notification-manual-dismissal-safety

# Notification Dismissal Safety ## Owner first; mirror second Read the current payload's `instructions`. If the notification mirrors producer- owned state, call that producer's read/dismiss verb first (for example, `email(action='read'|'dismiss', ...)`). A generic notification dismiss clears only the high-attention mirror; it never removes producer history, mailbox state, goal semantics, or a source queue. Choose one atomic target: ```text notification(action='dismiss_channel', input={'channel': 'nudge', 'force': null, 'reason': null}, reasoning='handled the notification') ``` `dismiss_channel` clears one whole mirror and requires `channel`. `dismiss_event` removes one matching `event_id`; `dismiss_ref` removes matching `ref_id` event(s). Event/ref calls default to `channel: 'system'`, target events in the system mirror; `channel: "daemon"` is also supported through the Core aggregate. Other channels refuse event/ref targets. Clearing the final system event removes that mirror, not producer records. The schema is the exact field source; cross-action fields are rejected before I/O. All three verbs delegate to the canonical Core policy path. ## Stale versions and `force` A non-force clear compares the delivered version with the current on-disk version. If the producer updated the channel after delivery, it refuses with `reason: 'stale_channel_version'`. Read the newly delivered state before making a decision. `force=true` is only for a confirmed stale **mirror**, never a routine retry and never a substitute for handling the producer. A successful clear still changes the mirror only. ## Protected and acknowledgement-sensitive channels `goal` is protected source of truth: generic `dismiss_channel` refuses even with `force=true`; use `../../../system-manual/reference/goal-manual/SKILL.md` to cancel or complete goal state. `post-molt` requires a non-empty decision reason in the form `continue: ...`, `defer: ...`, or `obsolete: ...`: ```text notification(action='dismiss_channel', input={'channel': 'post-molt', 'force': null, 'reason': 'continue: recovered the pending work'}, reasoning='acknowledge continuation') ``` A refusal is a route, not a retry: producer guard → producer verb; stale version → reread, then decide whether confirmed mirror-only `force` is justified; protected channel → owning manual; missing post-molt reason → record the real decision; one system event → use `dismiss_event` or `dismiss_ref`, not a whole channel clear. ## Hooks and legacy reminders A registered hook channel uses the same atomic dismissal and producer guards. Registration widens only this agent's allowlist, never producer-dismiss policy. `drop` removes the manifest and revokes the channel; it does **not** kill the hook process. Stop it using the manifest's `how_to_cancel`. If the process keeps publishing after drop, the kernel may emit the blocked-channel warning again. New large tool results are not notification events. They belong to `_meta.agent_meta.agent_state.current_tool_result_chars` and `context(action='summarize')`; `../../../context-manual/reference/summarize-manual/SKILL.md` owns digest, recovery, and summarize-versus-molt procedure. A persisted legacy `source='large_tool_result'` event can be cleared by successful summarization of the matching `tool_call_id`; if that is impossible, prefer `dismiss_ref(ref_id='large_tool_result:<tool_call_id>')` over a whole-channel system clear, which may remove unrelated events. The original result remains in chat history and `events.jsonl`. No dismissal mutates producer state. Never bypass these guards by deleting files.