Exhibit 12 — Agent tooling / compatibility← All work
A local-first harness that checks whether a repository reads the same to Codex, Claude Code, Gemini CLI and OpenCode.
Stack
JavaScript · SARIFYear
2026Last updated
Why it exists
Teams now point several coding agents at one repository. Each agent reads different files with different precedence, so the same repo quietly gives each of them different instructions.
How it’s built
Agent Compat Lab inspects the repository-facing contract every coding agent depends on — instruction files, nested scope, skills and MCP configuration — and scores how portable it is.
- Discover AGENTS.md, CLAUDE.md, GEMINI.md and override files, including nested scope.
- Check skills frontmatter and every MCP configuration shape each client expects.
- Flag oversized or destructive always-on instructions and likely inline credentials.
- Emit a portable 0–100 compatibility score as text, JSON or SARIF.
Capabilities
Instruction precedence analysisNested-scope drift detectionSKILL.md validationMCP config shape checksCredential smell detectionSARIF 2.1.0 output
Principles
Deterministic checks need no model
One repo, one truth, many agents
Score what can be measured
Fit into CI, not a dashboard
Questions
Does it call a model?
No. The default checks are deterministic and run locally.
Which agents does it cover?
Codex, Claude Code, Gemini CLI and OpenCode.
Can it run in CI?
Yes. SARIF output plugs into code-scanning views.
Next exhibit — 01
OpenRetention →