mirror of
https://github.com/diegosouzapw/OmniRoute.git
synced 2026-08-27 01:22:10 +03:00
2dfb730241
A universal handoff whose summary comes back unparseable persists nothing, so shouldGenerateUniversalHandoff keeps answering "generate" and the very next model switch in the same session re-issues the same full-history summarization call and discards the answer again — forever. With a switch-heavy combo strategy (weighted, random, round-robin, p2c) the models alternate on almost every turn, so that background call lands on a large fraction of requests: an upstream call whose response nobody reads is real money on a paid provider and real quota on a metered one. Measured on the weighted 70/30 matrix, that inflated the observed openai share to 0.895 where routing actually delivered 0.70, and at the unit level 199 of 200 model switches issued a fresh discarded summarization call. Back off per (session, combo) after an answer that is not a usable handoff: exponential 5min -> 1h, cleared on the first successful generation, capped at 500 tracked keys. Deliberately narrow — a transient upstream failure (!response.ok) is NOT tracked, so it still retries on the next switch, which is the behavior the context-relay path already depends on. After the fix, same harness at n=200: 201 upstream calls for 200 requests (1 extra, 0.5%), and the measured openai share equals the delivered share (0.725). The unit guard drops 199 discarded calls to 1. Refs #11552