0042 — Prompt schedules run read-only AND store-reads-only (v83-F5)
Status
Accepted (planned v83; amended by the v83 plan review, item 4).
Question
Hermes's cronjob runs a prompt at a time — the daily-briefing shape
skep lacked: every schedule caste either dispatched a worker or posted
static store-composed text. Can a scheduled Queen turn exist without
creating an unattended acting surface?
Decision
A prompt schedule caste: at tick time the serve daemon runs the
schedule's instructions as a fresh Queen turn in the chat that created
the schedule (the v43-F6 binding). Two hard properties, both by
construction rather than by policy:
- Read-only. The turn runs on the
/btwmachinery (read_only=True): mutations are refused, never carded. A card nobody is watching would expire denied anyway (ADR 0032); refusing up front means an unattended turn cannot even queue an action — the model never holds a trigger while the operator is away (I6). - Store-reads-only. The turn also refuses network-read tools
(
search_web;read_urlis already a mutation and refuses under read-only). The injection surface, named: a recurring unattended turn that could fetch granted-domain pages would let an injected page steer every morning's briefing and chain further granted fetches — query strings are an egress channel, and the v83-F1 budget raise makes ingested pages large. Scheduled turns therefore read the store (runs, approvals, chats, notes, memory) and nothing else; web reads belong in a live chat where the operator is present.
The CLI tick (skep tick) has no chat engine; a prompt tick there
fails honestly ("prompt schedules run inside the serve daemon") rather
than silently skipping (I8). The turn's transcript lands in the bound
chat like any conversation; the tick records prompt_posted /
prompt_failed in schedule health, and the reply pushes outbound
through the same channel funnel as note ticks.
Consequences
- "Every morning summarize yesterday's runs and approvals" works end-to-end and can never act or fetch unattended.
- The blogwatcher-style skill cannot run unattended web fetches; it teaches a live-chat pattern (a scheduled reminder to run it live) until a reviewed-unattended-fetch design exists — a deliberate narrowing recorded here, not a bug.
- A future loosening (e.g. pre-listed URLs reviewed at propose_schedule card time) must amend this ADR explicitly.