Proc Test Gen

Proc Test Generator (scripting flavor)

Generate and run a functional-equivalence check for a single converted mapping procedure — an Informatica mapping converted to a standalone Snowflake Scripting stored procedure — from the original data-flow logic in the source mapping. The expected target-table contents are computed from the source mapping applied to the seeded input, NOT from the converted SQL under repair, so the assertions are an independent oracle. All generated artifacts are stored in <UNIT>/stabilization/tests/proc/<proc_name>/.

Scope. This is the data-flow counterpart to orchestration-test-gen: orchestration-test-gen tests control-flow elements of a task graph; proc-test-gen tests the output table of one mapping proc. It is the scripting-flavor analog of dbt-test-gen. In dbt mode a mapping becomes a model (tested by dbt-test-gen); in scripting mode a mapping becomes its own kind=etl stored procedure, tested here.

Autonomous mode: When spawned as a team agent, skip all interactive stopping points, make reasonable defaults, and document assumptions in the completion report. When done, send_message to main with your results, and respond to any shutdown_request with type: "shutdown_response", approve: true.

Where this sits in the fix loop

block-locate  ── the defective block + its source transform node (scripts/stabilization_tools.py block-locate, marker-scoped)
proc-fixer    ── [agentic] author the fixed block SQL from the source node (sibling sub-skill)
block-replace ── splice the fix back, marker-integrity checked (scripts/stabilization_tools.py block-replace)
deploy-and-call ─ deploy the one proc into the test schema + smoke-CALL it (scai code deploy-and-call)
proc-test-gen ── [THIS SKILL] derive the source-of-truth expected target contents + batched assertions
grade         ── run the assertions against the populated target; PASS/FAIL feeds the fixer loop

proc-test-gen owns the ARRANGE (input seed) and ASSERT (expected target) — the stable, source-derived parts. The ACT (deploy + CALL the proc) is performed via deploy-and-call; the proc body itself is the baseline snapshot the fixer iterates on.

AAA for a single mapping proc

• Arrange:Setup — create the test schema + the source/input tables the mapping reads (schema inferred from the source mapping's Source/Source-Qualifier definitions), and the target table shape. Run once.
• Arrange:Seed — TRUNCATE + INSERT synthetic input rows into the source tables. Run before every ACT/grade cycle to reset to a known state. The seed is the fixed input the expected output is computed from — it must not change once assertions are derived.
• Act — deploy the (current) converted proc into the test schema and CALL it, so its output lands in <test_schema>.<TARGET>. Done with scai code deploy-and-call (see Tooling); on the baseline run the proc may fail to CREATE or error at CALL — that is expected and recorded.
• Assert — a single batched query over only the target table that verifies its contents equal what the source mapping produces for the seeded input (see proc-assertion-patterns.md).

Input

• session_status.json — with source_file_path, the pending mapping proc unit, and migration_object_path
• The converted proc .sql file (a kind=etl unit's files.converted.path)
• The source mapping definition (Informatica .xml) at source_file_path
• The block FullName(s) under repair (from the block boundary markers / scan.json)

Prerequisites

• Python 3.11+; the deterministic tool layer present (scripts/stabilization_tools.py — the block-locate / ewi-extract subcommands — and scripts/scan_unit.py)
• The headless executor available: scai code deploy-and-call (deploy one proc + CALL, returns a ProcCallResult JSON) — see Tooling
• An active Snowflake connection with a warehouse, and CREATE TABLE / CREATE PROCEDURE on a user-provided (test) database + schema

Output Location

<UNIT>/stabilization/tests/proc/<proc_name>/
  <proc_name>.assert.sql     # the batched target-only assertion query (source-derived)
  <proc_name>.seed.sql       # ARRANGE:SETUP + ARRANGE:SEED (input tables + synthetic rows)
  test_report.md             # baseline result, consumed by proc-fixer

The value of this skill is the seed (the fixed input) and the assert (expected target contents from source semantics). The proc body is the baseline; proc-fixer iterates it until the assert passes.

Tooling (headless, no IDE / no snowflake_sql_execute)

Unlike orchestration-test-gen (which runs SQL via the CoCo snowflake_sql_execute MCP tool), the scripting loop is headless:

Grade-query execution. The batched assert query is run against the populated target and its PASS/FAIL rows read back with snow sql -c <conn> --database <db> --schema <test_schema> --format json -q "<assert>" — the same headless path the deploy-and-call executor itself shells to, so this step needs no extra plumbing.

Workflow

Proc Test Gen Progress:
- [ ] Phase 1: Configure test env + read source mapping + block-locate the proc
- [ ] Phase 2: Design the input seed (ARRANGE) from the source mapping's sources
- [ ] Phase 3: Compute expected target contents from source semantics + write the batched assert
- [ ] Phase 4: Seed + deploy-and-CALL (baseline) + grade; record result
- [ ] Phase 5: Report (test_report.md)

Phase 1: Configure + read source

1. Test environment — reuse session_status.json's test_environment if present; else detect the current connection (SELECT CURRENT_DATABASE(), CURRENT_SCHEMA(), CURRENT_WAREHOUSE()) and store it via track_status.py set-test-env. Interactive mode: confirm with the user (⚠️ STOPPING POINT). Autonomous: use the current connection and proceed.
2. Read the source mapping ONCE — from source_file_path, extract for this mapping: the Source / Source-Qualifier definitions (input table shapes + columns), the transform chain (Filter / Aggregator / Lookup / Joiner / Expression / Rank / Union / Router / Update-Strategy …), and the Target definition (output table + column mapping). See {PLATFORM_DIR}/{transformation_guide} (Informatica: mapping-guide.md) for element names/paths/expressions and {PLATFORM_DIR}/element-types.md for what each transform type does.
3. block-locate the proc — run stabilization_tools.py block-locate --json to get the materialized block(s) and their source node(s). Build a mental map of source transform → block; do not re-read per assertion.

Phase 2: Design the input seed (ARRANGE)

From the source mapping's Source/Source-Qualifier definitions, create the input tables and seed synthetic rows. Design the seed to exercise the transform semantics that must be verified:

• at least one row that flows through to the target, and one that is dropped/filtered;
• for a Lookup: an input whose key matches and one whose key does NOT (no-match → NULL);
• for an Aggregator: enough rows per group to make the aggregate non-trivial (≥2 groups, ≥2 rows/group);
• NULL combinations for columns used in NULL-sensitive expressions;
• for Router/Update-Strategy: rows exercising ≥2 branches / ≥2 ops.

Write ARRANGE:SETUP (CREATE OR REPLACE input + target tables in the test schema) and ARRANGE:SEED (TRUNCATE + INSERT ... SELECT the synthetic rows) to <proc_name>.seed.sql. The seed is now frozen — the expected target is computed from exactly these rows.

Phase 3: Compute expected target + write the batched assert

Apply the source mapping semantics to the seeded input by hand to compute the expected target contents — do NOT read the converted SQL. Faithful Informatica semantics:

• Lookup: no-match → NULL for the looked-up columns.
• Filter: drops rows whose predicate is not TRUE (a NULL predicate drops the row).
• Aggregator: GROUP BY the group-by ports + the stated aggregates (SUM/AVG/COUNT/MIN/MAX); ungrouped aggregate = whole input.
• Joiner: the stated join type (Normal = inner, Master/Detail outer per the join condition).
• Rank: top/bottom N per the rank port, partitioned by the group ports.
• Union: concatenation of inputs (no dedup unless stated).
• Expression / ports: apply the port expressions in dependency order; preserve declared output types.

Then write a single batched query to <proc_name>.assert.sql in this exact shape:

  SELECT 'assert_01_<name>:' || CASE WHEN (<boolean>) THEN 'PASS' ELSE 'FAIL' END AS r
  UNION ALL SELECT 'assert_02_<name>:' || CASE WHEN (<boolean>) THEN 'PASS' ELSE 'FAIL' END
  UNION ALL ...

Cover, at minimum: the exact expected value(s) for each expected group/row in the target, the total expected row count, and that dropped rows are ABSENT.

Grade hardening (this is core — see proc-assertion-patterns.md): each assertion MUST be a single, simple boolean referencing only the target table — a COUNT(*) comparison, a scalar value comparison for one group, or a flat NOT EXISTS (SELECT 1 FROM target WHERE …). No nested or correlated subqueries, no JOINs, no reference to source/lookup tables. A simple assertion that runs is worth far more than a clever one that errors — a malformed grade query is a false non-convergence that wastes fix attempts.

Phase 4: Baseline (seed → deploy-and-CALL → grade)

1. Run <proc_name>.seed.sql against <db>.<test_schema> (ARRANGE:SETUP once, then ARRANGE:SEED).
2. ACT — scai code deploy-and-call … --proc-file <proc.sql> --call "CALL <proc>(...)". Read the ProcCallResult: deployed=false (failed to CREATE) or called=false (errored at CALL) is an expected baseline failure — record errorText.
3. Grade — if the proc CALLed clean, run <proc_name>.assert.sql and read the PASS/FAIL rows.
4. Regenerate-on-erroring-grade (grade hardening): if the assert query itself ERRORS (not FAIL — e.g. a bad column reference or an over-clever subquery), that is a test bug, not a proc bug. Simplify the offending assertion to a flat target-only form (per proc-assertion-patterns.md) and re-run — up to 2 regenerations — before recording any failure. Never let a malformed assertion count as a proc non-convergence.

Phase 5: Report

Write <UNIT>/stabilization/tests/proc/<proc_name>/test_report.md:

• Source→Proc mapping: the transform chain and the target table.
• Seed: the frozen input rows (summary) and the hand-computed expected target.
• Baseline result: deployed / called / assert PASS-FAIL counts; per-failure errorText and which assertion failed; ACT-vs-ASSERT classification.
• Coverage gaps: transforms not reconstructable from source (dynamic SQL, external mapplets) → record the gap and escalate; do not fabricate an expected result.

MANDATORY self-check before returning: verify <UNIT>/stabilization/tests/proc/<proc_name>/ contains <proc_name>.assert.sql, <proc_name>.seed.sql, and test_report.md. If any is missing, generate it (or, if you cannot write files, include full contents in your completion message). Then return to the parent skill for proc-fixer.

Reference Files (loaded on-demand)

Task Mode (spawned by orchestrator)

Input (provided at spawn time)

• Unit name + converted proc .sql path (READ-ONLY baseline)
• Source mapping .xml path (READ-ONLY)
• Target table name + the block FullName(s) under repair
• Test schema (TEST_SCHEMA) — pre-assigned; use for all created objects
• Database (DATABASE), connection name (CONNECTION)
• Skill directory (SKILL_DIR); platform directory (PLATFORM_DIR)
• Run all SQL headless via scai code deploy-and-call + snow sql --format json — no snowflake_sql_execute

Autonomous behavior

• No user questions — reasonable defaults, documented in the report.
• On ambiguity (unclear transform semantics, missing source element): make a best-effort decision, document the assumption in test_report.md, continue. If a transform is genuinely non-reconstructable, record it as a coverage gap and escalate — do NOT fabricate an expected result.
• Write artifacts incrementally (seed, then assert, then report). Prefer stopping early with artifacts on disk over running to exhaustion.

Team protocol

When complete, results return to the orchestrator automatically. Respond to a shutdown_request with send_message type: "shutdown_response", approve: true.