Comments on a card
Every card has a comment thread. It is for remarks about the work — a question, a clarification, “also cover the expired-token path” — and it is the one place a person and the agent working a card can talk without deciding a gate.
A comment is not a decision. To send work back with instructions, use request changes on the card’s gate: that feedback becomes the rework instruction the next run is handed. A comment is read, not enforced.
Who may comment
Section titled “Who may comment”| who | read the thread | post | delete |
|---|---|---|---|
a person in the workspace (viewer and up) |
✓ | ✓ | their own comments |
| a person who is not a member | refused (403) | refused (403) | — |
an agent token with read |
✓ | — | — |
an agent token with run, while its live run holds this card |
✓ | ✓, as itself | — |
| an agent on any other card, or after its run ended | ✓ with read |
refused (403) | — |
Posting needs only board read access: a comment changes nothing about the card. An agent may post only on the card it is working — the board checks that the agent holds a live run on that card, so a token cannot be used to talk on cards it is not working.
Nobody edits a comment. The thread is append-only. The person who wrote a comment may delete it; the thread then shows “Comment deleted” with the author and time, and the text is gone. The board’s event log records that a comment was added or deleted, never what it said, so deleting one does remove its text. An agent’s comments are never deleted.
A comment is Markdown text of at most 8 KB. It is shown as the text it is — the card drawer never renders it as HTML.
Where comments appear
Section titled “Where comments appear”- The card drawer, under Comments, with the author, whether a person or an agent wrote it, and when. A comment posted anywhere — another tab, the terminal, an agent — appears there live.
- The terminal:
supi comments <boardId> <cardId>reads the thread andsupi comment <boardId> <cardId> <text|->adds to it. See the reference. - The API:
GETandPOST /v1/boards/:id/cards/:cardId/comments, andDELETE /v1/boards/:id/cards/:cardId/comments/:commentId.
When an agent sees a comment
Section titled “When an agent sees a comment”At claim. The run context an agent reads when it claims a card (GET /v1/boards/:id/runs/:runId,
or superpipeline_get_run) carries the card’s newest comments — at most 20, and at most 16 KB of
text, oldest first — and commentsOmitted, the number of older ones left out. An agent dispatched
through AgentPod has them written into its prompt, under Comments on this card.
Mid-run. A comment does not interrupt an agent that is already working. Nothing pushes it
into a running session. Instead the agent is told — in its prompt and in the MCP server’s
instructions — to re-read the thread with superpipeline_list_comments before it finishes the
stage, and to answer with superpipeline_post_comment when a comment asks it something. So a
comment posted mid-run is seen when the agent checks, which it is told to do before it completes —
not the moment it is posted. If what you have to say must stop the work, block or move the card
instead.
The next stage. The next agent to claim the card is handed the thread as it is then.