Part of #448 ## Question About 55 candidate rows are feature requests, not defects (`tmp/upstream-review/` lists them: issues 259, 345, 392, 420, 433, 523, 524, 586, 595, 605, 620, 695, 730, 732, 765–770, 850, 880, 884, 899, 907, 919, 948, 981, 985, 1007, 1009, 1018, 1274, 1316, 1370, 1390, 1434, 1564, 1566; PRs 1247, 223, 467, 1011, 194, 241, 267, 376, 583, 617, 672, 679, 1779). Which address a cost the Operator actually pays, judged against the Goal and their Session history? Those get a verification ticket; the rest are rejected with the reason.
Activity
possiblyneal commentedon Nov 18, 2026
Resolution
52 distinct items triaged (39 issues, 13 PRs); full table with evidence per item in
tmp/verify/470/triage.md(local, quotes no transcripts).docs/experiments.mdalready rejected (Add option to include low-score results in semantic search rohitg00/agentmemory#948 relevance floor, fix(summarize): tell a chunk's retry why its first reply was rejected #345 graph, per Measure the entity graph's Recall value; turn off graph extraction if it adds none #420), falls in a CLAUDE.md wont-fix class (multi-user/LAN, another host, third-party provider, Windows, viewer cosmetics), or shows no measurable cost (transcripts use 8 of 43 MCP tools; no wrong-tool picks for Reduce MCP tool surface area rohitg00/agentmemory#620/MCP tool descriptions lack 'when to use' guidance — LLM struggles to choose the right tool among 53 rohitg00/agentmemory#884; feat: allow a cheaper model for compression than for summarization rohitg00/agentmemory#899/feat(budget): per-session token cap with soft warn + hard block rohitg00/agentmemory#767 save no billed cost on a local broker).Retry if: the Operator starts using a record kind listed above as empty, or a missed-injection or replay measurement points at one of these features.