Invert the agent model: the agent IS the computer. The message pump
becomes the kernel, handlers are sandboxed apps, and all access is
mediated by the platform.
Phase 1 — Container foundation:
- Multi-stage Dockerfile (python:3.12-slim, non-root user, /data volume)
- deploy/entrypoint.py with --dry-run config validation
- docker-compose.yml (cap_drop ALL, read_only, no-new-privileges)
- docker-compose.dev.yml overlay for development
- CI Docker build smoke test
Phase 2 — Security hardening:
- xml_pipeline/security/ module with default-deny container mode
- Permission gate: per-listener tool allowlist enforcement
- Network policy: egress control (only declared LLM backend domains)
- Shell tool: disabled in container mode
- File tool: restricted to /data and /config in container mode
- Fetch tool: integrates network egress policy
- Config loader: parses security and network YAML sections
Phase 3 — Management plane:
- Agent app (port 8080): minimal /health, /inject, /ws only
- Management app (port 9090): full API, audit log, dashboard
- SQLite-backed audit log for tool invocations and security events
- Static web dashboard (no framework, WebSocket-driven)
- CLI --split flag for dual-port serving
All 439 existing tests pass with zero regressions.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
