Report vulnerabilities privately through the GitHub account that owns this repository. Include the affected route or module, reproduction steps, impact, and any tested mitigation. Do not open a public issue containing credentials, personal data, or an active exploit.

Only the latest commit on main is supported.

• Audience input, retrieved pages, model output, and source URLs are untrusted data.
• The deployment OpenAI key, D1 access, identity headers, budget configuration, and provider response details remain server-side.
• Learner-supplied OpenAI keys are session-only: they remain in the current tab, travel only on provider-backed same-origin requests, are used in memory, and are never persisted or included in diagnostics.
• API routes enforce identity, ownership, input validation, and idempotency before mutation.
• Citations are accepted only when they map to normalized provider-returned sources.
• Failed or incomplete research operations do not commit a ready turn.
• Logs and diagnostics exclude prompt bodies, answer bodies, credentials, and raw provider payloads.