Inbox triage

Work through the user's unanswered comments, propose replies, and post only the replies the user approves.

Ground rules

• Replies are public and permanent from Artima's side. reply_to_comment posts under the comment on the social network. Artima cannot delete the reply afterwards; only the platform can. Send a reply only when the user has approved that exact text, either as shown or with their own edits. Approving the plan is not approving text you write later.
• Report the real count. In list_comments, count is the number of matching comments. comments is only a sample, 25 by default and at most 100. Lead with count, and when the sample is smaller, say "showing X of Y".
• Role. Anyone in the workspace can read the inbox. Sending replies, and drafting them with Artima's AI, needs the Editor role or higher. Check with list_workspaces before you promise to send anything.
• Credits. draft_comment_reply uses Artima's AI and spends credits (coinsCost). Drafting replies yourself costs nothing, so do that unless the user asks for Artima's drafts.

Workflow

1. Load the inbox

1. Call list_comments. It defaults to unanswered comments. Add platform if the user named one, and limit up to 100 when there are many.
2. If it returns {"error": ...}, report the error and stop.
3. If count is 0, say that the inbox has no unanswered comments. Do not guess at broken connections.
4. If the user wants the latest comments, or the data looks stale, call refresh_inbox (optionally with socialAccountIds). It only starts a background fetch and returns estimatedDurationMinutes. Tell the user, carry on with what you have, and offer to check again once that time has passed. Do not loop waiting for it.

2. Triage

Put every comment in the sample into one bucket:

Recommend leaving spam unanswered. Artima cannot hide or delete comments, so the user has to do that on the platform.

3. Draft the replies

• When the comment depends on what the post said, call get_post_detail with the comment's postId.
• Reply in the commenter's language, briefly and like a person. When the workspace has a default assistant (list_assistants, then get_assistant), follow its voice.
• For complaints: acknowledge, apologise where it fits, and move the conversation to a private channel. Never promise refunds, discounts or facts the user has not given you. Leave placeholders such as [email] instead of inventing contact details.
• For leads, answer only with facts you have and invite them to message the business.
• Use draft_comment_reply (commentText, optional originalPostText, tone of friendly, professional or casual, and language) only if the user wants Artima's AI to draft. Before more than a few, check get_credit_balance.

4. Present for approval

Group the drafts by priority. For each draft show the author, platform, the comment, your proposed reply, and its commentId. List the spam you suggest skipping. Then ask which replies to send. The user can approve all of them, some of them, or edit any of them.

5. Send what was approved

For each approved reply, call reply_to_comment once:

• postId: the Artima post ID from that comment;
• commentId: from list_comments;
• message: the approved text, character for character.

If the user's own message already gives the exact reply and tells you to send it, that counts as approval. Send that text verbatim, without rewording it.

After each call:

• status: "Sent": done.
• An error: report it. Do not resend blindly, because a duplicate reply is public. If the call timed out, tell the user it may have gone through and ask them to check on the platform before you try again. The comment shows as Replied in list_comments only after the next inbox sync.

6. Wrap up

Say what was sent, what was skipped and why, and what is still waiting (count minus what you handled). Remind the user that a sent reply can be removed only on the platform itself.