The attention queue

The attention queue is one list of things that need a person, ordered by what it costs to ignore them. It is not a feed. Every item on it is computed from something recorded — a symbol that moved, a clause that was broken, a session briefed before a decision existed.

You reach it in two places:

• Workspace → Attention (/workspaces/<id>/control) — one workspace's queue.
• Organization → Attention (/organizations/<id>/attention) — every workspace you can reach, grouped by workspace, with filters for workspace and disposition.

Why it works the way it does

The failure mode of an attention system is not missing an item. It is crying wolf until people stop reading it. A queue that is wrong a fifth of the time is worse than no queue at all: it costs the same attention it was built to save, and adds the insult of having promised otherwise.

Three rules hold throughout.

1. Five dispositions, and most things are not the top three. A system that only knows "notify" ends up notifying about everything.
2. Nothing reaches review, decide or intervene on a guess. Those three are for facts Madebook computed — a collision found by a join, a clause the verifier found violated, a deviation matched against a package list. A model's suspicion can produce an inform, and the item says so.
3. Every item names the thing and the next move. "Agent 04 created a second authorization pattern" is actionable. "Review needed" is a notification, and notifications are what people already drown in.

The five dispositions

Disposition	Meaning	Order
intervene	Something is wrong now and a person has to step in.	1st
decide	A question is waiting on a person's answer.	2nd
review	Worth a person's eyes before it goes further.	3rd
inform	True and relevant. Nobody has to act.	4th
ignore	Recorded and deliberately not surfaced.	5th

The queue sorts by disposition first — intervene always first — then by the item's urgency score.

Every rule that can raise an item

Each item carries the name of the rule that produced it, so a noisy queue is diagnosable rather than merely annoying. These are all of them.

Rule	Disposition	Urgency	Fires when
source_done	review	62	The tracker moved a source item to done or cancelled while the mission here is still open.
architecture_conflict	decide (confident) / inform	78 / 40	A change introduced something that departs from an architecture rule. decide only when the rule is confident.
cross_agent_collision	(see below)	95 high / 65	One session changed a symbol another session is standing on.
ready_to_ship	review	55	A mission's contract is satisfied and its evidence gate is met.
contract_violation	intervene / decide	90 / 72	Verification found violated contract clauses. intervene when a forbidden clause is among them.
context_stale	intervene	88	A live session's context health has fallen far enough that it is working from a stale picture.
high_risk_change	review	85 critical / 68	A change set's risk assessment landed in a high or critical band.
constitution_unapproved	decide / inform	70 / 30	The current architecture rules are a draft nobody has approved. decide when sessions are live against them.
policy_violation	intervene / decide	92 / 75	An open policy violation. intervene at critical severity.
session_failed	inform	30	A session stopped without finishing.
agent_question	decide	80	A session is blocked and waiting on a person.
decision_required	decide	86 / 74	An open decision gate. 74 when the agent still has allowed scope to work on; 86 when it has nothing left.
remediation_outstanding	inform / review	35 / 60	A remediation request is outstanding. review once the session is no longer live to act on it.
evidence_missing	review / inform	55 / 35	The contract is satisfied but evidence requirements are not. review when a requirement needs a person.
provider_refused	inform	30	A session tried a model the organization has not approved.
decision_proposed	decide	45	A decision proposed through the MCP is waiting for a lead.
source_changed	decide / inform	60 / 30	The source item behind a mission changed or was deleted after the contract was written.
mission_questions	decide	45	A mission has open product questions that must be answered before its contract can be approved.
plan_proposed	decide	70	An agent proposed an execution plan for a high-risk mission.

A collision's disposition depends on its severity: a high severity collision carries urgency 95, everything else 65.

What an item contains

Field	What it holds
headline	The thing, named. For example: "DEC-002: which expiry applies to invitation tokens?"
body	Why it fired and what the next move is.
disposition	One of the five.
urgency	0–100. Defaults to 50 when a rule does not set one.
kind	The rule's category, for filtering.
rule	The rule's name, exactly as in the table above.
determined_by	computed for a fact. Anything a model suggested says so.
subject_type / subject_id	What it is about — a Mission, a Session, a ChangeSet.
occurrences	How many sweeps have seen it.
fingerprint	The identity that makes repeat detections one row instead of many.

Sweeps, fingerprints and supersession

The engine runs as a sweep over one workspace. Rules are pure functions of state, so running the engine twice changes nothing.

• Every candidate item carries a fingerprint derived from what it is about. A collision detected on six consecutive sweeps is one row with occurrences = 6, not six rows.
• The sweep is a reconciliation, not an insertion. An item whose rule no longer fires is marked superseded rather than left to rot. A queue that only ever grows is a queue people stop opening, and stale items are worse than none because they make the reader distrust the fresh ones.
• Escalation is allowed; de-escalation of an open item is not. If a rule fires again with a more severe disposition, the item is raised and its headline and body are replaced. If it fires with a less severe one, nothing changes — something a person has already been told is serious must not quietly become less serious underneath them.
• One broken rule cannot take the queue down. A rule that throws is logged and skipped. A queue missing one class of item is degraded; a queue that throws is absent, and absent is what people plan around.

Clearing an item

Resolving an item in Madebook does not only hide the row. Where resolving should change the underlying fact, it changes the fact — otherwise the next sweep raises the same item again. A resolution that cannot change the fact leaves the item until the fact changes.

The honest way to clear most items is to deal with the thing they name:

Item	What actually clears it
decision_required	Answer the decision gate with a resolution and an instruction.
plan_proposed	Approve or reject the execution plan.
constitution_unapproved	Approve the architecture rules.
decision_proposed	Record or reject the proposed decision.
policy_violation	Resolve or waive the violation.
evidence_missing	Supply the missing evidence.
source_done	Ship or abandon the mission.
ready_to_ship	Ship it.

Related

• The execution protocol — what an agent is told while an item sits open.
• Changes and review capsules — where a high_risk_change item sends you.
• Violations — where a policy_violation item sends you.