Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2.4 KiB
Vikunja: #460 — Retry budget: track all seen gate fingerprints, charge on repeat (cycle detection)
Problem
DefaultRetryCoordinator.decide compares the current failure fingerprint against only the previous one (core/kernel/src/main/kotlin/com/correx/core/kernel/retry/DefaultRetryCoordinator.kt:29):
val prevFingerprint = state.gateFailureFingerprints[gate]
val progressed = prevFingerprint != fingerprint
val charged = !progressed
A single slot cannot see a cycle. Whack-a-mole (fix A breaks B, fix B breaks A) alternates two fingerprints forever, every round reads as progress, nothing is ever charged, and the per-gate budget never triggers. DefaultSessionOrchestratorStep.kt:296 already acknowledges this ("can be fooled into treating as still progressing indefinitely") but guards only inside the recovery stage. A normal implementer stage has no equivalent guard.
Fix
Make OrchestrationState.gateFailureFingerprints[gate] a Set<String> of every fingerprint seen for that gate, and charge when the current one is a repeat:
val seen = state.gateFailureFingerprints[gate]
val charged = fingerprint in seen
Touches: OrchestrationState field type, the reducer that writes it, DefaultRetryCoordinator.decide.
Why not count-based progress
Rejected: "charge unless the failing-item count went down". A syntax error masks later errors — fix '}' expected and tsc parses further, reporting 5 real type errors where there was 1. Count goes 1 to 5 on the most productive edit of the run, and the guard would charge the budget for genuine progress.
Cases
- Unmasking,
{syntax}to{5 type errors}: new fingerprint, unseen, free. Correct. - Grinding down, 3 to 2 to 1 errors: each state distinct, all free. Correct.
- Whack-a-mole, A/B/A: round 3 repeats round 1, charged. This is the case the single slot misses today.
Known limit
Detects cycles, not progress. A stage emitting a fresh distinct failure every round is still unbounded; only stageCount catches it. Worth measuring with scripts/artread.py over a few failed sessions before assuming that pattern matters.
Acceptance criteria
- (fill in before starting)
Quality gate
# the command that must pass, e.g. make check
Rules
- Do not edit this file.
- One task, one session, one PR. Keep the diff under ~300 lines.
- Finish with
task pr.