Skip to content

codex_install.md

Latest commit

 

History

History
57 lines (42 loc) · 1.63 KB

codex_install.md

File metadata and controls

57 lines (42 loc) · 1.63 KB
Raw

Installing SentrySkills for Codex

Installation

git clone https://github.com/AI45Lab/SentrySkills.git ~/.codex/sentryskills
mkdir -p ~/.agents/skills
ln -s ~/.codex/sentryskills ~/.agents/skills/sentryskills
cp ~/.codex/sentryskills/AGENTS.template.md ~/.codex/AGENTS.md

Windows PowerShell:

New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.agents\skills"
cmd /c mklink /J "$env:USERPROFILE\.agents\skills\sentryskills" "$env:USERPROFILE\.codex\sentryskills"
Copy-Item "$env:USERPROFILE\.codex\sentryskills\AGENTS.template.md" "$env:USERPROFILE\.codex\AGENTS.md" -Force

Restart Codex after installation.

Integration model

Codex does not provide the same hard enforcement as a hook-driven integration. Treat this as:

  • SKILL.md defines the method
  • AGENTS.md enforces agent behavior as much as the framework allows

Recommended policy:

  • always run the rule frontend first
  • if rule_stage_action == block, stop immediately
  • assign framework_risk_level after rule gating
  • only low-risk turns may use subagents for model_stage
  • if risk is high or subagent behavior is not reliable, fall back to synchronous model_stage
  • run one proposal sweep at the end of each main-agent task
  • proposal sweep only affects subsequent turns

What to expect

The runtime writes workspace-local state into:

  • .sentryskills/base/
  • .sentryskills/extra/

Key fields in results:

  • base_rule_action
  • extra_rule_action
  • rule_stage_action
  • framework_risk_level
  • model_dispatch_mode
  • model_stage_status
  • model_executor
  • model_stage_result_available
  • proposal_sweep_effect
  • final_action