Snowflake Access Guardian

Overview

Turn a sanitized Snowflake authorization inventory into an evidence-backed effective-access trace and a dry-run remediation packet. This is the focused Snowflake counterpart to a generic RBAC explainer: it catches the failure modes that make enterprise reviews expensive—role inheritance that was not followed, direct user and PUBLIC grants, abandoned grantees, ownership control, managed access semantics, secondary-role assumptions, and future-grant precedence.

Prerequisites

• Receipted, sanitized outputs from the bundled historical, session, role, user, database-role, and relevant database/schema future-grant collectors.
• A named principal/object/privilege question, account/role identity, UTC collection timestamp, observation window, and explicit freshness bound. Live Snowflake checks remain the operator's responsibility.
• Timestamped positive (allowed action) and negative (denied action) receipts captured under the same primary/secondary-role context; missing proof is NOT_PROVEN, never an inferred denial.
• Python 3.10+ for the bundled stdlib analyzer. No Snowflake driver or network access is required.

Authentication

This skill's analyzer is offline and deliberately has no authentication flow. If live Snowflake evidence is collected, use the organization's approved Snowflake session/authentication process; never put its credentials in the inventory or report. Use Write only to save a sanitized report or approved local change packet; never to apply Snowflake mutations.

Safety contract

• Read-only by default. The analyzer does not connect to Snowflake and never executes GRANT, REVOKE, GRANT OWNERSHIP, ALTER USER, or policy changes.
• Accept sanitized metadata only. Do not provide passwords, tokens, private keys, raw connection strings, or access-history payloads containing sensitive data.
• Do not infer denial from a missing historical row. Account Usage can lag by up to 120 minutes and has documented object/shared-role omissions.
• Never broaden to ACCOUNTADMIN or automatically grant MANAGE GRANTS to get a more complete snapshot. MANAGE GRANTS can administer grants and is not a read-only privilege, even when this collector executes only SHOW statements.
• Do not infer that a role is unused from one telemetry source. Name the review period, object coverage, and evidence gaps.
• Treat ownership as control-plane authority and future OWNERSHIP as a separate high-risk decision. Never auto-generate executable mutation SQL.

Workflow

1. Establish the principal, target object, privilege, evidence timestamp, review period, and whether the question is about a primary-role or secondar…

Snowflake Cost Leak Hunter

Overview

Produce a read-only, evidence-first Snowflake cost investigation. The deterministic analyzer emits a typed ledger whose total, attribution, context, estimate, and invoice-only roles cannot silently collapse into one savings claim.

Problem: Cost views answer different questions, arrive at different times, and do not equal invoice truth. Generic advice easily turns observed credits into unsupported prices or promised savings.

Outcome: A reproducible review packet that identifies what the supplied evidence confirms, what was estimated from an approved rate, what remains at risk, and what is unknown.

Prerequisites

• An exact account and UTC analysis window.
• Sanitized output collected by an operator through an approved read-only Snowflake session, or an equivalent exported evidence bundle.
• A role already authorized to read the required SNOWFLAKE.ACCOUNT_USAGE surfaces.
• A writable local working directory for report artifacts. Use Write only to create new local evidence/report files; never use it to alter Snowflake configuration.
• A customer-supplied rate-card record if currency estimates are requested.

Safety and evidence contract

• Read-only Snowflake work only. Do not execute DDL/DML, change warehouse state or size, alter auto-suspend, assign monitors, create budgets, change tags, or cancel queries. Emit proposed changes for a named owner to review.
• Do not invoke Snowflake authentication from this skill. The operator runs the bounded collection queries through an approved read-only session and supplies only sanitized results. Never request connection files, private keys, tokens, passwords, or environment values.
• Do not require ACCOUNTADMIN. Use a role already authorized to read the needed SNOWFLAKE.ACCOUNT_USAGE views. Visibility differs by database role and account configuration; report missing access rather than escalating privileges.
• Bind every history query to an explicit UTC half-open start and end time. Record session UTC offset, account/organization identity, role, query IDs, collection time, SQL and result hashes, row cap state, and the maximum activity timestamp returned. A single bundle may cover at most seven days; partition longer audits.
• Use the fixed official latency matrix. Apply the reviewed per-view and per-field cutoffs in references/cost-ledger-and-surfaces.md; caller-supplied latency or maximum activity time cannot make a recent window settled. Compute each cutoff from the same-statement execution_context.observed_at, never the later CLI …

Snowflake Data Quality Sentinel

Purpose

Produce separate configuration, history-observation, and history-completeness verdicts from an owner-approved requirement denominator and trusted schema-2 collector receipts. Missing or untrusted evidence never becomes a pass, and this history surface never supports an unqualified present-tense quality pass.

Read the input contract before assembly and consult the source notes when interpreting provider semantics.

Prerequisites

Use Python 3.10+, an owner-approved policy, and Snowflake CLI with an existing least-privilege profile. Never accept credentials or request customer rows, failed-row payloads, metric values, or SQL text.

Workflow

1. Collect the exact per-object receipt set.
2. Record independently trusted evidence and owner-policy digests.
3. Analyze at the policy-bound time and preserve every non-claim.

Step 1: Collect evidence

Collect one bounded history receipt. For every distinct governed object, collect one selector-bound live association receipt and one selector-bound live expectation receipt. Also collect one notification receipt for every distinct governed object whose notification_required is true:

# Example fixed UTC window; replace with the audited interval.
python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface data-quality --connection readonly-observer \
  --window-start 2026-09-01T00:00:00Z --window-end 2026-09-02T00:00:00Z \
  --output ./dq-history.json

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface data-quality-associations-current --connection readonly-observer \
  --data-quality-object GOVERNED_DB.GOVERNED_SCHEMA.GOVERNED_TABLE \
  --data-quality-domain TABLE \
  --output ./dq-associations.json

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface data-quality-expectations-current --connection readonly-observer \
  --data-quality-object GOVERNED_DB.GOVERNED_SCHEMA.GOVERNED_TABLE \
  --data-quality-domain TABLE \
  --output ./dq-expectations.json

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface data-quality-notification-current --connection readonly-observer \
  --data-quality-object GOVERNED_DB.GOVERNED_SCHEMA.GOVERNED_TABLE \
  --data-quality-domain TABLE --output ./dq-notification.json

The object selector is used only in reviewed local SQL. The receipt retains its scoped object hash and domain, never the raw selector. Do not escalate roles when a source is permission-blocked.

Step 2: Establish trust and analyze

Maintain the owner-approved policy as a separate policy.json. Assemble exactly schema_version, the byte-equivalent parsed …

Snowflake Deploy Medic

Overview

Snowflake deployment failures hide in the seams: a grant resource is declared as new although the remote grant already exists; provider state is truncated; a versioned migration was edited after application; a repeatable checksum change reruns unexpectedly; or a current CLI/driver behavior change is mistaken for a database defect. This skill creates one evidence-backed gate across those seams.

The deterministic classifier is scripts/analyze_deploy_evidence.py. It accepts a bounded schema-v2 projection plus its digest from an independent CI or artifact channel. It emits privacy-safe findings, a point-in-time zero-change status, ordered read-only checks, and post-deploy invariants. Nested receipt hashes detect internal inconsistency; they never substitute for trusted origin. Read references/terraform-provider-2x.md, references/schemachange-integrity.md, and references/zero-change-rollback.md for the relevant surface. Always verify current primary docs and release notes; the versions in a local receipt are observations, not timeless recommendations.

Hard boundaries

• Never run terraform apply, terraform destroy, state editing, schemachange deploy, mutating snow sql, or Snowflake DDL/DML automatically.
• Never hand-edit terraform.tfstate or CHANGE_HISTORY to make a plan green.
• Never treat a valid plan with changes as a zero-change adoption; inspect every grant, ownership, replacement, destroy, and preview feature.
• Never edit an applied versioned migration in place. A checksum mismatch is a release-blocking integrity signal; choose a new version, restore the exact applied content, or explicitly review a compensating path.
• Never treat a repeatable checksum change as either an error or a free rerun: verify idempotence, intended scope, and current schemachange behavior.
• Never freeze provider, CLI, driver, or BCR guidance to a version copied from this skill. Record current versions and check live release notes for the target account release window.
• Redact backend credentials, tokens, private keys, passphrases, passwords, sensitive plan values, customer data, and presigned URLs from receipts.
• Never execute Terraform, Snowflake, schemachange, a network client, or a shell from the classifier. It is a pure JSON reader and report writer.
• Never treat a generic query-history receipt as deployment proof. It cannot establish a complete plan, state, migration, BCR, or affected-object denominator.

Prerequisites

Collect a timestamped receipt from the exact account, role…

Snowflake Failover Readiness Drill

Purpose

Produces a deterministic readiness or attestation result from exact, independently trusted evidence. It observes and verifies but never performs or authorizes a Snowflake control-plane change.

Prerequisites

• Python 3.10+ and this complete skill directory, including its reviewed SQL.
• An owner-approved scope and separately retained digests for policy, collection, and operator/validation evidence.
• Existing least-privilege Snowflake CLI profiles in source and target accounts when live collection is required. Profiles used for Information Schema history, progress, or dangling functions must select an approved current database; the skill does not configure authentication or session state.
• Business Critical Edition or higher for failover groups.

Workflow

1. Have the recovery owner approve an exact policy containing every in-scope group, dependency, validation, RPO, and RTO. Record its digest at a trusted boundary before evidence is transported.
2. Collect replication-current in both source and target accounts. Collect replication and replication-progress in each target account over the same current, explicit half-open UTC window. A full failback drill also requires both surfaces in the original source account to prove the reverse refresh leg. Collect replication-dangling in both accounts for every in-scope local group.
3. Build the exact schema-2 wrapper in references/evidence-contract.md. Record independent digests for the collector bundle, policy, and operator/validation receipts. A digest recomputed from the delivered file is not independent.
4. Run the analyzer; it writes only to stdout:

   python3 "${CLAUDE_SKILL_DIR}/scripts/analyze_failover_readiness.py" \
     --input ./snowflake-failover-evidence.json \
     --evaluated-at "$EVALUATED_AT_UTC" \
     --trusted-input-sha256 sha256:... \
     --trusted-policy-sha256 sha256:... \
     --trusted-operator-sha256 sha256:... \
     > ./snowflake-failover-report.json

An INCONCLUSIVE or NOT_READY exit is 1; malformed input exits 2.

Collection pattern

Use an existing least-privilege Snowflake CLI profile. Never pass credentials to the collector and never switch to ACCOUNTADMIN because evidence is missing.

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface replication-current --connection source-observer \
  --output ./source-current.json

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface replication --connection target-observer \
  --window-start "$WIND…

Snowflake Governance Coverage Auditor

Purpose

Compare an owner-approved hashed denominator with trusted schema-2 current receipts and separately trusted, sanitized POLICY_CONTEXT simulations. Missing, stale, capped, privilege-filtered, unsupported, or context-mismatched evidence is never a pass; the result distinguishes observable coverage from evidence gaps without exposing governed data or claiming compliance.

Read the input contract and the source notes before assembling evidence.

Prerequisites

Use Python 3.10+, an owner-approved hashed denominator, and an existing Snowflake CLI read-only profile. Establish independent evidence and policy trust boundaries before analysis; do not accept credentials, raw identifiers, policy text, tag values, customer rows, or ad hoc SQL.

Workflow

1. Have the governance owner approve the exact asset and simulation denominator.
2. Collect one classification receipt per database and one tag plus one policy receipt per governed object. Use fixed unquoted selectors only:

   python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
     --surface governance-classification-current --connection readonly-observer \
     --governance-database GOVERNED_DB --output ./classification.json

   python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
     --surface governance-tags-current --connection readonly-observer \
     --governance-object GOVERNED_DB.GOVERNED_SCHEMA.GOVERNED_TABLE \
     --governance-domain TABLE --output ./tags.json

   python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
     --surface governance-policies-current --connection readonly-observer \
     --governance-object GOVERNED_DB.GOVERNED_SCHEMA.GOVERNED_TABLE \
     --governance-domain TABLE --output ./policies.json

1. Independently verify the collection role's complete visibility over exactly those hashes and produce the scope receipt. Execute the approved POLICY_CONTEXT cases outside this collector; retain only the strict hash-only receipt contract. Never add EXECUTE USING to the shared collector.
2. Record evidence and policy digests at their independent trusted boundaries:

   # Record only at trusted local boundaries.
   python3 "${CLAUDE_SKILL_DIR}/scripts/analyze_governance.py" evidence.json \
     --print-input-sha256
   python3 "${CLAUDE_SKILL_DIR}/scripts/analyze_governance.py" evidence.json \
     --policy-file policy.json --print-policy-sha256

1. Analyze only with the previously recorded digests and policy-bound clock:

   # Replace the quoted placeholder with the owner-policy tim…

Snowflake Native App Release Sheriff

Purpose

Produce a deterministic, as-of provider-side release preflight. A clean result is READY_FOR_OPERATOR_RELEASE_AS_OF, never permission or an instruction to publish, set a release directive, upgrade an application, or alter privileges.

Prerequisites

Read both reference files before analysis. The input must use the exact schema-2 contract in references/evidence-contract.md and must include independently retained hashes for the whole bundle, manifest, setup script, normalized cohort rows/denominators, lifecycle evidence, and the complete rollback receipt including its artifact digest. Every array carries an explicit count; zero is evidence only when the owner-approved denominator is exactly zero.

Python 3.10+ is required. A Snowflake CLI profile is optional for collection and must already exist; authentication is configured outside this skill. The package manifest and setup script remain local inputs. Never send source SQL, package names, consumer names, account names, free-text scan failures, or App Spec definitions through the evidence packet. Use stable, account-scoped hashes.

Instructions

1. Collect all three live provider surfaces for one exact package.
2. Build the schema-2 packet with owner-approved denominators and trusted hashes.
3. Run the analyzer at an explicit UTC evaluation time.
4. Stop on invalid, blocked, stale, capped, or incomplete evidence; hand a clean as-of report to the separately authorized release owner.

Collect current provider evidence

Use an existing least-privilege Snowflake CLI profile with visibility to the selected package. The collector uppercases and strictly validates the unquoted one-part package selector, then binds the rendered-query receipt to the Snowflake-produced selected-package hash. An error receipt contains no package fingerprint.

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface native-app-versions-current --application-package APP_PACKAGE \
  --connection native-app-observer --output ./versions.json

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface native-app-release-directives-current --application-package APP_PACKAGE \
  --connection native-app-observer --output ./directives.json

python3 "${CLAUDE_SKILL_DIR}/scripts/collect_snowflake_evidence.py" \
  --surface native-app-upgrade-cohorts-current --application-package APP_PACKAGE \
  --connection native-app-observer --output ./cohorts.json

SHOW VERSIONS requires package visibility. SHOW RELEASE DIRECTIVES requires package ownership or the documented package release/version management privilege. SNOWFLAKE.DATA_SHARING_USAGE.APPLICATION_STATE is provide…

Snowflake Pipeline Guardian

Overview

Pipeline symptoms cross object boundaries, but dependency order does not prove causality. This skill classifies a bounded evidence bundle, walks supplied dependency paths, and keeps observed facts, derived findings, hypotheses, and unknowns separate.

Use scripts/analyze_pipeline_state.py for deterministic classification. Read references/current-state.md before collecting live evidence and references/recovery-matrix.md before writing a recovery plan.

Hard boundaries

• Read-only diagnosis only. Never execute or emit runnable DDL, DML, task execution, refresh, resume, stream replacement, pipe refresh, or replay SQL.
• Never call an empty, privilege-hidden, stale, capped, context-mismatched, or untrusted result healthy or complete.
• Never treat current control-plane state as historical run proof, or lagged Account Usage history as current state.
• Never infer causality from graph order or a current/history disagreement.
• Never request or retain raw SQL text, object names, query IDs, file or stage paths, notification endpoints, integration names, free-text errors, or raw SYSTEM$PIPE_STATUS JSON. Use the reviewed hashed projections only.
• Never use SYSTEM$STREAM_HAS_DATA for diagnosis; calling it can affect a stream's staleness behavior.
• Treat hashes as pseudonymous operational data, not anonymization.

Trusted evidence contract

The live collector has six reviewed schema-2 surface types:

1. pipeline: explicit half-open UTC history for task runs, dynamic-table refreshes, and copy loads.
2. pipeline-task-current: current role-visible task inventory.
3. pipeline-stream-current: current role-visible stream inventory.
4. pipeline-dynamic-table-current: current role-visible dynamic-table inventory.
5. pipeline-pipe-current: current role-visible pipe inventory.
6. pipeline-pipe-status: one selector-bound, privacy-projected status receipt for each pipe in the current pipe inventory; zero status receipts are correct only when that inventory is empty.

Each receipt must be live CLI output with exactly one same-statement execution_context, the reviewed SQL/template/result hashes, exact datasets, finite documented state/status domains, the exact fixed non_claims, the declared cap, no collector error, and a valid self-hash. Invalid receipts never reach finding classification. All receipts must agree on organization/account, collector user, primary role and role type, secondary roles, and UTC timezone. Their observations may span at most 15 minutes and may …

Snowflake Query Forensics

Overview

Build a read-only query root-cause packet from platform evidence. Distinguish observed facts, derived metrics, and hypotheses; isolate one variable for any later experiment.

Problem: A slow elapsed time does not identify whether compilation, execution, queueing, blocking, spill, pruning, or query shape is responsible, and some diagnostic surfaces are delayed or unavailable for particular query classes.

Outcome: A query-scoped packet that ranks observed operator evidence, preserves competing explanations, and defines the next read-only check without changing SQL or compute.

Prerequisites

• An exact query ID, bounded candidate set, or sanitized exported evidence file.
• Sanitized output collected by an operator through an approved read-only Snowflake session, or an equivalent exported evidence bundle.
• A role with visibility to the selected history surface; OPERATE or MONITOR on the warehouse when operator statistics are required.
• A completed query within the documented operator-stat retrieval window for operator analysis.
• One Snowflake UUID query ID repeated on the anchor history, operator-statistics, and Query Insights rows, plus a positive source-freshness bound in seconds.
• A writable local working directory. Use Write only for new local evidence and report artifacts; never use it to alter SQL or Snowflake state.

Safety and evidence contract

• Read-only only. Do not cancel queries, resize/resume/suspend warehouses, alter clustering, enable acceleration/search optimization, modify SQL, or change session or account policy on the user's behalf.
• Do not invoke Snowflake authentication from this skill. The operator runs the bounded collection queries through an approved read-only session and supplies only sanitized results. Never request environment variables, connection files, tokens, passwords, or keys.
• Do not require ACCOUNTADMIN. Query-history visibility and operator-stat access depend on the approved role. GET_QUERY_OPERATOR_STATS requires OPERATE or MONITOR on the warehouse. Report missing access; do not grant it.
• Operator evidence exists only for completed queries and only within the platform's documented retrieval window. Do not invent operator findings for running, too-old, or inaccessible queries. Every nonterminal query is a partial packet even when no operator or insight rows were supplied. A terminal full-evidence packet requires at least one bound operator row; otherwise binding is incomplete, completeness is blocked, and confirmed/derived/hypothesis sections stay empty.
• History surfaces have different windows and latency…

Snowflake Strong-Auth Migration Pilot

Overview

Convert an uncertain Snowflake identity estate into an owner-backed, least-privilege authentication pilot. The useful unit is a named workload and its runtime—not a blanket account-wide password deadline. The pilot classifies PERSON, SERVICE, LEGACY_SERVICE, and SERVICE_AGENT principals, maps each bound workload to a supported target, and checks that managed MCP/OAuth primary-role scopes, client behavior, and secondary-role controls cannot silently broaden access.

Prerequisites

• Three live schema-2 collector receipts: near-current SHOW USERS, delayed Account Usage USERS, and the latency-settled portion of a bounded trailing seven-day LOGIN_HISTORY horizon. Record the final bundle SHA-256 outside the bundle when it crosses the controlled local collection boundary; an embedded self-checksum is not provenance.
• Named identity/workload owner, security approver, executor, recovery identity, and approved canary/change window. The inventory must include a separately tested break-glass identity and a canary receipt with positive and negative outcomes. Keep those operational receipts outside this posture bundle; the analyzer intentionally rejects and never echoes embedded canary or recovery payloads.
• Python 3.10+ for the bundled stdlib analyzer. No Snowflake driver or network access is required.

Authentication

This skill's analyzer is offline and intentionally has no credential or token authentication flow. If live Snowflake evidence is collected, use the organization's approved session/authentication process and record method names only; never place credentials in the inventory or report. This skill has no Edit authority and never edits local source or Snowflake objects. The Python analyzer may write only the explicitly requested sanitized report path via --out; never use that capability for credentials or mutation commands.

This package does not provide an MCP server, OAuth client, or token broker. Any connector is configured separately. Supply only sanitized read-only evidence and verify account, edition, connector, client behavior, and feature availability.

Non-negotiable boundaries

• Read-only planning. This skill never disables users, rotates keys, creates or alters integrations, changes authentication/network policies, or handles passwords, PATs, tokens, private keys, or client secrets.
• Do not invent a universal retirement date. Dates depend on the account, connector, runtime, feature availability, owner, and approved change window.
• A service without a named workload is an ownership gap, not permission to disable it. A LEGACY_SERVICE must be bound and tested before retirement.
• Snowflake-managed SNOWFLAKE_SERVICE rows remain in receipt cap accounting but are excluded from th…