A real process, not a chat simulation

Each Cockpit conversation is backed by a long-lived coding-agent process. The chat-style transcript and terminal pane are two views of the same underlying session.

Sessions and reattachment

  • One conversation maps to one long-lived terminal-backed agent process.
  • Returning to a session rehydrates the conversation from the agent’s native transcript.
  • The terminal view replays a bounded amount of scrollback; it is not a complete permanent log.
  • Starting a new conversation creates a new session rather than silently reusing an old process.

Models and effort

ChoiceUse whenNotes
HaikuSpeed and narrow, low-risk tasks matter mostDoes not use the same thinking behaviour as the other Claude choices
SonnetBalanced coding, explanation, and everyday debuggingSupports Claude effort levels
OpusComplex architecture, difficult debugging, or broad refactoringSupports Claude effort levels
CodexProvider-specific coding workflowsRepresented in product but disabled by default in the current release

Claude effort levels are low, medium, high, xhigh, and max. The default is high. Use higher effort when deeper analysis is worth additional time; use lower effort for narrow, reversible tasks.

Current integration boundary

Cockpit does not currently expose SOCK-specific slash commands or directly consume a capture. You can manually bring a cited SOCK answer into a Cockpit prompt, but describe that as a user-mediated workflow rather than built-in integration.