The extra-context layer for AI agents — L1 eviction, sticky L2/L3 context pools, context paging, cold-tier compression.
Registered before these solver and review calls. Four previously unattempted open issues from four repositories have complete nonempty public discussions and different retrieved comment-ID sets for the two arms. The manifest, SHA-256 01e432b3ab398c6124b0c9bc2d8c67ef18cfbe8d1d3462fdff4e549d6f997393, freezes source commits, contexts, partial checks, paired assignment order, models, budgets and failure policy. Selection followed prior trial results and the batch 28 timeout; several projects also appear in earlier batches. New issue identities do not establish statistical independence.
| Open issue | Frozen partial acceptance | Proposed reference |
|---|---|---|
| Koa #827 | With proxying enabled, forwarded IPv4 and bracketed IPv6 addresses with ports return addresses without ports; plain addresses and the proxy-disabled socket address remain valid. | Narrow manual port of PR #1205 to the current parser. Malformed inputs and broader security properties are outside the check. |
| Day.js #2179 | Duration strings containing hours and minutes, seconds, or day and millisecond parts yield expected values; ISO and numeric durations remain valid. | Source-only PR #3193 is used as a reference; its association with this issue is not asserted. |
| date-fns #1134 | W, WW and Wo format the week within a month for selected dates and week starts; ordinary date formatting remains valid. |
Path-translated PR #4186, manually corrected to handle Wo. The original reference passed W but failed Wo; that failed check is preserved. |
| Express #7362 | res.send handles ArrayBuffer and SharedArrayBuffer bytes with binary content type, while ordinary Buffer and JSON paths still work. |
Source-only PR #7363. The check invokes the real response prototype with an in-memory response rather than a network server. The issue itself describes a proposed fix. |
The qualification ledger preserves baseline failures, reference passes, controls and original and adapted reference patches. These checks are narrow and the proposed references do not establish maintainer acceptance. The TypeScript check uses a test-only source transpiler. Solver, two AI reviewers, a separate AI adjudicator, bounded tool-overflow handling, 503-only retries and stage stop rules follow batch 28. The estimated budget is below $1, not a hard cap. All labels use the same proxy, whose true upstream model provenance is not verified. Pre-seeded public issue discussion is not a real long-running production session or a test of autonomous memory management.