ai-email-triage

name: ai-email-triage description: Apply when changing AI summary, draft, or priority behavior. Keeps Anthropic-path output and deterministic-fallback-path output structurally identical so the UI cannot tell them apart. Triggers on edits to src/server/ai/, src/app/api/ai/.

ai-email-triage

Source of truth: specs/ai-features.md.

When to invoke

• A cooperation plan changes summary tone, draft tone, priority scoring, or any AI-emitted field shape.
• The Anthropic call path changes (model, parameters, system prompt).
• The deterministic-fallback path changes.

Invariants

1. Summary returns concise sender-intent + deadline statement.
2. Reply draft is review-only. The API route MUST NOT call any send/reply/forward method on MailProviderAdapter. It returns a draft for the compose sheet.
3. Priority values are exactly urgent, high, normal, low. No silent extension.
4. Output parity: the response shape is identical regardless of whether Anthropic was called or the fallback ran. The UI must not need to branch on which path produced the result.
5. Demo mode: when ANTHROPIC_API_KEY is unset OR DEMO_MODE=true, the deterministic path runs. The app must remain usable without an Anthropic key.
6. Secrets: ANTHROPIC_API_KEY is server-only. Never echo it back, never include it in logs, never expose the model identifier to the client unless the spec says to.

Checklist

• Both paths (real Anthropic + deterministic fallback) updated.
• src/server/ai/priority.test.ts exercises the new behavior.
• No new direct provider adapter calls from src/server/ai/* or src/app/api/ai/*.
• ANTHROPIC_MODEL env handling unchanged (default lives in .env.example).
• Response JSON shape verified identical between paths.

Anti-patterns

• Returning different field names depending on whether the fallback ran.
• Logging the prompt or response body in production routes.
• Routing AI output directly into a send action.
• Treating priority as a free-form string.