fix(gateway): compress subscription bridge requests

packages/gateway/src/routes/completion.ts

@@ -728,13 +728,40 @@ function isCodexBridgeModel(model: string | undefined): boolean {
   return /^(codex-default|codex-|gpt-5(?:\.|-|$))/i.test(model ?? '');
 }
 
-function chatMessagesToPrompt(messages: OpenAIChatCompletionRequest['messages']): string {
+function chatMessagesToPrompt(
+  messages: OpenAIChatCompletionRequest['messages'],
+  options: { includeSystem?: boolean } = {}
+): string {
   return messages
+    .filter((message) => options.includeSystem !== false || message.role !== 'system')
     .map((message) => `${message.role}: ${contentToText(message.content)}`)
     .join('\n\n')
     .trim();
 }
 
+function compressChatMessagesForBridge(body: OpenAIChatCompletionRequest): {
+  body: OpenAIChatCompletionRequest;
+  inputText: string;
+  compression: CompressionResult;
+} {
+  const inputText = chatMessagesToPrompt(body.messages, { includeSystem: false });
+  const compression = compressContext(inputText, { enabled: true, mode: 'auto' });
+  if (!compression.applied) return { body, inputText, compression };
+
+  const systemMessages = body.messages.filter((message) => message.role === 'system');
+  return {
+    body: {
+      ...body,
+      messages: [
+        ...systemMessages,
+        { role: 'user', content: compression.input },
+      ],
+    },
+    inputText: compression.input,
+    compression,
+  };
+}
+
 interface SubscriptionBridgeTarget {
   subscriptionId: string;
   provider: string;
@@ -857,12 +884,18 @@ async function postSubscriptionBridgeCompletion(
     return { statusCode: chatResponse.status, ok: chatResponse.ok, data: chatData };
   }
 
+  const system = body.messages
+    .filter((message) => message.role === 'system')
+    .map((message) => contentToText(message.content))
+    .filter(Boolean)
+    .join('\n\n');
   const generateResponse = await fetch(`${bridgeUrl}/api/generate`, {
     method: 'POST',
     headers: { 'Content-Type': 'application/json' },
     body: JSON.stringify({
       model: body.model,
       prompt: inputText,
+      system: system || undefined,
       stream: false,
       temperature: body.temperature ?? 0,
       max_tokens: body.max_tokens,
@@ -880,9 +913,10 @@ async function callSubscriptionBridgeChatCompletion(
   const target = subscriptionBridgeTargetForModel(body.model);
   if (!target) return null;
 
-  const inputText = chatMessagesToPrompt(body.messages);
+  const compressed = compressChatMessagesForBridge(body);
+  const inputText = compressed.inputText;
   try {
-    const upstream = await postSubscriptionBridgeCompletion(target, body, inputText);
+    const upstream = await postSubscriptionBridgeCompletion(target, compressed.body, inputText);
     if (!upstream.ok || upstream.data?.success === false) {
       const error = bridgeErrorMessage(upstream.data, upstream.statusCode);
       logger.warn(
@@ -920,6 +954,7 @@ async function callSubscriptionBridgeChatCompletion(
           subscription_id: target.subscriptionId,
           latency_ms: Date.now() - startMs,
           passthrough: true,
+          compression: buildCompressionResponse(compressed.compression),
         },
       },
     };
@@ -1400,10 +1435,12 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
     if (isCodexBridgeModel(parsed.data.model)) {
       try {
         const bridgeUrl = process.env['CODEX_BRIDGE_URL'] ?? 'http://127.0.0.1:3253';
-        const inputText = typeof parsed.data.input === 'string'
+        const originalInputText = typeof parsed.data.input === 'string'
           ? parsed.data.input
           : (Array.isArray(parsed.data.input)
               ? parsed.data.input.map((p: any) => typeof p?.content === 'string' ? p.content : (Array.isArray(p?.content) ? p.content.map((c: any) => c?.text ?? '').join(' ') : '')).join(' ')              : '');
+        const compression = compressContext(originalInputText, { enabled: true, mode: 'auto' });
+        const inputText = compression.input;
         const upstream = await fetch(`${bridgeUrl}/v1/chat/completions`, {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
@@ -1421,11 +1458,17 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
         }
         if (upstream.ok && upstreamJson?.success !== false) {
           const text = upstreamJson?.content ?? upstreamJson?.response ?? upstreamJson?.choices?.[0]?.message?.content ?? '';
-          const respBody = toOpenAIResponsesResponse({ output: text, model: parsed.data.model, status: 'approved' }, parsed.data.model);
           logger.info({ callId, model: parsed.data.model, len: text.length }, 'codex-bridge passthrough OK');
           const bridgeCaller = (request.headers['x-llm-interceptor-caller'] as string) || 'codex-app';
           const bridgeTokensIn = (upstreamJson?.usage?.prompt_tokens as number | undefined) ?? Math.ceil(inputText.length / 4);
           const bridgeTokensOut = (upstreamJson?.usage?.completion_tokens as number | undefined) ?? Math.ceil(text.length / 4);
+          const respBody = toOpenAIResponsesResponse({
+            output: text,
+            model: parsed.data.model,
+            status: 'approved',
+            tokens: { in: bridgeTokensIn, out: bridgeTokensOut },
+            compression: buildCompressionResponse(compression),
+          }, parsed.data.model);
           // Track against the merged OpenAI (ChatGPT+Codex) subscription pool.
           try {
             const subId = modelToSubscriptionId(parsed.data.model ?? '') ?? 'codex';
@@ -1447,7 +1490,7 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
               0,
               Date.now() - startMs,
               0,
-              false,
+              compression.applied,
               undefined,
             );
           } catch (e) {
@@ -1465,7 +1508,8 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
               latency_ms: Date.now() - startMs,
               confidence: 0,
               cost_usd: 0,
-              compression_applied: false,
+              compression_applied: compression.applied,
+              metadata: { compression: buildCompressionResponse(compression) },
               model: parsed.data.model ?? 'gpt-5.5',
             } as any);
           } catch (e) {