/observe — NL → observation policy (MVP step 1)

The owner (or a scoped user) says what they want watched, in natural language:

/observe watch doc changes in this room and ping me daily

The core agent COMPILES that into a structured draft; everything deterministic around the compilation lives in observe_policy.py. Design + resolved decisions: workspace notes/observe-mvp-slice-design.md.

Processing convention (core agent, on an /observe <text> owner task)

1. Compile (LLM-side — you): produce a draft dict: {room_id, event_types[], mode, cost_cap?, created_by, source_text}
   • room_id: the task's channel_id unless the text names another room.
   • event_types: map the intent to plane types (message.created, reaction.added, artifact.updated, member.joined, …). Unknown intent → ask, don't guess.
   • mode: observe (context only) | record (journal) | notify (ping the owner) | taskify (promote batches to ambient tasks — never standing-approved).
2. Validate: observe_policy.validate_draft(draft) — errors go back to the user verbatim; do not "fix" a draft silently.
3. Standing approval: evaluate_standing_approval(rec, owner_mxid=…, owner_rooms=…). owner_rooms = rooms the owner OWNS — created by the owner or where the owner holds PL≥50 (checked via the room state the core already has; when ownership is unknown, the room is NOT in scope). Familiarity — "the owner sent a task from here" — is NOT ownership: a shared room must never enter standing-approval scope (001 review; the server's four-way authz still gates the subscribe, but the standing-approval semantic is stricter by design). True → save + transition(id, "active") + subscribe + post the AUTO-ACTIVATED card (visibility is mandatory — never activate silently). False → save draft + post the CONFIRM card and wait for the decision.
4. Decision grammar (same infrastructure as human-action cards): policy <id> activate | edit <new text> | cancel — typed reply, reaction on the card, or (when custom-event fan-out lands) an A2UI button. edit recompiles with the new text into the SAME policy id.
5. Subscribe on activate: events client subscribe(room_id, event_types) — the events plane's four-way authz is the permission evaluator; a subscribe rejection (e.g. PL too low) goes back to the user as the card's failure line, not swallowed.
6. Enforcement: active notify/record/taskify policies are consumed by the sparrow drain handlers (taskify today; notify/record handlers are the next slice). cost_cap shows the CAP on cards; metering is a later slice.

Files

• observe_policy.py — validation, standing-approval evaluator, SubscriptionStore (<workspace>/state/observe/), effect-card renderer (plain text always; A2UI choice-group in the real renderer contract behind include_a2ui).
• Store records: obs_*.json, states draft → active → cancelled/expired, terminal states immutable, every transition audited.

Test: python3 tests/observe-policy.test.py (99% module coverage).