Add a server-side tool for launching same-backend child conversations

Problem or Motivation

Agent Canvas currently implements launch_child_conversation as a client-defined tool. When the agent invokes it, every connected Canvas client can receive the action and the browser performs the conversation-creation side effect. Canvas therefore maintains a browser-local claim ledger to prevent replayed or multiply delivered actions from launching duplicate children.

The browser should not own this operation when a child is intended to run on the same backend as its parent. The agent server already owns the parent conversation, workspace, worktree behavior, and parent/child relationship.

Desired Behavior

Provide a server-side tool that launches a child conversation on the same agent-server backend as its parent and returns the child conversation identity and launch status to the parent agent. Canvas should only render the tool action/result; it should not execute the launch.

The tool should support the child-launch inputs that are meaningful within one backend, including the task brief, optional title, and local isolation/worktree behavior. Cross-backend target selection is intentionally out of scope.

Acceptance Criteria

• An agent can invoke a server-executed tool to create a child conversation on its own backend.
• The child is persisted with its parent conversation relationship and inherits the parent workspace as required by the existing agent-server relationship rules.
• The tool returns the child conversation ID and initial launch status to the parent agent without requiring a browser-side callback.
• Local worktree and shared-workspace isolation behavior is covered by focused agent-server tests.
• The tool is not exposed when the active backend cannot support child creation.

Related

• OpenHands/OpenHands#16769
• OpenHands/OpenHands#16969

This issue was created by an AI agent (OpenHands) on behalf of Gregory Neubig.

Start with the existing client-defined launch_child_conversation tool and the agent server's conversation, workspace, and parent/child relationship rules. Trace how child-launch inputs and backend capability checks are handled, then add focused agent-server tests for local worktree and shared-workspace behavior. Done means the server returns the child identity and launch status without a browser callback.