devpit

Chat

A conversation with your agent, with its cost, its questions and its checklist on screen.

Last updated

A chat is a conversation with your agent in a tab of its own: its cost, its questions and its checklist on screen. It runs through Claude Code, the only CLI devpit drives end to end. Open one from the sidebar's New Task → Chat, or with ⌘N (Ctrl on Linux and Windows). Finding, naming and exporting conversations is in Conversations.

What the turn runs as

The chips under the composer decide the next turn.

  • Account and model. One picker for both, because a model belongs to an account, and picking one from another account would be changing the account. Search it, or star the models you reach for. Once the conversation has spoken, its account is fixed.
  • Reasoning, when the CLI offers levels.
  • What the agent may do without asking: Plan (it plans and changes nothing), Supervised (ask before commands and file changes), Accept edits (edits go through; ask before anything else) or Full access. The mode you pick stays picked for that conversation, and a new conversation starts in the last mode picked on that account — Supervised if you never picked one.

A model or mode picked in the middle of a conversation applies to the next message, without starting the agent again.

Type / at the start of a message and the CLI's own commands are offered, read from what it printed when a session of that account last started. + Skill adds a skill to this turn, from the catalogue in Skills and MCP.

One agent for the whole conversation

The agent keeps running between your messages, so the next one starts its answer at once. After ten minutes with nothing said it stops, and your next message picks the conversation up where it was. If it ends on its own before answering, your message goes to a new one and is not lost.

While it runs it also hears your account's other sessions and its own background tasks: what it says then shows in the chat, marked as not in answer to you. Supervised still starts the agent for each message.

A long conversation left idle can be compacted before its next turn; see Profiles.

Questions from the agent

In Supervised, a tool call the agent wants to make appears above the composer with its input and three buttons: Refuse, Always in this chat and Allow. Always in this chat allows, for the rest of this conversation only, every file edit, or that exact command, or that same tool — the button's tip says which. A long command sits in a box that scrolls, with how many lines it has beside the tool's name, so the buttons stay on screen.

A refusal can say why: type it in the field beside Refuse, and the agent reads it as the tool's answer. A refusal goes back as a tool result rather than killing the turn — an agent told no can try something else. Unanswered for two minutes, it is refused. The same question can be answered from the island.

Only a conversation that asked to be asked is held. Every other session, board steps included, is answered at once and left to the CLI's own permission mode, so a step never hangs on a window nobody is watching.

A plan to approve

In Plan, the agent ends with a plan for you, as a card with three buttons:

  • Approve leaves plan mode and goes on in Supervised, so it asks before changing anything.
  • Edit turns the plan into text you can change, and approving then sends your version.
  • Refuse, with the reason typed beside it when there is one.

While the turn runs

  • Cost — a running total in the chat's header: what this conversation has cost.
  • Checklist — the agent's own task list, replayed from its calls, as n of m done.
  • Subagents started side by side are one row — 2 subagents · 1 working — that opens to each one with its state (Working, Done, Failed, Stopped) and how long it ran. While only subagents are working, the live line says Waiting on N subagents.
  • Thinking shows how long it took: Thought for 12s.
  • In background — a chip by the composer, N in background, for work the turn left running: subagents, shells and other tasks, each with how long it has run and a Stop, and a wakeup the turn set, with when it fires.
  • Queued — what you send while it works waits above the composer and goes in as soon as it has answered, one at a time. Stopping the turn puts the queue back in the composer.
  • Steered — a message sent into the running turn instead. If the agent never took it in, it says Didn't reach the agent when the turn ends, with Resend.
  • Esc twice stops the turn. The first press arms it and the button says so; the second stops. An open menu or dialog takes the Escape instead. Stop works even before the agent has started, and if the turn has not ended ten seconds later, Force stop ends it with everything it started.

Switch project or reload the window in the middle of an answer, and coming back, the answer so far is there and the rest keeps arriving. Unsent text is kept per conversation until you send or clear it.

Settings → Sessions → Send with chooses Enter or ⌘Enter (Ctrl on Linux and Windows) to send; with ⌘Enter, Enter writes a new line. While a turn runs chooses whether a message sent then is Queued or Steered. While Enter sends, ⌘Enter sends it the other way.

A side question

/btw <question> asks something while a turn runs, without stopping it, steering it or waiting behind it. The answer comes from the conversation so far and shows in a bubble of its own above the composer. Neither the question nor its answer joins the conversation.

How much of the work you see

The Activity menu in the chat's header chooses how much of each turn is drawn:

Option What shows
Show all every tool call
Hide successful only calls that failed or are still running
Hide all the replies, without the work
Zen your messages and each turn's final answer

The choice is kept per conversation; the default for new ones is Settings → Sessions → Chat activity. Failures, changed files and permission questions always show, and find shows everything while it is open.

A closed tool row is one line. Opened, it shows its whole input and its output, each with a copy button, and an MCP tool names its server.

Files and voice

The paperclip, or ⇧⌘A (Ctrl on Linux and Windows), attaches files from the system's picker. Dropping files onto the chat and pasting a picture work too. Each reaches the agent as a path it can open. Click an attachment to take it off again.

The microphone beside it records a voice message: click it, or hold it and let go. The words land in the composer to read and edit before sending. They come from a whisper on this machine — whisper.cpp or whisper — or from a transcription service only if you choose one in Settings → Voice → Voice messages and keep its key there. Without an engine, the recording goes as a file.

Your message and the answer

Point at a message for its actions:

  • Your message: its time, Copy (your words as typed, with the files they went with listed after), Edit from here and Send again.
  • An answer: Copy, Edit from here and Fork.

Edit from here goes back to just before that message and puts it in the composer to change. When the turns from there on changed files, it asks whether to put the files back as that turn found them; that touches only your working tree, never what is staged. Fork goes on from an answer in a new conversation, and the one it came from keeps every turn it had.

Each turn ends with one quiet line: when it ended, the model and reasoning it ran on, how long it took, how full the context was, and what it cost.

Retry, beside Continue, sends a message again when its turn failed or never answered. Select words in an answer and Quote adds them to the composer as a quote.

In an answer

Web addresses are links, and open in your browser. A file path — written as code, or as file:line in the text or in a tool's input and output — opens beside the chat; in a card's chat, from the card's own checkout. Open as a tab expands it. A picture the answer points at on this machine, by its full path or one relative to the conversation's folder, is drawn in place, and any picture — in your messages, an answer or a tool's result — opens large with a click, with Open and Save.

With Settings → Appearance → LaTeX maths on, $…$, $$…$$, \(…\) and \[…\] in answers and Markdown files are drawn as formulas. It is off by default, and on, a price such as "$5 and $10" still stays text.

A tool that brings an interactive page — a Jira issue, a Datadog chart — shows it under the answer that called it, sandboxed; when the page wants to run one of its server's tools, it asks first: Deny, Allow once or Always here.

When the provider says no

A turn refused by the provider gets a banner under it: Usage limit reached, or The provider is not answering. Try again sends the message again now. When Claude says when the limit resets, Resume when it resets shows that time and sends the message again by itself then; it can be cancelled. A reopened conversation still shows the banner.

For a screen reader

The thread is a log named Conversation, and each message says who spoke: You or Agent. A turn is announced when it starts — The agent is working. — and when it ends — The agent answered. Worked for 12s — never word by word. Every chat control reached with the keyboard draws a focus ring.

Going elsewhere

Copy session ID, in the chat's header, copies the CLI's own id for this conversation — what claude --resume takes. The phone button turns on Remote Control, so the same conversation can be reached from claude.ai and the Claude app as devpit-<project>; Open on claude.ai opens its page. It needs Accept edits or Full access: Supervised runs a process per turn.