#348: drop a chrome skip that cannot fire, and pin the live pattern shape
CI / contract (push) Successful in 2m10s
CI / build (push) Successful in 2m12s

This commit is contained in:
Dai Ha
2026-09-04 16:57:36 +07:00
parent c1c8794c48
commit 4a5030a5c6
2 changed files with 66 additions and 17 deletions
@@ -618,30 +618,38 @@ public final class CompletionResolver implements TurnListener {
}
/**
* True when the pattern appears before the first sentence ending in the matched pane line, rather
* than after a member has started prose about it.
* True when nothing before the match on this pane line ends a sentence — that is, the match is
* still inside the line's first sentence rather than inside prose a member wrote about it.
* Used to decide whether an exhaustion match may quarantine a credential (fleetd #348).
*
* <p>Leading terminal chrome is skipped first — box-drawing characters, bullets, gutter bars and
* spaces. Exhaustion patterns often name only the decisive words in a provider message, such as
* {@code "usage limit has been reached"}; they do not include its leading {@code "The"}. A bare
* {@code lookingAt} would therefore reject that genuine refusal, including one behind terminal
* chrome.
* <p><strong>Why this is looser than {@link #startsWithBackendError}.</strong> An
* {@code exhaustedPattern} is written per profile and may name only the decisive words of a
* provider message — {@code "usage limit has been reached"} without its leading {@code "The"}.
* A start-of-line check would then reject the genuine refusal. That is the false negative
* fleetd #348's invariant 1 calls the worse direction: an unrecorded exhaustion leaves the
* fleet spawning into a credential with no capacity, and a quarantine runs 1800s against the
* backend-error cooldown's fixed 60s.
*
* <p>A member's prose about a refusal normally follows a completed sentence. That sentence ending
* is enough to keep it from reaching a cooldown sink while the send still fails with the full pane
* tail. This is deliberately less strict than the backend-error check because exhausted patterns
* may start after a provider's leading words.
* <p>This rule accepts a superset of what a start-of-line check accepts: if the match begins
* right after the chrome, there is nothing in front of it, so there is no sentence ending
* either. So moving to it cannot add a false negative.
*
* <p><strong>No chrome skipping here, deliberately.</strong> The first version of this method
* copied {@code startsWithBackendError}'s leading-chrome loop. Measured on merge: deleting that
* loop left all 1369 tests green, and it must — the scan only looks for {@code . ! ?}, and no
* terminal chrome character is one of those. A step that cannot change the result is worse than
* no step, because the next reader takes it as evidence that chrome was handled.
*
* <p>It stays a heuristic. Prose whose <em>first</em> sentence carries the pattern still
* notifies the sink, and a genuine refusal behind an earlier full stop (a hostname, a version
* number) still does not. Both are known and neither is fixed here.
*/
private static boolean startsWithExhaustion(String line, Pattern pattern) {
int i = 0;
while (i < line.length() && !Character.isLetterOrDigit(line.charAt(i))) {
i++;
}
var matcher = pattern.matcher(line);
if (!matcher.find()) {
return false;
}
for (int prefix = i; prefix < matcher.start(); prefix++) {
for (int prefix = 0; prefix < matcher.start(); prefix++) {
if (".!?".indexOf(line.charAt(prefix)) >= 0) {
return false;
}
@@ -649,7 +657,20 @@ public final class CompletionResolver implements TurnListener {
return true;
}
/** True when an error pattern begins the matched pane line after optional terminal chrome. */
/**
* True when the error pattern begins the matched pane line, rather than appearing in prose.
*
* <p>Leading terminal chrome is skipped first — box-drawing characters, bullets, gutter bars and
* spaces. #339 introduced this check with a bare {@code lookingAt}, and that rejected a genuine
* error line rendered as {@code "| 503 Service Unavailable: ..."}: the send still failed, but the
* credential outage was never recorded. That is the false negative #339's own invariant 3 called
* worse than the false positive it set out to fix — measured with a throwaway probe on the
* raw-scrape path, which is exactly the path whose comment says to expect leading chrome.
*
* <p>Skipping only a leading run of non-letter, non-digit characters keeps the fix's intent. A
* member's prose ({@code "I checked the retry path. An API Error: makes it back off."}) still
* does not match, because there the pattern sits after words, not after chrome.
*/
private static boolean startsWithBackendError(String line, Pattern pattern) {
int i = 0;
while (i < line.length() && !Character.isLetterOrDigit(line.charAt(i))) {
@@ -574,6 +574,34 @@ class CompletionResolverTest {
"a real exhaustion behind terminal chrome must reach the sink");
}
/**
* The live fleet configures {@code exhaustedPattern: "The usage limit has been reached"} — with
* the leading {@code "The"}. Every other test here uses a pattern without it, which is the shape
* that made fleetd #348 need a looser rule than a start-of-line check. This pins the deployed
* shape as well, so a later tightening of {@link CompletionResolver} cannot silently stop
* recording the exhaustion this fleet actually reports.
*
* <p>What it does not prove: that this is the only pattern shape an operator will write.
*/
@Test
void anExhaustionPatternCarryingItsLeadingWordsStillNotifiesTheSink() {
String block = "⏺ │ The usage limit has been reached. Try again later.\n❯ ";
FakeHerdr herdr = new FakeHerdr().readText(block);
Rendezvous rendezvous = new Rendezvous();
ExhaustedPatternLookup patterns = target -> Pattern.compile("The usage limit has been reached");
java.util.List<String> notified = new java.util.ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
var waiter = rendezvous.open("term_a");
resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
"the live pattern shape must still fail the send as exhausted");
assertEquals(1, notified.size(),
"the live pattern shape must still reach the sink");
}
@Test
void aLosingBackendExhaustedClassificationNeverNotifiesTheExhaustionSink() {
// The waiter was already resolved (e.g. by the worker's own reply) before this scrape landed —