Build a production AI assistant frontend with streaming, citations, approvals, cancellation, error states, and accessibility.
Scenario
A user works with an AI assistant that streams its response, shows sources, and may propose a risky action. The interface must explain the current state, support cancellation, and require explicit confirmation.
Outcome: The UI has explicit idle, loading, streaming, error, and approval states.
Tasks
Checks
Outcome: Responses update reliably and expose their evidence.
Tasks
Checks
Outcome: Risky actions cannot execute without consent.
Tasks
Checks
Outcome: The interface is accessible and can recover from failures.
Tasks
Checks
Assessment rubric
Streaming, error, cancellation, and approval states are explicit.
Insufficient
States are mixed together or race conditions remain.
Competent
Core states are explicit and tested.
Strong
Deterministic replay, concurrency tests, and recovery paths are implemented.
Evidence required
Sources are validated and untrusted content cannot become a command.
Insufficient
An unknown source is displayed as valid.
Competent
Citations are validated and have a fallback.
Strong
Provenance, schema validation, and security tests are implemented.
Evidence required
Risky actions expose context, confirm or deny controls, and an audit trail.
Insufficient
An action can execute without explicit consent.
Competent
Approval blocks execution and shows the consequences.
Strong
Undo, delegated approval, and a durable audit trail are implemented.
Evidence required
Keyboard access, focus, announcements, and retry behavior work reliably.
Insufficient
Focus is lost or streaming updates are not announced.
Competent
The core flow is accessible and recoverable.
Strong
Automated accessibility tests and progressive disclosure are implemented.
Evidence required