UnitAutogen

Author(s)

• Munaf Ibrahim Khatri

Tags

SQL SQLServer tSQLt Coverage CI-CD Cobertura JUnit AzureDevOps Testing UnitTest CodeCoverage BranchCoverage DatabaseTesting AutomatedTesting

Functions

Install-UnitAutogenDatabase Invoke-UnitAutogen Export-CoverageCoberturaXml Export-TestResultsJunitXml Export-CoverageHtmlReport

Dependencies

This module has no dependencies.

Release Notes

## v0.12.0 (beta) — 2026-06-07

WHERE-AWARE BOUNDARY SEEDING. The in-database SQLCLR parser now lifts a guarded UPDATE/DELETE''s
WHERE equalities (col = literal, col IN (...), AND-chained) into seed overrides, so the boundary
test pre-seeds rows that actually satisfy the filter. A selective UPDATE/DELETE behind a loosened
comparison operator (e.g. count > N silently changed to count >= N) is now caught even when the
filter uses specific literal values -- not just unfiltered or sample-matching DML. Strictly
additive and conservative: anything the parser cannot lift cleanly (OR, ranges, functions,
subqueries, INSERT, multi-gate) seeds exactly as v0.11.1 -- never an error, never a false failure.
Validated on AdventureWorks2025 including a 3-conjunct WHERE (Region = ''US'' AND Status IN (3,4)
AND Priority = 9). NOTE: this release rebuilds the SQLCLR assembly (new trust hash) and adds a
PredicateInbox column -- install BOTH Install_UnitAutogen.sql AND Install-UnitAutogenClr.SSMS.sql.

## v0.11.1 (beta) — 2026-06-07

BOUNDARY-VALUE MUTATION DETECTION. Predicate-branch tests now catch comparison-operator
loosening (e.g. count > N silently changed to count >= N). Two parts: (1) a COUNT_CMP ''>'' gate''s
non-satisfy case now seeds the boundary (N rows) instead of 0, so the discriminating value is
actually exercised; (2) the FALSE/boundary test pre-seeds the guarded write target and compares its
full content before/after the procedure runs — a row-count check catches INSERT/DELETE and
AssertEqualsTable catches UPDATE. On correct code the gate is false and the target is unchanged
(pass); loosen the operator and at the boundary the write fires, the content differs, and the test
fails with a clear message. Strictly conservative: emitted only for single-gate / single-fakeable-
target procedures, so every other shape generates exactly as before. Regression-clean on
AdventureWorks2025, HighValueCustomer and WideWorldImporters (0 fail / 0 err).

## v0.11.0 (beta) — 2026-06-06

SEARCH-BASED GATE SEEDING — the headline feature. Branch gates whose controlling value the
static reverse-seeder cannot invert (loop accumulators, per-row products, coupled cross-gate
conditions, aggregate bands) are now covered by a numeric oracle: the generator instruments the
procedure, drives candidate seeds, reads the operand''s live value through an XEvent probe
(rollback-immune, fires per loop iteration), and measure-and-interpolates the seed that lands
each arm. A verified witness becomes a real tSQLt test; an unreachable / environmental gate is
an honest NOT_TESTABLE. It runs automatically as part of the normal sweep.

Archetypes auto-derived: aggregate-over-table, scalar-from-table, IS [NOT] NULL, bare parameter,
per-row value, per-row categorical, COUPLED cross-gate (with a recursive prefix that reuses an
ancestor gate''s witness to establish the flag), and LOOP-COUNT (a loop accumulator driven by
the trip count over a seedable table — both arms seeded).

Measured impact on real procedures: a coupled/per-row reconciliation procedure went from 1 of 13
to 13 of 13 branch gates handled (36.7% -> 83.3% line); a loop-accumulator gate went 0% -> 100%.
No regressions on already-covered procedures. The whole search layer is collation-safe (works on
databases whose collation differs from the server default).

## v0.10.0 (beta) — 2026-06-05

Major capability + honesty release. New capabilities:

- DML EFFECT ASSERTIONS. A modifying procedure's per-table write effect is now asserted
 with EXACT row counts, measured at generation time by running the procedure under its
 own reverse-predicate seed (rolled back). Covers INSERT (VALUES and SELECT), UPDATE and
 DELETE; when the seed drives no write, the test is skipped and names the untouched table.
- DELETE and INSERT...SELECT branch coverage. The snapshot-and-replay branch assertions
 now handle a single-target DELETE and INSERT...SELECT, not just UPDATE / INSERT...VALUES.
- TABLE-VALUED PARAMETERS. A procedure with a TVP is now auto-tested: the generator builds
 and seeds a table variable of the type and passes it (such procedures were previously
 NOT_TESTABLE). The coverage instrumenter emits table parameters schema-qualified + READONLY.
- ROW-LEVEL SECURITY. SafeFakeTable now drops a SECURITY POLICY (and its predicate-function
 chain) before faking, so RLS-protected tables can be faked instead of dooming the test
 with "participates in enforced dependencies".

Honesty fixes (no false failures — every non-pass is a pass, a labelled skip, or a clear
NOT_TESTABLE):

- Error-expectation tests are no longer over-generated. A CATCH-block re-throw is no longer
 mistaken for input validation, and a parenthesis-less THROW (THROW n,'msg',state;) is now
 detected — so "must reject" / "raises error" tests fire only for genuine validation guards.
- When the generated happy / boundary / NULL inputs cannot satisfy a procedure's own
 validation, the affected tests are SkipTest-annotated with a clear reason instead of
 failing by construction.
- A FOR JSON / FOR XML result set no longer errors the row-baseline test (it cannot be
 captured via INSERT...EXEC); that single test is skipped for such procedures.

## Earlier releases

Full release history (v0.9.14 back to v0.9.0) is in CHANGES.md and on the GitHub
Releases page: https://github.com/unitautogen/unitautogen-public-repo/releases