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.

Creating, switching and watching sessions is in Running several agents. This page covers the skills and permission flags.

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):

• 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 colors its row. See Work-state colors.

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: a separate classifier model reviews each action before it runs. See Claude Code's permission modes. Codex has no equivalent reviewed mode here: its flags turn approvals and the sandbox off.

These permission 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.