dry_run_skill compiles a skill and shows every step's risk classification before anything runs.
Plan, adapt,
and stop when the ground moves.
Two real mechanisms, combined: invoke_skill compiles a skill to a risk-classified plan and gates risky steps on human approval; hash-based drift detection flags a skill the moment its source changes. Together, that is the engine—not a single monolithic component, but two verified pieces that already work.
AVAILABLE NOWTwo mechanisms.
Two mechanisms.
Working together.
A step classified as requiring approval pauses the run at "awaiting approval" until a person clears it—not a suggestion, an enforced gate.
Hash-based change detection opens a drift alert the moment a source document changes, on a 15–30 minute cycle or near-real-time via webhook.
LIVE TODAYHash-based drift detection,
Hash-based drift detection,
running against a real skill.
This is the same mechanism that watches every bound source today: a content hash changes, the skill it grounds gets flagged. Nothing about this state is simulated engine behaviour — it's the actual detection path.
INTERACTIVE DEMONSTRATION
DRIFT DETECTION / REFUND AUTHORITYANSWER LIVE
Manager approval required over
$300Finance Operations / approvedManager approval required over
$750Prepared for named owner reviewAGENT ANSWERCan I approve a $500 refund?
LIVENo — manager approval required
RECEIPTSOURCE → OWNER → ANSWER IMPACTHash matched, no alert
Automatic pausing
is not built yet.
A drift alert flags the affected skill for review today; it does not yet automatically block that skill from being served while the review is open. Automatic pausing is on the roadmap—see the governance page for the full, current status of that specific gap.
See the current status →