Memex

Logo

The extra-context layer for AI agents — L1 eviction, sticky L2/L3 context pools, context paging, cold-tier compression.

View the Project on GitHub QiuXu-Tech/memex

Previously unstarted Chart.js pair — registered follow-up

Registered after a transport failure, before these new calls. Batch 25 stopped after a solver connection error in a different issue. Both assignments for Chart.js #12191 were recorded NOT_RUN_AFTER_STOP; the selection check confirms neither slot has a solver-call start. The new manifest, SHA-256 b93eee9f75f63e37db0b609b63f73ab47f452911629565474c18c0f8e236cae4, binds the previous manifest and solver-log hashes, unstarted result hashes, unchanged checks and a newly randomized arm order. The failed request in batch 25 is not retried, and all original statuses remain visible.

The same source, test-only TypeScript hook, dependency lock, narrow mocked-controller checks, AI rubric, solver, reviewers, non-independent adjudicator and 503-only retry policy follow the original registration. Stop the stage after provider failure or missing usage; do not retry timeouts. A new three-model short-request health probe succeeded after the transport error, which does not establish long-request reliability.

This is one previously unstarted issue, not an additional distinct issue. It has no historical discussion records, so both conditions receive different empty-result messages and cannot demonstrate a memory effect. Selection after failure, source disclosure, partial tests and common-proxy AI limit inference. Estimated model-call target is below $1, not a hard cap; unknown charges remain unknown.