Mail Auto Reply（多通道：Agent Mail + Gmail）

Overview

Drives a complete email auto-reply cycle across two channels:

• Agent Mail — via the agent-mail connector (MCP tools: ListMessages / SearchMessages / GetMessage / ReplyMessage / SendMessage).
• Gmail — via the gog CLI (Google Workspace CLI, binary gog).

For each selected channel it verifies login, reads new/unread mail, replies per a user-supplied persona + policy, and emits a 邮件往来简报.md briefing plus a chat summary. Persona/tone is NEVER hardcoded — collected at runtime so each user applies their own voice.

When To Use

• The user asks to "处理邮件 / 自动回复 / 回复新邮件 / 出邮件简报".
• The user opens the 邮件自动回复管家 expert and invokes this skill.
• The user wants a safe, auditable mailbox pass over Agent Mail and/or Gmail.

Prerequisites

Agent Mail

• agent-mail connector connected; user has logged into the mailbox at https://agent.qq.com.
• If GetMe returns not_bound, guide the user to WorkBuddy 「更多 → 我的邮箱」to enable the mailbox first.

Gmail (requires the gog CLI)

• gog binary installed and on PATH. Install (pick one):
• macOS / Linux: brew install gogcli — or go install github.com/openclaw/gogcli/cmd/gog@latest (needs Go).
• Windows: download gogcli_*_windows_amd64.zip from https://github.com/openclaw/gogcli/releases, extract gog.exe, put it on PATH.
• One-time OAuth (Google requires this; no way around it):
  1. Google Cloud Console → create a "Desktop app" OAuth client → download client_secret.json.
  2. gog auth credentials ~/Downloads/client_secret.json
  3. gog auth add you@gmail.com --services gmail (opens browser for consent).
  4. export GOG_ACCOUNT=you@gmail.com (or pass --account on every call).
• Verify: gog auth list and gog gmail search 'is:unread' --max 1 --json.

Workflow

0. Channel selection (new)

Ask which channel(s) to process if not already stated: Agent Mail / Gmail / Both. Apply the matching branch below for each selected channel.

1. Login pre-check (per channel, mandatory)

• Agent Mail: open https://agent.qq.com, confirm the inbox is visible and logged in. If not, say "请用微信扫码重新登录 Agent Mail" and STOP until the user confirms. (You may also probe via agent-mail GetMe; on not_bound, guide "更多 → 我的邮箱" to enable.)
• Gmail: run gog auth list (or gog gmail search 'is:unread' --max 1 --json). If no authed account or an auth error appears, STOP and tell the user to run gog auth credentials + gog auth add (see Prerequisites). Never proceed without a working Gmail auth.

2. Collect runtime parameters (ask if missing)

Confirm in one go when not already provided:

1. Persona/tone — e.g. "热情专业的拉新顾问", "简洁客服助手". (Never assume a fixed persona.)
2. Reply policy — full auto-reply / FAQ-template only / draft-only.
3. Whitelist — only reply to which senders / subjects (optional).
4. Send mode — send directly / draft for review.

3. Read emails (per channel)

• Agent Mail: pull new/unread via ListMessages / SearchMessages; use GetMessage for full body.
• Gmail: gog gmail search 'is:unread newer_than:7d' --max 50 --json (parse threads/messages; prefer --json + --no-input). Read a thread body with the appropriate get/read subcommand — check gog gmail --help (commonly gog gmail get <messageId> --json).
• Filter out no-reply, subscription pushes, and automated bounce/return receipts on both channels.
• Maintain a "replied IDs" record in a local file, keyed by channel:messageId, to prevent duplicate replies.

4. Reply handling (per channel)

• Agent Mail: ReplyMessage / SendMessage.
• Gmail:
• Check gog gmail --help for a reply subcommand; if present, prefer it for correct threading.
• Otherwise gog gmail send --to <from> --subject "Re: <original subject>" --body "<reply>" (briefly quote the original).
• Draft mode: create a draft via gog gmail drafts create ... (or --draft if supported) and tell the user; do NOT send.
• Direct send: execute the send.
• Draft for review: produce the draft and clearly tell the user; do NOT send.
• Sensitive / undecidable / VIP: escalate to human (draft + notify user).

5. Produce briefing

• Generate 邮件往来简报.md into the user's working directory.
• Include: run time, per-channel mailbox address, received mails (sender / subject / summary / replied?), sent or drafted (recipient / subject / summary), unhandled items with reasons, and per-channel quota usage:
• Agent Mail: 50 sends/mailbox/day.
• Gmail: Google has no fixed daily cap but enforces rate limits — note any send errors.
• Give a 3–5 point chat summary (no long-form dump).

Guardrails (must enforce)

• Dedupe: track channel:messageId locally; never reply twice.
• Loop prevention: never auto-reply to automated systems; filter no-reply / subscription / bounce mail on both channels.
• Quota: Agent Mail 50/day (warn + pause as the limit nears). Gmail: watch for rate-limit errors and back off.
• Escalate: sensitive / undecidable / VIP → draft + notify, never blind-send.
• Client must stay running: if later wired to a scheduled automation, the WorkBuddy client must remain open or it will not trigger.
• No hardcoded persona: every reply derives from the user-supplied tone for that run.
• Gmail auth safety: never attempt to send before gog auth list shows a valid account.

Sharing With Colleagues

• Copy this skill directory into the colleague's ~/.workbuddy/skills/ to install.
• Each person must use their OWN Agent Mail mailbox (max 2 per WeChat account) and their OWN Gmail OAuth — Google does not permit shared mailbox auth. Do not share one mailbox/account (quota + identity collision).
• Optionally bundle inside an expert package's skills/ folder so the expert ships with it.