Problem it solves: run the whole toolkit on a schedule without letting an autonomous loop write anything consequential unattended - the Python drivers own the deterministic pipeline, the skills only narrate and stage, and every judgment call waits for the EM's tick.
The flagship morning routine. The deterministic driver (lib/em_routines/morning.py) owns the whole pipeline - sweep every assigned account via em_db.morning.sweep, fuse cross-domain composite signals, emit persisted hygiene, risk and composite signals via em_db.signals.emit, stage judgment items as proposals via em_db.proposals.stage (it never writes risks or feedback directly), and upsert today's daily plan via em.daily_plan_upsert leading with what changed since yesterday. The skill only narrates the plan and proposal prose on top of it. The run ends with trust suggestions and factual maintenance (proposals.expire, signals.expire).
The evening closer. The deterministic driver (lib/em_routines/closer.py) auto-approves and applies only trust-ladder 'auto' proposals, applies every human-approved proposal via em_db.proposals.apply, turns apply failures into 'act' signals, and closes today's plan via em.daily_plan_close with a structured EOD summary. On top of the driver the skill auto-logs the day's factual meeting captures (delegating to the 1:1 / standup / retro / profile logger skills) and stages the day's client conversation logs and directed follow-ups as proposals (conversation.create / carry_forward.create) for the EM's tick. It never writes risks, feedback, or conversation logs outside apply().
A date-ranged historical replay of the closer's factual-capture step for one account: it replays each historical Fireflies meeting through the domain logger skills over its true date, bypassing all judgment. It never creates risks, feedback, tasks, daily plans, act-severity signals, or proposals - purely a ledgered, idempotent factual backfill.