# AI Intent Detection

## Objective

Work out what a message is asking for so the app can route it to the right place.

Classification of incoming free text against a fixed catalogue of supported intents, with an explicit unknown outcome.

## 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. Define the intent catalogue as data owned by the app, not as prose inside a prompt. Every result must resolve to an entry that already exists, and anything else becomes unknown.
2. Detect intent from the actionable part of the message only. Strip quoted replies, forwarded threads, signatures, and prior messages, because the intent of a two-month-old quoted paragraph is not the intent of the message in front of you.
3. Return a ranked shortlist with confidence rather than a single verdict, and treat everything below a configured threshold as ambiguous so the interface can ask instead of guessing.
4. Where intent triggers an action, show the detected intent and let the user confirm or correct it. Corrections must be stored, because they are the only honest measure of whether the classifier is working.
5. Sentiment scoring on the same text belongs to AI Sentiment Analysis; this feature owns what the user wants and that one owns how they feel about it. Share one text extraction and redaction step between them rather than sending the same message to the provider twice.

## 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

- Results must be constrained to the closed catalogue. An invented intent, however plausible, is not routable and must be collapsed to unknown rather than created on the fly.
- Ambiguous and multi-purpose messages are common. Unknown must be a real outcome with its own path through the interface, not a hidden state that silently picks the first option.
- Quoted, forwarded, and historical text will otherwise dominate the classification. Separate the new content before analysis and record which portion was used.
- Text in a message may itself contain instructions aimed at the model. Treat all incoming content as untrusted data, never as configuration, and reject any result that names an intent outside the catalogue as evidence of exactly this.
- Anything with a consequence — cancelling, refunding, deleting, escalating, notifying a customer — needs an explicit confirmation step showing what was detected and what will happen.
- When the provider is unavailable, messages must still arrive and land in a default queue for manual triage. Routing may degrade; delivery may not.
- Truncated or malformed output must be rejected outright. A partially returned intent name that happens to prefix-match a real one is a wrong answer, not a near miss.
- Changing the catalogue does not reclassify history. Store the catalogue version with each result so old routing decisions remain explicable.

## 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

- [ ] Every detected intent maps to an entry in the stored catalogue, with the catalogue version recorded.
- [ ] Unknown and ambiguous outcomes are supported end to end and visible in the interface.
- [ ] Quoted and historical text is excluded from the analysed content.
- [ ] Consequential actions require explicit user confirmation of the detected intent.
- [ ] User corrections are captured and reportable against detected intents.
- [ ] A provider outage routes messages to manual triage without dropping them.
- [ ] 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.