Knowledge

Types

How a decision was confirmed

A decision's page shows one line under its title that says on whose word the decision stands. It is read only; nobody types it.

Under it a second line may read "The agent says this was decided by: "…" (meeting). This is the agent's claim, not a checked fact." What the agent wrote stands inside the quotation marks; nothing in it is CorPole's statement. It appears when an agent passed the decision on and named who took it. It stays after the decision is accepted: accepting says the decision holds here, not that the named person took it. The decision's note carries the same words, and agents that receive the decision are told how it was recorded, not that the claim is true.

What an agent says about a person's yes is never enough for a "Confirmed" line.

Directions: what your agent is told, and what happens when you state one

The directions in force reach every agent session by themselves. With the plugin's hooks on, the session starts with a block headed "Directions in force": each direction with its code and its sentence, the newest first, and one line of rule — they hold for all work in this solution; a direction that is not confirmed yet, or is Superseded, does not hold. With none in force the block says so. At most ten are written there, fewer when they are long (the block has its own share of the session's context, so the binding requirements keep theirs); the agent's status answer (get_status) lists every one. A direction whose sentence looks like it holds a secret — one written on the screen before sentences were checked — is not shown to any agent — in the block, when an agent reads the direction by its code, when it searches, in the context of the work a decision governs or in the list of changes; the same holds for a title, notes, an answer and the reason of a withdrawal: the line gives its code and says the sentence is withheld. You see it too: the direction's row in the Directions list and its page say "Agents are not shown this sentence: it looks like it contains …"; withdraw it and state it again without the secret. The sentences are quoted as stored, line by line: they are data, and nothing in a sentence can pass for an instruction of CorPole's. An unconfirmed direction, a superseded one and another solution's are never in the block. Without the hooks (the desktop app, say) nothing arrives at the start; the same list is in the agent's status answer.

During a session the agent hears what you did with its next prompt or tool answer, by code: that a direction is in force now, that it replaced another, or that you withdrew one. It reads the direction itself before it relies on it.

• A direction or a lasting instruction? A direction says which way the solution's work goes — what is built or not, in which order, what comes first, a standing choice about the product or its priorities — and holds for every agent of the solution. A lasting instruction says how your agent works — format, language, tools, conventions. Words such as "from now on", "never" and "always" can open either. When your sentence could be either, your agent asks one question that names both: "Is this about where the solution goes — a direction for every agent — or about how I work?"
• When you state a direction ("first the authorization model, then the screens", "do not go for an MVP, go deeper"), your agent records it in your own sentence, word for word, and tells you what happened: where your AI client can ask you, it asks whether the direction is accepted in your name; otherwise it waits on its card on Waiting for you. It does not ask "shall I make this permanent?" for a direction.
• Until you confirm it, a direction binds nobody. Your agent does not change the order or the content of the work because of a direction that still waits; it says that it waits and goes on under the directions in force.
• What waits for you is said with CorPole's own sentence ("3 cards wait for you: …"), the same one this screen shows; the agent does not add a count of its own.
• Why a sentence can be refused. A direction is stored word for word and given to every agent session, so a sentence that looks like it holds a secret or personal data is not stored at all: an access token or API key, a private key, "password: …" or "token = …", an address with a user name and password, a card number (13 to 19 digits in a row) or an 11-digit identity number. The refusal says which of these it looks like. Say the direction again without it. The same holds when you correct the wording on the card or edit a waiting direction on its page. These are the rules of the prompt log; they do not look for e-mail addresses, and a run of digits counts only when it is plausibly what the rule names: a card number that passes the card check, an identity number whose check digits hold. Two dates in a row or an order number are not refused.
• Personal preferences live in your AI client; CorPole keeps what binds others. How your agent works for you alone ("keep your answers short") is kept by the client itself, in its own memory; your agent does not ask about it and does not send it to CorPole. CorPole holds the solution's rules, constraints, standards, terms and directions. For something that should hold for every agent of the solution your agent asks once whether to propose it as a rule for the whole solution; it is a Draft until a person approves it. A personal rule is kept in CorPole only when you ask for that (or your client has no memory of its own): it is then a Draft of yours — not in force — until you confirm it in a question your client shows you or switch it on under your initials › Personal rules. What the agent reports about your yes in chat is not read.

Directions: the list, replacing and withdrawing

A direction is a decision that says which way the work goes from now on. The directions of a solution are a filter of the list of decisions: open Knowledge → Decisions and choose Directions in the Kind filter (the address ends in ?kind=direction). The solution's home page links to it with a line that says how many directions are in force, or "No directions in force". An ordinary decision is never in this list.

Small choices your agent does not ask you. A choice that is yours but cheap to change — which of two pieces of work comes first, or a matter of taste such as which column leads a list (the decision classes priority and brand and taste) — is not put to you as a question. Your agent chooses, and records the choice as its own assumption, with the options it chose among, the one it took and one sentence saying why. When a direction in force plainly covers the choice, the agent names that direction; when none does, the record is marked as following no direction. Your agent is told to present the choice as its own, never as yours, and that you can read it and take it back.

Where you read them. Such an assumption is not a card: nothing opens on Waiting for you, its number and the number in the context row below the top bar do not change, no notice is sent and your agent does not stop. The foot line there reads "2 choices your agent made without asking — read them" and opens the Assumptions view of the Questions table (Knowledge → Questions, Kind: Assumptions). The line counts the choices nobody has read: those derived from a direction of yours and those no direction covered. An agent's working assumption on a decision of its own and an assumption on a preparation question are not among them: they are an open question and a preparation item, as before. The table lists them newest first; you can search it, and filter it by state and by "No direction".

The assumption's page starts with a form named The agent's choice: the choice itself as its first line, then The agent chose, Among (every option, the chosen one marked "chosen by the agent"), Why (the agent's own words), Derived from (the direction's code, its sentence and its state now, or that no direction in force covered it), Class, as the agent declared it — with the reminder that the class is the agent's word and CorPole did not check it — the agent, the day and what it is about. Under it, Your answer has two answers; nothing is chosen for you:

• Fine as chosen — "It stays the agent's choice; you have read it." Only one thing is recorded: that you read it, with your name and the time. Nothing else changes. It stays the agent's assumption and stays open; it is never shown as answered or decided by you, it closes no preparation question, and your agent is told nothing. Wherever a status is shown it reads Read — the agent's choice (before you read it: The agent's choice — not read), and the page's Answer cell reads "The agent's choice · read by … on …". Reading is not final: on a read assumption's page Your answer shows who read it and when, and still offers This was mine to decide — Fine as chosen is not offered a second time.
• This was mine to decide — "It becomes a question for you on Waiting for you, and your agent is told it was not its choice." The assumption becomes a question of yours with its card in Questions: the options stay, none is recommended, and the one the agent took is marked "Your agent had chosen this". The class the agent declared and the link to a direction are dropped. An assumption that had no options also gets its card, with the one answer Your answer, in your own words. The agent whose connection made the choice is told once, before its next answer, whoever of you took it back: "… says QST-… was theirs to decide, not a choice for you to make (you declared it as …). It waits for their answer on its card: do not build on your choice." While guidance is suspended no card can open, so taking back waits until it is on again.
• Save answer is the page's one primary button. The question's own details below it are read-only on an assumption's page — they are the agent's text — with no Save, no Status menu, no Answer panel and no Workflow section: an assumption is not answered, and its state moves only by these two answers. Answering needs the right to accept and reject decisions; without it you see the same form with the answers disabled and a line saying who can. A second Fine as chosen, or a save on an assumption that was taken back meanwhile, says "This one is already closed." and writes nothing. Delete on an assumption is for a person with the same right who may also edit knowledge: it removes the record of a choice made in their name; anyone else is told so and nothing is deleted.
• Which assumptions these are. Only the choices an agent made in your place: those derived from a direction of yours and those no direction covered. An agent's working assumption on a decision of its own, a question guided start reclassified as an assumption, and an assumption on a preparation question are not among them: they are ordinary open questions and preparation items, answered and counted as before, and the foot line does not count them.
• Mark as read on the Assumptions table does Fine as chosen for the rows you tick — nothing is ticked for you; "Select all n shown" ticks the unread rows in view, and "n selected" counts them. At most 50 in one save; a row that was closed meanwhile is skipped and named in the result. Taking a choice back is never done in bulk: open it.

On a direction's page the health strip has a cell Derived choices ("2 open / 3"): it opens a side panel that lists the choices agents derived from this direction — the choice and what was chosen, the agent, the day and the state — each linked to its page. Under the Directions table one quiet line appears while choices made without a direction are unread: "Chosen without a direction: 3 — read them". Agents → Health has the table Choices made without asking, by agent: how many each agent made, and how many of them are not read, read (fine as chosen) and taken back; the share taken back is shown from five reviewed choices on, and "too few reviewed" below that.

CorPole checks the direction the agent names: it must be a direction of this solution that is in force and whose sentence agents are shown. A direction that still waits for your confirmation, a Superseded or Rejected one, a candidate, a code of another solution and a code that is no direction are refused, and nothing is stored. CorPole does not judge whether the direction really covers the choice, nor whether the choice was really a small one — both are the agent's reading, which is why you can take a choice back. A choice is stored only with at least two options and the one that was taken, and not at all when its text looks like it holds a secret or personal data. What you already decided is not assumed again: when you answered a question with the same title, or closed its card, the agent is told your answer and nothing is stored. At most 20 such assumptions wait unread in a solution; the next one is not stored, and the agent is told to say so — reading one or taking it back frees its place. An assumption is never an open question for your agent to ask you, and it blocks nothing: the work it concerns is not held, and the agent sees it under "Your assumptions (not the person's decisions)" with whether you read it. These records also count against the same hourly limits as the agent's other questions. A choice about scope, money, law, risk or a promise to others is never handled this way — it is always asked.

Candidates — not binding. Sometimes you say something that sounds like a direction only in passing, inside a request for other work ("…and in reports the risk order always comes first"). Your agent does not interrupt the work for it: it notes the sentence, word for word, as a direction candidate, tells you so in one line and goes on with what you asked. A candidate binds nobody. It opens no card, asks you nothing in your AI client, sends no notice, and is in no count of what waits for you. It may shape only the task of the prompt you said it in — never a later session and never another agent. A direction you state on purpose is not a candidate: it is relayed as before and waits on its card.

Stated or said in passing — how your agent tells them apart. Your agent is told, with the directions at the start of a session, in the tool it records with and by the plugin: "Stated on purpose: record_decision (kind direction) — asked in the person's client where it can be, else it waits on its card. Said in passing inside a request for other work: add candidate: true — no card, no question; say in one line that it is noted as a candidate and do the work. Unconfirmed, neither binds anyone." and "A candidate noted in this prompt may shape only the task of this prompt; it never binds a later session or another agent." In a solution under guided start the block of directions an agent session gets ends with the open candidates — at most the three newest, each with its code and its sentence cut to 120 characters — under the line "Candidates — not binding, do not act on them:". They are shown so that the agent does not note them twice and does not take them for directions; a dismissed or lapsed candidate is never listed, and a connection that is not shown your directions is not shown the candidates either. When the block is too long, the candidates are left out before any direction in force is.

"Write that down as a decision." If you tell your agent to record a candidate on purpose after all, it sends the same sentence as a stated direction. No second record is made: the candidate itself becomes the relayed direction, with a card on Waiting for you where you confirm, correct or reject it, and it leaves the candidates section (its place among the ten is free again). A sentence you dismissed is still refused.

When a candidate contradicts a direction in force, your agent does not stop and no conflict card opens — a candidate binds nobody. The agent marks it instead: its row and its page read The agent says it contradicts DEC-… with a link to that direction. A marked candidate is never confirmed in one click: the row's button reads Confirm… and opens its page, where the marked direction stands in the choice — Confirm and replace (tick it) or Confirm, keep both — with nothing ticked for you. The mark is the agent's reading, not a finding of CorPole. Only the agent's own recording sets it, once; nobody edits or removes it on the screen, and a link you add by hand is not a mark. If the marked direction is replaced meanwhile, the mark is not dropped: the line reads "The agent says it contradicted DEC-…, which DEC-… has replaced", and the direction that replaced it stands in the choice. Only when nothing in force is left in that chain does the line say so, and Confirm is one click again. The mark stays with the sentence when you have it written down as a decision: its card on Waiting for you shows The agent's mark and the marked direction beside the new one, and your AI client does not ask about it — the choice is made on the card.

Open candidates are listed in the section Candidates — not binding above the Directions table; the section is there only while a candidate is open. Each row shows the sentence word for word, the agent connection that noted it, the day it was noted, the day it lapses ("lapses 2026-11-08"), and Similar to DEC-… when a direction in force resembles it. An open candidate is not repeated in the table below.

• Confirm makes the candidate a direction in force: Accepted, with your name, the time and "confirmed on the screen" — exactly as when you confirm a relayed direction. When no direction in force resembles it, Confirm in its row is one click. What you confirm is exactly what the row shows: a candidate's title is its sentence (a long sentence is cut at a word, with "…"); the agent's own title is not kept.
• When a direction in force resembles it, the row's button reads Confirm… and does not confirm: it opens the candidate's own page, and a quiet line under "Similar to DEC-…" says so ("Choose on its page whether it replaces that one or stands beside it."). There Confirm in the action row shows the similar directions with a checkbox each, none ticked for you, and two buttons: Confirm and replace (the ones you tick become Superseded and point to the new direction) or Confirm, keep both.
• Not a direction dismisses it: it is Rejected with your name, and the list and its page read Dismissed — not a direction. Your agent does not note that sentence again, and it cannot bring it back as a relayed direction either: both are refused, it is told you dismissed it, and no card opens. If it is yours after all, open the dismissed candidate (it stays in the table below) and choose Reopen as candidate at the top of its page. The same holds the other way round: a sentence you called Not mine on a relayed direction's card is not noted as a candidate for you.
• Reopen as candidate is on the page of a candidate that was dismissed or that lapsed, for people who may accept and reject decisions. It is an open candidate again, back in the section above the table: the day it was noted is kept, and its 30 days start on the day you reopen it. The limit of 10 open candidates holds, and a sentence that is meanwhile in force, waits as a relayed direction or is another open candidate is not reopened; the page says which. Only a person does this, on the screen: your agent cannot.
• After you act, a strip at the top says what happened: "DEC-… is a direction in force now: confirmed on the screen." or "DEC-… is dismissed — not a direction. Your agent does not note this sentence again." A refusal is shown the same way, as one sentence that says what to do instead.
• While guidance is suspended, a candidate that resembles a direction in force cannot be confirmed: the choice between replacing and keeping both waits until guidance is back on, and its page says so. A candidate that resembles none is confirmed as usual, and Not a direction always works.
• Lapse. A candidate nobody confirmed lapses 30 days after it was noted. From that moment it is no longer offered and no longer counted, and it reads Lapsed — not confirmed in 30 days; a daily job also sets it to Rejected, with nobody as its decider. If it was your direction after all, reopen it on its page (Reopen as candidate) or tell your agent the direction again.
• At most 10 candidates are open in a solution. The eleventh is not stored; your agent is told to send you here to confirm or dismiss the others. A lapsed or dismissed candidate frees its place.
• The candidate's own page says "A direction candidate your agent noted in passing — not binding" with the day it lapses, and offers the same two acts, Confirm and Not a direction, in the action row. The Status menu offers no move for a candidate, and Waiting for you does not list it.
• Who can. Confirming and dismissing need the right to accept and reject decisions (the same right as accepting a direction). Without it you see the section and the page read-only, with a line saying who can.
• The same sentence twice. If your agent notes a sentence that is already an open candidate, nothing new is stored. If the sentence already waits as a relayed direction, the agent is given that direction's card; if it is already in force, it is told so.

Replace or keep both is your choice, on the card. When your agent relays a direction that resembles one in force (their sentences share most of their meaningful words), the new direction's card on Waiting for you shows the two side by side: the one in force with its code, who confirmed it and when, and the new one. Up to three similar directions are listed, newest first. Instead of one Confirm there are two buttons. Confirm and replace: the new direction is Accepted and each direction you ticked becomes Superseded; the most similar one is ticked when the card opens, and you can tick others or untick it. Confirm, keep both: the new direction is Accepted and the others stay in force beside it. Correct offers the same two buttons for your own wording. This is not mine is as before.

• A replaced direction is not deleted. It stays in the list as Superseded, and its page and its row point to the direction that replaced it; the new one's page says which ones it replaces. Both histories keep who did it and when.
• Your AI client never asks about replacing. When a similar direction is in force, a new direction is not asked in the client at all: it waits on its card, and your agent is told why.
• If a direction the card shows was replaced or withdrawn by someone else meanwhile, your confirm is refused with a sentence on the card and nothing is written. Read the card again: it shows what is in force now.
• Your agent is told what you did with its next tool answer, and with your next prompt where the plugin's hooks are on, by code: "confirmed DEC-…, replacing DEC-…".

Withdrawing a direction. A direction in force that no longer holds, with nothing in its place, is withdrawn: Withdraw in the action row of the direction's page, or in its row of the Directions list. Write why (at least 5 characters); the reason is kept in the direction's history and shown in the list. The direction is not deleted: it becomes Superseded and says "nothing replaces it". Only people who may decide on knowledge can withdraw, and only a direction that is Accepted — one that still waits is settled on its card. The Status menu does not offer Superseded for a direction in force, so a direction never leaves force without a reason. Your agent is told "a person withdrew DEC-…" and reads the reason on the direction itself.

A direction changes only by its own acts. A direction comes into force when a person confirms it (on its card, or in a question their client shows), and leaves force when a new direction replaces it on that direction's card or when it is withdrawn with a reason. Nothing else moves it, and what is refused is not offered:

• In force (Accepted): the Status menu offers no move, Delete is not offered, and its title, sentence and code are shown read-only on its page, with a line that says how a direction changes (notes can still be edited). Under the Status field the page says what ends it. To change what it says, tell your agent the new direction and choose Confirm and replace on its card. To end it, Withdraw.
• Superseded (replaced or withdrawn): it is not reopened and not deleted — it is the history of what held, and other directions point to it. If it should hold again, say it again as a new direction.
• Proposed (relayed, not confirmed): you can accept it on its own page with Status → Accepted, or reject it there — unless a similar direction is in force. A new direction is never accepted beside a similar one without your choice. In a solution under guided start the page then refuses and gives the address of its card (if the card was missing it is opened); while guidance is suspended it says the choice waits until guidance is back on. Where you have no card — the solution is not under guided start, or guided start is not turned on for you in it — the direction's page has an Accept button in the action row that lists the similar directions and offers Accept and replace (the ones you tick become Superseded) and Accept, keep both. Like the card, that form holds what it showed you: if another similar direction was confirmed meanwhile, or the direction's text changed, nothing is written and the page asks you to read it and choose again. It can be deleted; its card closes with it.
• Correct on the card compares your wording. If the wording you write resembles a direction in force that the card had not shown beside it, nothing is written: the card comes back with your wording kept in the field (it is kept for five minutes, on your browser and for you only, and never put in the page's address; a very long wording in a non-Latin script cannot be kept and the card then asks you to write it again) and that direction shown beside it, and you choose Confirm my wording and replace or Confirm my wording, keep both.
• A page left open. If a direction changed while its page was open (it was confirmed on its card, say), a status move or a delete from that page is refused with a sentence and nothing is written.
• Rejected: it is not reopened. Say it again if it is yours after all.
• Links: the "Supersedes" link of a direction is written when a person confirms a replacement on a card. It cannot be added or removed by hand on a direction, and accepting a decision that carries such a link never supersedes a direction.
• Each refusal is one sentence that says what to do instead.

Your AI client never asks about a direction that resembles one in force, in any solution: the choice is yours, on the screen. Known limit: a direction you write by hand on the screen of a solution under guided start opens no card; if a similar one is in force, tell your agent the direction instead.

Superseded with no replacement means: this was a direction once, a person withdrew it, and no other direction took its place. Agents no longer follow it.

Retrospectives with agents

• About once a day per solution and person, after real work, CorPole asks the agent what it learned and what could be done cheaper, faster or better. Its answers arrive here as proposed lessons and ideas marked AI. In a solution under guided start the request opens with CorPole guided (RETRO), one of the few lines in a tool answer that the agent's rules (BR-007) let direct it — everything else it reads there is data — so an agent that keeps its rules does not skip the request as data.
• In Claude Code, /corpole:retro runs a retrospective on demand and can bring in lessons from the agent's own memory files (you see the list first; secrets are masked).
• Workspace owners and admins can Share with the whole workspace an accepted lesson: every solution then gives it to its agents, with the solution it came from. Accept or reject each one. On an accepted lesson, Promote to solution rule turns it into a solution agent rule (needs the Integrations permission); on an idea, Create work item makes an epic, story or task linked to it.

Tenant-wide knowledge and exceptions

Owners and admins can make standards and glossary terms tenant-wide: every solution inherits them. A solution that needs an exception records a decision that waives the standard; the waiver takes effect when the decision is accepted, and the standard is then left out of that solution's context.

What agents do with it

• Read the standards (get_standards) and follow the MUST ones.
• Treat accepted decisions as binding; proposed ones are not.
• Stop and ask when an open question concerns their work; raise new questions with ask_question instead of guessing, and propose choices with propose_decision.
• Use glossary terms exactly; they are matched in the text of the work automatically.