amabito/veronica-core
Runtime containment kernel for LLM agents. Enforces budget, step, retry, and circuit-breaker limits before the model call.
What's novel
Runtime containment kernel for LLM agents. Enforces budget, step, retry, and circuit-breaker limits before the model call.
Code Analysis
9 files read · 3 roundsA zero-dependency Python runtime containment kernel that enforces hard budget, step, retry, and circuit-breaker limits on LLM agent execution chains before calls reach the model, with adaptive budgeting, semantic loop detection, and multi-agent budget allocation.
Strengths
Exceptionally thorough engineering: thread-safe with documented lock discipline, nogil-audited for Python 3.13 free-threading, comprehensive edge case handling (NaN/Inf/zero-budget), and genuinely sophisticated algorithms (adaptive budget with anomaly detection, dynamic budget allocation with starvation prevention, Jaccard-based semantic loop detection). The architecture cleanly separates concerns across containment, shield, security, and adapter layers.
Weaknesses
The project is extremely broad (164 source files, 100+ public API names) which creates a large surface area to maintain; some modules like the 68KB execution_context.py could benefit from further decomposition. The sheer scope means some features (A2A trust, memory governance, compliance export) may be less battle-tested than the core containment path.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
402
Contributors
1
Files
547
Active weeks
4
Repository
Language
Python
Stars
3
Forks
0
License
Apache-2.0