inherited-scope-doc-names-may-not-exist

Pre-Dispatch Schema Probe

Problem

Scope docs and plan docs authored by predecessor sessions routinely cite specific BigQuery paths, table names, and column names as if they were verified facts. Often they're not — authors write from memory, from stale schema snapshots, or from hypothetical schemas that never landed. When you mechanically copy those references into 7 parallel agent prompts and fire them against production data, you burn the entire wallclock window discovering the errors one at a time, mid-run.

The failure mode is costly because:

1. Multi-agent dispatches amplify the blast radius. 1 wrong table path × 7 agents × 10 min before each tap-out = 70 agent-minutes of wasted compute before anyone notices.
2. Predecessor scope docs carry authority. A scope doc that went through multiple review rounds, PR approvals, and handoff sessions LOOKS authoritative. No one re-checks the dataset paths because "surely someone already did."
3. Plausible-wrong > obviously-wrong. target_2025 looks like exactly the kind of column name a training-features table would have. analytics.predictions_daily.predictions_daily looks like a sensible namespaced path. Both are wrong. Obviously-wrong paths (typos) fail loudly; plausible-wrong paths fail slowly inside agents.
4. Column-name prefixes are assumed. Predecessor sessions often describe engagement columns as "the evt_* family" or "the engagement_* columns" when the actual warehouse naming has no prefix at all (e.g., page_visit_8_30d, login_1d).
5. The fix is cheap but the miss is expensive. A 5-minute bq ls + INFORMATION_SCHEMA.COLUMNS probe costs <$1 and surfaces 100% of path/name errors before firing. Skipping it costs a full overnight window.

Context / Trigger Conditions

Use this skill when ALL of:

• You're about to dispatch ≥2 parallel agents, Cloud Run jobs, scheduled triggers, or long-running Python scripts that query a data warehouse.
• The prompts / SQL / Python are being generated from a scope doc, plan doc, or handoff document authored in a prior session (not freshly inspected by you).
• The dispatch wallclock is measured in hours (not minutes) — i.e., the cost of mid-run failure is not trivial.