[WRONG BRANCH] fix(cursor): bound output guard reasoning quarantine

src/adapters/cursor.ts

@@ -31,6 +31,7 @@ import { rememberCursorThreadConversation } from "./cursor/thread-continuity";
 import { runCursorTurnWithRetry } from "./cursor/transport-retry";
 import { cursorRequestHasShellAlias, cursorRequestUsesCodeMode } from "./cursor/tool-definitions";
 import {
+  CURSOR_OUTPUT_GUARD_MAX_HOLD_BYTES,
   CURSOR_ECHO_RETRY_CONTINUATION_TEXT,
   CURSOR_ROUTING_COMMENTARY_RETRY_TEXT,
   CursorEnvelopeEchoSniffer,
@@ -242,12 +243,26 @@ export function createCursorAdapter(provider: OcxProviderConfig, deps: CursorAda
             ? new CursorRoutingCommentarySniffer()
             : undefined;
           let guardHeld: AdapterEvent[] = [];
+          let guardHeldBytes = 0;
+          const guardEncoder = new TextEncoder();
           const releaseGuardHeld = () => {
             for (const held of guardHeld) {
               if (held.type !== "heartbeat") emittedOutput = true;
               emit(held);
             }
             guardHeld = [];
+            guardHeldBytes = 0;
+          };
+          const holdGuardEvent = (event: AdapterEvent) => {
+            guardHeld.push(event);
+            // Count the complete retained representation, including per-event overhead, so an
+            // upstream cannot evade the cap with empty or non-text reasoning frames.
+            guardHeldBytes += guardEncoder.encode(JSON.stringify(event)).byteLength;
+            if (guardHeldBytes <= CURSOR_OUTPUT_GUARD_MAX_HOLD_BYTES) return true;
+            echoSniffer?.finish();
+            routingCommentarySniffer?.finish();
+            releaseGuardHeld();
+            return false;
           };
           const guardsSettled = () =>
             (!echoSniffer || echoSniffer.settled)
@@ -286,7 +301,7 @@ export function createCursorAdapter(provider: OcxProviderConfig, deps: CursorAda
              for (const event of events) {
                 if (!guardsSettled()) {
                   if (event.type === "text_delta") {
-                    guardHeld.push(event);
+                    if (!holdGuardEvent(event)) continue;
                     if (echoSniffer && !echoSniffer.settled) {
                       const decision = echoSniffer.feed(event.text);
                       if (decision.kind === "echo") {
@@ -306,7 +321,7 @@ export function createCursorAdapter(provider: OcxProviderConfig, deps: CursorAda
                   } else if (event.type === "thinking_delta" || event.type === "heartbeat") {
                     // Reasoning before first text stays ordered; liveness still passes through.
                     if (event.type === "thinking_delta") {
-                      guardHeld.push(event);
+                      holdGuardEvent(event);
                       continue;
                     }
                   } else {

src/adapters/cursor/envelope-echo.ts

@@ -16,7 +16,7 @@ const ECHO_MARKERS = ["[Tool Result]", "[Tool Error]", "[tool_result]"] as const
 const MAX_SNIFF_BYTES = 40;
 const MAX_ROUTING_COMMENTARY_BYTES = 512;
 /** Aggregate quarantine cap: past this, flush and disarm. */
-const MAX_HOLD_BYTES = 8 * 1024;
+export const CURSOR_OUTPUT_GUARD_MAX_HOLD_BYTES = 8 * 1024;
 const encoder = new TextEncoder();
 
 export class CursorToolResultEchoError extends Error {
@@ -70,7 +70,11 @@ export class CursorEnvelopeEchoSniffer {
     const stillPrefix = ECHO_MARKERS.some(marker =>
       probe.length < marker.length && marker.startsWith(probe),
     );
-    if (stillPrefix && this.byteCount <= MAX_SNIFF_BYTES && this.buffered.length < MAX_HOLD_BYTES) {
+    if (
+      stillPrefix
+      && this.byteCount <= MAX_SNIFF_BYTES
+      && this.buffered.length < CURSOR_OUTPUT_GUARD_MAX_HOLD_BYTES
+    ) {
       return { kind: "hold" };
     }
     this.done = true;
@@ -129,7 +133,7 @@ export class CursorRoutingCommentarySniffer {
       && lineBreakCount < 2;
     if (
       this.byteCount < MAX_ROUTING_COMMENTARY_BYTES
-      && this.buffered.length < MAX_HOLD_BYTES
+      && this.buffered.length < CURSOR_OUTPUT_GUARD_MAX_HOLD_BYTES
       && (lineBreakCount === 0 || pendingFailureClaim)
       && (hasRoutingHint || this.byteCount < 64)
     ) {

tests/cursor-envelope-echo-retry.test.ts

@@ -154,6 +154,37 @@ describe("cursor external output quarantine + corrective retry (devlog 260826 ga
     expect(text).toBe("[note] leading bracket but not an envelope");
   });
 
+  test("reasoning-only quarantine is capped and disarms before unbounded retention", async () => {
+    let attempt = 0;
+    const factory = () => ({
+      async *run() {
+        attempt += 1;
+        for (let i = 0; i < 100; i += 1) {
+          yield { type: "thinking", thinking: "x".repeat(128) } satisfies CursorServerMessage;
+        }
+        // Once the aggregate hold cap flushes, later marker-like text is ordinary output rather
+        // than evidence for a retry whose preceding reasoning has already reached the client.
+        yield { type: "text", text: ECHO_TEXT } satisfies CursorServerMessage;
+        yield { type: "done", usage: { inputTokens: 1, outputTokens: 1 } } satisfies CursorServerMessage;
+      },
+      writeClient() {},
+    });
+    const adapter = createCursorAdapter(
+      { ...provider, apiKey: "cursor-token" },
+      { createTransport: factory as never },
+    );
+    const events: AdapterEvent[] = [];
+    await adapter.runTurn?.(
+      toolResultBody("cursor/kimi-k3"),
+      { headers: new Headers() },
+      event => events.push(event),
+    );
+
+    expect(attempt).toBe(1);
+    expect(events.filter(event => event.type === "thinking_delta")).toHaveLength(100);
+    expect(events.filter(event => event.type === "text_delta")).not.toHaveLength(0);
+  });
+
   test("plain user turns (no trailing toolResult) never arm the sniffer", async () => {
     let attempt = 0;
     const factory = () => ({