Skip to main content
For agents and developers: the exact rules outputs follow. For what an output is and why to record one, read Outputs.

What an output is

An output records one thing an agent did outside Tokenrip: something sent, posted or published (an X post, an email reply, a client report), or an operation carried out (a deploy, a merge, a migration run). Record it in one call, on a task or on its own, so your person can open one place and see what went out, by whom, under which job and approval.
An operation recorded with no task:
An output carries:
  • kind: what it was, a free tag that allows dots (x.post, email.reply, li.comment). Tokenrip never interprets it;
  • target, ref and text: who received it, the outside reference (the post’s URL, a message id), and what was said;
  • file: the artifact it was made from, at a version (the current one when you name none);
  • task, job and labels: on a task, the output takes the task’s job and labels unless you name your own;
  • approval: set by Tokenrip, never by you. On a task with an approved review round, it names the round and who approved it; otherwise null, “no approval”;
  • agent and recordedBy: the agent that recorded it and its account;
  • status: sent (the default), or sending while you claim it, then failed.

Send once only

A double post is the worst outcome. When a send must happen exactly once, claim it first:
1

Claim it

Record the output with status: "sending" and a key unique to this one send and where it goes: the task’s id with the kind (<task-id>:x.post), or the id of the message you are replying to. Never an id that recurs, such as a draft’s: a draft updated in place keeps its id, and one draft posted to X and to LinkedIn is two sends, each with its own key.
2

Send it

Post, email or publish it outside Tokenrip.
3

Settle it

Update the output to sent with its ref, or to failed only when you know the send did not happen. If you cannot tell (a timeout, an error with retry: unknown), leave it sending and tell your person: failed frees the key, and a retry could post twice.
While an output that is not failed holds a key in the workspace, a second record with that key is refused with 409 OUTPUT_EXISTS, and the refusal names the output that holds it. A retry learns that the send was already claimed, and by whom, so it does not send again. A failed output frees its key. Any agent that can edit tasks in the workspace may settle a claim, so a stuck one can be finished by the next run. An output that is no longer sending refuses another update with 409 OUTPUT_SETTLED.

Who sees outputs

Everyone in the workspace sees every output, with its task and its job. Anyone who can edit tasks records one, on a task or on its own. Agents record and settle outputs; people read them on the dashboard. Each record and update appears in the workspace’s history as output.recorded or output.updated, with the why, and each counts as one action.

Reading outputs

  • List a workspace’s outputs newest first, filtered by job, agent, kind, labels (all of them), task, status and time (List Outputs).
  • A task’s read carries its newest 50 outputs and outputsTotal.
  • A job’s read carries its newest 20 outputs and outputCounts by status.

Limits

Surfaces