));
+    expect(logSuccess).toHaveBeenCalledWith(expect.objectContaining({
+      promptTokens: null,
+      completionTokens: null,
+      totalTokens: null,
+      usageSource: 'unknown',
+    }));
+  });
+
   it('treats success metrics as best-effort when requested', async () => {
     resolveProxyUsageWithSelfLogFallbackMock.mockRejectedValueOnce(new Error('billing failed'));
     const logSuccess = vi.fn().mockResolvedValue(undefined);
@@ -980,6 +1047,7 @@ describe('selectSurfaceChannelForAttempt', () => {
         recoveredFromSelfLog: false,
         estimatedCostFromQuota: 0,
         selfLogBillingMeta: null,
+        usageSource: 'upstream',
       },
       estimatedCost: 0,
       billingDetails: null,

src/server/proxy-core/surfaces/sharedSurface.ts

@@ -98,6 +98,7 @@ type SurfaceResolvedUsageSummary = {
   recoveredFromSelfLog: boolean;
   estimatedCostFromQuota: number;
   selfLogBillingMeta: import('../../services/proxyUsageFallbackService.js').SelfLogBillingMeta | null;
+  usageSource: 'upstream' | 'self-log' | 'unknown';
 };
 
 export async function selectSurfaceChannelForAttempt(input: {
@@ -228,12 +229,13 @@ export async function writeSurfaceProxyLog(input: {
   errorMessage: string | null;
   retryCount: number;
   downstreamPath: string;
-  promptTokens?: number;
-  completionTokens?: number;
-  totalTokens?: number;
+  promptTokens?: number | null;
+  completionTokens?: number | null;
+  totalTokens?: number | null;
   estimatedCost?: number;
   billingDetails?: unknown;
   upstreamPath?: string | null;
+  usageSource?: 'upstream' | 'self-log' | 'unknown' | null;
   clientContext?: DownstreamClientContext | null;
   downstreamApiKeyId?: number | null;
 }): Promise<void> {
@@ -247,6 +249,7 @@ export async function writeSurfaceProxyLog(input: {
       traceHint: input.clientContext?.traceHint || null,
       downstreamPath: input.downstreamPath,
       upstreamPath: input.upstreamPath || null,
+      usageSource: input.usageSource || null,
       errorMessage: input.errorMessage,
     });
     await insertProxyLog({
@@ -259,9 +262,9 @@ export async function writeSurfaceProxyLog(input: {
       status: input.status,
       httpStatus: input.httpStatus,
       latencyMs: input.latencyMs,
-      promptTokens: input.promptTokens ?? 0,
-      completionTokens: input.completionTokens ?? 0,
-      totalTokens: input.totalTokens ?? 0,
+      promptTokens: input.promptTokens ?? null,
+      completionTokens: input.completionTokens ?? null,
+      totalTokens: input.totalTokens ?? null,
       estimatedCost: input.estimatedCost ?? 0,
       billingDetails: input.billingDetails ?? null,
       clientFamily: input.clientContext?.clientKind || null,
@@ -354,6 +357,7 @@ export async function recordSurfaceSuccess(input: {
   requestedModel: string;
   modelName: string;
   parsedUsage: SurfaceUsageSummary;
+  upstreamUsagePresent?: boolean;
   requestStartedAtMs: number;
   latencyMs: number;
   retryCount: number;
@@ -366,9 +370,10 @@ export async function recordSurfaceSuccess(input: {
     latencyMs: number;
     errorMessage: string | null;
     retryCount: number;
-    promptTokens?: number;
-    completionTokens?: number;
-    totalTokens?: number;
+    promptTokens?: number | null;
+    completionTokens?: number | null;
+    totalTokens?: number | null;
+    usageSource?: 'upstream' | 'self-log' | 'unknown';
     estimatedCost?: number;
     billingDetails?: unknown;
     upstreamPath?: string | null;
@@ -382,6 +387,11 @@ export async function recordSurfaceSuccess(input: {
   estimatedCost: number;
   billingDetails: unknown;
 }> {
+  const hasUpstreamUsage = input.upstreamUsagePresent ?? (
+    input.parsedUsage.totalTokens > 0
+    || input.parsedUsage.promptTokens > 0
+    || input.parsedUsage.completionTokens > 0
+  );
   let resolvedUsage: SurfaceResolvedUsageSummary = {
     promptTokens: input.parsedUsage.promptTokens,
     completionTokens: input.parsedUsage.completionTokens,
@@ -389,6 +399,7 @@ export async function recordSurfaceSuccess(input: {
     recoveredFromSelfLog: false,
     estimatedCostFromQuota: 0,
     selfLogBillingMeta: null,
+    usageSource: hasUpstreamUsage ? 'upstream' : 'unknown',
   };
   let estimatedCost = 0;
   let billingDetails: unknown = null;
@@ -403,6 +414,7 @@ export async function recordSurfaceSuccess(input: {
       requestStartedAtMs: input.requestStartedAtMs,
       requestEndedAtMs: input.requestStartedAtMs + input.latencyMs,
       localLatencyMs: input.latencyMs,
+      upstreamUsagePresent: hasUpstreamUsage,
       usage: {
         promptTokens: input.parsedUsage.promptTokens,
         completionTokens: input.parsedUsage.completionTokens,
@@ -432,6 +444,17 @@ export async function recordSurfaceSuccess(input: {
     input.modelName,
   );
   input.recordDownstreamCost?.(estimatedCost);
+  const logTokens = resolvedUsage.usageSource === 'unknown'
+    ? {
+      promptTokens: null,
+      completionTokens: null,
+      totalTokens: null,
+    }
+    : {
+      promptTokens: resolvedUsage.promptTokens,
+      completionTokens: resolvedUsage.completionTokens,
+      totalTokens: resolvedUsage.totalTokens,
+    };
   await input.logSuccess({
     selected: input.selected,
     modelRequested: input.requestedModel,
@@ -440,9 +463,10 @@ export async function recordSurfaceSuccess(input: {
     latencyMs: input.latencyMs,
     errorMessage: null,
     retryCount: input.retryCount,
-    promptTokens: resolvedUsage.promptTokens,
-    completionTokens: resolvedUsage.completionTokens,
-    totalTokens: resolvedUsage.totalTokens,
+    promptTokens: logTokens.promptTokens,
+    completionTokens: logTokens.completionTokens,
+    totalTokens: logTokens.totalTokens,
+    usageSource: resolvedUsage.usageSource,
     estimatedCost,
     billingDetails,
     upstreamPath: input.upstreamPath,
@@ -470,9 +494,9 @@ export function createSurfaceFailureToolkit(input: {
     latencyMs: number;
     errorMessage: string | null;
     retryCount: number;
-    promptTokens?: number;
-    completionTokens?: number;
-    totalTokens?: number;
+    promptTokens?: number | null;
+    completionTokens?: number | null;
+    totalTokens?: number | null;
     estimatedCost?: number;
     billingDetails?: unknown;
     upstreamPath?: string | null;
@@ -581,9 +605,9 @@ export function createSurfaceFailureToolkit(input: {
       failure: { status: number; reason: string };
       latencyMs: number;
       retryCount: number;
-      promptTokens?: number;
-      completionTokens?: number;
-      totalTokens?: number;
+      promptTokens?: number | null;
+      completionTokens?: number | null;
+      totalTokens?: number | null;
       upstreamPath?: string | null;
     }): Promise<SurfaceFailureOutcome> {
       await tokenRouter.recordFailure(args.selected.channel.id, {
@@ -676,9 +700,9 @@ export function createSurfaceFailureToolkit(input: {
       errorMessage: string | null;
       latencyMs: number;
       retryCount: number;
-      promptTokens?: number;
-      completionTokens?: number;
-      totalTokens?: number;
+      promptTokens?: number | null;
+      completionTokens?: number | null;
+      totalTokens?: number | null;
       upstreamPath?: string | null;
       httpStatus?: number;
       runtimeFailureStatus?: number | null;

src/server/routes/api/stats.proxy-logs.test.ts

@@ -534,4 +534,54 @@ describe('stats proxy logs routes', () => {
     expect(detailBody.clientAppId).toBe(null);
     expect(detailBody.clientAppName).toBe(null);
   });
+
+  it('returns unknown usage source and nullable token fields for logs without recovered usage', async () => {
+    const site = await db.insert(schema.sites).values({
+      name: 'unknown-usage-site',
+      url: 'https://unknown-usage.example.com',
+      platform: 'new-api',
+    }).returning().get();
+
+    const account = await db.insert(schema.accounts).values({
+      siteId: site.id,
+      username: 'unknown-usage-user',
+      accessToken: 'unknown-usage-token',
+      status: 'active',
+    }).returning().get();
+
+    await db.insert(schema.proxyLogs).values({
+      accountId: account.id,
+      modelRequested: 'gpt-5',
+      modelActual: 'gpt-5',
+      status: 'success',
+      promptTokens: null,
+      completionTokens: null,
+      totalTokens: null,
+      estimatedCost: 0,
+      errorMessage: '[downstream:/v1/chat/completions] [upstream:/v1/chat/completions] [usage:unknown]',
+      createdAt: formatUtcSqlDateTime(new Date('2026-03-09T12:30:00.000Z')),
+    }).run();
+
+    const response = await app.inject({
+      method: 'GET',
+      url: '/api/stats/proxy-logs',
+    });
+
+    expect(response.statusCode).toBe(200);
+    const body = response.json() as {
+      items: Array<{
+        usageSource: string | null;
+        promptTokens: number | null;
+        completionTokens: number | null;
+        totalTokens: number | null;
+      }>;
+    };
+
+    expect(body.items[0]).toMatchObject({
+      usageSource: 'unknown',
+      promptTokens: null,
+      completionTokens: null,
+      totalTokens: null,
+    });
+  });
 });

src/server/routes/api/stats.ts

@@ -275,6 +275,14 @@ function resolveProxyLogClientMeta(proxyLog: Record<string, unknown>) {
   };
 }
 
+function normalizeProxyLogUsageSource(value: unknown): 'upstream' | 'self-log' | 'unknown' | null {
+  const normalized = normalizeNullableText(value)?.toLowerCase() || null;
+  if (normalized === 'upstream' || normalized === 'self-log' || normalized === 'unknown') {
+    return normalized;
+  }
+  return null;
+}
+
 function buildProxyLogClientOptions(rows: Array<{
   clientFamily?: string | null;
   clientAppId?: string | null;
@@ -472,6 +480,7 @@ function mapProxyLogRow(
   options?: { includeBillingDetails?: boolean },
 ) {
   const clientMeta = resolveProxyLogClientMeta(row.proxy_logs);
+  const legacyMeta = parseProxyLogMessageMeta(typeof row.proxy_logs.errorMessage === 'string' ? row.proxy_logs.errorMessage : '');
   return {
     ...row.proxy_logs,
     ...(options?.includeBillingDetails
@@ -481,6 +490,7 @@ function mapProxyLogRow(
     clientAppId: clientMeta.clientAppId,
     clientAppName: clientMeta.clientAppName,
     clientConfidence: clientMeta.clientConfidence,
+    usageSource: normalizeProxyLogUsageSource(legacyMeta.usageSource),
     username: row.accounts?.username || null,
     siteId: row.sites?.id || null,
     siteName: row.sites?.name || null,

src/server/routes/proxy/chat.stream.test.ts

@@ -169,6 +169,8 @@ describe('chat proxy stream behavior', () => {
     expect(response.body).toContain('"chat.completion.chunk"');
     expect(response.body).toContain('hello from upstream');
     expect(response.body).toContain('data: [DONE]');
+    expect(recordSuccessMock).toHaveBeenCalledTimes(1);
+    expect(recordFailureMock).not.toHaveBeenCalled();
   });
 
   it('decodes zstd-compressed non-stream chat responses before serializing downstream JSON', async () => {
@@ -347,6 +349,148 @@ describe('chat proxy stream behavior', () => {
     expect(recordFailureMock).toHaveBeenCalledTimes(1);
   });
 
+  it('returns HTTP upstream_error when streamed chat SSE yields only empty deltas before DONE', async () => {
+    config.proxyEmptyContentFailEnabled = true;
+
+    const encoder = new TextEncoder();
+    const upstreamBody = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(encoder.encode('data: {"id":"chatcmpl-empty-sse","choices":[{"delta":{}}]}\n\n'));
+        controller.enqueue(encoder.encode('data: [DONE]\n\n'));
+        controller.close();
+      },
+    });
+
+    fetchMock.mockResolvedValue(new Response(upstreamBody, {
+      status: 200,
+      headers: { 'content-type': 'text/event-stream; charset=utf-8' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-4o-mini',
+        stream: true,
+        messages: [{ role: 'user', content: 'hi' }],
+      },
+    });
+
+    expect(response.statusCode).toBe(502);
+    expect(response.headers['content-type']).not.toContain('text/event-stream');
+    expect(response.json()?.error?.type).toBe('upstream_error');
+    expect(response.json()?.error?.message).toContain('empty content');
+    expect(recordSuccessMock).not.toHaveBeenCalled();
+    expect(recordFailureMock).toHaveBeenCalledTimes(1);
+  });
+
+  it('returns HTTP upstream_error when streamed chat SSE carries prompt usage but no assistant output', async () => {
+    config.proxyEmptyContentFailEnabled = true;
+
+    const encoder = new TextEncoder();
+    const upstreamBody = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(encoder.encode('data: {"id":"chatcmpl-empty-usage","choices":[{"delta":{}}],"usage":{"prompt_tokens":42203,"completion_tokens":0,"total_tokens":42203}}\n\n'));
+        controller.enqueue(encoder.encode('data: [DONE]\n\n'));
+        controller.close();
+      },
+    });
+
+    fetchMock.mockResolvedValue(new Response(upstreamBody, {
+      status: 200,
+      headers: { 'content-type': 'text/event-stream; charset=utf-8' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.4',
+        stream: true,
+        messages: [{ role: 'user', content: 'long prompt' }],
+      },
+    });
+
+    expect(response.statusCode).toBe(502);
+    expect(response.headers['content-type']).not.toContain('text/event-stream');
+    expect(response.json()?.error?.type).toBe('upstream_error');
+    expect(response.json()?.error?.message).toContain('empty content');
+    expect(recordSuccessMock).not.toHaveBeenCalled();
+    expect(recordFailureMock).toHaveBeenCalledTimes(1);
+  });
+
+  it('returns HTTP upstream_error when streamed chat SSE carries completion usage but no assistant output', async () => {
+    config.proxyEmptyContentFailEnabled = true;
+
+    const encoder = new TextEncoder();
+    const upstreamBody = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(encoder.encode('data: {"id":"chatcmpl-empty-completion-usage","choices":[{"delta":{}}],"usage":{"prompt_tokens":12,"completion_tokens":3,"total_tokens":15}}\n\n'));
+        controller.enqueue(encoder.encode('data: [DONE]\n\n'));
+        controller.close();
+      },
+    });
+
+    fetchMock.mockResolvedValue(new Response(upstreamBody, {
+      status: 200,
+      headers: { 'content-type': 'text/event-stream; charset=utf-8' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.4',
+        stream: true,
+        messages: [{ role: 'user', content: 'long prompt' }],
+      },
+    });
+
+    expect(response.statusCode).toBe(502);
+    expect(response.headers['content-type']).not.toContain('text/event-stream');
+    expect(response.json()?.error?.type).toBe('upstream_error');
+    expect(response.json()?.error?.message).toContain('empty content');
+    expect(recordSuccessMock).not.toHaveBeenCalled();
+    expect(recordFailureMock).toHaveBeenCalledTimes(1);
+  });
+
+  it('keeps streamed non-SSE chat fallback successful when the final payload has visible output but zero completion usage', async () => {
+    config.proxyEmptyContentFailEnabled = true;
+
+    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      id: 'chatcmpl-visible-zero-usage',
+      object: 'chat.completion',
+      created: 1_706_000_000,
+      model: 'upstream-gpt',
+      choices: [{
+        index: 0,
+        message: { role: 'assistant', content: 'visible answer despite zero output usage' },
+        finish_reason: 'stop',
+      }],
+      usage: { prompt_tokens: 11, completion_tokens: 0, total_tokens: 11 },
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.4',
+        stream: true,
+        messages: [{ role: 'user', content: 'hi' }],
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+    expect(response.body).toContain('visible answer despite zero output usage');
+    expect(response.body).toContain('data: [DONE]');
+    expect(recordSuccessMock).toHaveBeenCalledTimes(1);
+    expect(recordFailureMock).not.toHaveBeenCalled();
+  });
+
   it('returns clear 400 when /v1/chat/completions receives responses-style input without messages', async () => {
     const response = await app.inject({
       method: 'POST',

src/server/routes/proxy/logPathMeta.ts

@@ -1,4 +1,8 @@
-import { parseProxyLogMetadata, type ParsedProxyLogMetadata } from '../../../shared/proxyLogMeta.js';
+import {
+  parseProxyLogMetadata,
+  type ParsedProxyLogMetadata,
+  type ProxyLogUsageSource,
+} from '../../../shared/proxyLogMeta.js';
 
 type ComposeProxyLogMessageArgs = {
   clientKind?: string | null;
@@ -6,6 +10,7 @@ type ComposeProxyLogMessageArgs = {
   traceHint?: string | null;
   downstreamPath?: string | null;
   upstreamPath?: string | null;
+  usageSource?: ProxyLogUsageSource;
   errorMessage?: string | null;
 };
 
@@ -21,6 +26,7 @@ export function composeProxyLogMessage({
   traceHint,
   downstreamPath,
   upstreamPath,
+  usageSource,
   errorMessage,
 }: ComposeProxyLogMessageArgs): string | null {
   const rawMessage = typeof errorMessage === 'string' ? errorMessage.trim() : '';
@@ -29,6 +35,7 @@ export function composeProxyLogMessage({
   const finalSessionId = (sessionId || traceHint || parsed.sessionId || '').trim();
   const finalDownstreamPath = (downstreamPath || parsed.downstreamPath || '').trim();
   const finalUpstreamPath = (upstreamPath || parsed.upstreamPath || '').trim();
+  const finalUsageSource = (usageSource || parsed.usageSource || '').trim();
   const finalMessageText = parsed.messageText.trim();
 
   const prefixParts: string[] = [];
@@ -36,6 +43,7 @@ export function composeProxyLogMessage({
   if (finalSessionId) prefixParts.push(`[session:${finalSessionId}]`);
   if (finalDownstreamPath) prefixParts.push(`[downstream:${finalDownstreamPath}]`);
   if (finalUpstreamPath) prefixParts.push(`[upstream:${finalUpstreamPath}]`);
+  if (finalUsageSource) prefixParts.push(`[usage:${finalUsageSource}]`);
 
   if (prefixParts.length === 0 && !finalMessageText) return null;
   if (finalMessageText) return `${prefixParts.join(' ')} ${finalMessageText}`.trim();

src/server/routes/proxy/responses.codex-oauth.test.ts

@@ -7,6 +7,7 @@ import { resetCodexHttpSessionQueue } from '../../proxy-core/runtime/codexHttpSe
 const fetchMock = vi.fn();
 const selectChannelMock = vi.fn();
 const selectNextChannelMock = vi.fn();
+const selectPreferredChannelMock = vi.fn();
 const recordSuccessMock = vi.fn();
 const recordFailureMock = vi.fn();
 const refreshModelsAndRebuildRoutesMock = vi.fn();
@@ -41,6 +42,7 @@ vi.mock('../../services/tokenRouter.js', () => ({
   tokenRouter: {
     selectChannel: (...args: unknown[]) => selectChannelMock(...args),
     selectNextChannel: (...args: unknown[]) => selectNextChannelMock(...args),
+    selectPreferredChannel: (...args: unknown[]) => selectPreferredChannelMock(...args),
     recordSuccess: (...args: unknown[]) => recordSuccessMock(...args),
     recordFailure: (...args: unknown[]) => recordFailureMock(...args),
   },
@@ -157,6 +159,7 @@ describe('responses proxy codex oauth refresh', () => {
     fetchMock.mockReset();
     selectChannelMock.mockReset();
     selectNextChannelMock.mockReset();
+    selectPreferredChannelMock.mockReset();
     recordSuccessMock.mockReset();
     recordFailureMock.mockReset();
     refreshModelsAndRebuildRoutesMock.mockReset();
@@ -188,6 +191,7 @@ describe('responses proxy codex oauth refresh', () => {
       tokenValue: 'expired-access-token',
       actualModel: 'gpt-5.2-codex',
     });
+    selectPreferredChannelMock.mockReturnValue(null);
     selectNextChannelMock.mockReturnValue(null);
     refreshOauthAccessTokenSingleflightMock.mockResolvedValue({
       accessToken: 'fresh-access-token',
@@ -724,6 +728,77 @@ describe('responses proxy codex oauth refresh', () => {
     });
   });
 
+  it('rebinds sticky channels after websocket transport fast-path successes', async () => {
+    config.proxyStickySessionEnabled = true;
+
+    const selected = {
+      channel: { id: 11, routeId: 22 },
+      site: { name: 'codex-site', url: 'https://chatgpt.com/backend-api/codex', platform: 'codex' },
+      account: {
+        id: 33,
+        username: 'codex-user@example.com',
+        extraConfig: JSON.stringify({
+          credentialMode: 'session',
+          oauth: {
+            provider: 'codex',
+            accountId: 'chatgpt-account-123',
+            email: 'codex-user@example.com',
+            planType: 'plus',
+          },
+        }),
+      },
+      tokenName: 'default',
+      tokenValue: 'expired-access-token',
+      actualModel: 'gpt-5.2-codex',
+    };
+    selectChannelMock.mockReturnValue(selected);
+    selectPreferredChannelMock.mockReturnValue(selected);
+
+    const stickyHeaders = {
+      'x-metapi-responses-websocket-transport': '1',
+      session_id: 'session-sticky-fast-path-1',
+    };
+    const ssePayload = createSseResponse([
+      'event: response.created\n',
+      'data: {"type":"response.created","response":{"id":"resp_codex_fast_path","model":"gpt-5.4","created_at":1706000000,"status":"in_progress","output":[]}}\n\n',
+      'event: response.completed\n',
+      'data: {"type":"response.completed","response":{"id":"resp_codex_fast_path","model":"gpt-5.4","status":"completed","usage":{"input_tokens":3,"output_tokens":1,"total_tokens":4}}}\n\n',
+      'data: [DONE]\n\n',
+    ]);
+    fetchMock.mockResolvedValue(ssePayload);
+
+    const firstResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: stickyHeaders,
+      payload: {
+        model: 'gpt-5.4',
+        input: 'hello codex',
+        stream: true,
+      },
+    });
+
+    expect(firstResponse.statusCode).toBe(200);
+    expect(firstResponse.body).toContain('event: response.completed');
+    expect(selectPreferredChannelMock).not.toHaveBeenCalled();
+
+    const secondResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: stickyHeaders,
+      payload: {
+        model: 'gpt-5.4',
+        input: 'hello again',
+        stream: true,
+      },
+    });
+
+    expect(secondResponse.statusCode).toBe(200);
+    expect(selectPreferredChannelMock).toHaveBeenCalledTimes(1);
+    expect(selectPreferredChannelMock.mock.calls[0]?.[0]).toBe('gpt-5.4');
+    expect(selectPreferredChannelMock.mock.calls[0]?.[1]).toBe(11);
+  });
+
   it('decodes zstd-compressed codex responses SSE before relaying native downstream streams', async () => {
     fetchMock.mockResolvedValue(createCompressedSseResponse([
       'event: response.created\n',
@@ -887,6 +962,39 @@ describe('responses proxy codex oauth refresh', () => {
     expect(String(insertedProxyLogs.at(-1)?.errorMessage || '')).toContain('empty content');
   });
 
+  it('does not record success when a native responses stream completes with prompt tokens only and no output', async () => {
+    config.proxyEmptyContentFailEnabled = true;
+
+    fetchMock.mockResolvedValue(createSseResponse([
+      'event: response.created\n',
+      'data: {"type":"response.created","response":{"id":"resp_codex_prompt_only","model":"gpt-5.4","created_at":1706000000,"status":"in_progress","output":[]}}\n\n',
+      'event: response.completed\n',
+      'data: {"type":"response.completed","response":{"id":"resp_codex_prompt_only","model":"gpt-5.4","status":"completed","output":[],"output_text":"","usage":{"input_tokens":5,"output_tokens":0,"total_tokens":5}}}\n\n',
+      'data: [DONE]\n\n',
+    ]));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      payload: {
+        model: 'gpt-5.4',
+        input: 'hello codex',
+        stream: true,
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.body).toContain('event: response.failed');
+    expect(response.body).not.toContain('event: response.completed');
+    expect(recordSuccessMock).not.toHaveBeenCalled();
+    expect(recordFailureMock).toHaveBeenCalledTimes(1);
+    expect(insertedProxyLogs.at(-1)).toMatchObject({
+      status: 'failed',
+      httpStatus: 200,
+    });
+    expect(String(insertedProxyLogs.at(-1)?.errorMessage || '')).toContain('empty content');
+  });
+
   it('does not retry or mark failure after converting a non-stream upstream payload into SSE when post-stream usage accounting fails', async () => {
     resolveProxyUsageWithSelfLogFallbackMock.mockRejectedValueOnce(new Error('usage accounting failed'));
     fetchMock.mockResolvedValue(new Response(JSON.stringify({

src/server/services/proxyLogStore.test.ts

@@ -173,6 +173,23 @@ describe('proxyLogStore', () => {
     });
   });
 
+  it('preserves null token fields instead of coercing unknown usage to zero', async () => {
+    await insertProxyLog({
+      modelRequested: 'gpt-5',
+      promptTokens: null,
+      completionTokens: null,
+      totalTokens: null,
+    });
+
+    expect(dbInsertValuesMock).toHaveBeenCalledTimes(1);
+    expect(dbInsertValuesMock.mock.calls[0][0]).toMatchObject({
+      modelRequested: 'gpt-5',
+      promptTokens: null,
+      completionTokens: null,
+      totalTokens: null,
+    });
+  });
+
   it('retries proxy log inserts without structured client fields when those columns are missing', async () => {
     hasProxyLogClientColumnsMock.mockResolvedValue(true);
     dbInsertRunMock

src/server/services/proxyLogStore.ts

@@ -214,9 +214,9 @@ export async function insertProxyLog(input: ProxyLogInsertInput): Promise<void>
     status: input.status ?? null,
     httpStatus: input.httpStatus ?? null,
     latencyMs: input.latencyMs ?? null,
-    promptTokens: input.promptTokens ?? 0,
-    completionTokens: input.completionTokens ?? 0,
-    totalTokens: input.totalTokens ?? 0,
+    promptTokens: input.promptTokens ?? null,
+    completionTokens: input.completionTokens ?? null,
+    totalTokens: input.totalTokens ?? null,
     estimatedCost: input.estimatedCost ?? 0,
     errorMessage: input.errorMessage ?? null,
     retryCount: input.retryCount ?? 0,

src/server/services/proxyUsageFallbackService.sub2api.test.ts

@@ -129,4 +129,36 @@ describe('proxyUsageFallbackService sub2api', () => {
       estimatedCostFromQuota: 0.143004,
     });
   });
+
+  it('treats explicit zero upstream usage as upstream instead of unknown', async () => {
+    const result = await resolveProxyUsageWithSelfLogFallback({
+      site: {
+        url: 'https://newapi.example.com',
+        platform: 'new-api',
+      },
+      account: {
+        accessToken: 'jwt-access-token',
+      },
+      modelName: 'gpt-5.4',
+      requestStartedAtMs: Date.parse('2026-03-07T06:12:07.000Z'),
+      requestEndedAtMs: Date.parse('2026-03-07T06:12:30.000Z'),
+      localLatencyMs: 23000,
+      upstreamUsagePresent: true,
+      usage: {
+        promptTokens: 0,
+        completionTokens: 0,
+        totalTokens: 0,
+      },
+    });
+
+    expect(result).toMatchObject({
+      promptTokens: 0,
+      completionTokens: 0,
+      totalTokens: 0,
+      recoveredFromSelfLog: false,
+      usageSource: 'upstream',
+    });
+    expect(fetchMock).not.toHaveBeenCalled();
+    expect(fetchJsonWithShieldCookieRetryMock).not.toHaveBeenCalled();
+  });
 });

src/server/services/proxyUsageFallbackService.ts

@@ -44,6 +44,7 @@ interface ProxyUsageFallbackInput {
   requestStartedAtMs: number;
   requestEndedAtMs: number;
   localLatencyMs: number;
+  upstreamUsagePresent?: boolean;
   usage: ProxyUsage;
 }
 
@@ -51,6 +52,7 @@ interface ProxyUsageFallbackResult extends ProxyUsage {
   recoveredFromSelfLog: boolean;
   estimatedCostFromQuota: number;
   selfLogBillingMeta: SelfLogBillingMeta | null;
+  usageSource: 'upstream' | 'self-log' | 'unknown';
 }
 
 export interface SelfLogBillingMeta {
@@ -149,11 +151,14 @@ function isUsageMissing(usage: ProxyUsage): boolean {
 export function shouldLookupSelfLog(
   platform: string,
   usage: ProxyUsage,
+  upstreamUsagePresent?: boolean,
 ): boolean {
   const normalizedPlatform = String(platform || '').toLowerCase();
   if (!SUPPORTED_USAGE_FALLBACK_PLATFORMS.has(normalizedPlatform)) return false;
   if (ALWAYS_LOOKUP_SELF_LOG_PLATFORMS.has(normalizedPlatform)) return true;
-  return isUsageMissing(usage);
+  const normalizedUsage = normalizeUsage(usage);
+  const hasUpstreamUsage = upstreamUsagePresent ?? !isUsageMissing(normalizedUsage);
+  return !hasUpstreamUsage;
 }
 
 function normalizeUrl(url: string): string {
@@ -534,15 +539,18 @@ export async function resolveProxyUsageWithSelfLogFallback(
   input: ProxyUsageFallbackInput,
 ): Promise<ProxyUsageFallbackResult> {
   const normalizedUsage = normalizeUsage(input.usage);
+  const hasUpstreamUsage = input.upstreamUsagePresent ?? !isUsageMissing(normalizedUsage);
+  const fallbackUsageSource = hasUpstreamUsage ? 'upstream' : 'unknown';
   const fallback: ProxyUsageFallbackResult = {
     ...normalizedUsage,
     recoveredFromSelfLog: false,
     estimatedCostFromQuota: 0,
     selfLogBillingMeta: null,
+    usageSource: fallbackUsageSource,
   };
 
   const platform = String(input.site.platform || '').toLowerCase();
-  if (!shouldLookupSelfLog(platform, normalizedUsage)) {
+  if (!shouldLookupSelfLog(platform, normalizedUsage, hasUpstreamUsage)) {
     return fallback;
   }
 
@@ -577,6 +585,7 @@ export async function resolveProxyUsageWithSelfLogFallback(
       recoveredFromSelfLog: true,
       estimatedCostFromQuota: toRecoveredCost(matched),
       selfLogBillingMeta: matched.billingMeta,
+      usageSource: useMatchedTokens ? 'self-log' : fallbackUsageSource,
     };
   } catch {
     return fallback;

src/server/services/proxyUsageParser.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it } from 'vitest';
 import {
+  hasProxyUsagePayload,
   mergeProxyUsage,
   parseProxyUsage,
   pullSseDataEvents,
@@ -160,6 +161,30 @@ describe('proxyUsageParser', () => {
     });
   });
 
+  it('does not treat null placeholders as explicit upstream usage', () => {
+    expect(hasProxyUsagePayload({
+      usage: {
+        total_tokens: null,
+        cache_creation: {},
+      },
+    })).toBe(false);
+  });
+
+  it('treats explicit zero usage values as upstream usage', () => {
+    expect(hasProxyUsagePayload({
+      usage: {
+        total_tokens: 0,
+      },
+    })).toBe(true);
+    expect(hasProxyUsagePayload({
+      usage: {
+        prompt_tokens_details: {
+          cached_tokens: 0,
+        },
+      },
+    })).toBe(true);
+  });
+
   it('pulls SSE data events across chunk boundaries', () => {
     const first = pullSseDataEvents('data: {"a":1}\n\ndata: {"b":');
     expect(first.events).toEqual(['{"a":1}']);

src/server/services/proxyUsageParser.ts

@@ -16,6 +16,60 @@ const ZERO_USAGE: ParsedProxyUsage = {
   promptTokensIncludeCache: null,
 };
 
+const USAGE_DIRECT_KEYS = [
+  'prompt_tokens',
+  'promptTokens',
+  'prompt_token_count',
+  'promptTokenCount',
+  'input_tokens',
+  'inputTokens',
+  'input_token_count',
+  'inputTokenCount',
+  'completion_tokens',
+  'completionTokens',
+  'completion_token_count',
+  'completionTokenCount',
+  'candidates_token_count',
+  'candidatesTokenCount',
+  'output_tokens',
+  'outputTokens',
+  'output_token_count',
+  'outputTokenCount',
+  'total_tokens',
+  'totalTokens',
+  'total_token_count',
+  'totalTokenCount',
+  'cache_read_input_tokens',
+  'cacheReadInputTokens',
+  'prompt_cache_hit_tokens',
+  'promptCacheHitTokens',
+  'cached_tokens',
+  'cachedTokens',
+  'cache_read_tokens',
+  'cacheReadTokens',
+  'cache_creation_input_tokens',
+  'cacheCreationInputTokens',
+  'cache_creation_tokens',
+  'cacheCreationTokens',
+  'claude_cache_creation_5_m_tokens',
+  'claudeCacheCreation5mTokens',
+  'claude_cache_creation_1_h_tokens',
+  'claudeCacheCreation1hTokens',
+] as const;
+
+const USAGE_DETAIL_KEYS = [
+  'prompt_tokens_details',
+  'promptTokensDetails',
+  'input_tokens_details',
+  'inputTokensDetails',
+  'completion_tokens_details',
+  'completionTokensDetails',
+  'output_tokens_details',
+  'outputTokensDetails',
+  'cache_creation',
+  'cacheCreation',
+] as const;
+
 function toPositiveInt(value: unknown): number {
   const n = typeof value === 'number' ? value : Number(value);
   if (!Number.isFinite(n)) return 0;
@@ -26,6 +80,24 @@ function isRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === 'object';
 }
 
+function hasOwn(record: Record<string, unknown>, key: string): boolean {
+  return Object.prototype.hasOwnProperty.call(record, key);
+}
+
+function hasExplicitUsageValue(value: unknown): boolean {
+  if (value === null || value === undefined) return false;
+  if (typeof value === 'number') return Number.isFinite(value);
+  if (typeof value === 'string') {
+    const trimmed = value.trim();
+    return trimmed.length > 0 && Number.isFinite(Number(trimmed));
+  }
+  if (Array.isArray(value)) {
+    return value.some((entry) => hasExplicitUsageValue(entry));
+  }
+  if (!isRecord(value)) return false;
+  return Object.values(value).some((entry) => hasExplicitUsageValue(entry));
+}
+
 function collectUsageCandidates(payload: unknown): Array<Record<string, unknown>> {
   const candidates: Array<Record<string, unknown>> = [];
   const visited = new Set<object>();
@@ -253,6 +325,15 @@ export function parseProxyUsage(payload: unknown): ParsedProxyUsage {
   return best;
 }
 
+export function hasProxyUsagePayload(payload: unknown): boolean {
+  if (!payload || typeof payload !== 'object') return false;
+  const candidates = collectUsageCandidates(payload);
+  return candidates.some((candidate) => (
+    USAGE_DIRECT_KEYS.some((key) => hasOwn(candidate, key) && hasExplicitUsageValue(candidate[key]))
+    || USAGE_DETAIL_KEYS.some((key) => hasOwn(candidate, key) && hasExplicitUsageValue(candidate[key]))
+  ));
+}
+
 export function mergeProxyUsage(base: ParsedProxyUsage, incoming: ParsedProxyUsage): ParsedProxyUsage {
   const normalizeUsage = (usage: ParsedProxyUsage): ParsedProxyUsage => ({
     promptTokens: toPositiveInt(usage.promptTokens),

src/server/transformers/openai/chat/proxyStream.ts

@@ -4,6 +4,7 @@ import { type DownstreamFormat, type ParsedSseEvent } from '../../shared/normali
 import { createOpenAiChatAggregateState, applyOpenAiChatStreamEvent, finalizeOpenAiChatAggregate } from './aggregator.js';
 import { openAiChatOutbound } from './outbound.js';
 import { openAiChatStream } from './stream.js';
+import { config } from '../../../config.js';
 
 type StreamReader = {
   read(): Promise<{ done: boolean; value?: Uint8Array }>;
@@ -49,6 +50,9 @@ export function createChatProxyStreamSession(input: ChatProxyStreamSessionInput)
     status: 'completed',
     errorMessage: null,
   };
+  let terminalNormalizedFinal: ReturnType<typeof openAiChatOutbound.normalizeFinal> | null = null;
+  let forwardedDownstreamOutput = false;
+  const pendingWrites: string[] = [];
 
   const extractFailureMessage = (payload: unknown, fallback = 'upstream stream failed'): string => {
     if (payload && typeof payload === 'object' && !Array.isArray(payload)) {
@@ -76,10 +80,116 @@ export function createChatProxyStreamSession(input: ChatProxyStreamSessionInput)
     };
   };
 
+  const hasMeaningfulChatAggregateOutput = (): boolean => {
+    if (input.downstreamFormat !== 'openai' || !chatAggregateState) return false;
+    for (const choice of chatAggregateState.choices.values()) {
+      if (choice.content.length > 0) return true;
+      if (choice.reasoning.length > 0) return true;
+      if (choice.toolCalls.some((item) => item.id || item.name || item.arguments)) return true;
+    }
+    return false;
+  };
+
+  const hasMeaningfulNormalizedFinalOutput = (): boolean => {
+    if (!terminalNormalizedFinal) return false;
+    const choices = Array.isArray(terminalNormalizedFinal.choices)
+      ? terminalNormalizedFinal.choices
+      : [];
+    if (choices.some((choice) => (
+      choice.content.length > 0
+      || choice.reasoningContent.length > 0
+      || choice.toolCalls.some((toolCall) => toolCall.id || toolCall.name || toolCall.arguments)
+    ))) {
+      return true;
+    }
+    if (terminalNormalizedFinal.content.length > 0) return true;
+    if (terminalNormalizedFinal.reasoningContent.length > 0) return true;
+    return terminalNormalizedFinal.toolCalls.some((toolCall) => toolCall.id || toolCall.name || toolCall.arguments);
+  };
+
+  const flushPendingWrites = () => {
+    if (pendingWrites.length <= 0) return;
+    input.writeLines([...pendingWrites]);
+    pendingWrites.length = 0;
+  };
+
+  const emitLines = (lines: string[], options?: { meaningful?: boolean; force?: boolean }) => {
+    if (lines.length <= 0) return;
+    if (input.downstreamFormat !== 'openai') {
+      input.writeLines(lines);
+      return;
+    }
+    if (forwardedDownstreamOutput) {
+      input.writeLines(lines);
+      return;
+    }
+    if (options?.force) {
+      pendingWrites.length = 0;
+      forwardedDownstreamOutput = true;
+      input.writeLines(lines);
+      return;
+    }
+    if (options?.meaningful) {
+      forwardedDownstreamOutput = true;
+      flushPendingWrites();
+      input.writeLines(lines);
+      return;
+    }
+    pendingWrites.push(...lines);
+  };
+
+  const emitRaw = (chunk: string, options?: { meaningful?: boolean; force?: boolean }) => {
+    if (!chunk) return;
+    if (input.downstreamFormat !== 'openai') {
+      input.writeRaw(chunk);
+      return;
+    }
+    if (forwardedDownstreamOutput) {
+      input.writeRaw(chunk);
+      return;
+    }
+    if (options?.force) {
+      pendingWrites.length = 0;
+      forwardedDownstreamOutput = true;
+      input.writeRaw(chunk);
+      return;
+    }
+    if (options?.meaningful) {
+      forwardedDownstreamOutput = true;
+      flushPendingWrites();
+      input.writeRaw(chunk);
+      return;
+    }
+    pendingWrites.push(chunk);
+  };
+
+  const shouldFailEmptyChatCompletion = (): boolean => {
+    if (!config.proxyEmptyContentFailEnabled) return false;
+    if (input.downstreamFormat !== 'openai') return false;
+    if (terminalResult.status === 'failed') return false;
+    if (hasMeaningfulChatAggregateOutput()) return false;
+    if (hasMeaningfulNormalizedFinalOutput()) return false;
+    return true;
+  };
+
   const finalize = () => {
     if (finalized) return;
     finalized = true;
 
+    if (shouldFailEmptyChatCompletion()) {
+      markFailed({
+        error: {
+          message: 'Upstream returned empty content',
+        },
+      }, 'Upstream returned empty content');
+      return;
+    }
+
+    if (input.downstreamFormat === 'openai' && !forwardedDownstreamOutput) {
+      forwardedDownstreamOutput = true;
+      flushPendingWrites();
+    }
+
     // For native Anthropic streams, EOF without message_stop is not a clean
     // completion. Forward the partial stream as-is instead of fabricating an
     // end_turn/message_stop pair that makes clients think the run finished.
@@ -108,12 +218,12 @@ export function createChatProxyStreamSession(input: ChatProxyStreamSessionInput)
           }),
         ).slice(-1)[0];
         if (terminalChunk) {
-          input.writeLines([`data: ${JSON.stringify(terminalChunk)}\n\n`]);
+          emitLines([`data: ${JSON.stringify(terminalChunk)}\n\n`], { meaningful: true });
         }
       }
     }
 
-    input.writeLines(downstreamTransformer.serializeDone(streamContext, claudeContext));
+    emitLines(downstreamTransformer.serializeDone(streamContext, claudeContext), { meaningful: true });
   };
 
   const handleEventBlock = async (eventBlock: ParsedSseEvent): Promise<boolean> => {
@@ -153,19 +263,26 @@ export function createChatProxyStreamSession(input: ChatProxyStreamSessionInput)
       const payloadType = typeof (parsedPayload as Record<string, unknown>).type === 'string'
         ? String((parsedPayload as Record<string, unknown>).type)
         : '';
-      if (payloadType === 'response.failed' || payloadType === 'error') {
+      const isFailurePayload = payloadType === 'response.failed' || payloadType === 'error';
+      if (isFailurePayload) {
         markFailed(parsedPayload);
       }
       const normalizedEvent = downstreamTransformer.transformStreamEvent(parsedPayload, streamContext, input.modelName);
       if (input.downstreamFormat === 'openai' && chatAggregateState) {
         applyOpenAiChatStreamEvent(chatAggregateState, normalizedEvent);
       }
-      input.writeLines(downstreamTransformer.serializeStreamEvent(normalizedEvent, streamContext, claudeContext));
+      emitLines(
+        downstreamTransformer.serializeStreamEvent(normalizedEvent, streamContext, claudeContext),
+        {
+          meaningful: hasMeaningfulChatAggregateOutput(),
+          force: isFailurePayload,
+        },
+      );
       return input.downstreamFormat === 'claude' && claudeContext.doneSent;
     }
 
     if (input.downstreamFormat === 'openai') {
-      input.writeRaw(`data: ${eventBlock.data}\n\n`);
+      emitRaw(`data: ${eventBlock.data}\n\n`, { meaningful: true });
       return false;
     }
 
@@ -190,16 +307,18 @@ export function createChatProxyStreamSession(input: ChatProxyStreamSessionInput)
       }
       if (input.downstreamFormat === 'openai') {
         const normalizedFinal = openAiChatOutbound.normalizeFinal(payload, input.modelName, fallbackText);
+        terminalNormalizedFinal = normalizedFinal;
         streamContext.id = normalizedFinal.id;
         streamContext.model = normalizedFinal.model;
         streamContext.created = normalizedFinal.created;
-        input.writeLines(
+        emitLines(
           openAiChatOutbound
             .buildSyntheticChunks(normalizedFinal)
             .map((chunk) => `data: ${JSON.stringify(chunk)}\n\n`),
+          { meaningful: true },
         );
       } else {
-        input.writeLines(
+        emitLines(
           anthropicMessagesTransformer.serializeUpstreamFinalAsStream(
             payload,
             input.modelName,
@@ -207,6 +326,7 @@ export function createChatProxyStreamSession(input: ChatProxyStreamSessionInput)
             streamContext,
             claudeContext,
           ),
+          { meaningful: true },
         );
       }
       finalize();

src/server/transformers/openai/responses/proxyStream.ts

@@ -98,7 +98,7 @@ function shouldFailEmptyResponsesCompletion(input: {
     : null;
   if (hasMeaningfulAggregateOutput(input.state)) return false;
   if (responsePayload && hasMeaningfulResponsesPayloadOutput(responsePayload)) return false;
-  return input.usage.completionTokens <= 0 && input.usage.totalTokens <= 0;
+  return input.usage.completionTokens <= 0;
 }
 
 function getResponsesStreamFailureMessage(payload: unknown, fallback = 'upstream stream failed'): string {

src/shared/proxyLogMeta.d.ts

@@ -1,8 +1,11 @@
+export type ProxyLogUsageSource = 'upstream' | 'self-log' | 'unknown' | null;
+
 export type ParsedProxyLogMetadata = {
   clientKind: string | null;
   sessionId: string | null;
   downstreamPath: string | null;
   upstreamPath: string | null;
+  usageSource: ProxyLogUsageSource;
   messageText: string;
 };
 

src/shared/proxyLogMeta.js

@@ -3,8 +3,9 @@ export function parseProxyLogMetadata(rawMessage) {
   const sessionMatch = rawMessage.match(/\[session:([^\]]+)\]/i);
   const downstreamMatch = rawMessage.match(/\[downstream:([^\]]+)\]/i);
   const upstreamMatch = rawMessage.match(/\[upstream:([^\]]+)\]/i);
+  const usageMatch = rawMessage.match(/\[usage:([^\]]+)\]/i);
   const messageText = rawMessage.replace(
-    /^\s*(?:\[(?:client|session|downstream|upstream):[^\]]+\]\s*)+/i,
+    /^\s*(?:\[(?:client|session|downstream|upstream|usage):[^\]]+\]\s*)+/i,
     '',
   ).trim();
 
@@ -13,6 +14,7 @@ export function parseProxyLogMetadata(rawMessage) {
     sessionId: sessionMatch?.[1]?.trim() || null,
     downstreamPath: downstreamMatch?.[1]?.trim() || null,
     upstreamPath: upstreamMatch?.[1]?.trim() || null,
+    usageSource: usageMatch?.[1]?.trim() || null,
     messageText,
   };
 }

src/shared/proxyLogMeta.test.ts

@@ -5,12 +5,13 @@ import { parseProxyLogMetadata } from './proxyLogMeta.js';
 describe('proxyLogMeta', () => {
   it('parses all supported proxy log prefixes', () => {
     expect(parseProxyLogMetadata(
-      '[client:codex] [session:turn-1] [downstream:/v1/responses] [upstream:/responses] boom',
+      '[client:codex] [session:turn-1] [downstream:/v1/responses] [upstream:/responses] [usage:self-log] boom',
     )).toEqual({
       clientKind: 'codex',
       sessionId: 'turn-1',
       downstreamPath: '/v1/responses',
       upstreamPath: '/responses',
+      usageSource: 'self-log',
       messageText: 'boom',
     });
   });
@@ -21,6 +22,7 @@ describe('proxyLogMeta', () => {
       sessionId: null,
       downstreamPath: null,
       upstreamPath: null,
+      usageSource: null,
       messageText: 'network timeout',
     });
   });
@@ -31,6 +33,7 @@ describe('proxyLogMeta', () => {
       sessionId: 'turn-1',
       downstreamPath: null,
       upstreamPath: null,
+      usageSource: null,
       messageText: 'boom',
     });
 
@@ -39,6 +42,7 @@ describe('proxyLogMeta', () => {
       sessionId: null,
       downstreamPath: null,
       upstreamPath: null,
+      usageSource: null,
       messageText: 'boom',
     });
 
@@ -47,6 +51,7 @@ describe('proxyLogMeta', () => {
       sessionId: '1',
       downstreamPath: null,
       upstreamPath: '/x',
+      usageSource: null,
       messageText: 'msg',
     });
 
@@ -55,6 +60,7 @@ describe('proxyLogMeta', () => {
       sessionId: null,
       downstreamPath: null,
       upstreamPath: null,
+      usageSource: null,
       messageText: '',
     });
   });

src/web/api.ts

@@ -393,6 +393,7 @@ export type RuntimeSettingsPayload = {
 
 export type ProxyLogStatusFilter = 'all' | 'success' | 'failed';
 export type ProxyLogClientConfidence = 'exact' | 'heuristic' | 'unknown' | null;
+export type ProxyLogUsageSource = 'upstream' | 'self-log' | 'unknown' | null;
 
 export type ProxyLogBillingDetails = {
   quotaType: number;
@@ -448,6 +449,7 @@ export type ProxyLogListItem = {
   clientAppId?: string | null;
   clientAppName?: string | null;
   clientConfidence?: ProxyLogClientConfidence;
+  usageSource?: ProxyLogUsageSource;
   promptTokens?: number | null;
   completionTokens?: number | null;
   estimatedCost?: number | null;

src/web/pages/ProxyLogs.server-driven.test.tsx

@@ -955,6 +955,66 @@ describe('ProxyLogs server-driven page', () => {
     }
   });
 
+  it('renders unknown usage as -- instead of 0 in the server-driven table', async () => {
+    apiMock.getProxyLogs.mockResolvedValue(buildListResponse({
+      items: [
+        {
+          id: 101,
+          createdAt: '2026-03-09 16:00:00',
+          modelRequested: 'gpt-5',
+          modelActual: 'gpt-5',
+          status: 'success',
+          latencyMs: 120,
+          promptTokens: null,
+          completionTokens: null,
+          totalTokens: null,
+          usageSource: 'unknown',
+          retryCount: 0,
+          estimatedCost: 0,
+          errorMessage: '[downstream:/v1/chat/completions] [upstream:/v1/chat/completions] [usage:unknown]',
+          username: 'tester',
+          siteName: 'main-site',
+          siteUrl: 'https://main-site.example.com',
+          clientFamily: 'codex',
+          clientAppId: 'cherry_studio',
+          clientAppName: 'Cherry Studio',
+          clientConfidence: 'heuristic',
+        },
+      ],
+      summary: {
+        totalCount: 1,
+        successCount: 1,
+        failedCount: 0,
+        totalCost: 0,
+        totalTokensAll: 0,
+      },
+    }));
+
+    let root!: WebTestRenderer;
+
+    try {
+      await act(async () => {
+        root = create(
+          <MemoryRouter initialEntries={['/logs']}>
+            <ToastProvider>
+              <ProxyLogs />
+            </ToastProvider>
+          </MemoryRouter>,
+        );
+      });
+      await flushMicrotasks();
+
+      const row = root!.root.find((node) => (
+        node.type === 'tr' && node.props['data-testid'] === 'proxy-log-row-101'
+      ));
+      const rowText = collectText(row);
+      expect(rowText).toContain('--');
+      expect(rowText).not.toContain('输入0');
+    } finally {
+      root?.unmount();
+    }
+  });
+
   it('hydrates site and time filters from the route query', async () => {
     let root!: WebTestRenderer;
 

src/web/pages/ProxyLogs.tsx

@@ -11,6 +11,7 @@ import {
   type ProxyLogListItem,
   type ProxyLogsSummary,
   type ProxyLogStatusFilter,
+  type ProxyLogUsageSource,
 } from '../api.js';
 import { useToast } from '../components/Toast.js';
 import { ModelBadge } from '../components/BrandIcon.js';
@@ -280,6 +281,17 @@ function formatBillingDetailSummary(log: ProxyLogRenderItem) {
   return `模型倍率 ${formatCompactNumber(detail.pricing.modelRatio)}，输出倍率 ${formatCompactNumber(detail.pricing.completionRatio)}，缓存倍率 ${formatCompactNumber(detail.pricing.cacheRatio)}，缓存创建倍率 ${formatCompactNumber(detail.pricing.cacheCreationRatio)}，分组倍率 ${formatCompactNumber(detail.pricing.groupRatio)}`;
 }
 
+function formatProxyLogUsageSource(source: ProxyLogUsageSource | undefined): string | null {
+  if (source === 'upstream') return '上游返回';
+  if (source === 'self-log') return '站点日志回填';
+  if (source === 'unknown') return '未知';
+  return null;
+}
+
+function formatProxyLogTokenValue(value: number | null | undefined): string {
+  return typeof value === 'number' ? value.toLocaleString() : '--';
+}
+
 function renderDownstreamKeySummary(log: ProxyLogRenderItem) {
   const parts = [
     log.downstreamKeyName ? `下游 Key: ${log.downstreamKeyName}` : null,
@@ -1905,13 +1917,13 @@ export default function ProxyLogs() {
                       <div className="mobile-summary-metric-label">用时</div>
                       <div className="mobile-summary-metric-value">{formatLatency(log.latencyMs)}</div>
                     </div>
-                    <div className="mobile-summary-metric">
-                      <div className="mobile-summary-metric-label">输入</div>
-                      <div className="mobile-summary-metric-value">{log.promptTokens?.toLocaleString() || '-'}</div>
+                      <div className="mobile-summary-metric">
+                        <div className="mobile-summary-metric-label">输入</div>
+                      <div className="mobile-summary-metric-value">{formatProxyLogTokenValue(log.promptTokens)}</div>
                     </div>
                     <div className="mobile-summary-metric">
                       <div className="mobile-summary-metric-label">输出</div>
-                      <div className="mobile-summary-metric-value">{log.completionTokens?.toLocaleString() || '-'}</div>
+                      <div className="mobile-summary-metric-value">{formatProxyLogTokenValue(log.completionTokens)}</div>
                     </div>
                     <div className="mobile-summary-metric">
                       <div className="mobile-summary-metric-label">花费</div>
@@ -1923,6 +1935,7 @@ export default function ProxyLogs() {
                       <MobileField label="时间" value={formatDateTimeLocal(log.createdAt)} />
                       <MobileField label="站点" value={<SiteBadgeLink siteId={siteIdByName.get(String(log.siteName || '').trim())} siteName={log.siteName} badgeStyle={{ fontSize: 11 }} />} />
                       <MobileField label="重试" value={log.retryCount > 0 ? log.retryCount : 0} />
+                      <MobileField label="用量来源" value={formatProxyLogUsageSource(detailLog.usageSource ?? pathMeta.usageSource) || '--'} />
                       {detailState?.loading && <div style={{ color: 'var(--color-text-muted)' }}>加载详情中...</div>}
                       {detailState?.error && <div style={{ color: 'var(--color-danger)' }}>{detailState.error}</div>}
                       {billingDetailSummary && <div style={{ color: 'var(--color-text-muted)' }}>{billingDetailSummary}</div>}
@@ -2033,10 +2046,10 @@ export default function ProxyLogs() {
                         </span>
                       </td>
                       <td style={{ textAlign: 'right', fontSize: 12, fontVariantNumeric: 'tabular-nums', color: 'var(--color-text-secondary)' }}>
-                        {log.promptTokens?.toLocaleString() || '-'}
+                        {formatProxyLogTokenValue(log.promptTokens)}
                       </td>
                       <td style={{ textAlign: 'right', fontSize: 12, fontVariantNumeric: 'tabular-nums', color: 'var(--color-text-secondary)' }}>
-                        {log.completionTokens?.toLocaleString() || '-'}
+                        {formatProxyLogTokenValue(log.completionTokens)}
                       </td>
                       <td style={{ textAlign: 'right', fontSize: 12, fontVariantNumeric: 'tabular-nums', fontWeight: 500 }}>
                         {typeof log.estimatedCost === 'number' ? `$${log.estimatedCost.toFixed(6)}` : '-'}
@@ -2084,6 +2097,9 @@ export default function ProxyLogs() {
                                     {billingDetailSummary && (
                                       <div style={{ color: 'var(--color-text-muted)' }}>{billingDetailSummary}</div>
                                     )}
+                                    <div style={{ color: 'var(--color-text-muted)' }}>
+                                      用量来源：{formatProxyLogUsageSource(detailLog.usageSource ?? pathMeta.usageSource) || '未知'}
+                                    </div>
                                     <div style={{ display: 'flex', gap: 6, alignItems: 'flex-start' }}>
                                       <span style={{ color: 'var(--color-text-muted)', flexShrink: 0 }}>客户端</span>
                                       <div style={{ minWidth: 0 }}>
@@ -2121,9 +2137,9 @@ export default function ProxyLogs() {
                                     </div>
                                   ) : (
                                     <span>
-                                      输入 {detailLog.promptTokens?.toLocaleString() || 0} tokens
-                                      {' + '}输出 {detailLog.completionTokens?.toLocaleString() || 0} tokens
-                                      {' = '}总计 {detailLog.totalTokens?.toLocaleString() || 0} tokens
+                                      输入 {formatProxyLogTokenValue(detailLog.promptTokens)} tokens
+                                      {' + '}输出 {formatProxyLogTokenValue(detailLog.completionTokens)} tokens
+                                      {' = '}总计 {formatProxyLogTokenValue(detailLog.totalTokens)} tokens
                                       {typeof detailLog.estimatedCost === 'number' && (
                                         <>，预估费用 <strong style={{ color: 'var(--color-text-primary)' }}>${detailLog.estimatedCost.toFixed(6)}</strong></>
                                       )}

src/web/pages/helpers/proxyLogPathMeta.ts

@@ -1,10 +1,11 @@
-import { parseProxyLogMetadata } from '../../../shared/proxyLogMeta.js';
+import { parseProxyLogMetadata, type ParsedProxyLogMetadata } from '../../../shared/proxyLogMeta.js';
 
 type ProxyLogPathMeta = {
   clientFamily: string | null;
   sessionId: string | null;
   downstreamPath: string | null;
   upstreamPath: string | null;
+  usageSource: ParsedProxyLogMetadata['usageSource'];
   errorMessage: string;
 };
 
@@ -17,6 +18,7 @@ export function parseProxyLogPathMeta(message?: string): ProxyLogPathMeta {
     sessionId: parsed.sessionId,
     downstreamPath: parsed.downstreamPath,
     upstreamPath: parsed.upstreamPath,
+    usageSource: parsed.usageSource,
     errorMessage: parsed.messageText,
   };
 }