Synthetic Seeder

Synthetic Seeder Skill

Generate two-sided synthetic data tests for migrated stored procedures and functions. Tests are derived from source SQL analysis — the source is the oracle, not the converted code. Each test is a step-based YAML artifact directly consumable by scai test capture and scai test validate.

YAML schema reference. See ../../references/step-based-yaml.md for the step types (SQL Query / Output Parameter / Table Read / Cursor Read), the validate: true | false | [list] semantics, and the modifies_data / affected_tables inference rules. This file describes what to generate for each procedure; the cheat sheet is the source of truth for the shape of each step. For special cases (multi-RS, OUT params, cursor reads, before/after capture) see ../../migrate-object/EDIT_TEST_YAML.md.

Autonomous mode: When spawned as a subagent, skip interactive stopping points, apply reasonable defaults, document assumptions in the completion report, and send a send_message to main when done.

Scope and Boundaries

This skill does:

• Analyze source SQL to identify branches, NULL paths, boundary values
• Generate two-sided INSERT / CALL / SELECT steps in both source and Snowflake dialects
• Execute the Snowflake side to validate the test SQL (not the object)
• Write YAML artifacts with test_cases: [[]] for test-runner compatibility
• Enrich YAMLs with modifies_data and affected_tables for clone isolation
• Fix test SQL if Snowflake execution fails (up to 3 retries)

This skill does NOT:

• Repair the migrated object — delegate failures caused by object bugs to migrate-object
• Run uv run migrate or run_migration_tests — test execution is handled by scai test capture + scai test validate
• Generate tests for TABLEs or VIEWs (procedures and functions only)
• Generate tests for BTEQ scripts. Units with kind: "script" (BTEQ) author a validation.steps[].script block per ../seed-script/SKILL.md from user-provided bindings + import fixtures — they are not synthesized here. Do not invent test_cases: for them.

Available MCP Tools

Output Format

Each test is a YAML file following the ValidationMetadata schema:

validation:
  test_cases:
  - []
  steps:
  # Setup steps — validate: false
  - source_query: <source-dialect INSERT or DDL>
    target_query: <Snowflake INSERT or DDL>
    validate: false
  # Call steps — validate: true (default, key omitted)
  - source_query: <EXEC / source CALL>
    target_query: <Snowflake CALL>
  # Assertion steps — validate: true (default, key omitted)
  - source_query: <SELECT from modified table, source dialect>
    target_query: <SELECT from modified table, Snowflake>

IMPORTANT: Every YAML must include test_cases: [[]] (a list containing one empty list). This is required by test-runner discovery. Step-based tests embed all data inline in the steps, so the test case list contains a single empty-parameter entry.

File layout: The object's test directory is whatever files.artifacts.path returns from the registry (Phase 0.4) with /test/ appended — read it, do not construct it. It may use a placeholder database segment (e.g. database_unspecified) and lowercased names for some dialects; that is correct and must be preserved. files.artifacts.path is PROJECT-RELATIVE, so prefix it with <project_dir>/ for all filesystem writes/reads (capture resolves it as project_root / files.artifacts.path / test).

<project_dir>/<files.artifacts.path>/test/
  <sanitized_object_name>.0.yml   ← one data arrangement
  <sanitized_object_name>.1.yml   ← another data arrangement

<sanitized_object_name> (for the FILE name only) replaces non-alphanumeric characters (except . and -) with _.

YAML granularity rule:

• One YAML per data arrangement — scenarios that require distinct INSERT data get their own file
• Multiple calls/selects in one YAML — scenarios exercisable against the same data state are additional steps within the same file

Qualities of a Good Test

Before generating, internalize these rules:

1. Deterministic — no CURRENT_TIMESTAMP, NOW(), GETDATE(), RANDOM() or similar. Use fixed constant dates and values only.
2. Non-trivially non-empty — each CALL or SELECT step must produce at least 3 distinct, non-empty output rows. Tests producing empty or single-row output rarely catch behavioral differences.
3. Covers all modified tables — include SELECT assertions for every table the procedure writes to, not just the return value or the primary output table.
4. Branch-driven — read the source SQL for IF/CASE/WHILE branches, NULL-handling paths, boundary values, and parameter variations. Each distinct data arrangement that exercises a different branch gets its own YAML file.
5. Minimal data — populate only columns required by the test logic plus NOT NULL constraints. Use short strings, small numbers, and brief identifiers.
6. High-numbered keys — use values like 9901, 9902 to avoid colliding with production data.
7. Idempotent setup — prefer INSERT ... SELECT ... WHERE NOT EXISTS (SELECT 1 FROM t WHERE pk = ?) so tests are safe to re-run.
8. Fully qualified names — always use SCHEMA.TABLE in all SQL, both source and Snowflake sides.
9. No object creation — never CREATE TABLE or CREATE VIEW in test steps. Only insert into tables that already exist as dependencies. If a required table is missing, that is a dependency bug — report it and stop.
10. Source is the oracle — derive expected behavior by reading the source SQL logic, not the Snowflake conversion.

Phase 0: Resolve Object

0.1 — Get session config

Call configure(). Extract:

• project_dir — root of the migration project
• source_dialect — e.g. Teradata, MS SQL Server
• snowflake_database — target Snowflake database

0.2 — Identify the object

This skill expects an <object_name>.

0.3 — Check skip condition

We skip objects that already have a seeded test YAML. The test directory is whatever files.artifacts.path reports (fetched in 0.4) — do not reconstruct it from the object name — so this check runs in 0.4 (Step 1.2), once that value is known.

0.4 — Locate SQL files

CUR is the source of truth for file paths — use query_registry instead of find so we do not rely on filename heuristics.

Step 1 — Get the object's files and dependency IDs. Call query_registry with:

• where = "source.canonicalName ilike '%<object_name>%'"
• fields = "id,source,target,files,dependencies"
• include_dependencies = true

From the response, read both SQL files (resolve paths relative to <project_dir>):

• files.source.path — original source SQL
• files.converted.path — converted Snowflake SQL
• files.artifacts.path — base directory for this object's test artifacts (used in Phase 4)

Collect every dependencies.dependsOn[*].id for Step 2.

Step 1.1 — Oracle package member check: if source.package is non-empty in the registry entry, this is a package member. Record:

• source_package = value of source.package (e.g. "ITEM_MGMT")
• source_schema = value of source.schema (e.g. "TEST")
• target_schema = value of target.schema (e.g. "TEST_ITEM_MGMT" — SnowConvert convention: {source_schema}_{PACKAGE_NAME})

These three values are used in Phase 3 to construct the correct call syntax.

Step 1.2 — Skip if already tested (the 0.3 check). Now that files.artifacts.path is known, check whether this object already has a seeded test YAML:

find <project_dir>/<files.artifacts.path>/test -name "*.yml" -type f

If any .yml is found: skip this object, report it as already tested, and (if looping) call next_object() for the next candidate — no need to fetch dependencies.

Step 2 — Get the dependency files in one call. Call query_registry again with:

• where = "id in ('<id1>', '<id2>', ...)" (comma-separated IDs from Step 1)
• fields = "id,source,files"

For each result, read files.source.path and files.converted.path. Skip entries where files is missing (built-in or external object).

You now have the full picture of what tables exist, their columns, constraints, and what the procedure modifies.

Phase 1: Analyze and Plan

1.1 — Identify modified tables

From the source SQL, list every table the procedure writes to (INSERT, UPDATE, DELETE, MERGE, TRUNCATE). These tables must have SELECT assertion steps in every YAML.

1.2 — Identify branches and scenarios

Read the source SQL carefully. For each logical branch, note:

• The condition that activates it (e.g. IF @status = 'ACTIVE', CASE WHEN amount > 1000)
• What tables are modified differently per branch
• NULL or empty-input edge cases
• Boundary values (zero, negative, maximum, date boundaries)

1.3 — Produce coverage matrix

Before generating any SQL, produce a table of planned YAMLs. Each row is one file:

YAML index | Scenario description          | Key distinguishing data         | Branches covered
---------- | ----------------------------- | --------------------------------| ----------------
0          | Happy path — status ACTIVE    | status='ACTIVE', amount=500     | main INSERT branch
1          | High-value path               | status='ACTIVE', amount=1500    | threshold CASE branch
2          | NULL input                    | status=NULL                     | NULL guard / COALESCE branch

STOPPING POINT (interactive mode): Present the coverage matrix and ask the user to confirm or adjust before generating. Autonomous agents: proceed and document the matrix in the completion report.

Phase 2: Arrange — Generate Insert Steps

For each YAML in the coverage matrix, generate the setup steps.

Rules for insert steps

• Each INSERT statement is one step with source_query and target_query
• validate: false on every setup step
• Source and target inserts carry the same data values; only the SQL dialect differs
• Batch multiple rows in a single INSERT where possible (saves steps):

  INSERT INTO SCHEMA.T (col1, col2) VALUES (9901, 'A'), (9902, 'B');
  

• For columns where source dialect uses type casting or date literals differently, adapt per side
• Do not use time/date conversion functions inside VALUES clauses on the Snowflake side — use SELECT ... AS col pattern instead:

  -- Invalid in Snowflake inside VALUES
  INSERT INTO T (dt) VALUES (TO_TIMESTAMP_NTZ('2024-01-01'));
  -- Correct
  INSERT INTO T (dt) SELECT TO_TIMESTAMP_NTZ('2024-01-01') AS dt;
  

• Respect NOT NULL constraints — always populate those columns
• For CHAR(N) columns in source dialect, pad values with trailing spaces as required

Phase 3: Act + Assert — Generate Call and Select Steps

Call steps

• One CALL (Snowflake) / EXEC or CALL (source) step per scenario or parameter combination
• For procedures with multiple output parameters, include SELECT statements immediately after the call to read their values
• Use :out_param syntax on the Snowflake side for OUT parameters
• If the source uses EXEC proc @param = value syntax, translate to positional CALL proc(value) on the Snowflake side

Dialect-specific call guidance:

• Oracle — load ./platforms/oracle.md for: anonymous PL/SQL block execution model, source_declare syntax, package member three-part FQN, artifact path conventions, OUT params / SYS_REFCURSOR.
• Teradata — load ./platforms/teradata.md for: cross-database GRANT steps required before each CALL in clone-isolated environments.

Assertion SELECT steps

For every table identified as modified in Phase 1, add a SELECT step:

SELECT <essential_columns> FROM SCHEMA.TABLE ORDER BY <pk_col>;

Rules:

• Always include ORDER BY for deterministic output comparison
• Select only columns relevant to the test scenario — avoid SELECT * when many columns are noise
• Do not add WHERE clauses that would hide the procedure's effects
• On the Snowflake side, align column names and aliases to match the source side (the runner does semantic comparison but matching names help)

Phase 4: Write Artifacts

4.1 — Write YAML files

For each successfully validated YAML (index 0, 1, 2, ...), write to:

<project_dir>/<files.artifacts.path>/test/<sanitized_name>.<idx>.yml

files.artifacts.path is the value fetched from query_registry in Phase 0.4, Step 1 — use it as-is rather than reconstructing it. <sanitized_name> (for the FILE name only) replaces non-alphanumeric characters (except . and -) with _. Write each file under <project_dir>/<files.artifacts.path>/test/.

Fail fast when the CUR has no artifacts path. If files.artifacts.path is missing or empty, stop and report a CUR/dependency bug. A database_unspecified database segment is NOT a bug — it is the legitimate source-DB placeholder emitted by the arranger; use it as-is.

Oracle package members: see ./platforms/oracle.md — the CUR files.artifacts.path already includes the package subdirectory; always read it from query_registry.

YAML content — always include test_cases: [[]]:

validation:
  test_cases:
  - []
  steps:
  - ...

Phase 5: Enrich YAMLs with Isolation Metadata

After writing YAMLs, populate modifies_data and affected_tables by combining DML tables extracted from the test steps with the code unit's transitive dependencies from CUR.

The bundled script handles only the YAML scan and rewrite — the agent must fetch CUR dependencies via the query_registry MCP tool and pass them to the script via --extra-tables.

5.1 — Fetch dependency tables from CUR

Call the query_registry MCP tool with:

• where = "source.canonicalName ilike '%<object_name>%'"
• fields = "id,source,target,dependencies"
• include_dependencies = true

From the returned JSON, take entries in dependencies.dependsOn[*] whose source.objectType == 'table', and use each entry's target.canonicalName as the FQN (e.g. ["DBO.CUSTOMERS", "DBO.ORDERS"]). Non-table dependencies (views, procedures, functions) are dropped — affected_tables must contain tables only.

If the unit has no table dependencies, this list is empty — the script will still enrich based on DML in the steps.

5.2 — Run the enrichment script

python <SKILL_DIR>/enrich_yamls.py <project_dir> "<object_name>" \
  --extra-tables "<fqn>,<fqn>,..."

Where <SKILL_DIR> is the directory containing this SKILL.md (use the absolute path), and --extra-tables is the comma-separated list of FQNs from Step 5.1. Omit --extra-tables (or pass an empty string) when CUR returned no dependencies.

The script:

1. Scans each YAML for INSERT/UPDATE/DELETE/MERGE in source/target query steps
2. Unions those with the FQNs from --extra-tables into affected_tables
3. Sets modifies_data: true if any affected tables found, false otherwise

Phase 6: Self-Check and Report

6.1 — Self-check

Before returning, verify every committed YAML exists on disk and contains isolation metadata:

find <project_dir>/<files.artifacts.path>/test -name "*.yml" -type f

Count must equal the number of YAMLs in the coverage matrix that passed validation. If any file is missing: generate it before returning.

6.2 — Report

Report to the user (or send_message to main in autonomous mode):

Test generation complete for <object_name>:

  YAMLs planned:    N  (coverage matrix)
  YAMLs validated:  M  (Snowflake execution passed)
  YAMLs written:    M
  YAMLs enriched:   M  (modifies_data + affected_tables added)
  Object bugs found: K  (if any — needs migrate-object)

  Scenarios covered:
    0: <scenario description>
    1: <scenario description>
    ...

  Artifacts: <project_dir>/<files.artifacts.path>/test/

  Next: run `scai test capture` and `scai test validate` to execute the tests.

Note (when called from migrate-objects/baseline-capture/SKILL.md): the caller loads baseline-capture/CAPTURE.md immediately after this skill returns to run scai test capture and upload baselines. The migrate-object loop then runs scai test validate each iteration. Do not run capture/validate yourself in that flow — just report completion so the parent can continue.

Snowflake SQL Dialect Rules

Apply these when writing target_query steps:

• Use CALL to invoke procedures — never SELECT or EXEC
• For table-returning functions: SELECT * FROM TABLE(fn(args))
• OUT parameters in CALL: CALL proc(in_val, :out_var)
• DECIMAL/NUMERIC/DEC are synonyms for NUMBER — do not replace them
• TRUNC(x) before CAST(x AS INT) when source T-SQL uses integer truncation
• Date/time values in INSERTs: use SELECT ... AS col pattern (not inside VALUES)
• Do not construct identifiers from string concatenation with IDENTIFIER()
• Fully qualify all object references: DATABASE.SCHEMA.OBJECT

Troubleshooting

INSERT fails: "column does not exist"

Read the dependency DDL file again — column name may differ between source and Snowflake sides. Use the Snowflake DDL column names in target_query, source DDL names in source_query.

INSERT fails: "NOT NULL constraint"

Identify the NOT NULL columns from the DDL and add them to the INSERT. Use minimal placeholder values (0, 'X', '1900-01-01').

CALL fails: "object does not exist"

The procedure or a dependency is not deployed. This is an object bug — stop and report. Do not retry.

CALL produces empty output

The test data does not satisfy the procedure's WHERE clauses or join conditions. Trace the source SQL logic and adjust the INSERT values so the rows match the procedure's filter predicates.

Row count less than 3

Add more INSERT rows covering additional input combinations. Each row should produce a distinct output row.