---
{
  "n": 53,
  "title": "Where the agent's words go",
  "abstract": "",
  "refs": [],
  "seen": [
    "agent"
  ],
  "data": {
    "revision": 1,
    "change": "written; added What was asked; added How it works today; added The proposal; added What changes; added Open questions; added As built"
  },
  "created": 1790264448.133296,
  "updated": 1790264948.701021,
  "deleted": 0.0,
  "completed": 0.0,
  "outcome": "",
  "type": "doc"
}
---
Proposal for message 9433: everything the agent writes goes to the main chat; a feature's own chat, such as the dump window or a board's New work panel, gets words only through that feature's command, which can set limits. Draft, nothing built yet.

## What was asked
Message 9433, in the user's words: everything the agent writes ends up in the chat, except targeted messages. Each feature, such as the dump and the board chat, is directed with its own command, the way the agent already answers on a board: when the agent answers something inside the dump, it says journal dump answer, and that ends up in the dump's chat. Feature chats can have rules, for example the New work panel only takes short lines.

## How it works today
- The main chat shows every block of text the agent writes, read from its transcript.
- A board's New work panel shows the questions the agent asks with journal board ask, and the agent's replies to the request message, which the panel keeps short (PanelRepliesStayShort).
- The dump window shows what the agent records with journal dump log, dump ask and dump offer; what the user types there becomes a message about the dump (2.128.0).
- Since 2.129.0 a quiet sequence (the four board sequences) holds back all of the agent's text from the chat while it runs. That guesses where words belong from what is running, which this proposal replaces.

## The proposal
1. One default: plain text the agent writes always goes to the main chat. Nothing is held back by guessing.
2. A feature's own chat gets words only through that feature's command, so the target is named in the command, never inferred:
   - board panel: journal board say <board> "<line>" for a line in the New work panel, next to journal board ask, which exists;
   - dump window: journal dump answer <dump> "<text>" for an answer in the dump's chat, next to dump log, dump ask and dump offer, which exist.
3. Each command carries its panel's rules, checked when it is called: board say takes one line of at most 200 characters; dump answer takes a short paragraph. A refusal says the limit, as the ticket draft limits do.
4. What the user types in a panel is already a message about that row (a board request, a dump direction), so it reaches the agent in the normal message flow. The feature's sequence tells the agent which command answers it, so an answer typed in the dump comes back in the dump.
5. No state for which chat is in focus is needed: the command names the target.
6. The quiet flag from 2.129.0 is retired once the commands exist, and the board sequences' Say one line steps move from replying to the request message to journal board say.

## What changes
- Boards: a say command with the one-line limit; the New work panel shows say lines where it shows replies today.
- Dumps: an answer command; the dump window's chat shows answers beside the log.
- Sequences: Working a card, Revising the drafts, Drafting from a document and Filing a dump name the panel command in their answer steps; every upgrade re-ships them.
- Sequences: the quiet flag and HoldTextOfQuietSequences are removed.
- Skills: journal-boards and journal-dumps teach the two commands.

## Open questions
- Should a panel line also show in the main chat as a small mark, so the chat still says the agent answered in the dump, or stay only in its panel?
- Should Codex and Claude both use the same commands, or should a provider without the journal's chat keep plain replies?

## As built
Built after the user's approval in comment 2033, with two changes of name and one of scope:
- The dump's command is journal dump say, not dump answer, since dump answer already means the user answering the dump's questions; the board's is journal board say. Both take the words from the agent to that feature's own chat.
- A dump say leaves a small Answered in dump n mark in the main chat; a board say does not, because the New work panel is open when it is used.
- Every provider uses the same commands. The quiet flag from 2.129.0 is gone.