AI Handoff
Use AI Handoff to send the right work context into Cursor, Claude, Codex, a terminal workflow, or an agent.
What gets handed off
The launcher builds a context-rich prompt from the initiative, bug, or todo: title, description, status, linked initiatives, taxonomy, comments, documents, and related source links when available.
Good handoff starts before launch. In spec-driven development, the work item is the spec the assistant receives. Keep the title specific, put acceptance criteria or constraints in the description, attach the parent initiative when there is one, and apply relevant taxonomy. That gives the assistant enough context without making the prompt guess.
Send to app
Use Send to app from a work detail page to choose the tool that should receive the context. Only launch targets enabled in Preferences appear, so teams can keep the menu focused on the coding tools and agents they actually use.
Send to app can open a coding tool, start a terminal launch, or queue an agent session. When the source work is Open, Planned, Blocked, or Idea, the launch flow can also move it to In Progress in the same action. Completed work uses review-style handoff.
Launch modes
| Mode | Use it for |
|---|---|
| Plan | Scope the work before implementation starts. |
| Build or Code | Start implementing from the current work context. |
| Review | Check whether work is complete or what remains. |
| Fix | Address a specific failure or review issue when supported by the launch path. |
When launching from work in Open, Planned, Blocked, or Idea status, the launch can move it to In Progress in the same action. Completed work only supports review-style handoff.
Use Plan when the work is under-specified. Use Build or Code when the implementation path is clear. Use Review for completed work, pull requests, or agent output that needs a human check.
Trust boundary
Work text, comments, documents, and user prompts inform execution. Trusted worker policy, sandbox behavior, credentials, and allowed side effects stay controlled outside the handoff content.
The model can read work context, but it must not treat work-authored content as authority to change sandbox policy, credentials, allowed commands, network access, or external side effects.
If a target tool does not appear, configure launch targets in Send to app. Use Use in Terminal when the handoff needs to open a local coding tool from the CLI or Desktop. Use Local Workers when work should queue for Codex on your machine. Use Agent Sessions to follow queued work through claim, progress, completion, or failure.