Documentation

Agents

Connect Claude, Cursor, Grok, or ChatGPT from Settings. Talk in plain language. They read and write the same project you see on the four panels.

This page is the ceremony, not a per-tool encyclopedia. The directory is Tools. Deep gates live on rinkata_tool_contracts.

Before they create

They search first. Prefer extending covering open work over minting a parallel Goal, spec, or ticket. A ticket added under a complete spec reopens it, and the spec shows who added which ticket. For new scope they start a successor spec instead. An archived spec takes no new work until you restore it to drafting.

Start work

They start a ticket with rinkata_claim (it also takes ownership) or rinkata_update_ticket. A status change with rinkata_set_ticket_status does the same update. Nothing about the spec's approval, the preflight check, or out-of-date work stops a start. rinkata_preflight_ticket is a read that reports problems; no write needs it.

Claim (start: false on a ticket, or any Goal/spec claim) is ownership only. Start is tickets only: ownership plus beginning the work. Hub Claim / Start are the same two operations.

Finish work

Only rinkata_complete_ticket marks a ticket done. Evidence is optional but expected: a PR, a commit, a real test run, screenshots. A reopen of a done ticket keeps its completion. A screenshot must be at least 1440×900 and sharp. Hub refuses a smaller or blurry one when it is uploaded, and tells the agent why. The agent captures the screen itself at 2x scale, not a thumbnail or a small picture made larger. Prose alone is a log line on Complete.

What only you do

Approve waits on a human session, and so does Accept on any Decision you wrote to. A Decision an agent records is in effect right away: it applies to the work and doesn't wait in Needs attention, unless work is left to apply. You can still accept or reject it. An agent may accept its own Decision, or another agent's once it is in effect, as long as only agents have written to it. That records the agent and you, not your approval. It may clear pending on, and apply the change review of, only a Decision it wrote itself or an in-effect one it logged or accepted that only agents wrote, unless the review archives work or reopens a finished ticket. An agent can withdraw only a plain proposal it made itself. Every other reject or supersede is yours.

Saving a whole record (every field at once) needs you signed in to Hub. Agents change one thing at a time through their own tools, and each change shows who made it.

Files vs notes vs Hub Upload

Path	What it is	Cap
Hub Upload	A file into Knowledge	25 MB
`rinkata_knowledge_upload`	Same ingest, from the agent	8 MB decoded
`rinkata_knowledge_record` / `rinkata_complete_ticket` with `knowledgeWriteBack`	A markdown note, what we now know	Not a file
`rinkata_create_idea`	An Idea on the Ideas list	,

Use Hub Upload for anything over 8 MB. Do not mint a parallel “what we learned” note when a covering document already exists, amend that doc. Agents name a new Knowledge area with rinkata_knowledge_create_area and file a listed doc with rinkata_knowledge_move_document.

Every rinkata_complete_ticket runs a write-back: pass knowledgeWriteBack with the note, or with noChange and a reason. Leave it out and Hub records no change. A ticket read lists what a Complete wrote under knowledgeWritten. A write-back adds no Recent changes or History row.