Merge branch 'main' into cl-6904-grokopenai-responses-adapters-omit-prompt_cache_key-xai

Author: TheGreatAxios

src/auth/codex/instructions.ts

@@ -21,7 +21,8 @@ const PROMPT_SENTINEL = "You are Codex";
 const MIN_PROMPT_LENGTH = 1000;
 
 // Bounds both network calls below so a black-holed connection can never hang
-// exec boot, which awaits refreshCodexInstructions before first inference.
+// exec boot or the TUI's first deliver, both of which wait on
+// refreshCodexInstructions settling before first inference.
 const CODEX_INSTRUCTIONS_TIMEOUT_MS = 10_000;
 
 export function isValidCodexPrompt(text: string): boolean {

src/tui/deliver-agent-message.test.ts

@@ -2,12 +2,12 @@ import { describe, expect, test } from "bun:test";
 import { deliverAgentMessage } from "./deliver-agent-message.js";
 
 describe("deliverAgentMessage", () => {
-  test("surfaces a not-delivered notice instead of throwing when the agent is mid-rebuild", () => {
+  test("surfaces a not-delivered notice instead of throwing when the agent is mid-rebuild", async () => {
     const notices: string[] = [];
     const fatal = new Error("agent rebuild failed: provider unreachable");
     let delivered = false;
 
-    deliverAgentMessage({
+    await deliverAgentMessage({
       getFatalBuildError: () => fatal,
       deliverToLiveAgent: () => {
         delivered = true;
@@ -24,10 +24,10 @@ describe("deliverAgentMessage", () => {
     expect(notices[0]).toContain("provider unreachable");
   });
 
-  test("surfaces a not-delivered notice when the live agent throws on delivery", () => {
+  test("surfaces a not-delivered notice when the live agent throws on delivery", async () => {
     const notices: string[] = [];
 
-    deliverAgentMessage({
+    await deliverAgentMessage({
       getFatalBuildError: () => null,
       deliverToLiveAgent: () => {
         throw new Error("agent is closed");
@@ -40,11 +40,11 @@ describe("deliverAgentMessage", () => {
     expect(notices[0]).toContain("agent is closed");
   });
 
-  test("delivers normally and stays silent when the agent is healthy", () => {
+  test("delivers normally and stays silent when the agent is healthy", async () => {
     const notices: string[] = [];
     let delivered = false;
 
-    deliverAgentMessage({
+    await deliverAgentMessage({
       getFatalBuildError: () => null,
       deliverToLiveAgent: () => {
         delivered = true;
@@ -55,4 +55,53 @@ describe("deliverAgentMessage", () => {
     expect(delivered).toBe(true);
     expect(notices).toHaveLength(0);
   });
+
+  test("holds delivery until ready settles so the first request cannot race startup work", async () => {
+    // Models the Codex instructions refresh: the request body prefix pins the
+    // instructions text, so delivery — and therefore request building — must
+    // wait for the in-flight refresh instead of letting it swap the value
+    // between turn 1 and turn 2.
+    let instructions = "cached";
+    let observedAtDelivery = "";
+    let resolveReady!: () => void;
+    const ready = new Promise<void>((resolve) => {
+      resolveReady = resolve;
+    });
+
+    const pending = deliverAgentMessage({
+      getFatalBuildError: () => null,
+      ready,
+      deliverToLiveAgent: () => {
+        observedAtDelivery = instructions;
+      },
+      onDeliverFailure: () => undefined,
+    });
+
+    // Refresh still in flight: nothing delivered yet.
+    await Promise.resolve();
+    expect(observedAtDelivery).toBe("");
+
+    instructions = "refreshed";
+    resolveReady();
+    await pending;
+
+    // The first request observed the post-refresh value, so a later-settling
+    // refresh can never change the prefix of an already-started session.
+    expect(observedAtDelivery).toBe("refreshed");
+  });
+
+  test("delivers immediately when ready is already settled", async () => {
+    let delivered = false;
+
+    await deliverAgentMessage({
+      getFatalBuildError: () => null,
+      ready: Promise.resolve(),
+      deliverToLiveAgent: () => {
+        delivered = true;
+      },
+      onDeliverFailure: () => undefined,
+    });
+
+    expect(delivered).toBe(true);
+  });
 });

src/tui/deliver-agent-message.ts

@@ -6,11 +6,20 @@
  */
 export interface DeliverAgentMessageDeps {
   getFatalBuildError: () => Error | null;
+  /**
+   * Settles before any delivery. Gates the session's first request on startup
+   * work that must not race request building — the Codex instructions refresh:
+   * a late in-memory swap would change the request prefix and forfeit the
+   * provider prompt cache for the rest of the session. Must never reject
+   * (callers attach their own fallback handling).
+   */
+  ready?: Promise<void>;
   deliverToLiveAgent: () => void;
   onDeliverFailure: (message: string) => void;
 }
 
-export function deliverAgentMessage(deps: DeliverAgentMessageDeps): void {
+export async function deliverAgentMessage(deps: DeliverAgentMessageDeps): Promise<void> {
+  if (deps.ready !== undefined) await deps.ready;
   const fatal = deps.getFatalBuildError();
   if (fatal !== null) {
     deps.onDeliverFailure(`Message not delivered: ${fatal.message}`);

src/tui/runner.ts

@@ -1319,6 +1319,25 @@ export async function runTUI(initialConfig: Config): Promise<number> {
         { providerName: config.providerName, model: config.model },
       );
 
+    const initialCodexProfile = codexProfileFromProviderName(config.providerName);
+    const initialXaiProfile = xaiProfileFromProviderName(config.providerName);
+
+    // Refresh the pinned Codex instructions without blocking TUI startup. Every
+    // deliver below awaits settlement, so the first request never races the
+    // refresh: codex-responses-adapter places the instructions at the top of
+    // every request body, and an in-memory swap after inference #1 would change
+    // the whole prefix and forfeit the provider prompt cache mid-session.
+    // Best-effort, same as exec boot: on failure the session runs on
+    // cached/bundled instructions.
+    const codexInstructionsRefreshed: Promise<void> =
+      initialCodexProfile === undefined
+        ? Promise.resolve()
+        : refreshCodexInstructions().catch((err: unknown) => {
+            tuiLogger.warn("Codex instructions refresh failed: {error}", {
+              error: err instanceof Error ? err.message : String(err),
+            });
+          });
+
     // Reload, interrupt, compaction continuation, and proxy deliver share one queue
     // so a rebuild never races an in-flight deliver.
     const sessionOps = createSessionOperationQueue();
@@ -1327,8 +1346,9 @@ export async function runTUI(initialConfig: Config): Promise<number> {
         // The shell already popped the queue item and painted it as delivered
         // by the time this runs, so a failed rebuild must be surfaced here —
         // otherwise the message silently never reaches the agent.
-        deliverAgentMessage({
+        await deliverAgentMessage({
           getFatalBuildError: () => fatalBuildError,
+          ready: codexInstructionsRefreshed,
           deliverToLiveAgent,
           onDeliverFailure: systemNotice,
         });
@@ -1381,16 +1401,6 @@ export async function runTUI(initialConfig: Config): Promise<number> {
     // When the session starts on a Codex profile, seed the agent with a Responses
     // source (account id pulled from the resolved catalog entry, session id from
     // the run) rather than the OpenAI-compatible one.
-    const initialCodexProfile = codexProfileFromProviderName(config.providerName);
-    const initialXaiProfile = xaiProfileFromProviderName(config.providerName);
-    if (initialCodexProfile !== undefined) {
-      void refreshCodexInstructions().catch((err: unknown) => {
-        // Best-effort; agent still starts with cached/default instructions.
-        tuiLogger.warn("Codex instructions refresh failed: {error}", {
-          error: err instanceof Error ? err.message : String(err),
-        });
-      });
-    }
     const initialCodexAccountId = config.providers.find(
       (p) => p.name === config.providerName,
     )?.codexAccountId;