cross-channel-triage

name	cross-channel-triage
description	Use when the user asks to triage messages, emails, chats, inboxes, unread messages, recent messages, open replies, who they owe replies to, outreach, or reply drafting across one or more channels such as WhatsApp, Telegram, Gmail/email, Slack, Signal, iMessage, browser portals, or files. For broad sweeps, first produce a scoped prioritized triage list; for one known target, use single-thread draft mode. Always require user approval before any external reply/send.
metadata	{"openclaw":{"emoji":"📬","displayName":"Cross-Channel Triage"}}

Cross-Channel Triage

Use this for messy inbox or communication triage across one channel or many: email, Telegram, WhatsApp, Signal, iMessage, Slack, browser portals, or files. The same protocol applies to a single inbox when there are multiple threads to prioritize.

For a one-off outreach or reply-draft request tied to one known person, thread, or support case, use single-thread draft mode instead of a broad inbox sweep.

Core Rule

Do not draft and send everything in one batch by default.

Default flow:

1. Scope the sweep: channel(s), timeframe, unread/recent/all, and any people or topics the user named.
2. Shortlist first: inspect snippets, previews, sender, timestamps, unread state, and labels before reading full bodies or long histories.
3. Filter noise early: newsletters, receipts, calendar churn, automated alerts, low-signal notifications, and routine FYI items should not consume deep-read budget unless the user asks.
4. Deep-read only the likely actionable items, active conversations, or threads where a snippet is not enough to classify urgency or reply-needed state.
5. Return explicit buckets and recommended next actions.
6. Handle items one by one, starting with the highest-impact item.
7. Require approval before any external send, purchase, deletion, archive, label action, calendar change, or other risky action.

Buckets

Use these buckets when they help the user scan:

• Urgent: direct asks with time pressure, blockers, deadlines, or operational risk if ignored.
• Needs reply soon: direct asks, active conversations, or follow-ups that will go stale if ignored.
• Waiting on them: the user already replied or the current blocker belongs to someone else.
• Schedule: the next useful action is a meeting, reminder, follow-up monitor, or calendar task.
• Delegate: the work should become a task for Jarvis, another agent, or a human rather than a simple reply.
• Archive / no action: receipts, newsletters, resolved threads, or low-value notifications that are safe to ignore or clean up after confirmation.
• FYI: useful context with no immediate action.

Treat "needs reply" as an inference, not a guaranteed state. Say "looks like" or "likely" when the channel does not expose a definitive next-responder state.

Triage Output

Keep the first pass short. Prefer the top 3-7 actionable items unless the user asks for everything.

For each item include:

• Contact/thread
• Channel
• Latest message date/time
• Latest sender/author
• Bucket
• What they want or why it matters
• Recommended action
• Needs user input? yes/no
• Source ref/id when available

Always state:

• Scope checked: channels, timeframe, filters, and approximate result count.
• Confidence: high / medium / low, with a short reason.
• Coverage limits: what was excluded or only sampled.

End with the next concrete step:

I recommend starting with #1. Want me to draft that one?

Reply Loop

For each selected item:

1. Treat trackers, memory, docs, and old chat context as stale indexes only. If live access exists, read the latest relevant source, thread, or person before drafting.
2. Fetch enough fresh context to understand the latest ask, participants, tone, and whether the conversation moved forward.
3. Include the exact latest relevant inbound text when available and useful. Summaries are helpful, but they must not replace the sender's wording when wording matters.
4. If live access is unavailable, or the only live path would require costly browser work, state that freshness is not verified. If you provide an optional sketch, label it stale/tracker-based and not ready to send, then ask whether to inspect the live source when freshness matters.
5. Draft in the user's voice.
6. Label every draft as new, already sent, optional, or do-not-send.
7. Ask: send, edit, skip, schedule, delegate, archive, or next?
8. If the user approves sending, refresh the same thread immediately before sending. Stop if newer relevant thread movement, inbound or outbound, changes or duplicates the reply.
9. Update compact status: pending, drafted, sent, skipped, scheduled, delegated, archived, waiting, or blocked.

Single-Thread Draft Mode

Use this mode when the user asks for a draft for one known contact, thread, case, or channel target rather than a broad triage sweep.

1. Identify the target and the channel/source you can inspect.
2. Skip broad inbox discovery unless the target is ambiguous.
3. Apply the Reply Loop freshness rules before drafting.
4. If live access is unavailable, or the only live path would require costly browser work, state that freshness is not verified. Label any optional sketch as stale/tracker-based and not ready to send, then ask whether to inspect the live source when freshness matters.
5. Do not send anything until the user approves a draft and you refresh the same thread immediately before sending.

Modes

Default supervised mode:

• Use for users, friends, investors, hiring, sales, conflict, negotiation, emotional threads, or anything reputational.
• Draft one at a time and wait.

Batch-draft mode:

• Use only if the user asks to review several drafts before action.
• Draft multiple replies, but send none until explicitly approved.

Low-risk batch-send mode:

• Use only if explicitly authorized and the messages are routine: confirmations, receipts, simple scheduling, or "thanks, received."
• Ask before the first batch send unless the user already authorized the batch.

Channel Adapters

Use the most specific available channel skill/tooling:

• WhatsApp-as-me: use wacli.
• Telegram-as-me: use telegram-user.
• Gmail / Google Workspace: use gog.
• Generic IMAP/SMTP email: use himalaya.

For monitors or waiting for replies, prefer deterministic helper scripts from the relevant channel skill. Do not make a recurring monitor vague; define the target, cadence, stop condition, expiry, and whether it may draft only or send.

Tracker

For more than about five actionable items, or when the user will review across turns, keep a compact tracker in the workspace or project directory.

Track only status and refs, not private transcripts:

• item number
• contact/thread
• channel
• latest message date/time
• bucket
• status
• source ref/id
• next action

Fetch fresh source context when acting.

Never treat the tracker itself as current enough for outreach or reply drafting.