Skip to main content

CaSH — Context-Aware Shell

Raksha AI — May 2026


Overview

CaSH is the shell, filesystem, and execution-layer governance component of the Context Governance architecture. It intercepts context-producing operations at the kernel and filesystem layers — before sensitive content enters the agent's reasoning state.

The core problem CaSH addresses is that tool-name governance is insufficient for shell-capable agents. An agent that can invoke a shell tool can execute arbitrary code, read any file, enumerate environment variables, and exfiltrate credentials — regardless of what tool-name policies say it is allowed to do. The MCP tool name is an abstraction. The execution surface is where actual context acquisition occurs.

CaSH operates at three interception layers:

Shell mediation intercepts shell command invocations before execution, applying policy against the command, its arguments, and the current session context model. Commands that would produce governed content — credential files, environment dumps, configuration reads — are evaluated before the shell executes them.

FUSE-based filesystem interception operates at the virtual filesystem layer, governing read operations on sensitive paths — .env files, cloud credential directories, SSH keys, deployment manifests, and other high-sensitivity locations — before file contents flow into agent output.

eBPF-based syscall governance operates at the kernel layer, providing a final enforcement boundary that no userspace agent implementation can bypass. Even an agent that spawns a subprocess, writes a Python script, or executes a compiled binary is subject to syscall-level policy enforcement.

This layered approach reflects a foundational principle: governance must be applied at the execution layer, not the tool abstraction layer. An agent that can execute code can bypass any tool-name policy. CaSH closes that gap by governing at the layer where execution actually occurs.

CaSH feeds all intercepted context into the Context Accumulator, which tracks what the agent has acquired, from which surfaces, under which policies, and for how long. The Context Firewall evaluates accumulated context before it reaches the model.

Details Coming Soon

Full architecture documentation for CaSH — including the session context model, FUSE interception design, eBPF policy engine, redaction pipeline, and integration with the Context Accumulator — will be published alongside the open-source reference implementation.