reviewing-model

name	reviewing-model
description	Use when the generate-model workflow needs a read-only review of generated dbt model artifacts before accepting the result or sending it back for fixes.
context	fork
user-invocable	false
argument-hint	<schema.table>

Reviewing Model

Read-only quality gate for one generated dbt artifact set. Review written SQL/YAML against the source routine, approved test spec, and shared standards, then return exactly one ModelReviewResult.

Quick Reference

• Input: $ARGUMENTS = target table FQN (schema.table)
• Guard: generating-model readiness via migrate-util ready generate --object <table_fqn> --project-root <project_root>
• Read: migrate context, generated SQL/YAML, test-specs/<item_id>.json
• Check: artifact fundamentals, correctness, test integration, standards
• Never: modify files, run dbt, override profile decisions
• Output: follow model-review-output.md
• Codes: follow review-codes.md

Workflow

1. Run the generating-model readiness guard.

   uv run --project "${CLAUDE_PLUGIN_ROOT}/packages/ad-migration-internal" migrate-util ready generate \
     --object <table_fqn> \
     --project-root <project_root>
   

   If ready is false, return a valid ModelReviewResult with status: "error" and the surfaced code and reason. Do not gather context, read generated artifacts, run review checks, or write feedback after a failed readiness check.

2. Gather context and locate artifacts.

   Read and apply review-inputs.md. Use selected_writer_ddl_slice as the correctness ground truth when present. Otherwise use proc_body. If both are empty, return a context error.

3. Verify artifact fundamentals.

   Read the written artifacts directly from disk and apply model-artifact-invariants.md plus the naming contract in model-naming.md. Do not trust generator self-reported booleans such as generated.*.

   Artifact invariant failures belong under checks.standards; never add a checks.artifact_fundamentals key. Missing blocker controls such as _dbt_run_id or _loaded_at must return revision_requested, not approved.

4. Review correctness.

   Apply the correctness checklist in review-checks.md. Use REVIEW_CORRECTNESS_GAP in both checks.correctness.issues[] and the matching feedback_for_model_generator[] item for any correctness failure.

5. Review test integration.

   Apply the test-integration checklist in review-checks.md. Use REVIEW_TEST_INTEGRATION_GAP in checks.test_integration.issues[] for any failure.

6. Review standards.

   Apply the standards checklist in review-checks.md and all shared standards it references. This is exhaustive: direct source() usage in mart models for confirmed source tables, and direct source/raw warehouse usage for seed dependencies, must kick back.

   Classify each dependency before assigning standards codes. If catalog metadata marks the relation is_seed: true, direct source() or raw warehouse references are MDL_017, not MDL_016.

   Use REVIEW_STANDARDS_VIOLATION in checks.standards.issues[]. Report every directly observable stable standards code in feedback_for_model_generator[].

Steps 3-6 always run unless a prerequisite/read/parse failure prevents review.

Verdict

Return:

• approved if all three review categories pass
• revision_requested if standards, correctness, or test-integration issues exist
• error only for prerequisite, IO/parse, or missing-artifact/test-spec failures

When returning revision_requested, populate feedback_for_model_generator[] and add REVIEW_KICKED_BACK to warnings[].

Do not emit approved_with_warnings; /generate-model owns max-review-loop and soft-approval policy.

Boundary Rules

This skill is read-only. It must not modify files, run dbt commands, or override profile decisions.