Merge remote-tracking branch 'origin/main' into feat/account-level-proxy

# Conflicts: # src/server/proxy-core/surfaces/chatSurface.ts # src/server/proxy-core/surfaces/openAiResponsesSurface.ts # src/server/routes/proxy/embeddings.ts # src/server/routes/proxy/images.ts # src/server/services/modelService.ts
2026-03-21 10:09:19 +08:00
parent 382c6b22d8 4ba7749391
commit 8df3b1431f
198 changed files with 20039 additions and 1105 deletions
@@ -1,7 +1,6 @@
 import { TextDecoder } from 'node:util';
 import type { FastifyRequest, FastifyReply } from 'fastify';
 import { tokenRouter } from '../../services/tokenRouter.js';
-import { db, hasProxyLogDownstreamApiKeyIdColumn, schema } from '../../db/index.js';
 import { refreshModelsAndRebuildRoutes } from '../../services/modelService.js';
 import { reportProxyAllFailed, reportTokenExpired } from '../../services/alertService.js';
 import { isTokenExpiredError } from '../../services/alertRules.js';
@@ -13,6 +12,8 @@ import { type DownstreamFormat } from '../../transformers/shared/normalized.js';
 import {
  buildClaudeCountTokensUpstreamRequest,
  buildUpstreamEndpointRequest,
+  recordUpstreamEndpointFailure,
+  recordUpstreamEndpointSuccess,
  resolveUpstreamEndpointCandidates,
 } from '../../routes/proxy/upstreamEndpoint.js';
 import {
@@ -23,6 +24,7 @@ import {
 import { composeProxyLogMessage } from '../../routes/proxy/logPathMeta.js';
 import { executeEndpointFlow, type BuiltEndpointRequest } from '../../routes/proxy/endpointFlow.js';
 import { detectProxyFailure } from '../../routes/proxy/proxyFailureJudge.js';
+import { buildUpstreamUrl } from '../../routes/proxy/upstreamUrl.js';
 import { formatUtcSqlDateTime } from '../../services/localTimeService.js';
 import { resolveProxyLogBilling } from '../../routes/proxy/proxyBilling.js';
 import { openAiChatTransformer } from '../../transformers/openai/chat/index.js';
@@ -30,7 +32,6 @@ import { anthropicMessagesTransformer } from '../../transformers/anthropic/messa
 import { getProxyAuthContext, getProxyResourceOwner } from '../../middleware/auth.js';
 import {
  ProxyInputFileResolutionError,
-  hasNonImageFileInputInOpenAiBody,
  resolveOpenAiBodyInputFiles,
 } from '../../services/proxyInputFileResolver.js';
 import {
@@ -39,14 +40,20 @@ import {
 import { getOauthInfoFromExtraConfig } from '../../services/oauth/oauthAccount.js';
 import { recordOauthQuotaResetHint } from '../../services/oauth/quota.js';
 import { refreshOauthAccessTokenSingleflight } from '../../services/oauth/refreshSingleflight.js';
-import { collectResponsesFinalPayloadFromSse } from '../../routes/proxy/responsesSseFinal.js';
+import {
+  collectResponsesFinalPayloadFromSse,
+  collectResponsesFinalPayloadFromSseText,
+  createSingleChunkStreamReader,
+  looksLikeResponsesSseText,
+} from '../../routes/proxy/responsesSseFinal.js';
 import {
  createGeminiCliStreamReader,
  unwrapGeminiCliPayload,
 } from '../../routes/proxy/geminiCliCompat.js';
 import { dispatchRuntimeRequest } from '../../routes/proxy/runtimeExecutor.js';
-import { detectCliProfile } from '../cliProfiles/registry.js';
-import type { CliProfileId } from '../cliProfiles/types.js';
+import { summarizeConversationFileInputsInOpenAiBody } from '../capabilities/conversationFileCapabilities.js';
+import { detectDownstreamClientContext, type DownstreamClientContext } from '../../routes/proxy/downstreamClientContext.js';
+import { insertProxyLog } from '../../services/proxyLogStore.js';

 const MAX_RETRIES = 2;

@@ -58,25 +65,6 @@ function asTrimmedString(value: unknown): string {
  return typeof value === 'string' ? value.trim() : '';
 }

-type DownstreamClientContext = {
-  clientKind: CliProfileId;
-  sessionId?: string;
-  traceHint?: string;
-};
-
-function detectDownstreamClientContext(input: {
-  downstreamPath: string;
-  headers?: Record<string, unknown>;
-  body?: unknown;
-}): DownstreamClientContext {
-  const detected = detectCliProfile(input);
-  return {
-    clientKind: detected.id,
-    ...(detected.sessionId ? { sessionId: detected.sessionId } : {}),
-    ...(detected.traceHint ? { traceHint: detected.traceHint } : {}),
-  };
-}
-
 export async function handleChatSurfaceRequest(
  request: FastifyRequest,
  reply: FastifyReply,
@@ -117,7 +105,8 @@ export async function handleChatSurfaceRequest(
      throw error;
    }
  }
-  const hasNonImageFileInput = hasNonImageFileInputInOpenAiBody(resolvedOpenAiBody);
+  const conversationFileSummary = summarizeConversationFileInputsInOpenAiBody(resolvedOpenAiBody);
+  const hasNonImageFileInput = conversationFileSummary.hasDocument;
  const codexSessionCacheKey = deriveCodexSessionCacheKey({
    downstreamFormat,
    body: downstreamFormat === 'claude' ? claudeOriginalBody : request.body,
@@ -125,8 +114,6 @@ export async function handleChatSurfaceRequest(
    proxyToken: getProxyAuthContext(request)?.token || null,
  });
  const downstreamApiKeyId = getProxyAuthContext(request)?.keyId ?? null;
-  const logDownstreamApiKeyId = downstreamApiKeyId !== null
-    && await hasProxyLogDownstreamApiKeyIdColumn();

  const excludeChannelIds: number[] = [];
  let retryCount = 0;
@@ -167,9 +154,20 @@ export async function handleChatSurfaceRequest(
        requestedModel,
        {
          hasNonImageFileInput,
+          conversationFileSummary,
        },
      ),
    ];
+    const endpointRuntimeContext = {
+      siteId: selected.site.id,
+      modelName,
+      downstreamFormat,
+      requestedModelHint: requestedModel,
+      requestCapabilities: {
+        hasNonImageFileInput,
+        conversationFileSummary,
+      },
+    };
    const buildProviderHeaders = () => (
      buildOauthProviderHeaders({
        extraConfig: typeof selected.account.extraConfig === 'string' ? selected.account.extraConfig : null,
@@ -245,7 +243,7 @@ export async function handleChatSurfaceRequest(
            extraConfig: refreshed.extraConfig ?? selected.account.extraConfig,
          };
          const refreshedRequest = buildEndpointRequest(ctx.request.endpoint);
-          const refreshedTargetUrl = `${selected.site.url}${refreshedRequest.path}`;
+          const refreshedTargetUrl = buildUpstreamUrl(selected.site.url, refreshedRequest.path);
          const refreshedResponse = await dispatchRequest(refreshedRequest, refreshedTargetUrl);
          if (refreshedResponse.ok) {
            return {
@@ -272,6 +270,20 @@ export async function handleChatSurfaceRequest(
          buildRequest: (endpoint) => buildEndpointRequest(endpoint),
          dispatchRequest,
          tryRecover,
+          onAttemptFailure: (ctx) => {
+            recordUpstreamEndpointFailure({
+              ...endpointRuntimeContext,
+              endpoint: ctx.request.endpoint,
+              status: ctx.response.status,
+              errorText: ctx.rawErrText,
+            });
+          },
+          onAttemptSuccess: (ctx) => {
+            recordUpstreamEndpointSuccess({
+              ...endpointRuntimeContext,
+              endpoint: ctx.request.endpoint,
+            });
+          },
          shouldDowngrade: endpointStrategy.shouldDowngrade,
          onDowngrade: (ctx) => {
          logProxy(
@@ -290,7 +302,7 @@ export async function handleChatSurfaceRequest(
          null,
          null,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );
      },
      });
@@ -299,7 +311,11 @@ export async function handleChatSurfaceRequest(
        const status = endpointResult.status || 502;
        const errText = endpointResult.errText || 'unknown error';
        const rawErrText = endpointResult.rawErrText || errText;
-        tokenRouter.recordFailure(selected.channel.id);
+        tokenRouter.recordFailure(selected.channel.id, {
+          status,
+          errorText: rawErrText,
+          modelName,
+        });
        logProxy(
          selected,
          requestedModel,
@@ -316,7 +332,7 @@ export async function handleChatSurfaceRequest(
          null,
          null,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );
        await recordOauthQuotaResetHint({
          accountId: selected.account.id,
@@ -394,6 +410,37 @@ export async function handleChatSurfaceRequest(
        if (!upstreamContentType.includes('text/event-stream')) {
          const fallbackText = await upstream.text();
          rawText = fallbackText;
+          if (looksLikeResponsesSseText(fallbackText)) {
+            startSseResponse();
+            const streamResult = await streamSession.run(
+              createSingleChunkStreamReader(fallbackText),
+              reply.raw,
+            );
+            const latency = Date.now() - startTime;
+            if (streamResult.status === 'failed') {
+              tokenRouter.recordFailure(selected.channel.id, modelName);
+              logProxy(
+                selected,
+                requestedModel,
+                'failed',
+                200,
+                latency,
+                streamResult.errorMessage,
+                retryCount,
+                downstreamPath,
+                parsedUsage.promptTokens,
+                parsedUsage.completionTokens,
+                parsedUsage.totalTokens,
+                0,
+                null,
+                successfulUpstreamPath,
+                clientContext,
+                downstreamApiKeyId,
+              );
+              return;
+            }
+            return;
+          }
          let fallbackData: unknown = null;
          try {
            fallbackData = JSON.parse(fallbackText);
@@ -407,7 +454,11 @@ export async function handleChatSurfaceRequest(
          const latency = Date.now() - startTime;
          const failure = detectProxyFailure({ rawText, usage: parsedUsage });
          if (failure) {
-            tokenRouter.recordFailure(selected.channel.id);
+            tokenRouter.recordFailure(selected.channel.id, {
+              status: failure.status,
+              errorText: failure.reason,
+              modelName,
+            });
            logProxy(
              selected,
              requestedModel,
@@ -424,7 +475,7 @@ export async function handleChatSurfaceRequest(
              null,
              successfulUpstreamPath,
              clientContext,
-              logDownstreamApiKeyId ? downstreamApiKeyId : null,
+              downstreamApiKeyId,
            );

            if (shouldRetryProxyRequest(failure.status, failure.reason) && retryCount < MAX_RETRIES) {
@@ -445,7 +496,11 @@ export async function handleChatSurfaceRequest(
          startSseResponse();
          const streamResult = streamSession.consumeUpstreamFinalPayload(fallbackData, fallbackText, reply.raw);
          if (streamResult.status === 'failed') {
-            tokenRouter.recordFailure(selected.channel.id);
+            tokenRouter.recordFailure(selected.channel.id, {
+              status: 502,
+              errorText: streamResult.errorMessage,
+              modelName,
+            });
            logProxy(
              selected,
              requestedModel,
@@ -462,7 +517,7 @@ export async function handleChatSurfaceRequest(
              null,
              successfulUpstreamPath,
              clientContext,
-              logDownstreamApiKeyId ? downstreamApiKeyId : null,
+              downstreamApiKeyId,
            );
            return;
          }
@@ -495,7 +550,11 @@ export async function handleChatSurfaceRequest(

          const latency = Date.now() - startTime;
          if (streamResult.status === 'failed') {
-            tokenRouter.recordFailure(selected.channel.id);
+            tokenRouter.recordFailure(selected.channel.id, {
+              status: 502,
+              errorText: streamResult.errorMessage,
+              modelName,
+            });
            logProxy(
              selected,
              requestedModel,
@@ -512,7 +571,7 @@ export async function handleChatSurfaceRequest(
              null,
              successfulUpstreamPath,
              clientContext,
-              logDownstreamApiKeyId ? downstreamApiKeyId : null,
+              downstreamApiKeyId,
            );
            return;
          }
@@ -548,7 +607,7 @@ export async function handleChatSurfaceRequest(
          resolvedUsage,
        });

-        tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost);
+        tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost, modelName);
        recordDownstreamCostUsage(request, estimatedCost);
        logProxy(
          selected,
@@ -566,7 +625,7 @@ export async function handleChatSurfaceRequest(
          billingDetails,
          successfulUpstreamPath,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );
        return;
      }
@@ -580,11 +639,15 @@ export async function handleChatSurfaceRequest(
        upstreamData = collected.payload;
      } else {
        rawText = await upstream.text();
-        upstreamData = rawText;
-        try {
-          upstreamData = JSON.parse(rawText);
-        } catch {
+        if (looksLikeResponsesSseText(rawText)) {
+          upstreamData = collectResponsesFinalPayloadFromSseText(rawText, modelName).payload;
+        } else {
          upstreamData = rawText;
+          try {
+            upstreamData = JSON.parse(rawText);
+          } catch {
+            upstreamData = rawText;
+          }
        }
      }
      if (String(selected.site.platform || '').trim().toLowerCase() === 'gemini-cli') {
@@ -595,7 +658,11 @@ export async function handleChatSurfaceRequest(
      const parsedUsage = parseProxyUsage(upstreamData);
      const failure = detectProxyFailure({ rawText, usage: parsedUsage });
      if (failure) {
-        tokenRouter.recordFailure(selected.channel.id);
+        tokenRouter.recordFailure(selected.channel.id, {
+          status: failure.status,
+          errorText: failure.reason,
+          modelName,
+        });
        logProxy(
          selected,
          requestedModel,
@@ -612,7 +679,7 @@ export async function handleChatSurfaceRequest(
          null,
          successfulUpstreamPath,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );

        if (shouldRetryProxyRequest(failure.status, failure.reason) && retryCount < MAX_RETRIES) {
@@ -656,7 +723,7 @@ export async function handleChatSurfaceRequest(
        resolvedUsage,
      });

-      tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost);
+      tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost, modelName);
      recordDownstreamCostUsage(request, estimatedCost);
      logProxy(
        selected,
@@ -674,12 +741,15 @@ export async function handleChatSurfaceRequest(
        billingDetails,
        successfulUpstreamPath,
        clientContext,
-        logDownstreamApiKeyId ? downstreamApiKeyId : null,
+        downstreamApiKeyId,
      );

      return reply.send(downstreamResponse);
    } catch (err: any) {
-      tokenRouter.recordFailure(selected.channel.id);
+      tokenRouter.recordFailure(selected.channel.id, {
+        errorText: err?.message,
+        modelName,
+      });
      logProxy(
        selected,
        requestedModel,
@@ -696,7 +766,7 @@ export async function handleChatSurfaceRequest(
        null,
        null,
        clientContext,
-        logDownstreamApiKeyId ? downstreamApiKeyId : null,
+        downstreamApiKeyId,
      );

      if (retryCount < MAX_RETRIES) {
@@ -775,8 +845,6 @@ export async function handleClaudeCountTokensSurfaceRequest(
  });
  const downstreamPolicy = getDownstreamRoutingPolicy(request);
  const downstreamApiKeyId = getProxyAuthContext(request)?.keyId ?? null;
-  const logDownstreamApiKeyId = downstreamApiKeyId !== null
-    && await hasProxyLogDownstreamApiKeyIdColumn();
  const excludeChannelIds: number[] = [];
  let retryCount = 0;

@@ -883,7 +951,11 @@ export async function handleClaudeCountTokensSurfaceRequest(
      }

      if (!upstream.ok) {
-        tokenRouter.recordFailure(selected.channel.id);
+        tokenRouter.recordFailure(selected.channel.id, {
+          status: upstream.status,
+          errorText: typeof payload === 'string' ? payload : text,
+          modelName,
+        });
        logProxy(
          selected,
          requestedModel,
@@ -900,7 +972,7 @@ export async function handleClaudeCountTokensSurfaceRequest(
          null,
          upstreamRequest.path,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );
        if (shouldRetryProxyRequest(upstream.status, typeof payload === 'string' ? payload : text) && retryCount < MAX_RETRIES) {
          retryCount += 1;
@@ -909,7 +981,7 @@ export async function handleClaudeCountTokensSurfaceRequest(
        return reply.code(upstream.status).type(contentType).send(payload);
      }

-      tokenRouter.recordSuccess(selected.channel.id, latency, 0);
+      tokenRouter.recordSuccess(selected.channel.id, latency, 0, modelName);
      recordDownstreamCostUsage(request, 0);
      logProxy(
        selected,
@@ -927,11 +999,14 @@ export async function handleClaudeCountTokensSurfaceRequest(
        null,
        upstreamRequest.path,
        clientContext,
-        logDownstreamApiKeyId ? downstreamApiKeyId : null,
+        downstreamApiKeyId,
      );
      return reply.code(upstream.status).type(contentType).send(payload);
    } catch (error: any) {
-      tokenRouter.recordFailure(selected.channel.id);
+      tokenRouter.recordFailure(selected.channel.id, {
+        errorText: error?.message,
+        modelName,
+      });
      logProxy(
        selected,
        requestedModel,
@@ -948,7 +1023,7 @@ export async function handleClaudeCountTokensSurfaceRequest(
        null,
        null,
        clientContext,
-        logDownstreamApiKeyId ? downstreamApiKeyId : null,
+        downstreamApiKeyId,
      );
      if (retryCount < MAX_RETRIES) {
        retryCount += 1;
@@ -994,11 +1069,11 @@ async function logProxy(
      upstreamPath,
      errorMessage,
    });
-    await db.insert(schema.proxyLogs).values({
+    await insertProxyLog({
      routeId: selected.channel.routeId,
      channelId: selected.channel.id,
      accountId: selected.account.id,
-      ...(downstreamApiKeyId !== null ? { downstreamApiKeyId } : {}),
+      downstreamApiKeyId,
      modelRequested,
      modelActual: selected.actualModel,
      status,
@@ -1008,11 +1083,15 @@ async function logProxy(
      completionTokens,
      totalTokens,
      estimatedCost,
-      billingDetails: billingDetails ? JSON.stringify(billingDetails) : null,
+      billingDetails,
+      clientFamily: clientContext?.clientKind || null,
+      clientAppId: clientContext?.clientAppId || null,
+      clientAppName: clientContext?.clientAppName || null,
+      clientConfidence: clientContext?.clientConfidence || null,
      errorMessage: normalizedErrorMessage,
      retryCount,
      createdAt,
-    }).run();
+    });
  } catch (error) {
    console.warn('[proxy/chat] failed to write proxy log', error);
  }
@@ -16,6 +16,8 @@ import { executeEndpointFlow, type BuiltEndpointRequest } from '../../routes/pro
 import { composeProxyLogMessage } from '../../routes/proxy/logPathMeta.js';
 import {
  buildUpstreamEndpointRequest,
+  recordUpstreamEndpointFailure,
+  recordUpstreamEndpointSuccess,
  resolveUpstreamEndpointCandidates,
 } from '../../routes/proxy/upstreamEndpoint.js';
 import {
@@ -29,6 +31,9 @@ import {
  wrapGeminiCliRequest,
 } from '../../routes/proxy/geminiCliCompat.js';
 import { dispatchRuntimeRequest } from '../../routes/proxy/runtimeExecutor.js';
+import { detectDownstreamClientContext, type DownstreamClientContext } from '../../routes/proxy/downstreamClientContext.js';
+import { insertProxyLog } from '../../services/proxyLogStore.js';
+import { summarizeConversationFileInputsInOpenAiBody } from '../capabilities/conversationFileCapabilities.js';

 const MAX_RETRIES = 2;
 const GEMINI_MODEL_PROBES = [
@@ -114,10 +119,10 @@ function resolveUpstreamPath(apiVersion: string, modelActionPath: string): strin
  return `/${normalizedVersion}/${normalizedAction}`;
 }

-function hasDownstreamModelRestrictions(policy: { supportedModels?: unknown; allowedRouteIds?: unknown }): boolean {
+function hasDownstreamModelRestrictions(policy: { supportedModels?: unknown; allowedRouteIds?: unknown; denyAllWhenEmpty?: unknown }): boolean {
  const supportedModels = Array.isArray(policy.supportedModels) ? policy.supportedModels : [];
  const allowedRouteIds = Array.isArray(policy.allowedRouteIds) ? policy.allowedRouteIds : [];
-  return supportedModels.length > 0 || allowedRouteIds.length > 0;
+  return supportedModels.length > 0 || allowedRouteIds.length > 0 || policy.denyAllWhenEmpty === true;
 }

 function extractGeminiListedModelName(item: unknown): string {
@@ -203,6 +208,7 @@ async function logProxy(
  retryCount: number,
  downstreamPath: string,
  upstreamPath: string | null,
+  clientContext: DownstreamClientContext | null = null,
  promptTokens = 0,
  completionTokens = 0,
  totalTokens = 0,
@@ -210,11 +216,16 @@ async function logProxy(
  try {
    const createdAt = formatUtcSqlDateTime(new Date());
    const normalizedErrorMessage = composeProxyLogMessage({
+      clientKind: clientContext?.clientKind && clientContext.clientKind !== 'generic'
+        ? clientContext.clientKind
+        : null,
+      sessionId: clientContext?.sessionId || null,
+      traceHint: clientContext?.traceHint || null,
      downstreamPath,
      upstreamPath,
      errorMessage,
    });
-    await db.insert(schema.proxyLogs).values({
+    await insertProxyLog({
      routeId: selected.channel.routeId,
      channelId: selected.channel.id,
      accountId: selected.account.id,
@@ -227,11 +238,14 @@ async function logProxy(
      completionTokens,
      totalTokens,
      estimatedCost: 0,
-      billingDetails: null,
+      clientFamily: clientContext?.clientKind || null,
+      clientAppId: clientContext?.clientAppId || null,
+      clientAppName: clientContext?.clientAppName || null,
+      clientConfidence: clientContext?.clientConfidence || null,
      errorMessage: normalizedErrorMessage,
      retryCount,
      createdAt,
-    }).run();
+    });
  } catch (error) {
    console.warn('[proxy/gemini] failed to write proxy log', error);
  }
@@ -285,7 +299,10 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          lastStatus = upstream.status;
          lastText = text;
          lastContentType = upstream.headers.get('content-type') || 'application/json';
-          await tokenRouter.recordFailure?.(selected.channel.id);
+          await tokenRouter.recordFailure?.(selected.channel.id, {
+            status: upstream.status,
+            errorText: text,
+          });
          if (retryCount < MAX_RETRIES) {
            retryCount += 1;
            continue;
@@ -300,7 +317,9 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          return reply.code(upstream.status).type(upstream.headers.get('content-type') || 'application/json').send(text);
        }
      } catch (error) {
-        await tokenRouter.recordFailure?.(selected.channel.id);
+        await tokenRouter.recordFailure?.(selected.channel.id, {
+          errorText: error instanceof Error ? error.message : 'Gemini upstream request failed',
+        });
        lastStatus = 502;
        lastContentType = 'application/json';
        lastText = JSON.stringify({
@@ -355,6 +374,11 @@ export async function geminiProxyRoute(app: FastifyInstance) {

    const policy = getDownstreamRoutingPolicy(request);
    const downstreamPath = resolveDownstreamPath(request);
+    const clientContext = detectDownstreamClientContext({
+      downstreamPath,
+      headers: request.headers as Record<string, unknown>,
+      body: request.body,
+    });
    const excludeChannelIds: number[] = [];
    let retryCount = 0;
    let lastStatus = 503;
@@ -394,7 +418,10 @@ export async function geminiProxyRoute(app: FastifyInstance) {
                type: 'server_error',
              },
            });
-            await tokenRouter.recordFailure?.(selected.channel.id);
+            await tokenRouter.recordFailure?.(selected.channel.id, {
+              status: 500,
+              errorText: 'Gemini CLI OAuth project is missing',
+            });
            if (retryCount < MAX_RETRIES) {
              retryCount += 1;
              continue;
@@ -478,7 +505,10 @@ export async function geminiProxyRoute(app: FastifyInstance) {
            lastStatus = upstream.status;
            lastContentType = contentType;
            lastText = await upstream.text();
-            await tokenRouter.recordFailure?.(selected.channel.id);
+            await tokenRouter.recordFailure?.(selected.channel.id, {
+              status: upstream.status,
+              errorText: lastText,
+            });
            await logProxy(
              selected,
              requestedModel,
@@ -489,6 +519,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
              retryCount,
              downstreamPath,
              upstreamPath,
+              clientContext,
            );
            if (retryCount < MAX_RETRIES) {
              retryCount += 1;
@@ -523,6 +554,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
                retryCount,
                downstreamPath,
                upstreamPath,
+                clientContext,
              );
              reply.raw.end();
              return;
@@ -572,6 +604,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
              retryCount,
              downstreamPath,
              upstreamPath,
+              clientContext,
              parsedUsage.promptTokens,
              parsedUsage.completionTokens,
              parsedUsage.totalTokens,
@@ -607,6 +640,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
              retryCount,
              downstreamPath,
              upstreamPath,
+              clientContext,
              parsedUsage.promptTokens,
              parsedUsage.completionTokens,
              parsedUsage.totalTokens,
@@ -629,6 +663,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
              retryCount,
              downstreamPath,
              upstreamPath,
+              clientContext,
            );
            return reply.code(upstream.status).type(contentType || 'application/json').send(text);
          }
@@ -651,6 +686,8 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          modelName: actualModel,
          stream: isStreamAction,
        });
+        const conversationFileSummary = summarizeConversationFileInputsInOpenAiBody(openAiBody);
+        const hasNonImageFileInput = conversationFileSummary.hasDocument;
        const endpointCandidates = await resolveUpstreamEndpointCandidates(
          {
            site: selected.site,
@@ -659,7 +696,21 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          actualModel,
          'openai',
          requestedModel,
+          {
+            hasNonImageFileInput,
+            conversationFileSummary,
+          },
        );
+        const endpointRuntimeContext = {
+          siteId: selected.site.id,
+          modelName: actualModel,
+          downstreamFormat: 'openai' as const,
+          requestedModelHint: requestedModel,
+          requestCapabilities: {
+            hasNonImageFileInput,
+            conversationFileSummary,
+          },
+        };
        const buildEndpointRequest = (
          endpoint: 'chat' | 'messages' | 'responses',
          requestOptions: { forceNormalizeClaudeBody?: boolean } = {},
@@ -721,6 +772,20 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          buildRequest: (endpoint) => buildEndpointRequest(endpoint),
          dispatchRequest,
          tryRecover: endpointStrategy.tryRecover,
+          onAttemptFailure: (ctx) => {
+            recordUpstreamEndpointFailure({
+              ...endpointRuntimeContext,
+              endpoint: ctx.request.endpoint,
+              status: ctx.response.status,
+              errorText: ctx.rawErrText,
+            });
+          },
+          onAttemptSuccess: (ctx) => {
+            recordUpstreamEndpointSuccess({
+              ...endpointRuntimeContext,
+              endpoint: ctx.request.endpoint,
+            });
+          },
          shouldDowngrade: endpointStrategy.shouldDowngrade,
        });
        if (!endpointResult.ok) {
@@ -732,7 +797,10 @@ export async function geminiProxyRoute(app: FastifyInstance) {
              type: 'upstream_error',
            },
          });
-          await tokenRouter.recordFailure?.(selected.channel.id);
+          await tokenRouter.recordFailure?.(selected.channel.id, {
+            status: endpointResult.status,
+            errorText: endpointResult.rawErrText || endpointResult.errText,
+          });
          await logProxy(
            selected,
            requestedModel,
@@ -743,6 +811,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
            retryCount,
            downstreamPath,
            null,
+            clientContext,
          );
          if (retryCount < MAX_RETRIES) {
            retryCount += 1;
@@ -780,6 +849,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          retryCount,
          downstreamPath,
          upstreamPath,
+          clientContext,
          parsedUsage.promptTokens,
          parsedUsage.completionTokens,
          parsedUsage.totalTokens,
@@ -798,7 +868,9 @@ export async function geminiProxyRoute(app: FastifyInstance) {
            type: 'upstream_error',
          },
        });
-        await tokenRouter.recordFailure?.(selected.channel.id);
+        await tokenRouter.recordFailure?.(selected.channel.id, {
+          errorText: error instanceof Error ? error.message : 'Gemini upstream request failed',
+        });
        await logProxy(
          selected,
          requestedModel,
@@ -809,6 +881,7 @@ export async function geminiProxyRoute(app: FastifyInstance) {
          retryCount,
          downstreamPath,
          upstreamPath || null,
+          clientContext,
        );
        if (retryCount < MAX_RETRIES) {
          retryCount += 1;
@@ -1,6 +1,5 @@
 import { TextDecoder } from 'node:util';
 import type { FastifyReply, FastifyRequest } from 'fastify';
-import { db, hasProxyLogDownstreamApiKeyIdColumn, schema } from '../../db/index.js';
 import { tokenRouter } from '../../services/tokenRouter.js';
 import { refreshModelsAndRebuildRoutes } from '../../services/modelService.js';
 import { reportProxyAllFailed, reportTokenExpired } from '../../services/alertService.js';
@@ -12,12 +11,15 @@ import { resolveChannelProxyUrl, withSiteRecordProxyRequestInit } from '../../se
 import { openAiResponsesTransformer } from '../../transformers/openai/responses/index.js';
 import {
  buildUpstreamEndpointRequest,
+  recordUpstreamEndpointFailure,
+  recordUpstreamEndpointSuccess,
  resolveUpstreamEndpointCandidates,
 } from '../../routes/proxy/upstreamEndpoint.js';
 import { ensureModelAllowedForDownstreamKey, getDownstreamRoutingPolicy, recordDownstreamCostUsage } from '../../routes/proxy/downstreamPolicy.js';
 import { composeProxyLogMessage } from '../../routes/proxy/logPathMeta.js';
 import { executeEndpointFlow, type BuiltEndpointRequest } from '../../routes/proxy/endpointFlow.js';
 import { detectProxyFailure } from '../../routes/proxy/proxyFailureJudge.js';
+import { buildUpstreamUrl } from '../../routes/proxy/upstreamUrl.js';
 import { formatUtcSqlDateTime } from '../../services/localTimeService.js';
 import { resolveProxyLogBilling } from '../../routes/proxy/proxyBilling.js';
 import { getProxyAuthContext, getProxyResourceOwner } from '../../middleware/auth.js';
@@ -25,7 +27,6 @@ import { dispatchRuntimeRequest } from '../../routes/proxy/runtimeExecutor.js';
 import { normalizeInputFileBlock } from '../../transformers/shared/inputFile.js';
 import {
  ProxyInputFileResolutionError,
-  hasNonImageFileInputInOpenAiBody,
  resolveResponsesBodyInputFiles,
 } from '../../services/proxyInputFileResolver.js';
 import {
@@ -34,14 +35,23 @@ import {
 import { getOauthInfoFromExtraConfig } from '../../services/oauth/oauthAccount.js';
 import { recordOauthQuotaResetHint } from '../../services/oauth/quota.js';
 import { refreshOauthAccessTokenSingleflight } from '../../services/oauth/refreshSingleflight.js';
-import { collectResponsesFinalPayloadFromSse } from '../../routes/proxy/responsesSseFinal.js';
+import {
+  collectResponsesFinalPayloadFromSse,
+  collectResponsesFinalPayloadFromSseText,
+  createSingleChunkStreamReader,
+  looksLikeResponsesSseText,
+} from '../../routes/proxy/responsesSseFinal.js';
 import {
  createGeminiCliStreamReader,
  unwrapGeminiCliPayload,
 } from '../../routes/proxy/geminiCliCompat.js';
 import { isCodexResponsesSurface } from '../cliProfiles/codexProfile.js';
-import { detectCliProfile } from '../cliProfiles/registry.js';
-import type { CliProfileId } from '../cliProfiles/types.js';
+import {
+  summarizeConversationFileInputsInOpenAiBody,
+  summarizeConversationFileInputsInResponsesBody,
+} from '../capabilities/conversationFileCapabilities.js';
+import { detectDownstreamClientContext, type DownstreamClientContext } from '../../routes/proxy/downstreamClientContext.js';
+import { insertProxyLog } from '../../services/proxyLogStore.js';

 const MAX_RETRIES = 2;

@@ -127,25 +137,6 @@ function carriesResponsesFileUrlInput(value: unknown): boolean {

 type UsageSummary = ReturnType<typeof parseProxyUsage>;

-type DownstreamClientContext = {
-  clientKind: CliProfileId;
-  sessionId?: string;
-  traceHint?: string;
-};
-
-function detectDownstreamClientContext(input: {
-  downstreamPath: string;
-  headers?: Record<string, unknown>;
-  body?: unknown;
-}): DownstreamClientContext {
-  const detected = detectCliProfile(input);
-  return {
-    clientKind: detected.id,
-    ...(detected.sessionId ? { sessionId: detected.sessionId } : {}),
-    ...(detected.traceHint ? { traceHint: detected.traceHint } : {}),
-  };
-}
-
 function deriveCodexExplicitSessionId(body: Record<string, unknown>): string | null {
  const promptCacheKey = typeof body.prompt_cache_key === 'string'
    ? body.prompt_cache_key.trim()
@@ -188,8 +179,6 @@ export async function handleOpenAiResponsesSurfaceRequest(
    if (!await ensureModelAllowedForDownstreamKey(request, reply, requestedModel)) return;
    const downstreamPolicy = getDownstreamRoutingPolicy(request);
    const downstreamApiKeyId = getProxyAuthContext(request)?.keyId ?? null;
-    const logDownstreamApiKeyId = downstreamApiKeyId !== null
-      && await hasProxyLogDownstreamApiKeyIdColumn();
    const excludeChannelIds: number[] = [];
    let retryCount = 0;

@@ -243,17 +232,12 @@ export async function handleOpenAiResponsesSurfaceRequest(
        isStream,
        { defaultEncryptedReasoningInclude },
      );
-      const hasNonImageFileInput = hasNonImageFileInputInOpenAiBody(openAiBody);
+      const conversationFileSummary = summarizeConversationFileInputsInOpenAiBody(openAiBody);
+      const hasNonImageFileInput = conversationFileSummary.hasDocument;
      const prefersNativeResponsesReasoning = wantsNativeResponsesReasoning(normalizedResponsesBody);
-      const requiresNativeResponsesFileUrl = carriesResponsesFileUrlInput(normalizedResponsesBody.input);
-      if (requiresNativeResponsesFileUrl && String(selected.site.platform || '').trim().toLowerCase() === 'claude') {
-        return reply.code(400).send({
-          error: {
-            message: 'Responses input_file.file_url requires an upstream /v1/responses endpoint; current site only supports /v1/messages.',
-            type: 'invalid_request_error',
-          },
-        });
-      }
+      const responsesConversationFileSummary = summarizeConversationFileInputsInResponsesBody(normalizedResponsesBody);
+      const requiresNativeResponsesFileUrl = responsesConversationFileSummary.hasRemoteDocumentUrl
+        || carriesResponsesFileUrlInput(normalizedResponsesBody.input);
      const endpointCandidates = await resolveUpstreamEndpointCandidates(
        {
          site: selected.site,
@@ -264,15 +248,24 @@ export async function handleOpenAiResponsesSurfaceRequest(
        requestedModel,
        {
          hasNonImageFileInput,
+          conversationFileSummary,
          wantsNativeResponsesReasoning: prefersNativeResponsesReasoning,
        },
      );
-      if (requiresNativeResponsesFileUrl) {
-        endpointCandidates.splice(0, endpointCandidates.length, 'responses');
-      }
      if (endpointCandidates.length === 0) {
        endpointCandidates.push('responses', 'chat', 'messages');
      }
+      const endpointRuntimeContext = {
+        siteId: selected.site.id,
+        modelName,
+        downstreamFormat: 'responses' as const,
+        requestedModelHint: requestedModel,
+        requestCapabilities: {
+          hasNonImageFileInput,
+          conversationFileSummary,
+          wantsNativeResponsesReasoning: prefersNativeResponsesReasoning,
+        },
+      };
      const buildProviderHeaders = () => (
        buildOauthProviderHeaders({
          extraConfig: typeof selected.account.extraConfig === 'string' ? selected.account.extraConfig : null,
@@ -339,7 +332,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
              extraConfig: refreshed.extraConfig ?? selected.account.extraConfig,
            };
            const refreshedRequest = buildEndpointRequest(ctx.request.endpoint);
-            const refreshedTargetUrl = `${selected.site.url}${refreshedRequest.path}`;
+            const refreshedTargetUrl = buildUpstreamUrl(selected.site.url, refreshedRequest.path);
            const refreshedResponse = await dispatchRequest(refreshedRequest, refreshedTargetUrl);
            if (refreshedResponse.ok) {
              return {
@@ -367,6 +360,20 @@ export async function handleOpenAiResponsesSurfaceRequest(
          buildRequest: (endpoint) => buildEndpointRequest(endpoint),
          dispatchRequest,
          tryRecover,
+          onAttemptFailure: (ctx) => {
+            recordUpstreamEndpointFailure({
+              ...endpointRuntimeContext,
+              endpoint: ctx.request.endpoint,
+              status: ctx.response.status,
+              errorText: ctx.rawErrText,
+            });
+          },
+          onAttemptSuccess: (ctx) => {
+            recordUpstreamEndpointSuccess({
+              ...endpointRuntimeContext,
+              endpoint: ctx.request.endpoint,
+            });
+          },
          shouldDowngrade: endpointStrategy.shouldDowngrade,
          onDowngrade: (ctx) => {
            logProxy(
@@ -385,7 +392,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
              null,
              null,
              clientContext,
-              logDownstreamApiKeyId ? downstreamApiKeyId : null,
+              downstreamApiKeyId,
            );
          },
        });
@@ -394,7 +401,11 @@ export async function handleOpenAiResponsesSurfaceRequest(
          const status = endpointResult.status || 502;
          const errText = endpointResult.errText || 'unknown error';
          const rawErrText = endpointResult.rawErrText || errText;
-          tokenRouter.recordFailure(selected.channel.id);
+          tokenRouter.recordFailure(selected.channel.id, {
+            status,
+            errorText: rawErrText,
+            modelName,
+          });
          logProxy(
            selected,
            requestedModel,
@@ -411,7 +422,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
            null,
            null,
            clientContext,
-            logDownstreamApiKeyId ? downstreamApiKeyId : null,
+            downstreamApiKeyId,
          );
          await recordOauthQuotaResetHint({
            accountId: selected.account.id,
@@ -486,14 +497,14 @@ export async function handleOpenAiResponsesSurfaceRequest(
        }

        try {
-          tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost);
+          tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost, modelName);
          recordDownstreamCostUsage(request, estimatedCost);
          logProxy(
            selected, requestedModel, 'success', 200, latency, null, retryCount, downstreamPath,
            usageForLog.promptTokens, usageForLog.completionTokens, usageForLog.totalTokens, estimatedCost, billingDetails,
            successfulUpstreamPath,
            clientContext,
-            logDownstreamApiKeyId ? downstreamApiKeyId : null,
+            downstreamApiKeyId,
          );
        } catch (error) {
          console.error('[responses] post-stream success logging failed:', error);
@@ -541,6 +552,39 @@ export async function handleOpenAiResponsesSurfaceRequest(
          });
          if (!upstreamContentType.includes('text/event-stream')) {
            const rawText = await upstream.text();
+            if (looksLikeResponsesSseText(rawText)) {
+              startSseResponse();
+              const streamResult = await streamSession.run(
+                createSingleChunkStreamReader(rawText),
+                reply.raw,
+              );
+              const latency = Date.now() - startTime;
+              if (streamResult.status === 'failed') {
+                tokenRouter.recordFailure(selected.channel.id, modelName);
+                logProxy(
+                  selected,
+                  requestedModel,
+                  'failed',
+                  200,
+                  latency,
+                  streamResult.errorMessage,
+                  retryCount,
+                  downstreamPath,
+                  parsedUsage.promptTokens,
+                  parsedUsage.completionTokens,
+                  parsedUsage.totalTokens,
+                  0,
+                  null,
+                  successfulUpstreamPath,
+                  clientContext,
+                  downstreamApiKeyId,
+                );
+                return;
+              }
+
+              await finalizeStreamSuccess(parsedUsage, latency);
+              return;
+            }
            let upstreamData: unknown = rawText;
            try {
              upstreamData = JSON.parse(rawText);
@@ -555,7 +599,11 @@ export async function handleOpenAiResponsesSurfaceRequest(
            const latency = Date.now() - startTime;
            const failure = detectProxyFailure({ rawText, usage: parsedUsage });
            if (failure) {
-              tokenRouter.recordFailure(selected.channel.id);
+              tokenRouter.recordFailure(selected.channel.id, {
+                status: failure.status,
+                errorText: failure.reason,
+                modelName,
+              });
              logProxy(
                selected,
                requestedModel,
@@ -572,7 +620,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
                null,
                successfulUpstreamPath,
                clientContext,
-                logDownstreamApiKeyId ? downstreamApiKeyId : null,
+                downstreamApiKeyId,
              );

              if (shouldRetryProxyRequest(failure.status, failure.reason) && retryCount < MAX_RETRIES) {
@@ -590,7 +638,11 @@ export async function handleOpenAiResponsesSurfaceRequest(
            startSseResponse();
            const streamResult = streamSession.consumeUpstreamFinalPayload(upstreamData, rawText, reply.raw);
            if (streamResult.status === 'failed') {
-              tokenRouter.recordFailure(selected.channel.id);
+              tokenRouter.recordFailure(selected.channel.id, {
+                status: 502,
+                errorText: streamResult.errorMessage,
+                modelName,
+              });
              logProxy(
                selected,
                requestedModel,
@@ -607,7 +659,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
                null,
                successfulUpstreamPath,
                clientContext,
-                logDownstreamApiKeyId ? downstreamApiKeyId : null,
+                downstreamApiKeyId,
              );
              return;
            }
@@ -646,7 +698,11 @@ export async function handleOpenAiResponsesSurfaceRequest(

          const latency = Date.now() - startTime;
          if (streamResult.status === 'failed') {
-            tokenRouter.recordFailure(selected.channel.id);
+            tokenRouter.recordFailure(selected.channel.id, {
+              status: 502,
+              errorText: streamResult.errorMessage,
+              modelName,
+            });
            logProxy(
              selected,
              requestedModel,
@@ -663,7 +719,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
              null,
              successfulUpstreamPath,
              clientContext,
-              logDownstreamApiKeyId ? downstreamApiKeyId : null,
+              downstreamApiKeyId,
            );
            return;
          }
@@ -692,11 +748,15 @@ export async function handleOpenAiResponsesSurfaceRequest(
          upstreamData = collected.payload;
        } else {
          rawText = await upstream.text();
-          upstreamData = rawText;
-          try {
-            upstreamData = JSON.parse(rawText);
-          } catch {
+          if (looksLikeResponsesSseText(rawText)) {
+            upstreamData = collectResponsesFinalPayloadFromSseText(rawText, modelName).payload;
+          } else {
            upstreamData = rawText;
+            try {
+              upstreamData = JSON.parse(rawText);
+            } catch {
+              upstreamData = rawText;
+            }
          }
        }
        if (String(selected.site.platform || '').trim().toLowerCase() === 'gemini-cli') {
@@ -706,7 +766,11 @@ export async function handleOpenAiResponsesSurfaceRequest(
        const parsedUsage = parseProxyUsage(upstreamData);
        const failure = detectProxyFailure({ rawText, usage: parsedUsage });
        if (failure) {
-          tokenRouter.recordFailure(selected.channel.id);
+          tokenRouter.recordFailure(selected.channel.id, {
+            status: failure.status,
+            errorText: failure.reason,
+            modelName,
+          });
          logProxy(
            selected,
            requestedModel,
@@ -723,7 +787,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
            null,
            successfulUpstreamPath,
            clientContext,
-            logDownstreamApiKeyId ? downstreamApiKeyId : null,
+            downstreamApiKeyId,
          );
          if (shouldRetryProxyRequest(failure.status, failure.reason) && retryCount < MAX_RETRIES) {
            retryCount += 1;
@@ -770,18 +834,21 @@ export async function handleOpenAiResponsesSurfaceRequest(
          resolvedUsage,
        });

-        tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost);
+        tokenRouter.recordSuccess(selected.channel.id, latency, estimatedCost, modelName);
        recordDownstreamCostUsage(request, estimatedCost);
        logProxy(
          selected, requestedModel, 'success', 200, latency, null, retryCount, downstreamPath,
          resolvedUsage.promptTokens, resolvedUsage.completionTokens, resolvedUsage.totalTokens, estimatedCost, billingDetails,
          successfulUpstreamPath,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );
        return reply.send(downstreamData);
      } catch (err: any) {
-        tokenRouter.recordFailure(selected.channel.id);
+        tokenRouter.recordFailure(selected.channel.id, {
+          errorText: err?.message,
+          modelName,
+        });
        logProxy(
          selected,
          requestedModel,
@@ -798,7 +865,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
          null,
          null,
          clientContext,
-          logDownstreamApiKeyId ? downstreamApiKeyId : null,
+          downstreamApiKeyId,
        );
        if (retryCount < MAX_RETRIES) {
          retryCount += 1;
@@ -845,11 +912,11 @@ async function logProxy(
      upstreamPath,
      errorMessage,
    });
-    await db.insert(schema.proxyLogs).values({
+    await insertProxyLog({
      routeId: selected.channel.routeId,
      channelId: selected.channel.id,
      accountId: selected.account.id,
-      ...(downstreamApiKeyId !== null ? { downstreamApiKeyId } : {}),
+      downstreamApiKeyId,
      modelRequested,
      modelActual: selected.actualModel,
      status,
@@ -859,11 +926,15 @@ async function logProxy(
      completionTokens,
      totalTokens,
      estimatedCost,
-      billingDetails: billingDetails ? JSON.stringify(billingDetails) : null,
+      billingDetails,
+      clientFamily: clientContext?.clientKind || null,
+      clientAppId: clientContext?.clientAppId || null,
+      clientAppName: clientContext?.clientAppName || null,
+      clientConfidence: clientContext?.clientConfidence || null,
      errorMessage: normalizedErrorMessage,
      retryCount,
      createdAt,
-    }).run();
+    });
  } catch (error) {
    console.warn('[proxy/responses] failed to write proxy log', error);
  }