agenter-message

name: agenter-message description: Read, query, send, edit, and recall durable room messages. Use this when room context, room history, or room-visible coordination may affect the next step.

Use this skill when you need to read or query room history, send, edit, or recall a durable room message, or decide whether room-visible coordination is part of the next step.

Quick start:

If the latest room message already contains the full task and no other room is involved yet, do not reread the room first.
If the task already shows the exact room chatId, use that literal chatId directly for message send, message edit, message recall, or message read.
Run message list once only when you truly need a chatId or need to discover another visible room.
Run message read with JSON stdin only when room history may actually change the decision.
Run message query when you need cross-message lookup, fielded search, or guarded SQL over the rooms you already hold.
For temporary cross-room search, prefer chatId:"*" so the runtime resolves only the rooms you already have grants for.
Decide whether a durable room message is useful now, and if so whether it should be an acknowledgement, a narrow follow-up question, a correction, or a final reply.
Send one durable message with the correct room scope and protocol. If you are replying to a specific earlier room message, include ref. If this acknowledgement or follow-up question may need a later silence check, you may add one-shot followUpAfterMs. Through root_bash, prefer command=message send plus JSON stdin; only use argv JSON for trivial single-line payloads. If message send --help marks compact as Suggested or Available, message send --compact is also available for positional payloads.
message send returns recentMessages. Inspect them immediately as a lightweight post-send sanity check.
If two of your recent messages look similar, run message read before changing history so you can also see referencedItems and the nearby reply context.
Similar wording alone is not enough to justify recall. Judge the room's actual conversation context first.
If the repetition was intentional in the room's protocol or request, keep both messages.
If your prior room message is invalid or stale and you already know its messageId, prefer message edit to correct that durable reply in place.
If the earlier durable reply should disappear rather than remain visible with corrected wording, use message recall on that messageId before sending a replacement.
If you cannot safely edit or recall the earlier durable reply, send a corrected replacement.
If that reply also resolves the related attention work, you can switch to attention and settle the same context.

Key laws:

The origin room owns the user-visible conversation.
For message read, message query, message send, message edit, and message recall through root_bash, default to JSON stdin; only use argv JSON when the payload is trivially small.
If message read --help, message query --help, message send --help, message edit --help, or message recall --help marks compact as Suggested or Available, --compact is an optional positional mode. If the positional array becomes unclear, switch back to standard object JSON immediately.
Use ref when the new room message is replying to a specific earlier durable message.
followUpAfterMs is an optional one-shot reminder for yourself, not shared room truth. Use it when you want the runtime to create later attention if this message is still the latest visible room fact after a delay.
followUpAfterMs never auto-sends another room reply. It only asks you to re-decide later whether another visible room message is needed.
message send returns recentMessages; treat them as a quick self-check, not a replacement for a full room read.
message read returns the visible items plus one-hop referencedItems, so you can inspect nearby reply context before deciding whether to revise room history.
If the task already gives the exact room chatId, that literal room id is enough to send the room-visible update; do not rediscover the same room through message list.
message query is message-history lookup, not a generic workspace search tool.
message query only sees the room scope you already hold. chatId:"*" means "all rooms currently granted to me", not "all rooms in the system".
Use mode:"match" for literal text lookup, mode:"query" for structured message filters, and mode:"sql" only when the analysis genuinely needs aggregation or projection.
mode:"sql" is read-only and runs on a pre-authorized temporary projection. If the runtime rejects your SQL, tighten the query instead of trying to mutate the dataset.
If you already have the prior durable reply's messageId, message edit is the cleanest correction path for your own earlier message.
If the earlier message should no longer remain visible in the room, message recall is the cleanest withdrawal path for your own earlier message.
If you do not have the messageId, or the earlier message belongs to someone else, send a corrected follow-up message instead of guessing.
Prefer message edit when the same room fact should stay visible with corrected content; prefer message recall when the prior room fact should be withdrawn before you post a replacement.
Similar wording alone is not enough to justify recall; decide from the actual room context and whether the repetition was intentional in the room's protocol or request.
If the current room message already fixes the task, URL, or reply token, you can usually move on to the next tool or room action instead of rereading the room.
If the room context already exposes visibleRooms, those rooms are real durable channels you can use now; a participant missing from the origin room is not automatically unavailable.
If the user asks you to ask or relay to another participant and a matching visible room already exists, relay there instead of telling the user that the participant is not here.
In relay flows, "I'm asking them now" is progress, not the delivered answer.
Room messages are durable shared truth.
Tool success alone does not change room-visible truth. If the room still needs a durable message, send that message explicitly.
If the room will wait through multi-step work, it often helps to send one short acknowledgement before the deeper tool work starts.
A good acknowledgement is brief: usually one or two short sentences, not a full requirement recap or numbered execution plan.
Keep room-visible updates concise and timely without narrating every internal step.
If another participant or room still owes the missing fact, keep the related attention context open; do not close it with attention commit ... done=true yet.
After a relay target replies, a common next step is to deliver that answer back into the origin room before settling the attention.
If the latest room-visible fact is only teammate progress and you do not yet have a new fact to deliver, keep doing the remaining private work instead of mirroring that progress back into the room.
Private reminders do not replace a durable room message; they usually mean there is still unresolved shared-room work.
Follow the room's requested protocol exactly.
If the room already fixed a URL, path, or other reply token, preserve that fact exactly instead of silently normalizing it.
Terminal success alone is not the room-visible result. Once the exact promised URL or path has been freshly verified, send the room message before you treat the chat work as done.
After the room-visible result is in place, a common cleanup step is attention list followed by attention commit ... done=true for the same context.
Later feedback can reopen the same context; keep it open until the updated room-visible result has actually been sent.

References:

references/room-protocols.md: room ownership, protocol prefixes, and correction behavior
references/chat-attention-items.md: acknowledgement timing, chat etiquette, and settling chat-related attention items