feat(email): multi-account audit — the chat assistant is now configured per account, not just scoped to one

[vjvarada]

Audit result

Checked the multi-account plumbing end to end. The data and automation layers were already correct; there was one real gap, in the assistant.

Verified per-account (no change needed)

• Schema — every email table is keyed by account_id and cascades from email_accounts: rules + actions, executed rules, assistant settings, knowledge, learned patterns, rule patterns, rule guidance, thread status, AI drafts, senders, newsletters, contacts, embeddings.
• Rules engine — rules are loaded WHERE account_id = :aid and evaluated per account, so each mailbox genuinely has its own rules engine. Same for the cleaner's label projection and the auto-categorize sweep.
• Drafting — already resolves About / signature / models (rule, draft, compose, chat) / voice profile / learned patterns / reply memories by account_id, with a per-account Mem0 scope (email_memory_scope).

The gap

The chat assistant knew which account to act on but nothing about how that mailbox should be handled. The persona carried the account list, the active account_id and the open email — never the account's About, standing instructions, or writing style. So the same assistant conversed identically on a work mailbox and a personal one, and only its tool calls differed. Drafts were already per-account (the server loads settings by account_id), which made the inconsistency easy to miss.

• buildEmailAssistantPersona now takes the active account's settings and emits them as a scoped block, stating they apply to the ACTIVE account only so the model doesn't carry them across a switch. Standing instructions are marked binding; the auto-derived learned style is marked advisory — matching the precedence the drafter already uses.
• Both entry points feed it: the email app (which already fetched these settings for the chat model and discarded the rest) and the chat app, which now also passes the resolved account id.
• /assistant/settings returns learned_writing_style so the chat can match the voice the drafter already writes in.

Left as-is deliberately

The chat agent's remember() namespace stays user-global. That ContextVar doubles as the agent's gateway-auth identity, so an account suffix would break every tool call — the per-account scope lives on the gateway side (email_memory_scope), where drafting uses it. Flagging it as a known, documented trade-off rather than changing it silently.

Testing

• Backend: 952 email tests passed.
• Frontend: new emailAssistantPersona.test.ts (7 cases) — the persona names the active account, carries its configuration, marks instructions binding and learned style advisory, swaps cleanly when the account changes, and degrades to account-awareness when settings haven't loaded or are blank. tsc --noEmit clean, next build clean, vitest 141/141, no new eslint issues.

🤖 Generated with Claude Code

https://claude.ai/code/session_0122zNo8kiZooLu49jsaSrQ6

---

Generated by Claude Code