| name | loopany-weekly-sweep |
| description | Use on weekly cadence or when the user says "what's slipping" / "weekly check" / "is the workspace healthy". Three problems it sweeps: drift (`loopany doctor`), slippage (overdue `check_at`), parking lots (running tasks ≥14d, active signals ≥7d, pending proposals). Forces closure on stuck artifacts. Distinct from the daily [[daily-followups/SKILL.md]]. |
weekly-sweep — drift + stuck-artifact sweep
Three problems loopany workspaces develop silently:
- Drift — artifacts reference a domain that isn't enabled, kinds
fail validation, references point nowhere.
loopany doctor catches
these.
- Slippage — items whose
check_at slid past today, never
addressed. Daily [[daily-followups/SKILL.md]] stays focused on today; this is
where last week's misses come back into view.
- Parking lots — artifacts sitting in non-terminal status with no
check_at and no recent activity. Running tasks, active signals,
pending skill-proposals. They don't slip (no date to slip past) but
they don't move either — and the daily digest can't see them.
Run once per week. Never daily — nothing stagnates in 24 hours.
When to run this
- Scheduled: once per week (agent platforms)
- On demand: user asks "what's slipping," "is the workspace healthy,"
"weekly check"
Step 1 — Workspace drift (doctor)
loopany doctor --format json
Exit 0 + empty findings → workspace is clean. Say so briefly — the
user wants to know you looked.
Findings present → summarize categories, not every item:
Workspace drift (3 findings):
- 2 artifacts use domain `fundraising` which isn't enabled
- 1 learning references a missing evidence artifact
For each category, propose a fix:
- unknown domain → enable it, or rewrite the artifact's domain
- dangling reference → remove the edge, or mark the referring artifact
- validation failure → re-read the kind and correct frontmatter
Don't auto-fix without the user saying yes. Drift is often the
fingerprint of a decision the user is mid-way through — a new domain
they haven't committed to, a rename that's half-done. Silent cleanup
erases that signal.
Step 2 — Overdue followups
loopany followups --due overdue
Everything whose check_at is past today. Classify each:
A. Still relevant — user forgot. Re-surface with a gentler framing
than the original.
B. Outdated — question no longer applies. Retire via the right
kind's terminal state: task cancelled --reason, learning superseded, etc. Use [[core-artifacts/SKILL.md]] for tasks.
C. Shouldn't have had a check_at. Remove the date with a brief
note. Repeated mis-scheduling is fodder for [[reflect/SKILL.md]].
D. Zombie — keeps coming back. Pushed ≥2 times with no progress,
or in overdue ≥2 consecutive sweeps. The user has seen it and chosen
not to act. Forced choice: commit to a specific next step with a
near-term check_at, or retire it entirely. No third drift.
Step 3 — Parking-lot sweeps
Three queries, same discipline: anything in non-terminal status
without recent activity gets forced to a closed state.
3a. Running tasks that stalled
loopany artifact list --kind task --status running
For any task running ≥14 days with no recent append:
- Still actively blocked → flip to
in_review if waiting on
external work, or set a near-term check_at with a concrete
unblock condition
- Abandoned →
failed or cancelled with Outcome per
[[core-artifacts/SKILL.md]]
Don't accept "still working on it" as closure — that's how running
turns into a parking lot.
3b. Active signals nobody acted on
loopany artifact list --kind signal --status active
For any signal active ≥7 days without upgrade or dismissal:
- Real, going to act → upgrade to task via [[core-artifacts/SKILL.md]] § Signal → Task promotion
- Real, not acting → dismiss with reason (so [[reflect/SKILL.md]] can
see it if it recurs)
- Wasn't really a signal → delete
An unresolved observation compounds as noise. Force a decision.
3c. Pending skill-proposals
loopany artifact list --kind skill-proposal --status pending
- Count > 0 → mention in the weekly report; these are waiting on
the user.
- Count > 5 → the self-improvement loop is blocked on triage. Nudge
the user to walk the pending queue before another
[[reflect/SKILL.md]] run adds more.
Step 4 — Report and close
Short digest, four sections:
Doctor: clean (or: 3 findings — see fixes above)
Slipped: 2 tasks past due, 1 learning 10 days overdue
Parking lot: 1 task running 21d, 3 active signals ≥1wk, 4 pending proposals
Per-item prompts for anything ambiguous. Same closure gate as
[[daily-followups/SKILL.md]]: every surfaced item ends this session in
resolved / re-scheduled / retired. Zombies are the failure mode
this skill exists to prevent; don't become the source of new ones.
Step 5 — Feed reflect
If this pass resolved ≥3 items (any mix of done / cancelled /
superseded / dismissed), that's a batch worth reflecting on. Either
nudge the user with "want to run a reflection?" or, on agent
platforms, make sure [[reflect/SKILL.md]] is on this week's cron.
Anti-patterns
❌ Running this daily
One week of overdue is a useful pattern. One day's worth is "the
session didn't happen yet." A daily weekly-sweep becomes a second
daily-followups digest and trains the user to ignore both.
❌ Treating doctor output as user-facing
JSON is for you. Categories + proposed fixes are for the user.
❌ Silently fixing drift
enable the domain, remove the edge — user decisions. Propose,
don't act.
❌ Accepting "still working on it" for running tasks
If it's really active, it earns a fresh check_at with an unblock
condition. If it can't earn that, it's abandoned.
❌ Rescheduling without a reason
Every check_at push needs a reason — same rule as
[[daily-followups/SKILL.md]].
Quick reference
DOCTOR: loopany doctor → categorize → propose fixes
OVERDUE: followups --due overdue → A forgot · B outdated · C mis-set · D zombie
PARKING: running ≥14d · active signals ≥7d · pending proposals >0
CLOSE: every surfaced item ends resolved / re-scheduled / retired
FEED: ≥3 resolutions → nudge reflect
