Repository navigation
Conversation
|
Current-main cache integration is local at Historical measured reads remain2816/8250 for a synthetic repeat-prefix workload; enabled was slower in both pairs. The original scripts explicitly prohibited tools, and the written issue does not require inventing a new tool-using live measurement. No provider/model experiment was repeated. Final source, findings-only follow-up, extracted/package acceptance and fresh exact-head CI remain pending; not merge-ready. |
|
Replacement draft #323 is published at The full initial review requested two changes; both now have committed native fixes and unchanged original overflow reproduction proof. Single findings-only completion was interrupted by a typed provider error after28events, without request replay, and remains unfinished. Full/package/native final acceptance is also running. #323 is a draft requesting fresh exact-head CI, not merge approval. Closing this superseded delivery in favor of #323 avoids two open PRs for the same cache scope. No assertion/workflow weakening, lost historical negative results, invented tool-using measurement requirement or merge. |
Recreated after the llxprt account suspension removed original PR #295. Head
66225a9fccf059a6d24626fd6d4b85153dfe8720on restored branchbranch4-issue79-prompt-cache.Contents: stateless prompt-cache routing/observation for the configured provider, vendor patch provenance, serialized-prefix coverage; matched live cache-on/off measurement recorded (off: 8250 input tokens/0 cached; enabled: 8250 input/2816 cached, 34.13% hit ratio) with the enabled arm slower in both final pairs and no latency-benefit claim; zero-hit and transient-503 evidence retained.
Verification recorded: 68 model-API, 85 profile, HTTP cache-integration, provenance/license tests pass. Remaining before ready: final shared #252/#286 integration, full/release/archive/extracted closure on final source, and independent native review. Do not merge.