# AI Reply Suggestions

## Objective

Offer a few short reply options that fit the current message or thread.

A small set of one-tap reply drafts generated from the visible thread, inserted into the composer for editing before sending.

## 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. Generate suggestions from the visible thread only, and only when the last message is from someone other than the user. Suggesting a reply to the user's own message is noise.
2. Insert the chosen suggestion into the composer as editable text. Do not send on tap, however short the reply, because a one-tap send makes every model error a message the user cannot recall.
3. Require the options to differ in intent, not in wording — for example accept, decline, and ask for more detail. Three paraphrases of the same sentence give the user nothing to choose between.
4. Recompute or clear suggestions when a new message arrives in the thread, so the user is never offered a reply to a message that has already been superseded.
5. Long-form composition from a record belongs to AI Email Drafting; this feature owns short in-thread responses. Reuse that feature's context allow-list rather than defining a second one.

## 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

- Distinct intent is the whole value. If the model returns options that are substantially the same, drop the duplicates and show fewer rather than padding the list.
- A suggestion must not commit the user to a delivery date, a refund, a discount, or an escalation the user has no authority to grant. Keep suggested commitments vague enough to be safe or omit them.
- Tone must follow the thread and the relationship. A breezy suggestion under a complaint, or a formal one in a running internal chat, reads worse than no suggestion at all.
- Internal comments, private notes, and messages from participants the current user cannot see must be excluded from the context, even when they sit in the same thread record.
- Stale suggestions are actively harmful. Clear them when the thread changes, when the composer already has typed text, and when the user switches threads.
- A refusal, a timeout, or malformed output must result in no suggestions and an unchanged composer, never an empty chip row or a placeholder the user can tap.
- Suggestions fire on every thread view, so cost accumulates quietly. Generate lazily on user intent or on thread open at most once, and cache against the thread's last message.
- With the model unavailable the composer must behave exactly as before, with the suggestion row absent rather than showing a persistent error.

## 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

- [ ] Suggestions appear only when the latest message is from another participant.
- [ ] Choosing a suggestion inserts editable text and never sends the message.
- [ ] The offered options differ in intent, and near-duplicates are removed.
- [ ] Suggestions are cleared or regenerated when the thread receives a new message.
- [ ] Context excludes internal notes and any message the current user cannot access.
- [ ] Generation is cached per thread state and capped per account.
- [ ] Model failure leaves the composer fully usable with no suggestion row.
- [ ] 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.