CB-118: clip the completion baseline so the CB-115 guard survives >cap blocks

captureBaseline stored the raw, unclipped last-assistant block while resolve()
compares against clip(...) capped at MAX_SCRAPE_CHARS. For a block longer than
4000 chars the two capped representations never match even when the pane is
unchanged, defeating the CB-115 misattribution guard and letting a stale
completion resolve a rapid back-to-back send. Clip the baseline identically.

Regression test: an unchanged >cap block stays suppressed.

Surfaced by the fan-out issue-hunt E2E (1 primary -> 3 concurrent workers,
e2e/issue_hunt_test.py, added here). The same hunt's WorkerService.stop() and
Rendezvous.complete() findings were verified as false positives (locatePane is
already guarded; the sender's finally-close already removes the waiter).

Closes #2
This commit is contained in:
Dai Ha
2026-07-16 09:11:20 +02:00
parent ff6aacdc78
commit 2a61fe69f1
4 changed files with 367 additions and 1 deletions
@@ -156,6 +156,31 @@ class CompletionResolverTest {
assertEquals("No, 391 = 17 × 23.", waiter.getNow(null).text());
}
@Test
void suppressesAnUnchangedCompletionEvenWhenTheBlockExceedsTheScrapeCap() {
// The fan-out issue-hunt finding: captureBaseline once stored the RAW (unclipped) assistant
// block while resolve compares against a clip()'d tail. For a block longer than MAX_SCRAPE_CHARS
// the two capped representations differ even when the pane never changed, so the CB-115
// byte-identical guard failed to fire and a stale completion could resolve the send. Both sides
// must clip identically; here an unchanged >cap block on rapid back-to-back turns stays suppressed.
String longBlock = "⏺ " + "x".repeat(CompletionResolver.MAX_SCRAPE_CHARS + 500) + "\n❯ ";
FakeHerdr herdr = new FakeHerdr().readText(longBlock);
Rendezvous rendezvous = new Rendezvous();
CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous);
var waiter = rendezvous.open("term_a"); // a send is blocked on this turn
resolver.captureBaseline("term_a"); // baseline is the clipped >cap block
var turn = resolver.inFlight("term_a");
assertEquals(CompletionResolver.MAX_SCRAPE_CHARS, turn.baseline().length(),
"the delivery baseline is clipped to the same cap resolve() applies to the tail");
resolver.resolve("term_a", turn); // scrape unchanged → clipped tail == baseline → suppress
assertFalse(waiter.isDone(),
"an unchanged >cap block must still be recognised as stale and suppressed");
assertTrue(rendezvous.isWaiting("term_a"), "the send stays waiting for a real reply");
}
@Test
void resolvesWhenThereIsNoBaseline() {
// No delivery baseline (e.g. the pre-turn read failed) ⇒ never suppress; the completion resolves.