Skip to content

Releases: anthonyonazure/agent-content-shield

v0.4.2 — full Python parity + shared-lexicon drift guard

20 Apr 13:30
@anthonyonazure anthonyonazure
v0.4.2
8d25dfc

Choose a tag to compare

View all tags
v0.4.2 — full Python parity + shared-lexicon drift guard Latest
Latest

Third release in the v0.4 line. Completes Python parity across the entire detection stack and eliminates inline lexicon drift between the JS and Python ports.

What's new

JS refactor — load from shared JSON (no behavior change)

  • core/semantic-detector.js loads INJECTION_SEEDS + THREAT_IDF from core/semantic-lexicon.json (−255 LOC inline).
  • core/nli-classifier.js loads THREAT_INTENTS + NLI_SYSTEM_PROMPT from core/nli-intents.json (−50 LOC inline).
  • Both fail loudly on missing JSON — defense modules must not silently load empty arrays.

Drift tests pinning the JSON contract

  • test/shared-lexicon.test.js (JS, 8 assertions)
  • test/test_shared_lexicon.py (Python, 10 assertions)
  • Both wired into the test runners + GitHub Actions CI. Future contributors can't re-inline data without the build breaking.

Python LLM classifier tiebreaker

  • classify_with_ollama() added to core/semantic_detector.py (stdlib urllib, no extra deps).
  • semantic_scan() now calls the classifier when the embedding layer returns a borderline alert verdict. Block/benign verdicts skip the call — no wasted round trip.
  • Final verdict ORs three layers: embedding ∨ offline ∨ (classifier confident).

Cumulative across v0.4.0 → v0.4.2

v0.3.0 v0.4.0 v0.4.1 v0.4.2
JS tests 52 83 83 91
Python tests 0 28 88 105
Shared JSON pattern files 1 1 3 3
Inline JS lexicon copies yes yes yes none

Upgrade notes

Nothing for existing users — JS behavior is byte-identical. Python users who want the LLM classifier path should have Ollama running with the model set via SHIELD_CLASSIFIER_MODEL (default deepseek-r1:8b).

Closes

  • #3 (v0.4.2 roadmap)
Assets 2
Loading