resume

Search briefs, erp tasks, erp notes and the wiki for prior context on a topic in one `search_all` call, then pick up a thread or start fresh. Trigger on:- resume, continue prior work, what do we know

SKILL.md

GitHub

Orient a session against everything already known about a topic, before doing work that may re-derive it.

When to use this

  • Starting substantive work on a system that has history — invoked explicitly, since the sweep costs a round-trip.
  • Continuing from a handoff brief (the original use; still the default with no argument) — no sweep, just the most recent brief.
  • Skip for a self-contained one-off, or a topic that plainly has no prior work.

Procedure

  1. Derive terms — from the argument, or from the user's opening request, take 2–3 search terms: the system or host name, the component, the problem word. search_all matching is typo- and word-order-tolerant but not synonym-aware, so include the obvious synonym — dnsmasq and DHCP are still two terms, not one (synonym expansion is deferred, erp task #167).

    No-argument path — with no topic, skip the sweep entirely: get_brief the most recent brief, confirm the pick-up in one line (step 6), and continue from its next step. Steps 2–5 are the topic-search path only.

  2. Sweep — one search_all(query, limit=15) call. It spans briefs, erp tasks, erp notes and the wiki, and returns ranked shortlist rows (never full bodies); each row's ref is the literal follow-up call to read that item. Pass get_upcoming alongside only when the topic is home-infra and you need due-date context the search rows lack.

  3. Dedupe — the same fact often sits in two stores (a task linked to a wiki doc). Collapse rows sharing a thread into one line. Ranking is server-side.

  4. Present a shortlist — at most 5–7 rows, numbered: store, title, date, one line of gist. Then stop and ask: pick a row, or start fresh.

  5. Branch:

    • Pick — read the item named by the row's ref (get_task / get_note / get_brief / get_doc / get_entity). Resolve the references it names (paths, URLs, commits) before acting. Continue from its stated next step; don't redo decisions it records. For a brief, follow its "suggested skills" section.
    • Nothing found, or fresh — say so in one line and start the work clean. An empty sweep is a valid outcome, not a dead end.
  6. Confirm before proceeding — one line on what you're picking up and the next action, so the user can redirect if it's the wrong thread.

Constraints

  • Read-only. The one exception: a brief that was picked up and finished is complete_briefd, as before.
  • Never act on a shortlist row without the user picking it — present, then wait.
  • Don't widen a found thread's scope. It says what it says; new work is new work.
  • No fabricated gist. If a row's content wasn't read, the gist is its title.

Done criteria

  • search_all swept and deduped by thread (topic mode), or the most recent brief taken directly (no-argument mode).
  • A shortlist presented and a thread picked, or "nothing relevant / starting fresh" stated explicitly.
  • The next action confirmed in one line before work begins.