Implemented
0
feat(signals): report research agent remembers judgments in the fleet scratchpad
Source: PostHog/posthog#91269 · opened by @andrewm4894
What Give the report research agent write access to the shared scout scratchpad (scout-scratchpad-remember / scout-scratchpad-forget) so per-report judgments persist — "report on X judged noise because Y", keyed on the entity — and scouts plus future research runs inherit them instead of re-deriving. Why Read-only memory (#91267) surfaces what the fleet knows, but the research stage still re-derives its own judgments on every run. Writes close that loop: a judgment lands once, entity-keyed, and the whole system (scouts, later research runs, implementation runs) can find it by searching the entity. Implementation sketch • Scopes. scout-scratchpad-remember/-forget require signal_scout_internal:write — server-mint-only, granted today only via the signals_scout / signals_scout_reports presets (posthog/temporal/oauth.py::resolve_scopes). Two options: - a new signals_research preset = the read posture plus SCOUT_INTERNAL_SCOPES, reporting write scopes so the MCP layer…
No pledges yet. Be the first to back this.
Comments
Similar requests
feat(signals): report research agent reads steering notes and fleet memory
0 votes · 0 comments
feat(signals): pipeline audience targets for steering notes
0 votes · 0 comments
feat(tasks): implementation runs read steering notes and fleet memory
0 votes · 0 comments
feat(signals): scheduled follow-up checks on a report
0 votes · 0 comments
feat(signals): report research steering eval suite
0 votes · 0 comments
No comments yet.