Outcome
Make the running interface the canvas described by VISION.md: a person asks the agent for a change, sees that change in the working UI as soon as it validates, and continues editing or using it without a mandatory Keep/Reject decision.
Product principles
- The agent is the programmable backbone, not a chatbot beside a static app.
- The interface stays alive; customization is not a separate design-and-export phase.
- Safety, attribution, and recovery remain deterministic but quiet.
- Normal UI edits must feel reversible, not dangerous.
Required behavior
- Keep one visible composer and one continuous user conversation while the agent uses, inspects, or edits the interface.
- Apply each valid UI-only change to the working canvas automatically.
- Never disable further shaping merely because a valid working revision exists.
- Preserve the last successful render while a new change builds or fails.
- Return build, runtime, and validation diagnostics to the same agent so it can repair its work.
- Create an automatic attributable checkpoint after a valid visible update.
- Keep capability grants, destructive product actions, sharing, and recovery behind explicit confirmation without imposing that ceremony on ordinary presentation changes.
- Keep internal App/Shaper/Guardian privilege separation out of the default user workflow.
Acceptance criteria
- A clean workspace can produce and run a functional web interface from one ordinary prompt.
- Ten consecutive UI-edit prompts can be completed without a Keep, Reject, Reset, branch, or mode-selection step.
- The canvas remains interactive and the composer remains usable while the agent works.
- A valid change becomes visible automatically; an invalid change leaves the last successful canvas intact and gives the agent actionable diagnostics.
- One-click Undo returns to the previous visible checkpoint without entering safe mode.
- Production Chromium and packaged macOS prove the complete prompt → edit → live result → continue → undo loop.
- User-visible copy describes the work and recovery action, not proposal-state-machine terminology.
Non-goals
- Automatically approving new external capabilities or destructive actions.
- Removing deterministic validation, last-known-good recovery, or attribution.
- Exposing Git internals as the normal editing workflow.
Outcome
Make the running interface the canvas described by VISION.md: a person asks the agent for a change, sees that change in the working UI as soon as it validates, and continues editing or using it without a mandatory Keep/Reject decision.
Product principles
Required behavior
Acceptance criteria
Non-goals