Work with sessions and context

Use a session to keep one assignment's conversation together. State the goal and allowed actions explicitly, then review each result before continuing. A session history is not a promise that every earlier message is sent verbatim to the model forever.

These controls describe the October 2026 development interface; labels and availability may differ in your installed release.

Start a bounded assignment

1. Open Main Agent for the introductory text-only exercise, or open the existing session for the assignment you are continuing.
2. Check the selected model.
3. In Message AgReady, supply the relevant input, requested output and limits. For example: “Compare these two supplied notes. Return differences in the conversation. Do not browse, change files or send messages.”
4. Use Send message once. Watch the displayed state and any error or approval request.
5. Review the response against your input before adding the next instruction.

When a separate agent session is appropriate and available, use New agent session, select an available agent, enter a session name and choose Create session. Enter the concrete assignment after opening it. If no suitable agent is available, do not assume one has been installed or configured.

Continue without losing the objective

For a correction to the same deliverable, stay in the session and specify exactly what should change. Quote a short source passage or identify the attached file when the correction depends on it. For a substantially different assignment, use a separate session where available and provide its required inputs explicitly.

Do not assume a new session inherits all earlier material, permissions or files. A useful opening brief includes:

• the outcome and output format;
• which supplied material is authoritative;
• actions allowed now and actions requiring approval;
• unresolved questions and a concrete acceptance check.

Understand history and context

The conversation you can browse and the material assembled for the next model request are related but not identical. Long conversations can be compressed into summaries while recent material is retained. Compression is not a guarantee that every detail survives unchanged.

Where context usage is shown, distinguish estimated input from provider-observed usage and reserved output. Treat it as a diagnostic aid, not proof of unlimited capacity. If the interface says context compression is in progress, wait until it finishes rather than repeating the operation.

For consequential facts, repeat the short authoritative statement or refer back to its source. If an answer contradicts an earlier instruction, point out that discrepancy rather than relying on the agent to infer the correction.

Stop, archive, restore or clear

Stop current run

Interrupt work that is currently running

It does not undo an external action or prove every queued message has been removed. Inspect the resulting state.

Archive

Put an available session aside as read-only

Archiving stops active execution; it is not deletion or a backup. Some managed sessions have separate lifecycle restrictions.

Restore

Make an archived session available again

Restoration does not automatically resume the runner or consume its queued work. Inspect the restored state before asking it to continue.

/clear-history

Deliberately reset the current session's conversation context

This resets the visible history boundary and related context state. It is not a secure-erasure command or a way to undo completed actions.

Do not clear history as routine troubleshooting. Preserve any information you still need first, and use the command only when you intend a reset. Main Agent is not a disposable session; deletion controls are limited to eligible sub-sessions. Review the displayed confirmation before deleting anything.

If nothing appears

Confirm the selected session and model, then check whether the message was actually submitted. An active run, pending approval, processing attachment or connection error needs a different response. Do not repeatedly resend while the previous outcome is unknown. Use the diagnostic checks, then retry a small text-only request once the state is clear.

All documentation · Get started