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
| Chord | Mode | Action |
|---|---|---|
<CR> | Insert | Send the prompt to the agent |
<C-j> | Insert | Insert a newline in the prompt |
<C-c> | Insert | Interrupt the active turn |
<C-t> | Normal | Toggle 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-modeprompt 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
opencode-mode— the opencode agent.ai-permission-mode— answering a permission request.ai-log-mode— the agent process's own log.