Skip to content

preview-first-batch

MANDATORY. Before any batch / bulk / long-running operation, run a small smoke-test sample first, surface the result, and obtain explicit user confirmation (via AskUserQuestion) before scaling to the full set.

Whichever comes first: estimated wall-clock time > 2 minutes, or item count > 10 items. If you cannot estimate, the trigger fires by default.

State the estimate before the first tool call (“smoke N=5 (~30s), then full N=285 (~17h) on your go-ahead”; unknown → say so) → smoke the smallest meaningful N (5 items / --limit 5 / lowest-traffic env) → surface the smoke output (wall-clock per item, values, anomalies) → AskUserQuestion to gate the full run even on a clean smoke (proceed / investigate / re-smoke / abort). Never silently scale — a clean smoke is not auto-permission. If the smoke wrote files in place, list them and the rollback command before proceeding.

docs/preview-first-batch.md (worked examples, the 2026-04-30 incident postmortem, rule interactions, self-test) · always-ask-on-unresolved.md (the strict-ask mandate this depends on) · coding-guidelines.md §4–§5 (before/after verification gates)