Give stall recovery a full timeout grace after the first nudge

Queued empty stall pings that drain right after the first nudge used
to count as a second consecutive stall and stop the leaf immediately.
Record stallNudgeAt on the nudge, wait through pings inside
stallTimeoutMs without treating them as activity or restarting grace,
and stop only once the grace elapses with no tool.done or turn reset.

Author: TheGreatAxios

src/subagent/nudge-director.test.ts

@@ -880,6 +880,88 @@ describe("SubAgentDirector post-complete terminalization (CL-7068)", () => {
   });
 });
 
+describe("SubAgentDirector stall nudge grace", () => {
+  test("queued pings inside grace wait; stop only after grace with no activity", async () => {
+    let now = 1_000_000;
+    const director = new SubAgentDirector(
+      "system",
+      [],
+      undefined,
+      1_000,
+      () => now,
+    );
+    const caps = capabilities();
+
+    await director.decide(inferenceDoneText("working"), state, caps);
+
+    now += 1_000;
+    const first = actions(
+      await director.decide(messageReceived(""), state, caps),
+    );
+    expect(first).toContainEqual({
+      type: "checkpoint",
+      message: "subagent-stall-nudge",
+    });
+
+    now += 200;
+    const midGrace = actions(
+      await director.decide(messageReceived(""), state, caps),
+    );
+    expect(midGrace).toEqual([{ type: "wait" }]);
+
+    now += 200;
+    const stillGrace = actions(
+      await director.decide(messageReceived(""), state, caps),
+    );
+    expect(stillGrace).toEqual([{ type: "wait" }]);
+
+    now += 600;
+    const stopped = actions(
+      await director.decide(messageReceived(""), state, caps),
+    );
+    expect(stopped).toContainEqual({
+      type: "checkpoint",
+      message: "subagent-stalled",
+    });
+    expect(stopped.some((action) => action.type === "reply")).toBe(true);
+  });
+
+  test("tool.done during grace clears stallNudgeAt so a later silence nudges again", async () => {
+    let now = 2_000_000;
+    const director = new SubAgentDirector(
+      "system",
+      [],
+      undefined,
+      1_000,
+      () => now,
+    );
+    const caps = capabilities();
+
+    await director.decide(inferenceDone(["read-1"]), state, caps);
+    now += 1_000;
+    const first = actions(
+      await director.decide(messageReceived(""), state, caps),
+    );
+    expect(first).toContainEqual({
+      type: "checkpoint",
+      message: "subagent-stall-nudge",
+    });
+
+    now += 100;
+    await director.decide(toolDone("read-1"), state, caps);
+
+    now += 1_000;
+    const afterActivity = actions(
+      await director.decide(messageReceived(""), state, caps),
+    );
+    expect(afterActivity).toContainEqual({
+      type: "checkpoint",
+      message: "subagent-stall-nudge",
+    });
+    expect(afterActivity.some((action) => action.type === "reply")).toBe(false);
+  });
+});
+
 function stubAdmission(
   notes: { provider: string; until: number }[],
 ): AdmissionQueue {

src/subagent/nudge-director.ts

@@ -172,7 +172,11 @@ export class SubAgentDirector extends DefaultDirector {
   private readonly stallTimeoutMs: number | undefined;
   private readonly now: () => number;
   private lastActivityAt: number;
-  private consecutiveStalls = 0;
+  // Wall clock when the first stall nudge was issued. Later empty pings inside
+  // stallTimeoutMs of this instant wait without stopping or restarting grace;
+  // stop only after the grace elapses with no activity. Cleared on real
+  // tool.done / turn-boundary activity.
+  private stallNudgeAt: number | undefined;
   private lastAssistantText = "";
   // Every stop and nudge is recorded with its measured value beside its
   // threshold, so a later threshold change can cite data instead of judgment
@@ -299,7 +303,7 @@ export class SubAgentDirector extends DefaultDirector {
     if (onTurnBoundary(event)) {
       this.lastConsumedNudgeText = null;
       this.lastActivityAt = this.now();
-      this.consecutiveStalls = 0;
+      this.stallNudgeAt = undefined;
       this.compaction.noteInferenceDone(event, state.turns);
       this.turnsCompleted++;
       const content = event.turn.content as readonly {
@@ -401,7 +405,7 @@ export class SubAgentDirector extends DefaultDirector {
     }
     if (event.type === "tool.done") {
       this.lastActivityAt = this.now();
-      this.consecutiveStalls = 0;
+      this.stallNudgeAt = undefined;
       if (event.result.isError === true) {
         // Failed-tool recovery guidance. Arm once; coalesce consecutive failure
         // audits until applyPendingNudge flushes a single counted record.
@@ -429,12 +433,12 @@ export class SubAgentDirector extends DefaultDirector {
    * "no pending harness-tracked work" falls out of when this method can run
    * at all rather than needing separate bookkeeping.
    *
-   * First stall past the timeout: one continuation nudge, asking the leaf to
-   * report status or keep going. A second consecutive stall (no activity
-   * since the nudge) escalates to the existing salvage path, same shape as
-   * the turn-boundary checks above. Returns null when this event is not
-   * a stall check the director should act on (let it fall through as an
-   * ordinary continuation).
+   * First silence past the timeout: one continuation nudge, and record
+   * stallNudgeAt. Queued pings that arrive inside the stallTimeoutMs grace
+   * after that nudge neither stop nor restart the grace (and do not count as
+   * activity). Stop only when a ping arrives after the grace with still no
+   * activity. Returns null when this event is not a stall check the director
+   * should act on (let it fall through as an ordinary continuation).
    */
   private checkStallPing(
     event: ReactorInboundEvent,
@@ -447,8 +451,8 @@ export class SubAgentDirector extends DefaultDirector {
     const elapsed = this.now() - this.lastActivityAt;
     if (elapsed < this.stallTimeoutMs) return null;
 
-    this.consecutiveStalls++;
-    if (this.consecutiveStalls === 1) {
+    if (this.stallNudgeAt === undefined) {
+      this.stallNudgeAt = this.now();
       this.interventions({
         id: "stall-nudge",
         class: "nudge",
@@ -464,6 +468,14 @@ export class SubAgentDirector extends DefaultDirector {
         inferWithSubAgentNudge(capabilities, SUBAGENT_STALL_NUDGE),
       ];
     }
+
+    const sinceNudge = this.now() - this.stallNudgeAt;
+    if (sinceNudge < this.stallTimeoutMs) {
+      // Still inside the post-nudge grace. Wait without faking activity or
+      // restarting the grace clock.
+      return [capabilities.wait()];
+    }
+
     this.interventions({
       id: "stalled",
       class: "stop",