Skip to content
Skribly Docs

Terminals and sessions

A session is one agent, or one plain shell, running in a pane. Start one from Agents → New session, with Ctrl+B a, from the split buttons in a pane header, or from the command palette (New session: …).

New session

Section titled “New session”
Field Meaning
Agent An agent card or Terminal for a plain shell. Agents that are not installed show their install command.
Project The directory the session starts in
Working copy Main, an existing worktree or + New worktree…
Machine This computer or an SSH machine from settings
Name for the CLI and other agents [a-z][a-z0-9_-]{0,31}, for example reviewer or impl-1
Mode Terminal (the interactive agent) or Chat (headless)
Can modify files Chat only: whether the agent may edit files
First message Sent right away; Enter starts the session

In terminal mode Skribly opens a login shell ($SHELL -l; on Windows pwsh or powershell.exe) and types the agent command into it with the first message as an argument. Agents that do not take a message as an argument get it typed in after launch. Every session has the environment variable SKRIBLY_TERMINAL=<id>.

Chat mode

Section titled “Chat mode”

Chat runs an agent headless: each message is a separate call with the session id from the previous reply, so the agent remembers the conversation while no process runs between messages. The pane shows text, tool calls and errors.

Key Action
Enter Send
Shift+Enter New line
Stop Interrupt the reply
Agent Can modify files Read-only
Claude Code --permission-mode acceptEdits --permission-mode plan
Codex --full-auto --sandbox read-only
Cursor Agent --force no flag
Gemini CLI --approval-mode auto_edit no flag
Copilot CLI --allow-all-tools no flag
Aider no flag --dry-run
OpenCode no toggle no toggle

Amp, Droid, Qwen Code, Goose and Crush also have a headless mode. Kimi CLI works in the terminal only. Chats are kept in memory and do not come back after a restart; continue the conversation in a terminal, for example with claude --resume. The same headless mode powers pipelines and the Agent button for commit messages.

Agent panel

Section titled “Agent panel”

The panel on the right opens over any view with ⌘⇧A (Ctrl+Shift+A). Each project keeps its own conversations, and they come back after a restart.

Tool What it does
Agent menu Chooses the agent, or installs a missing one. Changing agents starts a new conversation.
Model The default model, the agent’s list, or any name you type
Effort Claude Code only: --effort from low to max
/ Commands from the agent, the project, your home folder and plugins; /clear starts over
Plugins Turns Claude Code plugins on and off for the next message
Paperclip Attaches files; you can also drop files or paste an image
Edit / Plan Whether the agent may change files
Remote Starts a Claude Code session with Remote Control
Terminal icon Continues the same conversation in a terminal

Under the message box you see the session time, subagents, context and cost, as far as the agent reports them. Agent terminals get the same tools in a bar under the pane; the bar holds back while the agent is asking a question, because text sent then would answer it.

Resume after restart

Section titled “Resume after restart”

Settings → Sessions & control → Restore sessions after restart is on by default.

  • Tabs, splits and panes come back in their places with the same names, worktrees and tasks.
  • Agents with a native session id resume the conversation and get a resumed badge:
Agent Resume
Claude Code, Cursor Agent, Droid, Qwen Code --resume <id>
Codex codex resume <id>
Copilot CLI --resume=<id>
OpenCode, Crush, Kimi CLI --session <id>
Goose goose session --resume --session-id <id>
Custom agent Resume arguments from its definition ({id})
  • Other panes (shells, Gemini CLI, Aider, Amp, SSH sessions) start as a fresh shell in the same directory, with a hint showing the command that starts the agent again.

Quitting the app ends all terminal processes; there is no background server like tmux has. Continuity comes from the agents’ native resume and from state files (STATE.md, handoff.md).

SSH machines

Section titled “SSH machines”

An agent can work on a server while its pane stays in Skribly.

  1. Settings → Sessions & control → SSH machines: enter a name and an ssh target, user@host or an alias from ~/.ssh/config, then Add.
  2. In New session, choose the machine in the Machine field.

Skribly opens ssh -t <host>, changes to the session directory (or the home directory if it does not exist there), starts a login shell and runs the agent. The agent must be installed on the server. Authentication is handled by OpenSSH with keys in ssh-agent or ~/.ssh/config; Skribly stores no passwords.

Broadcast

Section titled “Broadcast”

Send one prompt to several agents at once, for example “run the tests and describe the failures”.

  1. Tick the checkbox in the pane headers.
  2. Type the text in the field on the tab bar (“To N selected…”).
  3. Press Enter.

The text reaches every running pane as a bracketed paste followed by Enter. Exited panes are skipped.

Handing off work

Section titled “Handing off work”

The hand-off button in an agent pane’s header passes the work to another agent without replaying the conversation.

  1. Send request: the current agent saves its state (what is done, the next step, decisions, open questions) to .skribly/tasks/<KEY>/STATE.md for a task, or .skribly/handoff.md otherwise.
  2. Run another agent: the new agent starts in the same directory and worktree with a prompt to read the state file and TASK.md, look at git status and git diff, summarise in three points and continue.

The new agent starts from a few dozen lines of state instead of a whole transcript.