analyst

List the agent's open findings, at most five: each one sentence, how many conversations it rests on, the ones you may read with two moments as its proof, and its one action -- offered only if you could do it yourself. Use when showing what the Analyst has flagged about this agent and what can be done about each one. Changes nothing.

Call back the people a finding rests on: a campaign of this agent, started now, holding the numbers of the conversations you may read whose numbers are still kept, with the retry policy you write. Use when a finding's offered action is to call the affected people back. The finding closes; refused or failed, it stays open and says why.

Dismiss a finding: it frees its slot and is never raised again. Use when a finding does not deserve any of its offered actions and should stop taking up one of the agent's five slots. Undo it by deleting the dismissal.

Undo a dismissal: the finding is open again until it expires as it would have. Use when a finding was dismissed by mistake. Refused while all five slots are taken.

Teach the agent the answer a finding says it lacked: hands the finding and the answer to the Architect in a new conversation with it, where the change is shown to you before it is published. Use when a finding's offered action is to teach the agent, and you have the answer it was missing. The finding closes.

Keep watch on a finding: starts a tracked question from the finding's question, offering the conversations the finding rests on that you may read for a verdict. Use when a finding's offered action is to keep watching for it going forward, as a tracked question. Carry it on with the question's proposal and confirmation; confirming it closes the finding. Refused, naming the bound, when the agent already has as many tracked questions as it may.

Draw each tracked question of this agent as the one chart its kind gives. Use when showing how a tracked question's answers moved over time. Changes nothing.

A yes-or-no question as the share of conversations saying yes in each of the last weeks weeks, a closed-set question as its answers ranked over the last four weeks, a number question as its average in each week still kept. Weeks are the reader's, in their own timezone. Every conversation is counted, whoever may open it, and each chart says how many it comes from; an agent never read answers with empty series.

List the questions the Analyst answers about every conversation of this agent, active and paused, each saying whether it is answered now and why not. Use when showing the agent's tracked questions, or checking what already exists before starting a new one. Changes nothing.

Confirm a proposed question, with its kind and short name as you leave them. Use as the last step after search_for_tracked_question and propose_tracked_question, once the proposal reads right. It keeps the words the search was made from, and is answered for every conversation that ends from now on; a yes-or-no question also keeps your verdicts as its first answers. Refused when the agent reached its bound of questions or reading was turned off since the search.

Say what counting a new question will cost before it starts: how many of the agent's readings from the last 30 days the search would read, and an estimate of what reading them costs -- the amount only to a person who may see consumption; anyone else is told it is charged by what the Analyst's model reads and writes. Use before starting a search, to warn what it will cost. Changes nothing.

Start counting a question you write in your own words, at most 120 characters: look through what the Analyst read of this agent's conversations over the last 30 days for those that may be about it, and offer those you may read for a verdict. Use when adding a tracked question to the agent — the first step, before proposing and confirming it. Refused, saying why, when the agent already has 10 questions, active and paused, or reading is off.

Charged as the Analyst's work for this agent; not started at a zero balance.

Say of the offered conversations whether each is about the question, and get the Analyst's proposed kind, short name and meaning. Use after search_for_tracked_question, before confirming with create_tracked_question. At least 3 verdicts are asked, unless fewer than 3 conversations were offered; then send none, and it proposes from the sentence alone. Refused, saying why and before anything is charged, when the agent reached its bound of questions or reading was turned off since the search.

Delete a tracked question and every answer it holds. Use when the question should stop existing, not merely stop running — there is no undo.

Rename a tracked question, pause it or resume it. Use when a question needs a clearer name, or should stop or restart being answered. Its meaning, kind and closed set never change: a different question is a new one.

Read whether the Analyst reads this agent's conversations. Use when showing the switch's current state before offering to flip it, or checking why it is refused.

Says why it cannot be turned on when it cannot, and what the Analyst's work for this agent cost over the last 30 days -- the amount only to a person who may see consumption.

Turn the Analyst's reading of this agent's conversations on or off. Use when the organization decides the Analyst should start or stop watching this agent's conversations.

Takes effect at once; publishing or restoring a version never moves it. Turning it on is refused, saying why, where the organization keeps no conversation (a retention window of zero).

Read the agent's summary: one or two sentences, in its own voice, about its outcomes on the organization's day -- or its week so far on a quiet day -- beside how many findings are open. Use when showing the agent's own take on how things are going right now. Nothing until the agent has a full week of readings; ready_at says when.