Tasks and Conversations

Conversations, turns, and runs

A “conversation” stores the continuous work history between you and the root agent. Each request you send begins a new turn. The model calls, tool calls, and approval process for that turn make up one run.

Use the same conversation to continue working toward the same goal. Start a new conversation—or choose “Continue in a new task” from a response—when the goal changes substantially, the history has become too long, or you want to preserve a branch of the work.

Create a conversation

Select “New conversation” in the left sidebar.
Select a project or “No project.”
Select a model and permissions, then add attachments and a skill as needed.
Enter the goal and press Enter. Use Shift+Enter for a new line.

Composer drafts are stored on your device. Each conversation and the new-conversation screen maintain separate drafts. A draft is not a backup system; do not leave important material only in an unsent state.

Write an executable task

A practical request usually includes:

Goal: What the final result should be.
Scope: Which files, services, or data may be viewed or modified.
Constraints: What must not be done and what requires confirmation first.
Acceptance criteria: Which tests, format, or facts determine completion.

For example:

Find out why the sign-in page loads slowly. Analyze only src/auth and do not change code yet. Provide the evidence, the impact, and two possible fixes.

This is more likely to produce a verifiable result than “Optimize the sign-in page.” See Writing Effective Tasks for more guidance.

Add messages during a run

While the agent is working, a new message enters the pending-message queue instead of starting a second root task. You can:

Drag messages to reorder them.
Edit or remove messages that have not been submitted.
Use “Steer,” when available, to pass a message into the current run.

“Open in side conversation” in the queue currently shows an information page only; it does not create a usable parallel conversation. For parallel work, ask the root agent to use Multi-Agent.

Stop a task

When the composer is empty and the agent is running, select the stop button. The stop request is passed to the model and tools as soon as possible, with three boundaries:

Completed file writes are not automatically reverted.
Requests already sent to a website, model, MCP server, or image service may already have taken effect or incurred charges.
If the app cannot determine the result of an external operation, it may report the outcome as unknown. Do not immediately repeat a step that has side effects.

Handle approvals

Under Default permissions, file edits and commands usually pause for your approval. The approval card shows the action, target, and necessary safety information. You can approve or deny the action. Some actions also accept an explanation for denial or let you remember similar permission only for the current run.

See Permissions and Approvals for complete rules.

Message actions

Depending on the message and run state, the timeline may offer:

Copy, bookmark, or remove bookmark.
Edit the last completed user message and rerun that turn.
Choose “Continue in a new task” from a completed assistant response.
Open the previous turn's Git changes or generated artifacts.
View token usage after enabling conversation token details under “Settings → Usage and cost estimates.”

Editing the last user message creates a new execution result. Do not treat both the old and new responses as current fact. Some state changes cannot be made safely while a task is still running or waiting for approval.

“Continue in a new task” creates a new conversation with a snapshot of the selected history. It does not continue synchronizing with the original conversation. If the original task still has commands running, wait for or stop them first.

Search and organize

The left sidebar provides:

Local text search across conversation titles and content.
Sorting by creation time or most recent update.
Pinning projects or conversations.
Renaming, marking unread, and archiving.
Bulk archiving of a project's conversations.

Archiving does not immediately delete content. Under “Settings → Archived conversations,” you can filter by project, restore records, or permanently delete one or all of them. Permanent deletion cannot be undone from the interface.

Relationship between projects and conversations

A project supplies a folder boundary for a conversation, but the conversation remains an independent history. Removing a project permanently deletes the project, related local conversations, and attachment records saved by Captain Who. It does not modify the actual files in the project directory.

A “new conversation” created by a scheduled automation also appears in the conversation list. Deleting the automation does not delete conversations it created previously.

Current limitations

Conversations do not synchronize across devices.
Search is local text retrieval, not general semantic search.
There is no unified undo for external side effects from a run.
A regular user cannot send messages directly into a subagent's read-only conversation.