weekly-sweep

Lists agents that have not run within their expected cadence window and renders the results as an OpenUI Lang Card for quick triage.

Steps

1. Load agent registry from state/agents/*.json - collect id, name, lastRun, and cadence_h for each.
2. For each agent, compute elapsed = now - lastRun in hours. Mark overdue if elapsed > cadence_h.
3. Sort overdue agents by elapsed descending (most overdue first).
4. Emit an OpenUI Lang Card:

• Title: "Overdue Agents"
• For each overdue agent: ListItem with name, elapsed time in human-readable form (e.g. "3 days 4 hours overdue"), how long it has been overdue (i.e. elapsed - cadence_h in hours/days), and a Tag showing status (overdue).
• Secondary line per item: "Overdue by: <elapsed - cadence_h formatted>" so the reader sees both total elapsed AND how far past cadence the agent is.
• Footer: total overdue count out of total agent count.

1. Write eval row to state/log/evals.ndjson with shape=lang-card, score based on whether the Lang card rendered without error.

Eval

Actor: the Steps above, executed by an agent. Auditor: shape validator - verifies Lang card renders (no unknown-component errors), list has at least one item.

Score convention:

• 1.0 - Lang card rendered, overdue list non-empty, eval row written.
• 0.5 - Card rendered but overdue list empty (no agents found or all on schedule).
• 0.0 - Render error or eval row missing.

weekly-sweep - loader

Per-turn rules for the weekly-sweep skill. Full reference: state/skills/weekly-sweep/SKILL.md.

Critical Rules

1. Data source is state/agents/*.json - each file is a JSON agent config with at minimum id, name, and optionally lastRun (ISO timestamp) and cadence_h (number). Missing cadence_h defaults to 168 (one week). Missing lastRun = treat as never run = always overdue.
2. Emit a Lang Card, not plain text. The output MUST be a [[TOOL:Lang]]...[[/TOOL]] block containing OpenUI primitives: Card, CardHeader, ListBlock, ListItem, Tag. No HTML, no markdown table.
3. Write an eval row. Before exiting, append a row to state/log/evals.ndjson:

   echo '{"ts":"<ISO>","slug":"weekly-sweep","score":<0|0.5|1.0>,"shape":"lang-card","note":"<brief>"}' >> state/log/evals.ndjson

1. Sort by most overdue first. elapsed = now - lastRun in hours. Sort descending.
2. Human-readable elapsed. Show days + hours, not raw hours (e.g. "3 days 4 hrs", not "76 hours").

Commands

Lang Card Template

[[TOOL:Lang]]
root = Card(
  CardHeader("Overdue Agents"),
  ListBlock([
    ListItem("Agent Name — 3 days 4 hrs", Tag("overdue", color="red")),
    ...
  ]),
  Text("<N> of <total> agents overdue")
)
[[/TOOL]]

Self-Test

1. [ ] Load agent registry from state/agents/*.json?
2. [ ] Compute elapsed and mark overdue where elapsed > cadence_h?
3. [ ] Sort descending by elapsed?
4. [ ] Emit [[TOOL:Lang]] block with Card + ListBlock + Tags?
5. [ ] Write eval row to state/log/evals.ndjson?

<!-- kernel-ok: historical attribution for the PID-loader pattern. -->

Self-correcting loader (PID feedback)

Before you finish, do two things:

1. Fix gaps (P - proportional). MANDATORY. If this loader didn't cover your case, edit it inline before logging.

2. Log the result.

echo "[$(date -u +%FT%TZ)] weekly-sweep: <what was missing or fixed> [FIXED|LOGGED] action_kind=<kind>" >> state/log/loader-feedback.log