Fluid Forge Docs Baseline: CLI 0.14.0

Release Date: July 28, 2026 Status: Current stable docs baseline (supersedes 0.13.0)

This baseline covers two CLI releases: 0.13.1 (July 26) and 0.14.0 (July 28). 0.13.1 did not receive its own docs pass, so both change sets are documented here.

Headline

0.14.0 is the live-verification hardening release, and 0.13.1 beneath it is the interoperability release that set it up.

The headline is the dbt Iceberg loop reaching all three cloud warehouses. 0.13.1 shipped both halves for Snowflake: fluid generate emits dbt's catalogs.yml for Iceberg exposes — Snowflake Horizon (built_in) plus the glue / polaris / unity / rest / nessie REST-family catalogs — and fluid apply provisions the prerequisites dbt refuses to create: the EXTERNAL VOLUME and the AWS Glue CATALOG INTEGRATION (#469, #470). 0.14.0 extends the loop to BigQuery: an iceberg expose on a GCP binding — which previously fell through the emit dispatch and produced nothing, silently — now emits catalogs.yml with catalog_type: biglake_metastore and provisions the GCS bucket dbt refuses to create (#474). The whole surface sits behind a new validate-time anti-no-op gate: every case where an IaC emitter would silently skip an Iceberg expose is now a fluid validate error naming the missing field, and the check↔skip-branch pairing is test-pinned in both directions (#475).

Alongside it, the 104-defect live-verification wave (#476, #478): a month of shipped features was exercised end-to-end against a real Snowflake account, every fix re-verified by a second agent re-running the original failing scenario. The headline finding — embedded-SQL builds on a platform: snowflake contract executed against local DuckDB and exited 0 either way. Snowflake builds now actually run on Snowflake; the rest of the wave is itemized below.

pip install --upgrade data-product-forge.

What changed in v0.14.0

Added — the dbt Iceberg loop reaches BigQuery

An iceberg expose on a GCP binding previously emitted nothing — the dispatch only handled bigquery_table / view / gcs_bucket / pubsub_topic, so it fell through silently. Now (#474):

• fluid generate emits catalogs.yml with catalog_type: biglake_metastore. BigQuery's shape genuinely differs from Snowflake's: external_volume is a bare gs:// URI (not the name of a pre-existing object), and file_format is required (Snowflake has no such key).
• fluid generate iac provisions the GCS bucket dbt refuses to create, reusing the plain gcs_bucket emit so settings, labels, and access-grant IAM stay identical.
• The bucket dbt loads into is always the bucket the IaC creates and governs. The IaC bucket name is derived from the same warehouse URI dbt writes into external_volume — one shared helper pair, so the two sides cannot diverge. A foreign scheme (s3:// under biglake_metastore) or a bucket-less URI makes both sides skip together.
• Whole-bucket force_destroy is dropped when the product owns only a prefix of a shared warehouse root.

Databricks stays out deliberately: its catalogs.yml shape is unverified, and emitting a guess is worse than emitting nothing.

Added — validate-time anti-no-op gate for Iceberg prerequisite bindings

The Snowflake and GCP IaC emitters skip an Iceberg expose missing a required input rather than emit a broken resource — which previously meant the user found out at dbt run. Now fluid validate errors on exactly those cases, naming the missing field (#475). Each validate check mirrors one emitter skip-branch, and the pairing is test-pinned in both directions — a new skip-branch without a matching check fails the suite, and vice versa.

Note the --strict consequence in the warning block above: Snowflake catalogs that authenticate with secrets now fail strict validation, since the emitted module is credential-free.

Fixed — the 104-defect live-verification wave

A month of shipped features exercised against a real Snowflake account; every fix re-verified by a second agent re-running the original failing scenario (#476, #478). The clusters:

• Snowflake builds actually run on Snowflake. Embedded-SQL builds on a platform: snowflake contract executed against local DuckDB (exit 0 either way); --model-contracts flattened every parameterized / alias type to VARCHAR(16777216), so contract enforcement passed vacuously; generated profiles.yml ignored the contract's schema and silently targeted PUBLIC; a freshly generated project failed its own contract on first dbt run; and a freshness dq rule made the generated project unparseable.
• fluid import dbt → apply → verify round-trips cleanly. Importing then applying no longer creates a duplicate lowercase shadow namespace (+2 ~2 -0 → +0 ~2 -0), verify can read the real objects through snowflake_view, and a p90 metric no longer silently round-trips into a median.
• fluid publish no longer silently skips contract-declared catalog targets (an ImportError swallowed by a bare except), and apply hooks now run on the OpenTofu path — making --env plumbing reachable for Snowflake / AWS / GCP applies.
• Governed access enforcement holds under aliasing. The governed query path now enforces policy.authz.columnRestrictions (a denied column was readable as a measure), and PII redaction can no longer be bypassed by aliasing a restricted column — policy was previously checked on the output column name, which the semantic layer aliases away.
• Data-quality test reporting is truthful. A failing severity: critical rule no longer reports PASS with exit 0; accuracy rules evaluate upper bounds instead of only MIN(); summary.checks_passed no longer counts warnings and failed criticals as passed; --no-data no longer asserts checks it never performed; and --engine soda can now actually produce checks (the schemas never defined the exposes[].quality.tests[] key it reads). See fluid test.
• ODCS export/import fidelity. Export no longer flattens every parameterized type to logicalType: string, round-trips no longer lose 88 of 114 leaf fields, and fluid odcs import no longer rewrites a Snowflake binding to bigquery while reporting success — now validated against the official ODCS schema.
• --adopt-shared-container respects per-exposure binding.packaging overrides. Flipping one binding no longer absorbs a shared platform pool into the tenant's state (which erased the pool's COMMENTs); transition scoping now keys on which container is nested where, not on resource type.

What changed in v0.13.1

Added — Snowflake Iceberg loop for dbt, both halves

fluid generate emits dbt's catalogs.yml (v1 schema, Snowflake adapter) for Iceberg exposes, and fluid apply provisions the prerequisites dbt refuses to create — the EXTERNAL VOLUME for Snowflake-managed catalogs and the AWS Glue CATALOG INTEGRATION (#469, #470). A single deterministic naming helper is shared by both emitters, so the volume fluid apply creates carries exactly the name catalogs.yml references (explicit override honoured via binding.icebergConfig.properties). Live-verified with a real tofu apply. See the Snowflake provider.

Added — fluid import dbt --split-by

fluid import dbt --split-by {project|folder|group} splits a dbt manifest into multiple data products along folder or dbt-group boundaries (project remains the byte-stable single-contract default). Cross-split ref()s become cross-product consumes[], --out becomes the output directory for multi-contract imports, and dbt ≥1.10 manifests with arguments:-nested test params now import correctly alongside legacy flat kwargs (#465). In passing, model-level expose descriptions are now scrubbed like their column/semantic siblings (closing a hostile-Jinja smuggling path into generated schema.yml).

Added — Bitol ODPS v1.1.0 and declared consumers

• Bitol ODPS v1.1.0 top-level type (approved RFC 0029): sourceAligned / aggregate / consumerAligned map 1:1 and bidirectionally to FLUID's SDP / ADP / CDP classification. The default emit target stays v1.0.0 until Bitol cuts the release — opt in via --api-version / ODPS_API_VERSION. Validation keys on the document's own apiVersion, and custom org types round-trip verbatim (#471). See fluid exporters and fluid odps-bitol.
• Declared consumers: block in the 0.7.6 preview schema. Contracts can now declare their downstream consumers — dashboards, notebooks, ML systems, applications — with a shape borrowed from dbt exposures (name / label / type / owner / url / maturity) plus FLUID-native exposeIds tying a consumer to specific output ports. Additive and optional; preview schema only, no GA change (#466).

Fixed — OpenLineage events real consumers can ingest

The emitter historically produced payloads no OpenLineage consumer would accept (missing required producer / schemaURL, flattened run / job structure, non-UUID runId) — and was never wired up, so zero events were ever sent. Emission now routes through openlineage-python at the acquisition-runner chokepoints (all six engines, no per-runner wiring), honours the standard OPENLINEAGE_URL so an existing Marquez / DataHub deployment just works, and redacts run facets and stream names before anything leaves the machine. ODCS contract publishing also moves onto OpenMetadata's first-class Data Contracts entity (#467).

Also in 0.13.1: fluid-schema-0.7.5.json no longer carries UTF-16 surrogate-pair escapes that broke YAML-based schema consumers such as datamodel-code-generator (#451), and the safety-gate override audit events moved to WARNING (#450 — see the warning block above).

Security — across both releases

• binding.location.dbFile is now confined to --readable-paths (0.13.1, #463). The DuckDB output-port driver gated binding.location.path and .attach against the operator's allowlist but passed dbFile raw to duckdb.connect, so a served contract could open and read any host database — a read-side sandbox escape. dbFile now flows through the same resolve-and-contain gate as its siblings (:memory: passes through; no-allowlist behaviour is unchanged).
• Document-controlled ids are contained in exporter filenames (0.13.1, #472). A contract with a traversal-shaped exposeId run through fluid generate artifacts --out dist could write an attacker-influenced file outside --out — and MANIFEST.json would bless the escaped path. New providers/_path_safety.py (the filename sibling of _sql_safety.py) passes schema-valid FLUID ids through verbatim, cleans-plus-digests anything else, and re-checks the resolved path against the output root at all three write sites. Verified against 204k+ separator/control-char cases and a full Unicode sweep with zero escapes and zero canonical-layout changes.
• fluid viz-graph escapes contract-authored text before it reaches generated DOT source (0.14.0, #481). Mesh node/edge labels were interpolated verbatim, so a product label like X" label="SPOOFED injected a second label attribute and let a contract display a name of its choosing — including another product's — in the rendered mesh graph; node IDs interpolated through a denylist that missed " and \ let a consumes[].ref inject phantom nodes into the lineage graph. Both shared helpers are fixed (allowlist IDs, backslash-then-quote label escaping).
• Secret redaction rewritten to redact by exact known value (0.14.0, #478). The previous regex-based approach guessed a secret's extent from its shape and leaked 17 known inputs; redaction is now delimiter-agnostic by construction, masking the literal values it knows.
• Commits carrying values from the committer's own environment are refused (0.14.0, #479). A new pre-commit/CI hook flags any diff line byte-identical to a non-trivial value in the committing machine's environment (SNOWFLAKE_*, AWS_*, *_TOKEN, *_PASSWORD, …) or a supplied --env-file — catching the low-entropy identifiers (account locators, usernames) that secret-shape scanners cannot see — and never prints the matched value. detect-secrets now also actually runs in CI, on changed files.
• Vulnerability reports route through GitHub Private Vulnerability Reporting (0.14.0, #473).

Compatibility

• No breaking changes to existing contracts. 0.7.5 remains the stable schema default; the consumers: block is 0.7.6 preview, opt-in via fluidVersion: "0.7.6".
• fluid validate --strict is stricter on Snowflake Iceberg. Secret-authenticating catalogs (polaris / unity / rest / nessie) now fail under --strict — see the warning block.
• Safety-gate override events log at WARNING. If your log pipeline alerted on these at INFO, adjust the filter; names and payloads are unchanged.
• Data-quality summaries may newly fail. The truthful-reporting fixes mean CI lanes that previously passed on vacuous checks (critical failures counted as passed, --no-data assertions, DuckDB-instead-of-Snowflake builds) can now go red — for real reasons.
• ODPS emit target unchanged. v1.0.0 remains the default; v1.1.0 type is opt-in via --api-version / ODPS_API_VERSION.
• SDK / custom-scaffold: unchanged (data-product-forge-sdk 0.10.0, data-product-forge-custom-scaffold 0.4.0).
• Install: pip install --upgrade data-product-forge → 0.14.0.