TALK mode

§ TALK

TALK is a focused conversation with the model of your choice, scoped to a single workspace. It is not a sidebar bolted onto an editor — it is the surface where you think out loud, hand work to the agent, and watch it read, write, and run inside the worktree.

Every response streams in as it is generated. Tool calls are surfaced as cards, every message carries its own token and cost, and you can switch models mid-session without losing the thread.

How to open it

Press ⌘⇧1 from any workspace, or click Talk in the mode switcher. TALK is always present — every workspace has it.

[screenshot: TALK mode with a streaming response and a tool-call card]

The interface

A streamed answer is rendered in two parts. The key phrase — the lead sentence — is set in the serif italic display face, so the gist of the reply reads as a single moment. The body follows in plain text below it.

When the agent uses a tool, it appears as a card prefixed with §:

• § READ — the agent read a file.
• § WRITE — the agent wrote or edited a file.
• § RUN — the agent ran a command.

Each message footer shows its token count (input, output, cache) and its cost for the active model.

The Companion panel on the right has two halves:

• Conversation history — tabs for past conversations in this workspace.
• Context — token usage against the active model’s context limit, the list of unstaged files the agent can see, and a live tool-call counter for the current turn.

A typical flow

1. Open a workspace and press ⌘⇧1.
2. Type a request and send it. The response streams in, key phrase first.
3. Watch the § READ / § WRITE / § RUN cards as the agent works.
4. Check the per-message cost and the Context panel’s token usage.
5. Need a stronger model for the next step? Open the model picker and switch — the new model picks up the same conversation.
6. Press ↑ to recall and re-send a previous prompt.

Shortcuts

Shortcut	Action
⌘⇧1	Switch to TALK
↑ / ↓	Cycle prompt history
⌘K	Command palette (model switch, templates, theme)

Tips

Switch models for the hard parts

The model picker changes provider and model in place. Draft with a cheaper or local model, then switch to a stronger one for the step that matters — the conversation continues uninterrupted, and the per-message cost updates to reflect the new model.

Note

The Context panel’s token usage is measured against the active model’s limit. If you switch to a model with a smaller context window, the same conversation may suddenly read as much closer to full.

Once the agent has made changes, head to REVIEW to see the diff, or RUN to execute commands in a real terminal.