feat(wechat): smart-reply panel — AI drafts, user approves on phone

[chauncygu]

Inbound messages from whitelisted contacts no longer auto-reply via the agent. Instead the auxiliary cheap model drafts 3 candidate replies and pushes a panel to the user's filehelper (文件传输助手) chat:

💬 [AA] Alice (大学同学) → 「周末有空吗」

[1] 有的，周六下午行
[2] 周末出差，下周可以吗
[3] 在忙，晚点回你

回 1/2/3 发送 · 直接打字自定义 · x 跳过 · q 看队列

The user replies on their phone with 1/2/3/x/freeform/queue commands; the bridge sends the chosen text to the original sender. Off by default; opt-in via wechat_smart_reply: true plus a contact whitelist.

Captures most of the "AI replies on my behalf" value while keeping the user as the last-button-pressed (so impersonation, regulatory, and liability concerns from the original full-auto idea all evaporate).

Five composable subsystems:

1. SQLite persistence — ~/.cheetahclaws/wx_smart_reply.db with three additive tables (wx_panels, wx_reply_history, wx_id_counter). Bridge restart no longer drops in-flight panels; reply history survives across daemon recycles so style mimicking has data to draw from immediately. Init failure auto-falls-back to an identical-API in-memory store; nothing crashes.

2. Multi-panel queue UI — every panel carries a 2-letter ID (AA, AB, … rolling at ZZ). q / queue / 队列 lists pending panels. <ID> <choice> addresses panels explicitly (e.g. AA 2, AB x, AA 我自己写的). Plain 1/2/3/x still defaults to the latest active panel for simple cases.

3. Style mimicking — every confirmed send (candidate or freeform) appends to wx_reply_history. The candidate-generation prompt reads the last 10 entries (excluding the current contact) and tells the auxiliary model to mirror the user's tone and length. Builds over a few days without any explicit user action. Janitor prunes rows older than 30 days.

4. Per-contact relationship context — ~/.cheetahclaws/wx_contacts.json maps wxid_* → {label, relationship, notes}; injected into the generation prompt when present. File-mtime watch reloads on edits without bridge restart. ContactsStore is the loader/saver.

5. Group @-only trigger — wechat_smart_reply_groups_at_only plus wechat_self_nickname makes group messages silent unless the text contains @<nickname> with a word boundary (@李 won't match @李明, etc.). Empty nickname blocks all group messages safely.

Behaviour change for existing WeChat bridge users: NONE — defaults preserve current auto-reply. Enabling requires explicit opt-in.

Wire-up: bridges/wechat.py gets a 45-line patch — initialize store + contacts in the poll loop; before the existing agent dispatch, check is_smart_reply_target(uid, config, text=...) and route through trigger_smart_reply instead. Filehelper inbound messages are intercepted early via handle_filehelper_message; falls through to normal dispatch when there's no active panel (so !jobs etc. from filehelper still work).

Cost: one auxiliary-model call per inbound smart-reply message, capped at 200 tokens, thinking=False. Whitelist gate keeps the billing surface small.

Tests: 75 unit tests covering identity heuristics, gating rules, panel-ID generation, both store backends (SQLite + in-memory), filehelper input parsing, panel + queue formatting, candidate extraction, prompt construction (style + contact), trigger / handle flows, contacts file roundtrip / mtime reload / corruption recovery, and the new config defaults. Full suite 734 passing on this branch.

Files:

• bridges/wechat_smart_reply.py (new, 473 LoC) — entry points,
  parsing, formatting, prompt building, generation
• bridges/wechat_smart_reply_store.py (new, 595 LoC) — SQLite +
  in-memory stores, contacts JSON loader, ID generation
• tests/test_wechat_smart_reply.py (new, 840 LoC) — 75 cases
• bridges/wechat.py (+45) — wire smart-reply
  into the poll loop
• cc_config.py (+13) — six new config keys

What's still out of v1 scope:

• No "send proactively when user is away" (smart-reply only triggers on inbound messages).
• No analytics / metrics on candidate acceptance rate (would need to track "1 vs freeform vs skip" ratios; happy to add if useful).
• No slash command UX for editing contacts (manual JSON edit; could add /wechat contact add/edit/list later).