Ultraplan: approving a plan in the web session executes it in the cloud instead of returning it to the originating CLI session

Environment

Claude Code CLI in VS Code on Windows 11, working in a local git checkout. Plan built locally in plan mode.

What happened

1. Built an implementation plan in plan mode in the CLI session (IDE-centric workflow).
2. At the ExitPlanMode approval prompt, the user chose to send the plan to Ultraplan for remote refinement, with the stated intent of iterating on it in the browser and teleporting it back for local implementation.
3. The cloud session refined the plan and reported "ready for approval" with a web link.
4. Approving the refined plan in the web UI immediately began executing it in the cloud session ("Results will land as a pull request... There is nothing to do here").

The originating CLI session received only passive status notifications. No execution-venue choice was ever surfaced in the CLI, and the approval action in the web UI carried an implicit venue decision the user did not want.

The problem

Plan approval and execution venue are orthogonal decisions, but the web approval flow welds them together: approve = execute in the cloud; the only path that returns the plan to the originating IDE session appears to be rejecting it. A user who wants to say "yes, this plan is good, now run it on my machine where I am working" has no affordance for that at the moment it matters.

For an IDE-centric developer, this converts a local edit-run-verify loop into a cloud-execution-plus-PR round trip they never asked for.

Expected behavior

• The approval screen should ask where execution happens (or offer both actions: "Approve & run in cloud" / "Approve & return to CLI").
• A plan that originated in a CLI session should default to returning there for execution.
• The originating CLI session should be able to reclaim the approved plan rather than only receiving read-only status notifications.

Impact

Severe workflow disruption and loss of trust in the handoff feature: the user explicitly wanted remote refinement with local execution, followed the flow as presented, and got cloud execution as a fait accompli.