Agent Sessions

An agent session is a Claude Code or Codex process running in a workspace on the backend daemon: you describe what you want in the agent pane, and it reads code, writes code, and runs code in the workspace's REPL to get there. For the pane itself, see The agent pane.

Each session runs Claude Code, Codex or a plain shell, and brings that agent's own tools and context management. A session belongs to a workspace on the backend, so it sits alongside project state and survives a frontend restart. Several sessions run concurrently, one per workspace, across machines, and message each other over the inter-agent communication system. You start and switch them from Sessions mode. A dedicated Agents mode (tasks, timeline, step detail) is not built.

What Ship of Tools adds to the agent

The installer gives each agent CLI a set of skills and hooks (listed under What the installer changes). For Claude Code they exist only in a session Ship of Tools starts, a row or ccb: a claude you start yourself in a terminal loads none of them.

• Skills that call small command-line tools: julia-repl runs code in the session's REPL (How agents use the REPL), show-result puts a file in your preview pane or opens a page in your browser, and sot-comm sends messages to other sessions. Claude Code gets all of them; Codex gets only sot-comm and sot-session-start, so Codex sessions have no REPL skill yet.

• Hooks that report the session's work state (working, blocked, done), which is what colours its row. See Work-state colours.

Permission mode

Sessions created from the window start their agent with:

• Claude Code: claude --permission-mode auto.

• Codex: codex --dangerously-bypass-approvals-and-sandbox --dangerously-bypass-hook-trust, through the ccx launcher, which also marks the session's directory as trusted in ~/.codex/config.toml.

Claude Code's auto mode is not a bypass. It runs without routine permission prompts, but a separate classifier model reviews each action before it runs and blocks anything that goes beyond your request, targets infrastructure it does not recognise, or looks driven by hostile content; your own ask rules still prompt. See Claude Code's permission modes. Codex has no equivalent reviewed mode here: its flags turn approvals and the sandbox off.

These flags are fixed: no setting or environment variable changes them. The Claude Code flags are built into the daemon, and the Codex flags are in the ccx script, which every update overwrites. Ship of Tools adds no permission tiers or action gating of its own on top.

See also

• The REPL — the session the agent runs code in.

• Running several agents — creating, switching and watching sessions.