Skip to main content

Agent Environments

SicarioSpec can bootstrap agent-native instructions for Claude Code, Codex, and GitHub Copilot coding agent.

Use the combined integration for teams that use multiple agents:

sicario init . --integration all --profile ai-system,agent-fleet,supply-chain

Claude Code​

--integration claude and --integration all create:

  • CLAUDE.md
  • .claude/skills/sicario-verify/SKILL.md
  • .claude/skills/sicario-governance-review/SKILL.md
  • .claude/skills/sicario-release-readiness/SKILL.md
  • .claude/agents/sicario-security-reviewer.md
  • .claude/agents/sicario-release-manager.md

The skills provide reusable SicarioSpec workflows. The subagents are read-first reviewers for security/governance and release readiness.

Codex​

--integration codex and --integration all create:

  • AGENTS.md
  • .agents/skills/sicario-verify/SKILL.md
  • .agents/skills/sicario-governance-review/SKILL.md
  • .agents/skills/sicario-release-readiness/SKILL.md

AGENTS.md is the durable repository instruction file. The skills are repo-scoped workflows that Codex can invoke explicitly or implicitly.

GitHub Copilot​

--integration copilot and --integration all create:

  • AGENTS.md
  • .github/copilot-instructions.md
  • .github/instructions/sicario-governance.instructions.md
  • .github/workflows/copilot-setup-steps.yml

The Copilot setup workflow installs the local project when Python package metadata exists, then runs the SicarioSpec verifier when available. The workflow is intentionally minimal so Copilot can run in a clean cloud environment without requiring private secrets.

Operating Rules​

  • Do not put secrets, private evidence, tenant identifiers, customer data, or unpublished vulnerability details into agent instructions or prompts.
  • Keep data classification and tagging current when prompts, specs, logs, evidence, workflows, or release artifacts change.
  • Use machine-user pull requests when available; document the maintainer fallback when a machine user is unavailable.
  • Treat verification as a gate, not approval. High-impact changes still need human review.