mirror of https://github.com/garrytan/gstack.git
750161bbbe
The sidebar-agent now runs a ML security check on the user message BEFORE
spawning claude. If the content classifier and (gated) transcript classifier
ensemble returns BLOCK, the session is refused with a security_event +
agent_error — the sidepanel renders the approved banner.
Two pieces:
1. On agent startup, loadTestsavant() warms the classifier in the background.
First run triggers a 112MB model download from HuggingFace (~30s on
average broadband). Non-blocking — sidebar stays functional during
cold-start, shield just reports 'off' until warmed.
2. preSpawnSecurityCheck() runs the ensemble against the user message:
- L4 (testsavant_content) always runs
- L4b (transcript_classifier via Haiku) runs only if L4 flagged at
>= LOG_ONLY — plan §E1 gating optimization, saves ~70% of Haiku spend
combineVerdict() applies the BLOCK-requires-both-layers rule, which
downgrades any single-layer high confidence to WARN. Stack Overflow-style
instruction-heavy writing false-positives on TestSavantAI alone are
caught by this degrade — Haiku corrects them when called.
Fail-open everywhere: any subprocess/load/inference error returns confidence=0
so the sidebar keeps working on architectural controls alone. Shield icon
reflects degraded state via getClassifierStatus().
BLOCK path emits both:
- security_event {verdict, reason, layer, confidence, domain} (for the
approved canary-leak banner UX mockup — variant A)
- agent_error "Session blocked — prompt injection detected..."
(backward-compat with existing error surface)
Regression test suite still passes (12/12 sidebar-security tests).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|---|---|---|
| .. | ||
| bin | ||
| scripts | ||
| src | ||
| test | ||
| PLAN-snapshot-dropdown-interactive.md | ||
| SKILL.md | ||
| SKILL.md.tmpl | ||