Second Brain

snowflake-mcp v22.0.3 upgrade quality audit: FULL PASS — Bible v19.1.7…

3 min read

snowflake-mcp v22.0.3 upgrade quality audit: FULL PASS — Bible v19.1.7…

Decision

snowflake-mcp v22.0.3 upgrade quality audit: FULL PASS — Bible v19.1.7 §07 §10 compliance fully restored after 7 same-day remediations from v22.0.0. Platform cleared for production use on Q1 FY26-27 intelligence push (Cortex Analyst tuning, Construction semantic views, Snowflake Intelligence rollout).

Rationale

Audit ran all 5 Bible v19.1.7 §07 §10 verification rituals against live server, then re-ran the exact same probes after AJ confirmed fixes. Every probe transitioned from RED to GREEN with empirical before/after evidence: (R1) R13 atomic-version-cascade — Prometheus mcp_server_info gauge now matches /health (v22.0.3 + new bible=“v19.1.7” label adds a second cross-check axis); (R2) circuit_breaker subsystem returns structured 4-breaker payload with G10 thresholds visible (snowflake fail_max=5, metadata=10, admin=3, external=5, all closed, 0 failures) — was crashing serializer with 'str' object has no attribute 'name'; (R3) list_cortex_models returns 20 models curated from vendor docs with honest “no catalog SQL exists, pinned 2026-04-25” note — replaces invalid SHOW MODELS SQL; (R4) show_semantic_views_v2 returns view_count=14 scope=ACCOUNT covering all PROD/DEV semantic views — was returning empty + Pydantic validation crash; (R5) list_capabilities envelope refreshed (version=22.0.3, architecture=4-tier, bible_version=v19.1.7, total_tools=880, guardrails_active list explicit) — replaces stale v4.0.0/three_tier/166-tool envelope; (R6) tool count math now closes (850 dynamic + 6 system + 24 capability adds = 880) reconciling skill frontmatter with dispatcher; (R7) check_account_edition returns ENTERPRISE via feature_inference fallback (cortex_complete + masking + row_access + tri_secret signals) — was returning UNKNOWN. Sanity probe: Gen 1 banned-tool gate still hot (-32006 with migration text). Active prod surface healthy: 14 semantic views including Construction layer (RNWL_COLAB_REPORTING 24-Apr, RNWL_CONSTRUCTION_REPORTING 23-Apr, RNWL_SALES_REPORTING 09-Apr), 2 prod Cortex agents (RUNWAL_BOT, RUNWAL_CRM_BOT), all 12 warehouses STANDARD_GEN_2, Cost Observer fully wired (30 Fivetran metadata tables + WH_FIVETRAN_COSTGUARD + COST_OBSERVATIONS audit log). Phase 4c retraction baseline verified — 37 Cortex SQL functions / 79 overloads present in SNOWFLAKE.CORTEX (AGENT_RUN, AI_COMPLETE, ANALYST_PREVIEW, DATA_AGENT_RUN, AI_PARSE_DOCUMENT). Only outstanding ritual is the daily audit_gen1_callers(7d) cron — operational, not interactive, must be confirmed via post-cutover ops trail separately. Full audit + follow-up appendix: outputs/reports/snowflake-mcp-quality-audit-25-apr-2026.md. Empirical confidence is high because every remediation was verified by the exact probe that surfaced it; rollback path is clean (patch-version downgrade to v22.0.0 if regression appears).

Alternatives Rejected

Outcome

Pending

Graph View

Backlinks