feat(email): voice & writing-style profile learned from past mail (mig 94)

[vjvarada]

What

AI Settings → Advanced Settings gains a Voice profile card (Your voice section): the assistant learns a writing-style profile — plus knowledge-base suggestions — from emails the user has already written, over a date range and source folders the user picks.

Also includes (second commit): mobile layout fixes for the email compose/reply UI, whose buttons could hide behind the bottom nav bar — see the last section.

UX flow

1. Builder — choose sources (Sent mail on by default; Drafts opt-in, with a caution that the folder can contain assistant-written drafts), pick a date range (Last 3/6/12 months, All time, or custom from→to dates), and see a live, debounced count of how many emails the build would study (capped at the 150 most recent). Optional toggle: also propose knowledge entries.
2. Building — the build runs as a background job; the dialog polls live progress (collecting → analyzing batches → synthesizing → knowledge) with a progress bar, and can be closed mid-build.
3. Overview — the finished profile: trait chips (tone, formality, typical length, emoji use), greeting/sign-off/recurring-phrase panels, an editable narrative style guide (the text the drafter actually follows), an enable/disable toggle, "Try it" (generate a sample email in the learned voice for preset or custom scenarios), a suggested-knowledge review list (approve/dismiss), plus Rebuild and Remove.

The Knowledge base manager also shows Suggested / Learned provenance badges with inline approve; suggested entries sort first and never feed drafts until approved.

How

DB (infra/postgres/94_email_voice_profile.sql, idempotent)

• email_voice_profiles — one row per account: status (EMPTY/BUILDING/READY/FAILED), enabled, traits JSONB, style_guide, sources, range_start/end, analyzed_count, last_error, built_at.
• email_knowledge gains source ('manual' | 'voice_profile') and status ('active' | 'suggested'); existing rows backfill as manual+active.

Backend (gateway/routes/email/automation/voice_profile.py)

• Build job: fetch messages in range → strip quoted chains (split_quoted_text) so only the owner's own prose is studied → per-batch observation extraction on tier-balanced → one synthesis call on tier-powerful producing whitelisted traits + style guide → knowledge candidates stored as suggested (never clobbering existing titles). Progress via the shared JobTracker; the profile row's status survives restarts, and a BUILDING row with no live job reads as FAILED instead of stuck.
• Endpoints: GET/PUT/DELETE /email/voice-profile, GET /voice-profile/preview (per-source counts), POST /voice-profile/build, GET /voice-profile/status, POST /voice-profile/sample, POST /knowledge/{id}/approve.
• Prompt contract: _load_assistant_about injects <voice_profile> between the explicit <writing_style> (outranks it) and the auto-derived <learned_writing_style>; both drafting system prompts (_llm_draft_reply, _llm_compose_assist) updated. The KB block now only includes status='active' entries. JSONB traits are parsed defensively (asyncpg may return str).
• voice-profile/sample added to the Next proxy's AI_SLOW_POST_PATHS (120s timeout).

Mobile fix: compose/reply UI vs bottom nav bar (commit 2)

Reported from phone use: while drafting, sending, and replying, buttons/sections sat behind or past the mobile bottom menu bar. Three layout causes, all fixed:

• ComposePanel had no height cap and no internal scrolling — on a phone (especially with the keyboard open) the card ran past the screen, taking Discard/Send with it. Now a flex column capped at calc(100dvh-6.5rem) with a scrollable middle and pinned header/footer. Verified in a Playwright harness at 390×500 (keyboard-open height): old structure puts Send at y=527 (off-screen), new keeps it at y=412.
• App shell reserved pb-14 but the nav also pads itself with the safe-area inset — on devices with a home indicator the last inset-height of page content (the inline reply composer's Send row) hid behind the bar. New .pb-nav utility reserves calc(3.5rem + env(safe-area-inset-bottom)).
• Modals capped at 85/88vh — mobile 100vh includes browser chrome, so a "fits in 85vh" card could overrun the visible viewport and clip its action row. The shared automation Modal, TaskCaptureModal (the commitment popup after sending a reply), EmailPreviewModal, and MessageTimelineModal now cap against 100dvh.
• The undo-send and captured-to-Tasks toasts are lifted above the bottom bar on mobile so "Undo" is tappable during the send window.

Desktop geometry is unchanged (sm: caps preserve the previous values).

Tests

• New tests/unit/test_email_voice_profile.py (13 tests): quote-strip sample prep, observation merge/dedupe, trait whitelisting, <voice_profile> block rendering (incl. str-JSONB), stale-BUILDING → FAILED, and _load_assistant_about ordering (<writing_style> → <voice_profile> → <learned_writing_style>).
• Updated test_email_knowledge.py fakes for the extra profile query.
• Full unit suite: 1620 passed (the 5 failing integration tests fail on main too — they need live Postgres/services). Frontend: tsc --noEmit clean; ESLint clean on all touched files; next build passes and the generated CSS was inspected for the new dvh/safe-area rules.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QYXnVMP3RLBNSwQd5taGZg