# Feedback Lab: upcoming triage feature

This Artifact is a static browser preview for exercise 3. Reports and results are synthetic. It has no API calls, authentication, credentials, real AI execution, or persistence. Acceptance and dismissal affect only this local demo. The report status remains unchanged.

## State guide

Use the labelled report selector and state controls, or append one of these hash URLs to the Artifact URL:

- `#report=login-button&state=idle`: no analysis yet.
- `#report=login-button&state=queued`: queued with duplicate submission disabled.
- `#report=login-button&state=running`: analysis in progress.
- `#report=login-button&state=completed`: a valid suggestion with accept/dismiss actions.
- `#report=login-button&state=error`: request failure with retry.
- `#report=login-button&state=invalid`: invalid schema output rejected; no acceptance action.
- `#report=login-button&state=accepted`: explicit human acceptance.
- `#report=login-button&state=dismissed`: explicit human dismissal.

Other report IDs: `csv-export`, `status-question`, `prompt-injection`. Report text, including malicious instructions, is rendered as text and is not executed. The preview does not prove a real agent's resistance to prompt injection.

Analyze feedback transitions through queued/running to a suggestion. Retry repeats this simulation. Jump controls stay deterministic until another action is taken. Reset restores the first report and initial state. Hash links work on direct open and refresh without SPA history fallback.

Revision 2 preserves the selected report/state when the keyboard skip link is activated. The scenario controls expose `aria-pressed` and a visible active state to identify the current preview scenario.

## Review request

With browser tools, inspect desktop and phone layouts. Complete analyze -> inspect -> accept/dismiss, then check failure, retry, and invalid output. Check keyboard access, labelled controls, visible failure messages and whether human review is clear. Report actual observations separately from assumptions. If you lack browser access, inspect screenshots and these notes and state that limitation.

## Handoff to exercise 4

Implement this panel on the existing feedback detail view, and adapt it to mobile. Replace fixtures with owned feedback records; invoke the separate cloud Agents service from authenticated Lite handlers. Persist attempts and review decisions; return promptly and poll for completion. Prevent duplicate runs with a persisted request key, validate structured output, scope every read/update to the signed-in user, and handle failures without fabricating success. Keep API keys server-side. The agent must not autonomously change report status.

The preview's classification, priority, summary and nextSteps shape matches exercise 4's planned structured response. Backend integration, real events, costs and agent behaviour are outside this static preview.