recall

Recall — retrieve what the graph knows

You answer one question: what is written about X and whether one can act on it. What is written is retrieved, not remembered: the model's memory and a single search result are not a source. Recall ends not with a result dump but with an answer with nodes, a check of each, and a named ceiling — or an honest "I don't see it" after all instruments and the brother.

When to call it, without the person's word: you are about to ask about something potentially written; after compression you rely on "I remember"; mid-code you needed the basis for a decision; before advice, a retelling, the words "it's not in the graph", the novelty gate writing or prose about meaning. An already open node answers directly — use it, don't start the ladder over.

Do not mix three sources. Your own memory, including the summary after compression and a subagent's retelling, hints at where to look but does not certify what is written. Reading the current node certifies the content of the record. What the author of the record says about the world is their testimony with a named instrument and limit, not your observation of the world (canon: "Distinction: one's own memory ≠ reading a record ≠ the author's testimony").

1. Name the question

In one line: what you are looking for and what counts as an answer — a decision, a rule, a responsible person, a trap, a person's permission, the state of work. A question with two concepts — two searches. Without a named "what counts as an answer" step 4 has nothing to measure against.

A past decision ("what was agreed?") — search. The present will ("are we taking it now?"), a new preference, or an intention not yet written down — go to the person straight away by the door question rules: the brother does not know their will.

2. Two instruments, several formulations

• Words — iskron_search(realm, q=<термин>): catches only the author's words. Go through the names by which the item could have been called: a project term, an everyday word, an English one, the name of a tool or file, the name of the role that deals with it.
• Meaning — iskron_semantic_search(realm, query=<одна мысль фразой>): on a conceptual question it comes first. Two or three phrases from different sides: what it is, where it is done, what would break without it.
• Narrow filter — node_type for both instruments, context_holon и attrs_filter — only for iskron_search, when the output is even or drowns in noise; raise limit — the output is a page, not the whole set.
• Cases — live ones on the subject: iskron_case(action="at", node=<узел предмета>), then read. A person's permission, given by word, lies in the case or in the node into which it was woven.

One instrument, one formulation give "I don't see", not "there is none".

Steps 2–3 can be given to the role recaller by a statement of the task: what we are looking for, what counts as an answer, where we have already looked. Its finding is "reported": you open the node yourself at step 4, the brother, the person and the answer-word in the case (it does not read case logs) — are yours.

3. Open the candidates and walk through to the supports

The output is leads, knowledge in the body and in the connections. Each candidate — iskron_look(realm, node_id): body, modes, arrows. From it — to the support of the question along the arrow: context → contour; upadhi → the rule through which they act; incoming vimarsha_of → questions and cases on it; addressed_by → what it was answered with; the surroundings — iskron_orient(focus=<узел>), the relay — lens="trace". Read the closed questions and risks at the subject's place too: they often hold a fact that is still in force.

The version history and the case log are not opened on this path — only by an audit for the named doubt (entry, "Re-entry"). The exception is a word that you are looking for as an answer: a permission or a person's answer given in a case and not yet woven into the node. Look for it in a targeted way — iskron_case(action="history", room=<№N>, key=<ключ> or since=<[N]>, keep_cursor=true) — as a receipt, without reading the node through the log.

If what was found does not answer — establish where the answer should lie: the subject's contour, its steward, neighboring inputs and outputs. Someone else's ownership does not by itself cancel an independent reading of the accessible graph. Only then — step 5 (canon: "Recall along the ladder").

4. Check what you found

For each node you intend to rely on — three questions:

• Does it answer this question — the body says this, not a neighboring one; the node's name and output do not count.
• Is the needed judgment still in force — read the body and the modes by kind. A closed question, a past case, or a released schema may hold a current fact, answer, or ground; closure does not make the whole body false. badhita holds a refutation, not permission to act on the refuted; do not pass a draft judgment off as execution. Check contradictions against neighbors and observation.
• Does it cover by scope — what, where, until when; permission for one does not permit another.

Nodes contradict each other or the world — that is a finding: name it, make the correction a question to the steward or via writing, if the node is of your contour.

5. Ask a brother

Your own search, connections, and expected place did not yield an answer, the source is probably at a neighbor's — ask the subject's steward to recall or weave so that it can be found. A light move is the subject's existing matter: iskron_case(action="at", node=<предмет>), read, with your share join, then say с to to its place; do not open a second matter for one question. An answer is needed — an expectation card via the door. No suitable matter — iskron_case(action="talk", with=<его место или роль>, about=<предмет>); this is a new matter, its conduct and closure — per the door's reference. Name the question, what you sought, where, and with which queries; a question on substance — via writing.

A brother can give a link, trace a path, or answer "I don't know"; his word remains testimony — open the node he found yourself. Depth is one: a brother does not start a chain of brothers. If there is no live steward — name a waiting period; after it, go to the human. Do not call on a brother about a person's will, a fresh unwritten intention, when he has nothing beyond the same graph, or when the human is answering here and it is cheaper to ask him (canon: "Request to a brother agent to recall", "A brother's answer to a recall request").

6. Answer

• Found — an answer, each statement with a node #N; act with the reference, do not ask the human.
• Not found — "I don't see", with a list: which formulations, with which instruments, which candidates you opened, whom you asked. Only after that — a question to the human via card (door iskron, "Question to the human"), and the card names this list.
• Ceiling — reading the graph sees the record, not the world: a claim about behavior is verified on the carrier (reality-audit).
• Seen for correction — something found that diverges from the world or from a neighbor — by question or correction per writing in the same move, not as an answer line.

7. Lay a path if the search was difficult

The question repeats by kind, the miss was in the place or name of a record, what was found is alive — per weaving link the place where you searched with what was found, refine the name or the meaning of the arrow; record the answer of a brother or the human where they will search. Across a boundary — only with firm knowledge, otherwise to the holder. Finishing a recall request, name the path or why it is not needed (canon: "Treading the found", "Across a holon boundary — by firmness of knowledge").

Do not tread a one-off question of the matter, an instrument failure, a whim of your own query, doubtful or outdated findings. Personal — in minding, not in the project. A finding does not obligate creating a new node: often a link or a name is enough.