Branches, Side Chat, and Subagents

Message revisions and branches

Editing an earlier message creates a revision. Previous/Next message revision moves through the history. Continuing from an earlier node with Branch creates a separate path while leaving the original branch available. Branches work well for method or model comparisons; don't use them to conceal a failed run.

Side chat

More send options → Side chat opens a parallel discussion. Use it to clarify a concept, draft supporting text, or discuss the next action without immediately changing the main session's execution order. Advice can reach a running main turn through the framework's native follow-up channel; if live injection is unavailable, it is relayed with the next user turn. Closing the panel alone does not guarantee that unsent conclusions enter the main path.

Specialists

Agent controls can select a Specialist. A Specialist is a reusable agent configuration with an identity, instructions, base prompt, and capability boundaries. During a switch, Open Science resolves the active-run barrier. A request to delete or switch the active specialist can display a dedicated confirmation with the affected details.

Delegation and subagents

With Delegation enabled, the main agent can create subagents. Each subagent has its own state, message row, and notebook runs, but remains attached to the current project and session. A delegation card reports the question, objective, and lifecycle. Returned results enter the main timeline.

• Configure Subagent model and its Reasoning effort under Settings → Model.
• Settings → Agent reports whether the current framework supports delegation.
• Subagent permission requests remain visible, and the permission panel shows the pending count.
• Release ends or frees the subtask resource. It doesn't delete the main session evidence.
• Use the Agent filter in Session Notebook to inspect one subagent's runs.

Context window

Composer + → Context reports the model limit, current use, and trend. When usage reaches the configured threshold, the agent can compact the conversation and add an explicit activity row. Changing to a larger-context model doesn't restore content already compressed. Save important inputs as project files and reference them with @.

The Context dialog can show stacked composition bars, compaction markers, and per-call input/cache/output details. The provider-reported total is authoritative; category attribution can be a local estimate. A coverage notice means some historical calls were not reported and must not be treated as zero usage.

Archive and delete

Archive is reversible organization; Settings → Archived can restore the item. Delete is permanent and requires confirmation. Before deletion, download required artifacts and notebooks, then check that no external conclusion depends on a file that exists only in this project.