Target: AEGIS Core runtime + Claude Code CLI
Status: RFC-0006 v0.2 (Draft) — active implementation target for Q2 2026
Spec: RFC-0006 — Claude Code Plug-in
Kind: AEGIS Core runtime extension

What it is

A plug-in that runs inside the Claude Code execution environment, intercepts the agent’s tool calls, evaluates each through the AEGIS Core decision pipeline, and returns approve / deny / escalate verdicts inline — without requiring the host application to embed AEGIS Core directly. The plug-in implements RDP-03 (Embedded Lightweight) from RFC-0005’s reference deployment patterns.

Why this scope

Claude Code is a representative high-trust environment: an autonomous agent operating directly on a developer’s filesystem and shell, with broad tool access. It is the kind of deployment where execution-boundary enforcement matters most and where the test of “does AEGIS hold up at the action boundary?” is most visible. RFC-0006 makes this the next implementation target after the v0.1.x runtime stabilization.

Adversarial validation

RFC-0006 adversarial testing on 2026-03-26 surfaced RC5 (No Environment Model) — the fifth ATX-1 Root Cause, which had not been documented by Agents of Chaos. RC5 captures the failure mode where an agent acts on stale or imagined environment state because no governance primitive is checking that the agent’s mental model of the environment matches reality. The Claude Code plug-in’s design now includes RC5-aware checks at the tool-invocation boundary.

Status

Relationship to other AEGIS work