Skip to main content
Use the terminal control beside the Task composer to continue its coding-agent session yourself. Cogno keeps the session, working directory, and Task connected while you work.

​
Prerequisites

  • Open the Task in the Cogno desktop app.
  • The Task has a known Claude Code or Codex executor.
  • The Task is connected to an Environment. A saved execution session provides its own Engine target; a Task without one needs a local directory.
The terminal icon stays in the composer rail while Cogno checks availability. It remains disabled when the Task has no terminal target.

​
Open the terminal menu

Hover the terminal icon to read Resume in CLI when a saved session exists, or Start in CLI when Cogno can start the known executor without one. Select the icon and choose an available destination: A Task on another machine offers only Open in this app. A native terminal or copied path would address the current computer’s disk.

​
Continue the session

For a saved session, Cogno asks the Engine for the exact working directory and resume command. The command shape is:
The embedded terminal reuses the Task’s agent terminal when one already exists. Reopening the control does not intentionally start a second copy of the same session. When Cogno starts a manual session from Task handoff data, it records manual work before or alongside opening the terminal. That keeps the Task in In Progress and blocks a second run on the same Task. Recording is best-effort; a terminal opening is not held hostage by an unrelated tracking failure. Terminals started outside Cogno are not tracked.

​
Read an agent run

The conversation keeps the prompt and final response visible. During a run, the activity header shows Working for … and leaves the current narration visible. After the run finishes, Worked for … folds the intermediate work so the response remains easy to scan. Consecutive file reads, searches, edits, and commands appear as one activity summary. Open the summary to see each tool, then open a tool to inspect its command, output, or file change. Cogno bounds long terminal output inside the detail view so an old run cannot fill or stall the conversation.

​
Understand permissions

Cogno-managed runs can show one-time tool approval requests. Native allow rules and tool metadata can allow some operations without a prompt. Closing the UI or losing the connection does not approve a pending request. The approval card stays inside the conversation and names the operation it is asking about. Terminal requests show the command. Cogno’s own integration tools show the service and the action, such as Slack · Send a message or GitHub · Create a pull request, with the message or title on the line below. A Sends outside Cogno badge marks an action other people will see; Read only marks a lookup. When the agent gives no reason of its own, the card asks about the action itself, for example “Send a message to Slack?”. Tools Cogno does not recognize keep their own name. Select Show details to review the remaining arguments, then choose Deny or Allow once. You can also press Esc to deny or Enter to allow once while the card has focus. A pending approval expires after 15 minutes. For autonomous runs, Cloud receives waiting metadata such as the count and deadline, without the tool arguments or credentials. Keep Cogno and Engine updated so execution-policy acknowledgements remain compatible. Connector preferences apply to Cogno’s invocation of Claude Code or Codex. A CLI you start yourself keeps its own global configuration.

​
Troubleshooting

For other failures, see Troubleshooting.