Capybala

CAPYBALA (Context-Aware Pruning for Yielding & Backtrack-Aligned Loop Avoidance) is an ultra-fast, zero-dependency in-memory runtime supervisor designed to intercept cognitive deadlocks, semantic oscillation traps, and context overflows in autonomous AI coding agents. THE PROBLEM: When autonomous agents powered by frontier models (such as Meta's muse-spark series and Llama open-weights models) tackle long-horizon software engineering tasks, they frequently encounter cognitive deadlocks. An agent gets trapped repeating failing bash commands, ping-ponging between incompatible file patches (A ↔ B), or circulating in paraphrased search loops. This floods context windows with 15,000+ redundant traceback tokens, destroys prompt cache efficiency, and eventually causes catastrophic context overflows ($0.00039+ wasted per deadlocked run). THE SOLUTION & ARCHITECTURE: Grounded in formal academic research (published as a preprint on Research Square with registered DOI: 10.21203/rs.3.rs-11597011/v1), Capybala operates as a real-time supervisor with O(1) space complexity and sub-millisecond execution latency (mean 0.64 ms, tail p99 < 5.8 ms): 1. Tier-1 Exact Ring-Buffer Matcher: Intercepts deterministic tool repetitions and multi-period oscillations in <0.01 ms using SHA-256 hash rings. 2. Tier-2 Phase-Space Manifold Engine: Projects sequential agent actions into a signed 1024-dimensional feature-hashed dynamical manifold. By tracking directional semantic velocity (v_t) and Markov directional entropy (H), it catches semantic attractor traps in <0.6 ms. 3. Cache-Preserving Atomic Context Rollback: Unlike naive pruners that break Anthropic/OpenAI/Meta prompt cache prefix hashes, Capybala guarantees strict tool-call protocol invariants (Δ_protocol = 0, N_tool_call = N_tool_response). It surgically prunes redundant error spam and injects grounded Doctor Reflection Directives, allowing the agent to pivot and solve the bug. EMPIRICAL & OPEN-SOURCE ARTIFACTS: • PyPI Distribution: Production package distributed globally via `pip install capybala` (v0.9.2), requiring 0 external dependencies (pure Python standard library) with 95/95 passing unit tests. • Kaggle Benchmark Dataset: Published the "AI Agent Cognitive Deadlock Benchmark" comprising 3,209 real-world SWE-bench execution turns across 200 agent trajectories, demonstrating that 59.7% of non-productive turns in failed runs are eliminated by Capybala. • Live Interactive Demonstrator: Fully deployed at https://capybala.onrender.com featuring a Live Head-to-Head Duel against Meta AI (muse-spark-1.3), an AST Context Healer, and live Phase-Space Telemetry. ECOSYSTEM LINKS: • GitHub: https://github.com/codesbykhairannoor/capybala • PyPI: https://pypi.org/project/capybala/ • Live App: https://capybala.onrender.com/ • Research Square Preprint: https://doi.org/10.21203/rs.3.rs-11597011/v1 • Kaggle Dataset: https://www.kaggle.com/datasets/khairan7/ai-agent-cognitive-deadlock-benchmark/data