ContractForge Snowflake Adapter

contractforge-snowflake is the Snowflake adapter for ContractForge.

The documented snowflake_sql_warehouse surface is stable-supported: real Snowflake validation has covered table, SQL and staged-file sources, hosted procedure execution, quality, schema policy, evidence/control tables, governance comments/tags, lineage/explain and cost reconciliation. Task graph live execution and the reference hash-diff production benchmark are also validated. SCD2/snapshot soft delete are excluded from stable-final. Continuous ingestion surfaces and account-feature-dependent access policy validation remain explicit review or final-certification boundaries.

The implementation is planning/publish/runtime-first:

• consumes ContractForge Core contracts through public core models;
• declares conservative Snowflake capabilities;
• builds publish bundles for a stable Snowflake library runner;
• runs supported contracts through the adapter runtime with Snowflake sessions;
• writes ContractForge evidence/control tables in Snowflake;
• renders project task graph deployment artifacts for schedule/timezone and step dependencies;
• keeps Snowflake connector and Snowpark dependencies optional.

pip install contractforge-core contractforge-snowflake

from contractforge_snowflake import plan_snowflake_contract

result = plan_snowflake_contract(contract)
print(result.status)

from contractforge_snowflake import build_snowflake_publish_bundle

bundle = build_snowflake_publish_bundle(contract)
print(bundle.artifacts["snowflake.publish_manifest.json"])

from contractforge_snowflake import run_snowflake_contract

result = run_snowflake_contract(
    contract_uri="@CONTRACTFORGE_ARTIFACTS/dev/runtime/ANALYTICS.BRONZE_CUSTOMERS.contract.json",
    environment_uri="@CONTRACTFORGE_ARTIFACTS/dev/runtime/ANALYTICS.BRONZE_CUSTOMERS.environment.json",
    session=snowflake_session,
)
print(result["status"])

The adapter does not generate per-contract ingestion SQL as the default execution model. Published contracts are consumed by the contractforge_snowflake runtime library. Project deployment artifacts may create Snowflake tasks, but those tasks only call the stable runner with contract and environment artifact URIs.

Governance annotations support table and column comments, plus Snowflake tags when tag objects are already provisioned. Use fully qualified tag names such as GOVERNANCE.PUBLIC.DOMAIN for deterministic live apply. To validate tag intent and record ctrl_ingestion_annotations evidence without executing tag DDL, set:

extensions:
  snowflake:
    annotation_tag_mode: validate_only

Access governance supports table grants, row access policy attachments and masking policy attachments, with each step recorded in ctrl_ingestion_access. Use access.mode: validate_only to verify grant and policy intent without applying DDL. revoke_unmanaged: true remains review-required because it can remove inherited or unmanaged Snowflake access.

Use contractforge-snowflake smoke-stage-publish --execute --execute-cleanup to live-test stage publication. The smoke creates a temporary internal stage, uploads the publish bundle, reloads the staged manifest and runtime artifacts, and runs the contract through the connector-backed library runner. Live smoke commands accept either --connect-options <yaml> or --connection <name> when the Snowflake Python connector can resolve that connection name, for example --connection cfingestsvc-pat.

Staged-file sources support CSV, JSON and Parquet batch reads from Snowflake stages. Provide a named Snowflake file format through source.options.file_format, or use a stage with a default file format. CSV sources can project positional fields with source.options.columns as either a list of names or a mapping of output names to Snowflake expressions. JSON and Parquet sources default to a payload VARIANT projection and can project typed columns with expressions such as $1:order_id::NUMBER.

source:
  type: staged_files
  path: '@RAW_STAGE/orders/orders.csv'
  format: csv
  options:
    file_format: RAW_CSV_FORMAT
    columns:
      order_id: '$1::NUMBER'
      status: '$2::STRING'
      amount: '$3::NUMBER(10,2)'

Use contractforge-snowflake smoke-procedure --execute --execute-cleanup to deploy and call the stable Snowpark runtime procedure. The service role needs CREATE PROCEDURE on the target schema. The smoke accepts the built core and adapter wheels locally, stages Snowflake-compatible ZIP copies, and imports those ZIP archives from the procedure:

GRANT CREATE PROCEDURE ON SCHEMA CONTRACTFORGE_TEST_DB.PUBLIC
  TO ROLE CONTRACTFORGE_INGEST_ROLE;

Use contractforge-snowflake smoke-task-graph --execute --execute-cleanup to deploy and manually execute a two-step task graph. In addition to the procedure grant, the service role needs task creation/execution privileges in the task schema; the live validation account now has these grants:

GRANT CREATE TASK ON SCHEMA CONTRACTFORGE_TEST_DB.PUBLIC
  TO ROLE CONTRACTFORGE_INGEST_ROLE;
GRANT EXECUTE TASK ON ACCOUNT TO ROLE CONTRACTFORGE_INGEST_ROLE;

Project CLI parity includes deploy-project, run-project, and cleanup-plan. run-project --dry-run renders the root task EXECUTE TASK commands without connecting; run-project --wait polls bounded INFORMATION_SCHEMA.TASK_HISTORY for terminal task states. cleanup-plan prints explicit drop commands for tasks and the runtime procedure but does not drop data target tables or staged artifacts.

Each runtime run records immediate lineage and explain evidence. The lineage row in ctrl_ingestion_lineage includes a ContractForge/OpenLineage-style event with the run id, source reference, target table, row count and Snowflake query ids. The explain row in ctrl_ingestion_explain captures EXPLAIN USING TEXT output for the rendered write statement. Disable plan capture for a contract with:

extensions:
  snowflake:
    explain_enabled: false

Native Snowflake lineage from SNOWFLAKE.ACCOUNT_USAGE.ACCESS_HISTORY is handled as delayed reconciliation because Account Usage can lag runtime execution. The helper probes by structured QUERY_TAG run_id and returns PENDING without inserting rows until matching Access History rows are visible.

contractforge-snowflake reconcile-lineage \
  --connect-options .tmp/snowflake-smoke/connect-options.yaml \
  --environment .tmp/snowflake-smoke/environment.json \
  --run-id "<contractforge-run-id>"

Use contractforge-snowflake reconcile-cost after a run to append delayed ctrl_ingestion_cost signals from Snowflake Account Usage. The command probes QUERY_HISTORY by structured QUERY_TAG run_id; if rows have not arrived yet or the service role cannot read Account Usage, it returns PENDING without inserting duplicate evidence and includes a warning when access is unavailable. When rows are available, it deletes prior adapter-owned cost signals for the same run_id/target table before inserting query-history and, when accessible, query-attribution signals.

To let the same service role query Account Usage, grant the Snowflake database role that covers QUERY_HISTORY and QUERY_ATTRIBUTION_HISTORY:

GRANT DATABASE ROLE SNOWFLAKE.GOVERNANCE_VIEWER
  TO ROLE CONTRACTFORGE_INGEST_ROLE;

contractforge-snowflake reconcile-cost \
  --connect-options .tmp/snowflake-smoke/connect-options.yaml \
  --environment .tmp/snowflake-smoke/environment.json \
  --run-id "<contractforge-run-id>" \
  --target-table '"CONTRACTFORGE_TEST_DB"."PUBLIC"."CF_SMOKE_APPEND_TARGET"' \
  --wait --max-wait-seconds 300

Architecture and implementation plan:

• docs/specs/snowflake-capability-parity.md
• docs/specs/snowflake-ga-criteria.md
• docs/specs/snowflake-ga-waivers.md
• docs/specs/snowflake-adapter-implementation-plan.md
• docs/specs/snowflake-adapter-parity-execution-plan.md