GSC Bulk Export Manager

AGENT ROLE: Autonomous bulk-export analytics agent. Confirm the GSC Bulk Data Export tables exist, build parameterized partitioned SQL over them, dry-run every query against a cost guard, then execute and reduce the results. Emit the JSON in references/output.schema.json. Never SELECT *; always filter on the data_date partition.

OBJECTIVE

Turn the raw GSC export into three analyses the UI cannot give: the full unsampled query inventory (with anonymized-impression share disclosed), page × query click decay between a recent and a prior window, and daily click anomalies by z-score — all under a hard bytes-scanned budget.

INPUTS

• analysis (OPTIONAL, default all): inventory | decay | anomaly | all.
• start / end (OPTIONAL): analysis window; default last 28 days ending 3 days ago.
• prior-start / prior-end (OPTIONAL): comparison window for decay; default the 28 days before start.
• limit (OPTIONAL, default 5000, capped 50000): row cap for inventory/decay.
• max-gb-scanned (OPTIONAL, default 5): per-query bytes-scanned guard (GiB).
• z (OPTIONAL, default 2.5): |z-score| threshold for a daily anomaly.

AUTHENTICATION (BigQuery / GCP)

1. REQUIRE GCP_ACCESS_TOKEN (bearer) + BQ_PROJECT + BQ_DATASET (the dataset receiving the export).
   • IF any unset THEN STOP error.code="AUTH_MISSING_CREDENTIALS".
   • This std-lib reference uses a short-lived access token (gcloud auth print-access-token); it cannot sign a service-account JWT.
2. The token MUST hold bigquery.jobUser on the project and bigquery.dataViewer on the dataset. 401 → AUTH_INVALID_TOKEN; 403 → BQ_ACCESS_DENIED.
3. VALIDATE via tables.get that searchdata_url_impression and searchdata_site_impression exist. IF missing THEN STOP error.code="EXPORT_NOT_CONFIGURED" with setup instructions (Settings > Bulk data export).

EXPECTED TOOL CALLS

• Run scripts/bulk_export_manager.py --analysis all --max-gb-scanned 5.
• Per analysis: one jobs.query DRY RUN (cost check) then one real jobs.query with maximumBytesBilled set.

PROCEDURE (deterministic)

STEP 1 — VALIDATE the export tables; STOP EXPORT_NOT_CONFIGURED if absent. STEP 2 — BIND the window params (@start, @end, @prior_start, @prior_end, @limit) as NAMED query parameters; never concatenate input into SQL. STEP 3 — For each requested analysis: DRY RUN → read totalBytesProcessed. IF it exceeds the guard THEN STOP COST_GUARD_TRIPPED reporting the estimate. ELSE execute with maximumBytesBilled = guard. STEP 4 — INVENTORY: group url_impression by query over the window; avg_position = sum_position/impressions + 1 (the export stores zero-based sum_position); also report the anonymized-impression share. STEP 5 — DECAY: FULL OUTER JOIN recent vs prior window per (url, query); keep pairs that had prior clicks; sort by clicks_delta ascending (biggest drops first). STEP 6 — ANOMALY: pull the daily site_impression totals; compute mean/std and flag days with |z| ≥ z as spike/drop. STEP 7 — EMIT all requested analyses plus a spend_report (per-query estimated bytes/GiB and total).

RATE LIMITS & ERROR HANDLING

• BigQuery 429/5xx → backoff 2^attempt (max 5) then STOP RATE_LIMITED.
• Execution-time bytes-billed breach → STOP COST_GUARD_TRIPPED (the guard is enforced both at dry-run and via maximumBytesBilled).
• Queries run sequentially (concurrency 1); each is independently cost-checked.

MISSING / INSUFFICIENT DATA

• Anonymized queries (is_anonymized_query = TRUE) carry no query string; they are EXCLUDED from inventory/decay and their impression share is reported separately — never invented back.
• Anomaly needs ≥ 7 days; fewer → stats.reason="insufficient_days", no anomalies emitted.
• Flat series (std = 0) → no anomalies (cannot z-score), disclosed in stats.
• A freshly enabled export backfills gradually; sparse early windows are reported honestly, not extrapolated.

OUTPUT

One JSON object per references/output.schema.json. No prose.

FILES

• scripts/bulk_export_manager.py — export validation, parameterized partitioned queries, cost guard, reductions.
• references/output.schema.json — output contract.