From fc185d9de538108263a478a53fd616337175d1fe Mon Sep 17 00:00:00 2001 From: Xiangzhe Date: Wed, 26 Aug 2026 01:26:19 -0300 Subject: [PATCH] docs(changelog): correct the #11552 waste figure to the measured rate MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The fragment claimed 199 of 200 model switches fired a discarded call. That number is the TDD proof — summarization calls counted with the new guard disabled — not the observed waste. The measured run was ~254 upstream calls for 200 requests, i.e. roughly one request in four, dropping to 201 after the fix. Also state explicitly that the delivered 70/30 share was correct all along, so the entry cannot be read as a routing defect. --- .../fixes/11552-handoff-discarded-summarization-calls.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/changelog.d/fixes/11552-handoff-discarded-summarization-calls.md b/changelog.d/fixes/11552-handoff-discarded-summarization-calls.md index 6eb583f55f..ad9c127779 100644 --- a/changelog.d/fixes/11552-handoff-discarded-summarization-calls.md +++ b/changelog.d/fixes/11552-handoff-discarded-summarization-calls.md @@ -1 +1 @@ -- **fix(sse):** a universal handoff whose summary comes back unusable is no longer regenerated on every single model switch, which was burning paid quota on upstream calls whose answers were thrown away ([#11552](https://github.com/diegosouzapw/OmniRoute/issues/11552)) — nothing is persisted when the summary does not parse, so the next switch in the same session re-issued the same full-history summarization request and discarded it again, forever. With a switch-heavy combo strategy (weighted, random, round-robin, p2c) that landed on a large share of requests: measured at n=200, **199 of 200 model switches fired a fresh discarded call**, and the extra traffic skewed a weighted 70/30 combo to an observed 0.895 share for one provider. There is now an exponential back-off per (session, combo) — 5 min up to 1 h, cleared on the first successful handoff, capped at 500 tracked keys. A transient upstream failure is deliberately **not** tracked, so it still retries immediately. After the fix: 201 upstream calls for 200 requests, and the measured share matches the delivered one. +- **fix(sse):** a universal handoff whose summary comes back unusable is no longer regenerated on every single model switch, which was burning paid quota on upstream calls whose answers were thrown away ([#11552](https://github.com/diegosouzapw/OmniRoute/issues/11552)) — nothing is persisted when the summary does not parse, so the next switch in the same session re-issued the same full-history summarization request and discarded it again, forever. With a switch-heavy combo strategy (weighted, random, round-robin, p2c) that landed on a large share of requests: measured at n=200, roughly **one request in four carried an extra discarded upstream call**, and that traffic skewed a weighted 70/30 combo to an observed 0.895 share for one provider even though the share actually delivered to the client was a correct 0.70. There is now an exponential back-off per (session, combo) — 5 min up to 1 h, cleared on the first successful handoff, capped at 500 tracked keys. A transient upstream failure is deliberately **not** tracked, so it still retries immediately. After the fix: 201 upstream calls for 200 requests, and the measured share matches the delivered one.