Agentic Cognitive Firewall SDK

LLM agents built with LangGraph and LangChain retrieve documents, call tools, and write to memory - but nothing validates what enters the model's context. A poisoned RAG document can hijack an agent's behaviour. A malicious tool parameter can exfiltrate data. A single memory write can silently alter every future response. There's no standard security layer for any of this. ACF-SDK is a Zero Trust firewall that sits between the agent and the LLM - intercepting every input across four hooks (prompt, context, tool call, memory) and enforcing policy-as-code decisions through an isolated Go sidecar. I want to take this pipeline from hardcoded ALLOW to real enforcement: upgrade the semantic scanner to catch paraphrased attacks, wire scanner signals into the real sidecar, build memory write protection with HMAC and provenance tracking, ship a Docker demo anyone can clone and run in 5 minutes, and stress-test the whole system against adversarial evasion techniques with a technical evaluation report. The foundation is already built. I've shipped 6 PRs - semantic scanner, benchmark harness, LangGraph adapter, Rego policy templates for all four hooks, audit logger, and a Go scan stage implementation. 155 tests passing. GSoC wires these components into a running system and proves they hold up under real attacks.