fix(reply): preserve preflight failure handling

steipete · steipete · commit dc6888fd6bda · 2026-07-06T03:56:41.000-04:00
diff --git a/src/auto-reply/reply/agent-runner-direct-runtime-config.test.ts b/src/auto-reply/reply/agent-runner-direct-runtime-config.test.ts
@@ -491,6 +491,29 @@ describe("runReplyAgent runtime config", () => {
     expect(runAgentTurnWithFallbackMock).toHaveBeenCalledOnce();
   });
 
+  it("surfaces unrelated preflight failures after an exhausted memory flush", async () => {
+    const { replyParams } = createDirectRuntimeReplyParams({
+      shouldFollowup: false,
+      isActive: false,
+    });
+    runMemoryFlushIfNeededMock.mockResolvedValue({
+      sessionEntry: { sessionId: "session-1", updatedAt: 1, compactionCount: 4 },
+      outcome: "exhausted",
+    });
+    runPreflightCompactionIfNeededMock.mockRejectedValue(
+      new Error("Preflight compaction required but failed: auth profile mismatch"),
+    );
+
+    const result = await runReplyAgent(replyParams);
+
+    if (!result || Array.isArray(result)) {
+      throw new Error("expected a single preflight compaction failure reply payload");
+    }
+    expect(result.text).toContain("auto-compaction could not recover");
+    expect(resetReplyRunSessionMock).not.toHaveBeenCalled();
+    expect(runAgentTurnWithFallbackMock).not.toHaveBeenCalled();
+  });
+
   it("does not start the main turn after cancellation during memory flush", async () => {
     const { replyParams } = createDirectRuntimeReplyParams({
       shouldFollowup: false,
diff --git a/src/auto-reply/reply/agent-runner.ts b/src/auto-reply/reply/agent-runner.ts
@@ -10,6 +10,7 @@ import {
 } from "../../agents/agent-scope.js";
 import { resolveContextTokensForModel } from "../../agents/context.js";
 import { DEFAULT_CONTEXT_TOKENS } from "../../agents/defaults.js";
+import { isLikelyContextOverflowError } from "../../agents/embedded-agent-helpers/errors.js";
 import {
   hasCommittedSourceReplyDeliveryEvidence,
   hasVisibleCommittedMessagingToolDeliveryEvidence,
@@ -1661,7 +1662,11 @@ export async function runReplyAgent(params: {
       preflightCompactionApplied =
         (activeSessionEntry?.compactionCount ?? 0) > prePreflightCompactionCount;
     } catch (err) {
-      if (memoryFlushResult.outcome !== "exhausted" || replyOperation.abortSignal.aborted) {
+      const canRotateAfterPreflightFailure =
+        memoryFlushResult.outcome === "exhausted" &&
+        !replyOperation.abortSignal.aborted &&
+        isLikelyContextOverflowError(String(err));
+      if (!canRotateAfterPreflightFailure) {
         throw err;
       }
       logVerbose(`Preflight compaction could not recover exhausted memory flush: ${String(err)}`);
