Skip to content

cortex-sentinel.md

Latest commit

 

History

History
71 lines (52 loc) · 3.26 KB

cortex-sentinel.md

File metadata and controls

71 lines (52 loc) · 3.26 KB
Raw

Cortex Sentinel Showpiece Brief

One-Sentence Story

Cortex Sentinel is a local safety console that lets a developer run autonomous coding agents with deterministic policy, human approval, protected-file rollback evidence, and replayable traces.

Why It Is A Showpiece

This repo is not just a wrapper script. It demonstrates a complete safety loop around local coding agents:

  • It launches real local commands through a supervised runner.
  • It distinguishes hard policy blocks from model-assisted review.
  • It has a terminal UI for approve, block, pause, kill, and scoped override actions.
  • It records decision evidence in local traces instead of relying on memory or chat summaries.
  • It includes a readiness matrix and disposable smokes that separate proven behavior from open gaps.
  • It keeps the limitation visible: this is supervision, not a hard sandbox.

The landing page should make that loop legible in the first screen: agent request, effect classification, policy decision, rollback event, and trace replay.

Demo Narrative

Use this sequence for a short walkthrough:

  1. Open docs/index.html.
  2. Explain that coding agents can edit files and run commands quickly, so they need a local control layer.
  3. Point to the hero console: the agent proposes a risky command, Sentinel classifies it, and policy blocks it.
  4. Point to the decision queue: ambiguous work waits for the human.
  5. Point to rollback evidence: a protected .env write is detected, suspended, and rolled back in a disposable smoke.
  6. Point to trace replay: the decision trail can be audited later.
  7. Close with the boundary: proven supervision, not guaranteed containment.

Verified Claims To Use

These are appropriate public claims for the README and landing page:

  • Root test suite has passed with 142 passed.
  • Nested MCP-Cortex tests have passed with 11 passed.
  • A fresh independent GitHub clone passed after the optional MCP-Cortex skip fix with 140 passed, 2 skipped.
  • Codex full-screen TUI smoke has proven workspace trust, command approval, and harmless output.
  • Claude Code startup trust-prompt control has been proven in a disposable workspace.
  • Continuous protected-file enforcement has proven suspend plus rollback for a no-prompt .env write.
  • The real MLX/Gemma auditor has produced structured output in a Metal-capable local session.

Claims To Avoid

Do not describe Cortex Sentinel as:

  • a hard sandbox;
  • a complete terminal emulator;
  • a production security boundary;
  • a cloud authorization service;
  • a full MCP proxy;
  • proven for full interactive Claude/Gemini tool-confirmation behavior.

Landing Page Checklist

  • First viewport says "Cortex Sentinel" clearly.
  • First viewport explains why someone would want it.
  • First viewport shows the agent supervision loop visually.
  • Proof section uses measured or explicitly named evidence.
  • Boundary section makes the limitation impossible to miss.
  • Try-it section uses real repo commands.

README Checklist

  • Explain the app to a non-specialist.
  • Link to the landing page.
  • Link to the GOAL proof contract.
  • Preserve tested sections: Quick Start, Installation, Configuration Reference, Safety Boundary, Troubleshooting, and Examples.
  • Preserve core command strings used by README tests.
  • Separate "what is proven" from "what is open."