cross-channel-messaging

name	cross-channel-messaging
description	Use when deciding how to deliver an AI result: reply now, put durable news, reports, recommendations, or articles in the NextClaw inbox, or send to an explicitly requested chat channel without guessing a route.
description_zh	当需要决定 AI 结果如何交付时使用：当前会话直接回复、把新闻简报/报告/推荐/文章投递到 NextClaw 收件箱，或发送到用户明确指定的聊天渠道且不猜测路由。

Result Delivery And Cross-Channel Messaging

This is a NextClaw built-in skill for the AI itself.

It teaches the AI how to choose among NextClaw's existing delivery primitives in a predictable way.

Use this skill when the user wants any of these:

• send a result to another conversation,
• deliver collected news, a briefing, report, recommendation, or article for later reading,
• notify them through Weixin, Lark, Telegram, Signal, or another channel,
• proactively send a message after a task finishes,
• forward or relay content across sessions or channels.

Strong triggers include user wording such as:

• "notify me when done",
• "collect the news and send it to me",
• "put this report somewhere I can read later",
• "send this to my Weixin",
• "message me on Telegram",
• "forward the result to another chat",
• "after you finish, send me a note elsewhere".

Do not wait for the user to name this skill. If the task is about delivery, routing, relaying, notifying, or proactive messaging, load this skill.

Core Rule

Pick the existing primitive that matches both the content shape and the destination:

1. Reply in the current conversation: Just reply normally when the user means the current session.
2. Deliver durable reading material: Use deliver_to_inbox for news digests, briefings, reports, recommendations, or articles meant to remain available for later reading and follow-up, unless the user explicitly names an external channel.
3. Send to another conversation or channel: Use message when the user wants a direct outbound send to a specific channel/chat/account route. If an existing session may already hold the route, use sessions_list narrowly to recover that route first, then call message.

"Send it to me" or "notify me" alone is not an external route. Do not infer Weixin, Lark, Telegram, or another channel from saved configuration, a previous route, or the mere availability of message.

Tool Choice

Normal reply

Use a normal assistant reply when:

• the user only means "tell me here",
• the target is clearly the current session,
• no cross-session or cross-channel delivery is needed.

deliver_to_inbox

Use deliver_to_inbox when:

• the result is durable reading material rather than a short conversational answer,
• common examples include collected news, daily/weekly briefings, research reports, recommendations, and generated articles,
• the user wants it sent, delivered, saved, or available to read later,
• and no external channel or other conversation was explicitly requested.

Give the delivery a concise title and useful summary. Use Markdown for normal documents and static HTML when the content benefits from a designed report surface. Do not also send the full result through message unless the user explicitly asks for both destinations.

message

Use message when:

• the user explicitly names another conversation or channel destination,
• the target is described as a channel route,
• you need to send to a specific channel + to/chatId,
• or the user asks for a channel action that belongs to the message tool.

For action=send, provide:

• message or content,
• to or chatId,
• channel when the destination channel is not the current one,
• accountId when the channel is multi-account and the account is known.

If you explicitly set channel to a different channel than the current session, to/chatId is mandatory. Do not rely on current-session fallback for cross-channel delivery.

If message is used to deliver the user-visible result for the turn, do not also send a normal assistant reply.

Route Resolution Order

Resolve the target in this order:

1. Explicit user input: sessionKey, label, channel, to/chatId, accountId, or a clearly named destination.
2. Current session route: Only when the user clearly means "here", "this chat", or the current conversation.
3. Existing known session: Use sessions_list to recover the saved route if the intended target already exists as a routable session.
4. Channel discovery command: If the exact channel id, default account, or bound self user is not already explicit, run nextclaw channels list --json and treat its output as authoritative channel state.
5. Authoritative context already exposed to the AI: existing session metadata, project docs, or a known local config file path/content.
6. Ask a narrow follow-up question.

Never guess channel, chatId, sessionKey, or accountId.

Current-session fallback only applies when the current session is already the intended delivery conversation. It does not authorize cross-channel sends such as channel=feishu from a UI/CLI/Weixin session without an explicit Feishu target.

Config And Local Files

Local config can be a useful source of truth, but only when it is already available to the AI in an explicit and auditable way.

Use config or local files when:

• the relevant path is already known,
• the current environment clearly exposes that file,
• and the file is meant to describe saved channel/account routing information.

Do not assume:

• that a hidden config file exists,
• that you know its path without evidence,
• or that a saved account in config automatically proves the destination is reachable right now.

Treat config as route data, not as delivery proof.

Authoritative Route Sources

When this skill is active, prefer route sources in this order:

1. A route the user explicitly typed.
2. Channel ids and account facts returned by nextclaw channels list --json.
3. A saved session route that the environment already surfaced.
4. A local config file only when its path/content is already available in the current environment.

Do not translate natural language channel names into guessed ids. If the exact id is not already known, discover it with the command first.

Multi-Account Channels

For channels that may have multiple logged-in accounts:

• if one account is explicitly specified, pass accountId,
• if the environment clearly exposes a default account, that may be enough,
• if there are multiple plausible accounts and no clear default, ask.

Do not silently choose an arbitrary account.

Bound Self Route

When the user asks to send to "me", "my Weixin", "my Telegram", or another self-owned channel destination, do not ask for a route id before checking channel discovery.

Use this generic lookup:

1. Run nextclaw channels list --json.
2. Pick the exact channel from channels[].id.
3. Read defaultAccountId.
4. In accounts, find the account whose id equals defaultAccountId.
5. If that account has userId, use it as to for message.
6. If defaultAccountId, the matching account, or userId is missing, ask only for the smallest missing value or tell the user to log in/bind the channel first.

The command output is state, not a sending recipe. The skill owns this workflow; message only receives the final channel, accountId, to/chatId, and content.

Weixin As An Example

Weixin is only one example of a channel handled through the same general rule set.

When the user asks for Weixin delivery:

• prefer an existing Weixin session if one already exists,
• otherwise use message with explicit route data,
• run nextclaw channels list --json if weixin, the default account, or the bound user route is not already known,
• use channel: "weixin" as the concrete channel id when the command or existing context exposes it; do not write channel: "wechat" even when the user says WeChat/微信,
• include accountId from defaultAccountId when the command exposes it,
• for "send to my Weixin", use accounts[].userId from the default account as to,
• do not claim that proactive delivery is guaranteed visible unless the environment already proves that.

Weixin Route Lookup Checklist

Before asking the user for a Weixin user_id, check whether nextclaw channels list --json already exposes:

• id: "weixin",
• defaultAccountId,
• accounts[] containing that account id,
• and accounts[].userId.

Rules:

• If the default account exposes userId, do not ask again for user_id.
• If only accountId is known but to/user_id is still unknown, ask only for the missing Weixin user_id or ask the user to log in/bind Weixin first.
• If multiple Weixin routes are exposed, ask the user which one to use instead of guessing.
• If you need to check whether a matching Weixin session already exists, prefer sessions_list({ channel: "weixin", to, accountId }) over a broad unfiltered session listing.

Weixin User ID Safety

Weixin [e-mail] is channel-specific route data.

• Do not call Feishu/Lark/Telegram contact lookup tools to guess a Weixin user id.
• Do not assume a user id from another channel can be reused on Weixin.
• Do not ask for a generic "user info" field when the only missing value is the Weixin user_id.
• Ask only for the exact missing field, for example: Please provide the Weixin user id in the form <[e-mail]>.

Do not turn this skill into a Weixin-only skill.

Feishu As Another Example

When the user asks for Feishu/Lark delivery:

• omitting target is only valid if the current session itself is already a Feishu conversation,
• for proactive sends from UI, CLI, or another channel, resolve an explicit Feishu route first,
• preferred explicit targets are user:<open_id> for direct messages and chat:<chat_id> for group chats,
• if a saved Feishu session already exists, prefer reusing that session route over guessing.

Feishu Route Lookup Checklist

Before asking the user again, check whether the environment already exposed any of these:

• a current Feishu session,
• an existing Feishu session from sessions_list,
• saved session metadata such as last_channel=feishu and last_to=ou_...,
• a known default Feishu account when multi-account routing matters.

Rules:

• If the current session is Feishu, replying there may omit target.
• If the current session is not Feishu, do not call message(channel=feishu) without to/chatId.
• If an existing Feishu session already gives you the route, recover that route and send with message.
• If only the Feishu destination is missing, ask only for the missing open_id or chat_id.

Failure Handling

If delivery fails because route information is missing or ambiguous:

• surface the missing field clearly,
• ask only for the smallest missing piece,
• do not silently fall back to another channel,
• do not silently send back into the current session instead.

If the user asks only to "notify me when done" with no external destination, finish in the current session; ask for route data only when an external channel is clearly requested but cannot be resolved. If message returns unknown channel, run nextclaw channels list --json, pick only an exact returned id, and retry only when the target route is otherwise unambiguous.

Common Failure Patterns To Avoid

• Seeing "Weixin" and then trying unrelated user lookup tools from another channel.
• Writing channel: "wechat" instead of discovering and using the exact runtime id.
• Skipping nextclaw channels list --json when the exact channel id is unknown.
• Asking for both accountId and user_id when only one of them is actually missing.
• Treating "send it to me" as permission to pick Weixin or another configured channel.
• Returning a durable news digest or report only as a transient chat message when deliver_to_inbox is available and no external channel was requested.
• Falling back to a normal reply in the current chat when the user explicitly requested proactive delivery.
• Claiming proactive send is impossible before checking channel discovery output, current session metadata, and known routes.

Success Criteria

This skill is working correctly when:

• the AI chooses correctly among a normal reply, deliver_to_inbox, and message,
• durable material without an explicit external destination lands in the NextClaw inbox,
• it reuses existing session routes when possible,
• it reads route/account information from explicit context when available,
• it asks for missing route data instead of guessing,
• and it keeps the solution lightweight by using existing NextClaw primitives only.