Skip to content

The inbox (triage)

When an agent captures knowledge, it doesn’t go straight into trusted memory. AI-generated entries land in the inbox — the INBOX stage of the knowledge lifecycle — where a human reviews them before they become part of the org’s living memory. The web app calls this surface the Inbox; against the MCP it’s the triage queue. It’s the same thing: the quality gate between “an agent wrote this down” and “the org believes this.”

Entries created by an agent are flagged generatedByAi and routed to the inbox with a review window — roughly 24 hours, carried on each item as review_due_at. A human-authored entry doesn’t need this gate; an AI-authored one does, because the org hasn’t vouched for it yet.

{
"id": "triage_…",
"entry_type": "WORK_OUTPUT",
"title": "",
"tags": [""],
"author": { "name": "" },
"source_provider": "claude",
"review_due_at": "2026-06-17T…",
"created_at": "2026-06-16T…"
}

An empty queue returns plainly:

{ "status": "success", "total": 0, "count": 0, "entries": [] }

A reviewer can do four things with an item. Every triage tool keys on triage_id (not id):

ActionToolEffect
Acceptaccept_triage(triage_id, …)Promote the entry into filed knowledge; optionally override title, content, type, or tags
Dismissdismiss_triage(triage_id, reason?)Reject the entry without deleting the record
Deletedelete_triage_item(triage_id)Remove the item entirely
Suggest editssuggest_triage_edits(triage_id, instruction) then apply_triage_suggestion(triage_id, message_id)Refine the entry before accepting

Browse the queue with list_triage(filter?) and open a single item with get_triage_item(triage_id). Non-admins see only their own pending items; admins see the whole org queue, scoped by filter (all / mine / overdue).

The inbox protects memory quality. Agents produce knowledge as a side-effect of work — fast and frequent — and not all of it is right, durable, or worth keeping. A short human review catches the entries that would otherwise pollute search for everyone who comes after. It’s the place where “every claim has a source” (principle P5) gets enforced in practice: a reviewer can see what an entry was derived from and accept, fix, or reject it accordingly.

The window is deliberately short. The goal is a light touch — accept the good, dismiss the noise — not a heavyweight approval process. Capture stays ceremony-free; the gate just keeps the bad entries from becoming trusted truth.