Skip to main content

Declarative agent governance

Runtime guardrails come too late. By the time the agent has typed a destructive command, the prompt is over and you're triaging the cleanup. StrayMark moves agent governance upstream: the rules are declared in versioned files the agent reads before it begins, and updates to those rules are visible in git log.

Why this matters​

  • Runtime ≠ workflow. Most LLM guardrails are prompt-injection defenses or output filters — important, but they don't decide whether the agent should write that DB migration in the first place. Workflow-level governance does.
  • Versioned rules are reviewable rules. STRAYMARK.md is a file with PR diffs. Tightening or loosening a rule shows up in git log with an author, a date, and a discussion thread. That's auditable in a way runtime configs aren't.
  • Same contract across CLIs. The same STRAYMARK.md directs Claude Code, Gemini CLI, Copilot CLI, Codex CLI. One source of truth, N agents. Switching models doesn't mean rewriting the operating contract.

The contract layers​

  1. STRAYMARK.md at the repo root (~600 lines, ~12 sections). The unified rules: hierarchies, prohibited operations, documentation policy, security stance, when to ask before acting. Agents are required to read it before any non-trivial action.
  2. AGENT-RULES.md under .straymark/00-governance/: stricter operational rules — when to flag drift, when to recommend a TDE, the "Be Proactive" clause that authorizes the agent to surface findings unprompted.
  3. Skill frontmatter augments the global rules for specific rituals. Each /straymark-* skill has its own preconditions, output contract, and post-conditions.

When the framework version bumps, the contract bumps. straymark update-framework shows you the diff before applying.

Learn more​