Memex
The extra-context layer for AI agents — L1 eviction, sticky L2/L3 context pools, context paging, cold-tier compression.
Two further Prettier task checks
Registered before these task calls and after inspecting prior batches. These two additional open issues are separate from earlier registrations, not a pre-registered 30–50-task study. Their frozen manifest has SHA-256 5e53c07679c582279738a4ef8d3d33044096706c9c9e9dcb47c14266ef805e5b and binds the randomized paired order, scripts, checks, issue snapshots and contexts.
| Issue | Frozen criterion | Positive reference used only for check validation |
|---|---|---|
| Prettier #19955 | Five parenthesized arrow/comment examples, including the four reported cases, must format idempotently and retain the indicated comments. | Source hunks of PR #20008, head be7c77345641a959fa4b4fccda3ba1ffd59ba936 |
| Prettier #19968 | Two reported TypeScript return-type examples and the discussion’s union-return example must format idempotently. | Source hunks of PR #19970, head 9e105df0f7cf034f63ae5a01094611e4e4aaec4a |
On pinned Prettier commit bed4edbb1cf2a9db4004215b0ce5707cdb02b5d9, the #19955 issue check failed four of five cases and the #19968 check failed all three. Each corresponding source-only reference patch passed its issue check; both baseline and reference passed the small custom regression controls. Public examples also failed on the baseline and passed on the reference. This does not establish full project behavior or upstream acceptance. The discussion of #19955 includes detailed proposed solutions, and the #19968 discussion compares existing PRs; the solver can see retrieved comments, so prior solution exposure is a material confound.
The protocol, explicit AI scoring rubric, restricted VM tools, stage stopping rules and cost limitations are the same as the five-issue batch, with a separately randomized order, two new issues, at most 12 tool turns per assignment and an estimated batch target below $1 rather than a hard cap. gpt-6-astra is the solver and adjudicator, deepseek-flash and gpt-6.1-sol are the blinded AI reviewers. Missing and invalid results remain missing; the human-review, production, cost-effectiveness and autonomous-memory claims remain unverified.