Test that a failed start keeps a waiting steer until its run is cleared

A steer waiting on that start is released only when the run is cleared.

packages/server/src/server/agent/agent-manager.test.ts

@@ -890,6 +890,131 @@ test("unavailable steer interrupts once and starts one replacement turn", async
   }
 });
 
+test("steers a send that arrives while the previous turn is still starting", async () => {
+  const startHeld = deferred<void>();
+  class SlowStartSteeringSession extends SteeringTestSession {
+    override async startTurn(prompt: AgentPromptInput): Promise<{ turnId: string }> {
+      await startHeld.promise;
+      return await super.startTurn(prompt);
+    }
+  }
+  const session = new SlowStartSteeringSession({
+    provider: "codex",
+    cwd: process.cwd(),
+  });
+  const workdir = mkdtempSync(join(tmpdir(), "agent-manager-steer-starting-"));
+  const manager = new AgentManager({
+    clients: {
+      codex: new (class extends TestAgentClient {
+        override async createSession(): Promise<AgentSession> {
+          return session;
+        }
+      })(),
+    },
+    logger,
+  });
+  let agentId: string | null = null;
+  try {
+    const agent = await manager.createAgent({ provider: "codex", cwd: workdir }, undefined, {
+      workspaceId: undefined,
+    });
+    agentId = agent.id;
+    const steerOptions = (clientMessageId: string) => ({
+      replaceRunning: true,
+      activeTurnBehavior: "steer" as const,
+      runOptions: { clientMessageId },
+    });
+    await startAgentRun(manager, agent.id, "FIRST", logger, steerOptions("first-client"));
+    const second = startAgentRun(
+      manager,
+      agent.id,
+      "SECOND",
+      logger,
+      steerOptions("second-client"),
+    );
+
+    startHeld.resolve();
+    await expect(second).resolves.toEqual({ disposition: "steered" });
+
+    expect(session.startPrompts).toEqual(["FIRST"]);
+    expect(session.interruptCount).toBe(0);
+    expect(
+      manager
+        .fetchTimeline(agent.id, { limit: 0 })
+        .rows.filter((row) => row.item.type === "user_message")
+        .map((row) => ({
+          text: row.item.type === "user_message" ? row.item.text : null,
+          turnId: row.turnId,
+        })),
+    ).toEqual([
+      { text: "FIRST", turnId: "active-turn-1" },
+      { text: "SECOND", turnId: "active-turn-1" },
+    ]);
+  } finally {
+    startHeld.resolve();
+    if (agentId) await manager.closeAgent(agentId).catch(() => undefined);
+    rmSync(workdir, { recursive: true, force: true });
+  }
+});
+
+test("starts its own turn when the previous turn fails to start under a steer send", async () => {
+  const startHeld = deferred<void>();
+  class FailingFirstStartSession extends SteeringTestSession {
+    override async startTurn(prompt: AgentPromptInput): Promise<{ turnId: string }> {
+      if (prompt === "FIRST") {
+        this.startPrompts.push(prompt);
+        await startHeld.promise;
+        throw new Error("first start failed");
+      }
+      return await super.startTurn(prompt);
+    }
+  }
+  const session = new FailingFirstStartSession({
+    provider: "codex",
+    cwd: process.cwd(),
+  });
+  const workdir = mkdtempSync(join(tmpdir(), "agent-manager-steer-failed-start-"));
+  const manager = new AgentManager({
+    clients: {
+      codex: new (class extends TestAgentClient {
+        override async createSession(): Promise<AgentSession> {
+          return session;
+        }
+      })(),
+    },
+    logger,
+  });
+  let agentId: string | null = null;
+  try {
+    const agent = await manager.createAgent({ provider: "codex", cwd: workdir }, undefined, {
+      workspaceId: undefined,
+    });
+    agentId = agent.id;
+    const steerOptions = (clientMessageId: string) => ({
+      replaceRunning: true,
+      activeTurnBehavior: "steer" as const,
+      runOptions: { clientMessageId },
+    });
+    await startAgentRun(manager, agent.id, "FIRST", logger, steerOptions("first-client"));
+    const second = startAgentRun(
+      manager,
+      agent.id,
+      "SECOND",
+      logger,
+      steerOptions("second-client"),
+    );
+
+    startHeld.resolve();
+    await expect(second).resolves.toEqual({ disposition: "turn_started" });
+    await vi.waitFor(() => expect(session.startPrompts).toEqual(["FIRST", "SECOND"]));
+    expect(session.interruptCount).toBe(0);
+  } finally {
+    startHeld.resolve();
+    if (agentId) await manager.closeAgent(agentId).catch(() => undefined);
+    rmSync(workdir, { recursive: true, force: true });
+  }
+});
+
 test("orders an accepted steer before output emitted while acknowledgement is pending", async () => {
   const entered = deferred<void>();
   const release = deferred<void>();

packages/server/src/server/agent/agent-manager.ts

@@ -2448,7 +2448,7 @@ export class AgentManager {
         throw error;
       }
       if (isStaleProviderSessionError(error)) {
-        pendingRun.start = { status: "failed", error: error.message };
+        this.runs.settleForegroundStart(pendingRun, { status: "failed", error: error.message });
         agent.pendingReplacement = false;
         if (!agent.activeForegroundTurnId) agent.lifecycle = "idle";
         this.runs.settleForegroundRun(agentId, pendingRun.token);
@@ -2456,14 +2456,18 @@ export class AgentManager {
       }
       agent.pendingReplacement = false;
       const errorMsg = error instanceof Error ? error.message : "Failed to start turn";
-      pendingRun.start = { status: "failed", error: errorMsg };
+      this.runs.settleForegroundStart(pendingRun, { status: "failed", error: errorMsg });
-      await this.handleStreamEvent(agent, {
+      try {
-        type: "turn_failed",
+        await this.handleStreamEvent(agent, {
-        provider: agent.provider,
+          type: "turn_failed",
-        error: errorMsg,
+          provider: agent.provider,
-      });
+          error: errorMsg,
-      this.finalizeForegroundTurn(agent);
+        });
-      this.runs.settleForegroundRun(agentId, pendingRun.token);
+        this.finalizeForegroundTurn(agent);
+      } finally {
+        // A steer waiting on this start is released only once the run is cleared.
+        this.runs.settleForegroundRun(agentId, pendingRun.token);
+      }
       throw error;
     }
   }
@@ -2524,7 +2528,7 @@ export class AgentManager {
         agent.pendingReplacement = false;
       }
       const turnStartedAt = new Date();
-      pendingRun.start = { status: "started", turnId };
+      this.runs.settleForegroundStart(pendingRun, { status: "started", turnId });
       agent.activeForegroundTurnId = turnId;
       this.openActiveTurn(agent, turnId, turnStartedAt);
       agent.lifecycle = "running";
@@ -2730,6 +2734,9 @@ export class AgentManager {
     prompt: AgentPromptInput,
     options?: AgentSteerOptions,
   ): Promise<ActiveTurnSteerDispatchResult> {
+    // A turn that is still starting has no id to steer yet. Replacing it would drop its
+    // prompt, so wait until it is accepted (or fails) and steer the turn it became.
+    await this.runs.waitForForegroundStart(agentId);
     const agent = this.requireSessionAgent(agentId);
     const expectedTurnId = agent.activeForegroundTurnId ?? agent.activeTurnId;
     if (!expectedTurnId) {

packages/server/src/server/agent/agent-run-state.test.ts

@@ -0,0 +1,51 @@
+import { describe, expect, test } from "vitest";
+
+import { AgentRunState } from "./agent-run-state.js";
+
+function track(promise: Promise<void>): { readonly resolved: boolean } {
+  const state = { resolved: false };
+  void (async () => {
+    await promise;
+    state.resolved = true;
+  })();
+  return state;
+}
+
+async function flushMicrotasks(): Promise<void> {
+  await new Promise<void>((resolve) => setImmediate(resolve));
+}
+
+describe("waitForForegroundStart", () => {
+  test("resolves once the starting turn has an id", async () => {
+    const runs = new AgentRunState();
+    const run = runs.createPendingRun("agent");
+    const wait = track(runs.waitForForegroundStart("agent"));
+
+    await flushMicrotasks();
+    expect(wait.resolved).toBe(false);
+
+    runs.settleForegroundStart(run, { status: "started", turnId: "turn-1" });
+    await flushMicrotasks();
+    expect(wait.resolved).toBe(true);
+  });
+
+  test("waits for a failed start's run to be cleared", async () => {
+    const runs = new AgentRunState();
+    const run = runs.createPendingRun("agent");
+    const wait = track(runs.waitForForegroundStart("agent"));
+
+    runs.settleForegroundStart(run, { status: "failed", error: "start failed" });
+    await flushMicrotasks();
+    expect(wait.resolved).toBe(false);
+    expect(runs.hasPendingRun("agent")).toBe(true);
+
+    runs.settleForegroundRun("agent", run.token);
+    await flushMicrotasks();
+    expect(wait.resolved).toBe(true);
+  });
+
+  test("resolves at once when no turn is starting", async () => {
+    const runs = new AgentRunState();
+    await expect(runs.waitForForegroundStart("agent")).resolves.toBeUndefined();
+  });
+});

packages/server/src/server/agent/agent-run-state.ts

@@ -10,14 +10,18 @@ export interface ForegroundTurnWaiter {
   resolveSettled: () => void;
 }
 
+type ForegroundRunStart =
+  | { status: "pending" }
+  | { status: "started"; turnId: string }
+  | { status: "failed"; error: string };
+
 export interface PendingForegroundRun {
   token: string;
   kind: "foreground";
   stagedEvents: AgentStreamEvent[];
-  start:
+  start: ForegroundRunStart;
-    | { status: "pending" }
+  turnStartedPromise: Promise<void>;
-    | { status: "started"; turnId: string }
+  resolveTurnStarted: () => void;
-    | { status: "failed"; error: string };
   settled: boolean;
   settledPromise: Promise<void>;
   resolveSettled: () => void;
@@ -54,6 +58,28 @@ export class AgentRunState {
     return run?.kind === "foreground" ? run : null;
   }
 
+  settleForegroundStart(
+    run: PendingForegroundRun,
+    start: Exclude<ForegroundRunStart, { status: "pending" }>,
+  ): void {
+    run.start = start;
+    if (start.status === "started") {
+      run.resolveTurnStarted();
+    }
+  }
+
+  /**
+   * Resolves once the agent's foreground run has a turn id or has ended. A failed start
+   * resolves only once its run is cleared, so callers never see a failed run still tracked.
+   */
+  async waitForForegroundStart(agentId: string): Promise<void> {
+    const run = this.getPendingRun(agentId);
+    if (run?.start.status !== "pending") {
+      return;
+    }
+    await Promise.race([run.turnStartedPromise, run.settledPromise]);
+  }
+
   hasPendingRun(agentId: string): boolean {
     return this.getPendingRun(agentId) !== null;
   }
@@ -277,10 +303,16 @@ export class ForegroundTurnStream {
 }
 
 function createPendingForegroundRun(): PendingForegroundRun {
+  let resolveTurnStarted!: () => void;
+  const turnStartedPromise = new Promise<void>((resolvePromise) => {
+    resolveTurnStarted = resolvePromise;
+  });
   return {
     ...createTrackedRunState(),
     kind: "foreground",
     start: { status: "pending" },
+    turnStartedPromise,
+    resolveTurnStarted,
     stagedEvents: [],
   };
 }