/observe

The proactive observer — a force-multiplier on your own vigilance. Instead of you being the one who eventually notices "this only handles the self case, not the others" or "this file will silently truncate as it grows," /observe goes looking for that class of gap first and hands you a ranked, already-verified digest.

It has two dimensions, and by default runs both:

• Trigger-watcher — walks the deferred backlog's _Trigger:_ conditions against what has actually accumulated (usage, recent work, the knowledge layer) and surfaces the deferred items that are now ripe — the "which one is ready?" watch, lifted off you.
• Latent-gap auditor (the load-bearing half) — a proactive audit for untracked gaps: shipped behavior that no longer matches design intent, things silently about to break as the project grows, decisions made but never shipped, and drift in your own global setup (~/.atl/).

Same discipline as /docs-audit: findings are grep-grounded (no claim without a verbatim source quote) and adversarially verified (each candidate is challenged before it's kept). Multi-agent audits hallucinate ~40% of the time — that guard is what keeps the digest signal, not noise.

When to run it

• When atl reports "a proactive observer sweep is due — run /observe" at session start.
• Any time you want to proactively check for ripe items or latent gaps.

Scope it with --triggers-only or --gaps-only to run just one dimension; the default is both.

What it does

1. Orient — read the project's CLAUDE.md, recent .atl/journal/, and the deferral surface (.atl/backlog.md, or a delivery board when one is configured).
2. Ripe triggers — judge each deferred item's _Trigger:_ against real evidence; an item is ripe only with a verbatim quote of the signal that fired it.
3. Latent-gap sweep — fan out finders across lenses (shipped-vs-designed, growth/scale, decided-but-unshipped, user setup), each grounding its findings with a source quote.
4. Verify adversarially — try to refute every candidate; keep only what survives, re-weigh severity.
5. Surface — a ranked digest of the verified flags, most-actionable first, each with its evidence and a suggested next step. An empty sweep is a valid result.
6. Record — atl observe --record stamps the cursor so the signal won't re-fire for ~1 day.

Boundaries

• It proposes, it doesn't auto-act. A latent-gap finding often needs a decision (a brainstorm) — /observe surfaces it and lets you choose; it doesn't silently open PRs or create work items.
• Honest bound. It catches a class of gaps reliably — shipped-vs-designed mismatches, growth/scale risks, unshipped decisions, ripe triggers — enough to stop the recurring "you caught it first." It is not a guarantee to catch everything; it's a multiplier on your vigilance, not a replacement.
• Advisor boundary. It audits the advisor's setup and output from the outside; it never runs inside, or as, the advisor — that conversation stays pure.

Related

• atl observe — the deterministic half: the "sweep due?" signal and the cursor this skill stamps.
• /docs-audit — the same grep-grounded + adversarial discipline, aimed at the docs site.