# AI Agent Actions with Approval

## Objective

Let the assistant draft app actions and run them only after the user approves the exact plan.

A plan-and-approve step between an assistant's proposed actions and their execution, showing every target and change before anything runs.

## Before You Begin

This feature is being added to an application that already exists and already
works. Do not scaffold a new project, and do not assume a blank slate.

Inspect the codebase first and establish:

- The existing application structure and where code of this kind already lives.
- The framework and version in use.
- The existing design system — colours, spacing, typography, and component conventions.
- Existing UI components you can reuse instead of writing new ones.
- The existing database structure, if this feature needs to persist anything.
- The existing authentication and authorization system, if this feature is user-scoped.
- Dependencies already installed, so you don't add a library that duplicates one.
- The existing test setup and conventions.

Only start writing code once you understand the above. If the application
already implements part of this feature, extend it rather than replacing it.

## Implementation Instructions

1. Split the flow in two. Producing a plan must have no side effects at all, and execution must run only from an approved, stored plan rather than from the conversation that produced it.
2. Render the plan in the app's own language: which records are affected, which fields change from what to what, and every external side effect such as an email, a charge, or a webhook. Nothing may execute that was not on the screen the user approved.
3. Re-check permissions and the current state of every target at execution time, not at planning time. A plan drafted two minutes ago may reference a record that has since been edited, deleted, or locked.
4. Invalidate the approval whenever the plan changes. Any edit, re-plan, or change in the underlying records requires a fresh approval covering the new plan, and an approved plan should expire after a short window.
5. The catalogue of actions the model may invoke, their argument schemas, and the call limits belong to AI Tool Calling Framework. Use that catalogue for the actions in a plan rather than defining a second, parallel set of capabilities here.

## UI and UX Requirements

Match the application's existing design system exactly. Reuse its components,
spacing, and typography. This feature should look like it was always there.

## Responsive Requirements

Works on mobile, tablet, and desktop. Touch targets are large enough to hit on a
phone, and nothing overflows horizontally at 320px.

## Accessibility Requirements

- Fully keyboard navigable.
- Correct semantic elements and ARIA roles.
- Visible focus states.
- Meets WCAG AA contrast.
- Dynamic changes are announced to screen readers.
- Respects prefers-reduced-motion.

## Edge Cases

- Every target, every field-level before-and-after, and every outbound side effect must be visible before approval. A plan that summarises itself as updating some records has not been approved in any meaningful sense.
- Permissions and record state are revalidated at execution. If a target moved out of scope between approval and execution, skip that step, mark it clearly, and do not silently substitute another record.
- A plan that changes for any reason loses its approval. Re-approving a modified plan must show what changed since the last approval, not present it as if it were new.
- Steps must be individually idempotent and keyed, so a retry after a partial failure does not send a second email or apply a change twice. On partial failure, report exactly which steps completed, which did not, and what state the data is now in.
- Arguments the model generated but the interface did not display must never be executable. If a value cannot be rendered for inspection, the plan is not approvable.
- Long-running plans need a visible progress state and a way to stop between steps. Stopping mid-plan must leave a clear record of the boundary reached.
- Approval must be attributable to a specific person and recorded in the audit trail alongside the plan as approved, so the plan text cannot be reconstructed differently later.
- If the model is unavailable or returns malformed structured output, the user must still be able to perform the same actions through the ordinary interface. Nothing should be reachable only through the assistant.

## Testing

Exercise the feature end to end in the running application. Cover every edge case
above, then run the existing test suite and confirm nothing regressed.

## Acceptance Criteria

- [ ] Planning produces no side effects, and execution runs only from a stored approved plan.
- [ ] The approval screen shows every target, every field change, and every external side effect.
- [ ] Permissions and target state are revalidated at execution time and out-of-scope steps are skipped and reported.
- [ ] Any change to a plan invalidates the approval and requires re-approval showing the diff.
- [ ] Steps are idempotent and partial failures report exactly what did and did not run.
- [ ] No argument can be executed that was not displayed to the approver.
- [ ] Approvals are attributed to a person and recorded with the plan as approved.
- [ ] The feature matches the existing design system.
- [ ] No existing functionality is broken.

## Adaptation Rules

- Match the existing design system. Do not introduce a new colour palette,
  spacing scale, or component library.
- Reuse existing components and utilities wherever they fit.
- Follow the naming, file layout, and code style already present.
- Do not upgrade, replace, or remove existing dependencies to make this
  feature fit. Adapt the feature to the app, not the app to the feature.
- Do not break existing functionality. If a change is genuinely required in
  existing code, make the smallest one that works and say so.
- If something in these instructions conflicts with how the application is
  built, follow the application and explain the deviation.

## Final Verification

Before you report the work as done:

1. Re-read the acceptance criteria above and check each one against what you
   actually built.
2. Run the application and exercise the feature end to end.
3. Run the existing test suite and confirm you have broken nothing.
4. Check the feature on mobile, tablet, and desktop widths.
5. Check keyboard navigation and focus handling.
6. Summarize what changed: files added, files modified, and anything you
   deliberately did differently because of how this application is built.

If any acceptance criterion is unmet, fix it before reporting completion.