mirror of
https://github.com/clash-verge-rev/clash-verge-rev.git
synced 2026-10-18 23:30:51 +08:00
Tunglies
f754ee53a0
Two months of usage data (56 evaluated runs, 1372 AI credits) show a median of 22 credits and 6 tool turns per review, with 97% of billed context being cache tokens anchored on the prompt itself. Cap each run at 12 turns and 60 credits so only runaway sessions are stopped, and stop triggering on reopened, which carries no new implementation evidence; workflow_dispatch remains available for manual re-reviews. Compact the policy body from 367 to 177 lines by merging the signal lists with the weighting/escalation sections that repeated them, with no verdict, label, rerun-stability, or output-contract semantics dropped. Add a tool-call budget for normal evaluations and a fast path for fork-sync and placeholder PRs, which were 29% of measured credits.