mirror of
https://github.com/garrytan/gstack.git
synced 2026-05-06 13:45:35 +02:00
27954de0b0
Pure-function tests for security-classifier.ts that don't need a model
download, claude CLI, or network. Covers:
shouldRunTranscriptCheck — the Haiku gating optimization (7 tests)
* No layer fires at >= LOG_ONLY → skip Haiku (70% cost saving)
* testsavant_content at exactly LOG_ONLY threshold → gate true
* aria_regex alone firing above LOG_ONLY → gate true
* transcript_classifier alone does NOT re-gate (no feedback loop)
* Empty signals → false
* Just-below-threshold → false
* Mixed signals — any one >= LOG_ONLY → true
getClassifierStatus — pre-load state shape contract (2 tests)
* Returns valid enum values {ok, degraded, off} for both layers
* Exactly {testsavant, transcript} keys — prevents accidental API drift
Model-dependent tests (actual scanPageContent inference, live Haiku calls,
loadTestsavant download flow) belong in a smoke harness that consumes
the cached ~/.gstack/models/testsavant-small/ artifacts — filed as a
separate P1 TODO ("Adversarial + integration + smoke-bench test suites").
Full security suite now 156 tests / 287 expectations, 112ms.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>