Role archetypes are agents dedicated to the various hats you need to wear. They are roles with sub-agents, managed by Tuner.
Codify weekly-agent backfill rule + plan to re-enable cloud schedules (laptop-sleep failsafe, 3rd consecutive Friday miss)
Friday Apr 25 weekly agents did not fire. Same root cause as W15 + W16: Ricky was off, laptop asleep, no session opened to register the cron. The "Step 2: Catchup" rule currently says ">12 hours missed → skip (stale)." For weekly agents this is too aggressive — the data is still meaningful Mon morning, and the backfill is the user's de-facto behavior already (we did it this morning). Three consecutive weeks of partial Friday miss now (W15, W16, W17). Codify the backfill so it doesn't keep being a manual ask. Part B (re-enable cloud schedules for weekly agents only) is the structural fix but requires RunLayer status changes — keeping that as a follow-up proposal.
Stream chat — add persist_context tool so topics survive across sessions
Three explicit user asks on Apr 22 in the chat agent: "persist that context moving forward", "save this for later", "get up to speed on X moving forward." Each turn was answered well in the moment, but the context evaporated by the next session. User filled the gap manually by re-pasting transcripts. Pattern: chat agent acts as a research scratchpad with no memory.
Stream chat — tighten response-length nudge (median ~700 chars over 600-char ceiling)
104 stream chat turns Apr 22-24. Median response_chars ≈ 700, over the stated 600 ceiling. Many longer responses were on legitimately heavy questions, but drift is real — a "tell me more on Security audit MGR-OS" turn returned 1623 chars when 600 would have done. The current prompt says "one sentence usually right, two often right, three is already long" — that's being interpreted loosely. Strengthen the nudge.
Stream chat — strengthen writing-style anchor (explicit Apr 24 user feedback: AI-tone leaking through)
Apr 24 explicit user feedback: "it's still giving AI like 'that's real'". One data point + direct user signal = act now (per CLAUDE.md "explicit human instruction is highest-confidence signal"). AI-tone phrases leaking through the writing-style anchor: "that's real", "you've got this", "totally fair", "appreciate you sharing", and similar pseudo-affirming filler. The current prompt cites the writing-style.yaml file as voice baseline but doesn't explicitly enumerate the AI-tone phrases to AVOID.