loopany-monthly-review

name	loopany-monthly-review
description	Use on monthly cadence or when the user asks "is this still the right mission?" / "should we reset?" / "anything structural worth locking in?". Detects two drifts - (1) mission drift: active `mission` no longer matches recent tasks; (2) structural drift: recurring usage that should crystallize into a new `domain` or `kind`. Surfaces evidence; user decides.

monthly-review — mission-drift + structural-drift detection

Two kinds of drift accumulate silently over a month:

1. Mission drift — the active mission no longer describes the user's actual work. Feeds ONBOARDING.md's re-run trigger.
2. Structural drift — a recurring theme hasn't crystallized into a domain or kind yet.

Neither auto-acts. Surface evidence; the user decides.

When to run this

• Scheduled: once per month (agent platforms)
• On demand: user asks "is this still the right mission," "should we reset," "anything structural worth locking in?"

Never weekly. One week of task data is too small to tell drift from noise.

Step 1 — Pull the mission and recent tasks

loopany artifact list --kind mission --status active
loopany artifact list --kind task   # filter by id prefix to last 30d

Step 2 — Compute alignment

For each active mission:

alignment = tasks mentioning this mission / total recent tasks
             (restricted to the mission's domain, if it has one)

Thresholds:

• ≥ 60% — healthy. No action.
• 30–60% — partial drift. Ask whether the stray tasks actually serve the mission (missing --mentions backfill) or a second mission is worth locking in.
• < 30% — clear drift. Propose re-running onboarding.

Weight toward recent originals. [backfill] tasks seeded during onboarding don't reflect current intent.

Step 3 — Surface to the user

All missions healthy → one-line note, done.

Drift detected → present concretely, with evidence:

Drift on `mis-mrr-growth` (active since 2026-01):
- 23 tasks in the last 30 days, 6 mention this mission (26%)
- The other 17: mostly hiring + infra work

Two options:
1. Hiring/infra actually serves MRR — backfill the mentions
2. MRR isn't the main thing anymore — re-run onboarding

Step 4 — Act

• Backfill → batch-update --mentions on affected tasks.
• Re-run onboarding → flip the old mission to abandoned, create the new one, add the supersedes edge. See ONBOARDING.md § "When to re-run onboarding."
• Neither → push this skill's check_at 2–4 weeks with a reason. "Revisit after launch" is a reason; "dunno yet" isn't.

Step 5 — Structural drift (new domain? new kind?)

Don't multiply entities beyond necessity. Most months both checks come back empty, and that's healthy.

5a. Emerging domain

loopany domain list            # shows `observed_only`
loopany doctor --format json   # `domain coverage` lists offenders

Proposal threshold — all three must hold:

• ≥ 5 artifacts use the same unenabled domain value
• Usage spans ≥ 2 weeks (not a single burst)
• The domain is distinguishable from every enabled one — if it's a re-label of an existing one, rename, don't add

If met, propose a task (see [[core-artifacts/SKILL.md]]). Suggested content:

• ## Why — treating xyz as its own domain should clear doctor's coverage warnings and make artifact list --domain xyz useful
• ## Current evidence — the N ids currently tagged domain: xyz
• ## Sweep plan — real sweep (domain is a tag, safe to retroactively reassign):
  1. loopany domain enable xyz
  2. For each id: loopany artifact set <id> --field domain --value xyz
  3. loopany doctor — confirm domain-coverage clean

5b. Emerging kind

loopany artifact list --kind note   # then read bodies

Looking for repeating body shape — same section headings, same fields at the top. Not "notes about the same topic" — notes that look structurally alike.

Proposal threshold — both must hold:

• ≥ 3 notes share a recognizable body skeleton, spanning ≥ 2 weeks
• The pattern passes the 4-question test in CLAUDE.md (state machine / identity / structured queries / required body shape). If none of the four hold, it stays a note.

If met, propose a task. Suggested content:

• ## Why — a dedicated experiment kind should make artifact list --kind experiment useful and enforce sections the notes already carry informally
• ## Current evidence — the N existing notes matching the shape, listed as evidence, not migration candidates
• ## Sweep plan — None. kind determines id prefix and storage path; retroactive change would invalidate references.jsonl and scramble audit.jsonl history. The new kind applies only to artifacts written after kinds/<name>.md lands. Past notes stay notes.
  1. Write kinds/<name>.md (schema + status machine + required sections)
  2. loopany doctor — new kind parses, old notes still validate

Surface both together

One monthly digest:

Mission drift: <see Step 3>
Structural:  domain `xyz` candidate (7 artifacts, 3wk) — propose enable+sweep?
             kind `experiment` candidate (4 notes, 2wk, passes 4Q) — propose?

One drift → one task. None → say so and stop; the user wants to know you looked.

Anti-patterns

❌ Running this weekly

One week is noise; false drift alerts follow.

❌ Auto-abandoning a mission, or auto-filing a structural task

Drift is a hypothesis about intent, not a fact. Surface evidence and wait for approval — both mission status flips and structural tasks need a human on the trigger.

❌ Counting [backfill] tasks equally

Seeded during onboarding; doesn't reflect current intent. Weight down or exclude from the denominator.

❌ Ignoring domain scope in alignment

If the mission has a domain, compute alignment within that domain only. Tasks in unrelated domains aren't expected to mention it.

❌ Proposing a new kind on first sighting

One note with a distinctive shape is a note. Three with the same skeleton is still interesting, not a kind. Run the 4-question test; if none hold, it stays a note. Kinds ship a schema the runtime enforces forever.

❌ Enabling a domain without committing to the sweep

If you enable xyz but leave existing domain: xyz artifacts untouched (or never rename near-misses), the task hasn't cleared anything. Commit to the sweep or don't file.

Quick reference

QUERY:     active missions + last-30d tasks
COMPUTE:   alignment ratio per mission (within domain scope if any)
SURFACE:   <30% drift · 30-60% check · ≥60% healthy
ACT:       backfill mentions OR re-run onboarding (supersedes)

STRUCTURE: ≥5 artifacts on unenabled domain, ≥2wk
             → propose task: enable + real sweep
           ≥3 notes sharing body shape, ≥2wk, passes 4Q test
             → propose task: add kind, forward-only (no migration)