eric-tramel/moraine/realtime-peek

Realtime Peek

Use this skill when the user asks about current or very recent agent activity, including other harnesses, sibling agents, subagents, or your own active session.

Workflow

1. Choose a recent explicit time window with timezone. For "now" questions, start with the last one to two hours and widen only if needed.
2. Call list_sessions for that window, sorted newest first.
3. Prefer sessions that are still updating, have incomplete final turns, or match the harness, branch, issue, file, or agent named by the user.
4. Open relevant sessions or turns with id-only open first. If the compact

summary is insufficient, request a bounded page with id plus limit and continue with { "cursor": next_cursor } only as far as needed. Open an individual event ID when exact content is required.

1. If the user asks about a specific file, call file_attention and then open the returned event, turn, or session handles.

What To Report

Summarize active work by source or harness, session title or ID, branch or working directory when visible, last update time, visible tool or terminal activity, and the current apparent state. Keep the summary compact and separate observed facts from inference.

If a record is still active, say that it is a live view and may change after the query. If the latest turn is incomplete, avoid treating it as a final decision.

Do not replay sensitive terminal or tool output into the current chat. Summarize what happened, redact secret-looking values such as tokens, keys, cookies, and credentials, and avoid quoting private data unless it is necessary for the user's requested debugging task.

Search Follow-Up

Use search_sessions with keyword queries when the recent session list is too broad or when the user gives a content clue such as an issue number, command, error, file path, or branch name. After finding a candidate, narrow with within_id rather than pulling whole sessions into context.