Agent chat
For Claude Code and Codex panes, Melta shows the conversation as a chat instead of a terminal screen. The terminal is one tap away with the switch at the top right.
Where the chat comes from
Claude Code and Codex write every turn to a session log on your computer. Melta asks Herdr which session the pane is running, finds that log (under ~/.claude/projects or ~/.codex/sessions) and follows it over SSH. The conversation comes from the log, not the screen, and does not pass through Melta's server.
Opening a chat loads up to the last 300 messages, then adds new ones as they are written. When the agent starts a new session, for example after /clear or /resume, Melta follows the new log.
Herdr only knows the session if its agent integration is installed. If the chat says No transcript yet, install it on the computer:
$ herdr integration install claude $ herdr integration install codex
A brand-new session has no log until its first turn, so Melta keeps checking every few seconds. For other agents, the chat screen cannot show the conversation and says so; use the terminal or Show screen.
What you see
- Your messages as blue bubbles on the right.
- The agent's replies as text, each with Copy.
- Thought: the agent's thinking, folded.
- Tool calls as one line, such as Ran, Edited, Read or Searched, with the command or file. Tap for the full input.
- Output or Error with a line count, folded.
Claude Code's background notices and sub-agent conversations are left out.
What the agent is doing
While the agent works, the bottom of the chat shows a typing indicator, a short description from its latest log entry such as Running npm test, and how long it has been working. After 90 seconds without new output it adds no output for and the time. Stop sends Esc. If the connection drops, the chat says so instead of guessing.
When the agent asks for permission, the chat shows Allow and Deny. See Approvals.
Send a message
Type in the box and tap the arrow. While the agent works, the box says Queue a message. Melta first clears anything half-typed in the agent's own input, then hands your text to Herdr and waits up to 20 seconds for the agent to start on it. Under your message you see:
- Sending…, then Delivered once the agent picks it up.
- Sent, but not found in the agent's log if the turn ends and the log never showed it.
- Not Delivered with a reason, for example when the agent is waiting for an approval.
Failed and unconfirmed messages offer Retry and Dismiss. Text starting with / is typed in as a command; see Slash commands.
Attach images
Tap Add image and pick up to four photos. Melta copies each to the computer over SSH, into ~/.melta/uploads, and adds a line Attached image: with its path to your message, so the agent can read it. Without text, the message says Look at the attached image.
Buttons and menu
Under the message box: Switch workspace or pane and ⇄ (Jump To and last pane), Add image, / for slash commands, esc and ⌃C, and the eye to show the live screen.
The ⋯ menu has Show screen, Search in chat, Share chat… (see Share a chat), Reconnect, Show on computer, Send Escape and Send Ctrl+C.