[stack 8/8] feat(trace): cover live call-create HTTP

docs-site/src/content/docs/reference/cli.md

@@ -546,10 +546,12 @@ Tracing is **off by default** and is configured by environment variable at proxy
 - `full` stores bodies verbatim. Use it deliberately and briefly.
 
 Outbound capture covers adapters that send through the shared upstream fetch helper; inbound request and
-response capture covers `/v1/responses`, `/v1/messages`, and `/v1/chat/completions`. Responses WebSocket
-`response.create` turns are traced as separate logical requests: the inbound hash covers the client WS frame,
-while the outbound hash covers the final provider wire body. Voice/realtime live-sideband WebSockets remain
-outside this trace lane because they are long-lived bidirectional sessions rather than request/response turns.
+response capture covers `/v1/responses`, `/v1/messages`, `/v1/chat/completions`, and live call-create HTTP
+(`/v1/live` and `/v1/realtime/calls`). Live call-create also records the rewritten provider wire body before
+its bounded upstream POST. Responses WebSocket `response.create` turns are traced as separate logical
+requests: the inbound hash covers the client WS frame, while the outbound hash covers the final provider
+wire body. Voice/realtime live-sideband WebSockets remain outside this trace lane because they are
+long-lived bidirectional sessions rather than request/response turns.
 Proxy response-cache hits are traced without duplicating the cached response body: the usage trace records
 `cacheHit`, a response hash/byte count, and the request id that originally populated the cache when known.
 WebSocket/live/realtime traffic is not yet covered.

src/server/index.ts

@@ -1461,7 +1461,10 @@ export function startServer(port?: number) {
           model: "gpt-live",
           provider: "unknown",
         };
-        const response = await handleLive(req, config, logCtx);
+        await beginTrace(logCtx, req);
+        const response = await runWithTrace(logCtx, () =>
+          handleLive(req, config, logCtx),
+        );
         addFinalRequestLog(
           requestId,
           start,

src/server/live.ts

@@ -32,6 +32,7 @@ import {
 import { formatCodexProviderForLog } from "../codex/routing";
 import { signalWithTimeout } from "../lib/abort";
 import { sidecarEnter } from "../lib/sidecar-tracker";
+import { appendTraceResponse, noteOutboundRequestBody } from "../trace/capture";
 import type { OcxConfig } from "../types";
 import { resolveFirstUsableOpenAiSidecar, selectOpenAiImagesProvider } from "../providers/openai-sidecar";
 import { ForwardAdmissionCredentialError, validateForwardAdmissionCredential } from "./auth-cors";
@@ -490,6 +491,7 @@ export async function handleLive(
   const linkedSignal = signalWithTimeout(LIVE_UPSTREAM_TIMEOUT_MS, req.signal);
   const sidecarExit = sidecarEnter("live");
   try {
+    noteOutboundRequestBody(new Uint8Array(outboundBody));
     const upstreamResponse = await fetch(url, {
       method: "POST",
       headers,
@@ -505,6 +507,7 @@ export async function handleLive(
       total => `live response too large (${total} bytes)`,
     );
     if (payload instanceof Response) return payload;
+    appendTraceResponse(logCtx.trace, new TextDecoder().decode(payload));
     const relayHeaders: Record<string, string> = {};
     for (const name of LIVE_RELAY_HEADERS) {
       const value = upstreamResponse.headers.get(name);

tests/server-live.test.ts

@@ -16,6 +16,7 @@ import {
 } from "../src/codex/routing";
 import { saveConfig } from "../src/config";
 import { startServer } from "../src/server";
+import { clearRequestLogsForTests, getRequestLogEntries } from "../src/server/request-log";
 import type { OcxConfig } from "../src/types";
 import { fakeChatGptJwt } from "./helpers/fake-chatgpt-jwt";
 import {
@@ -25,6 +26,7 @@ import {
 
 const previousApiToken = process.env.OPENCODEX_API_AUTH_TOKEN;
 const previousOpencodexHome = process.env.OPENCODEX_HOME;
+const previousTraceMode = process.env.OCX_TRACE;
 const originalFetch = globalThis.fetch;
 const TEST_DIR = join(import.meta.dir, ".tmp-server-live-test");
 let isolatedCodexHome: IsolatedCodexHome | null = null;
@@ -35,6 +37,8 @@ beforeEach(() => {
   mkdirSync(TEST_DIR, { recursive: true });
   process.env.OPENCODEX_HOME = TEST_DIR;
   delete process.env.OPENCODEX_API_AUTH_TOKEN;
+  delete process.env.OCX_TRACE;
+  clearRequestLogsForTests();
   isolatedCodexHome = installIsolatedCodexHome("ocx-server-live-codex-");
   clearCodexUpstreamHealth();
   clearThreadAccountMap();
@@ -50,6 +54,8 @@ afterEach(() => {
   else process.env.OPENCODEX_API_AUTH_TOKEN = previousApiToken;
   if (previousOpencodexHome === undefined) delete process.env.OPENCODEX_HOME;
   else process.env.OPENCODEX_HOME = previousOpencodexHome;
+  if (previousTraceMode === undefined) delete process.env.OCX_TRACE;
+  else process.env.OCX_TRACE = previousTraceMode;
   isolatedCodexHome?.restore();
   isolatedCodexHome = null;
   clearCodexUpstreamHealth();
@@ -167,6 +173,7 @@ function multipartLiveBody(
 }
 
 test("POST /v1/live rewrites ChatGPT multipart into backend realtime/calls JSON", async () => {
+  process.env.OCX_TRACE = "metadata";
   const captured: CapturedRequest[] = [];
   const upstream = fakeLiveUpstream(captured);
   saveConfig(forwardConfig());
@@ -205,6 +212,14 @@ test("POST /v1/live rewrites ChatGPT multipart into backend realtime/calls JSON"
       sdp: "v=0-offer",
       session: { model: "gpt-live", instructions: "hi" },
     });
+    const liveTrace = getRequestLogEntries().at(-1)?.trace;
+    expect(liveTrace).toMatchObject({ mode: "metadata", stored: false });
+    expect(liveTrace?.requestHash).toMatch(/^[0-9a-f]{32}$/);
+    expect(liveTrace?.outboundHash).toMatch(/^[0-9a-f]{32}$/);
+    expect(liveTrace?.responseHash).toMatch(/^[0-9a-f]{32}$/);
+    expect(liveTrace?.requestBytes).toBeGreaterThan(0);
+    expect(liveTrace?.outboundBytes).toBeGreaterThan(0);
+    expect(liveTrace?.responseBytes).toBeGreaterThan(0);
   } finally {
     await server.stop(true);
     await upstream.stop(true);