ai-conversation-mode

ai-conversation-mode: the agent conversation buffer — a read-only transcript with an editable prompt. CR sends, C-j newlines, C-c interrupts, C-t toggles trust.

On this page

The agent conversation buffer: the transcript above, an editable prompt at the bottom. This is the ACP-backed view — see opencode-mode for the agent that most commonly drives it.

Because it's a transcript-plus-prompt buffer, it also carries repl-mode: pressing i (or a / o / A / I / O) anywhere jumps you to the prompt and enters Insert, rather than trying to insert into the read-only transcript.

The headerline reports the session's token usage and cost, plus a ⌛ N queued count when prompts are waiting behind the running turn.

Chords

ChordModeAction
<CR>InsertSend the prompt to the agent
<C-j>InsertInsert a newline in the prompt
<C-c>InsertInterrupt the active turn
<C-t>NormalToggle trust mode (auto-accept vs review)

<CR> sends, which is why a newline needs its own key — a multi-line prompt is common enough (paste a stack trace, write a spec) that <C-j> is worth the muscle memory.

Trust mode

<C-t> switches between review and auto-accept for the agent's permission requests:

  • Review — each request the agent makes surfaces as an ai-permission-mode prompt you answer.
  • Auto-accept — requests are granted without asking.

Toggle deliberately. Auto-accept is the right setting when you're watching a long mechanical task and wrong when the agent is touching anything you'd want to see first.

The transcript

Streams in as the agent works — you can read and search it while a turn is still running. Tool calls and reasoning blocks are foldable, so a long transcript stays navigable; the usual fold chords apply (see folding).

Only the prompt region is editable. The transcript is a record, and the read-only gate is what keeps it one.

See also