harness-health-auditlisted

# Harness Health Audit A **harness** is everything that steers an agent but is not the model itself: `CLAUDE.md` / `AGENTS.md` at every level, `.claude/rules/*.md`, skills, subagents, hook wiring, slash commands, and tool/MCP descriptions — plus the `.codex` / `.opencode` equivalents when the repo targets several CLIs. LangChain frames it as `Agent = Model + Harness`, defining the harness as "every piece of code, configuration, and execution logic that isn't the model itself" ([LangChain](https://www.langchain.com/blog/the-anatomy-of-an-agent-harness)); Addy Osmani puts the stakes plainly: "A decent model with a great harness beats a great model with a bad harness" ([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)). Harnesses fail in three different ways, and most audits only catch the first: - **They rot.** The stack migrates, a command is renamed, a plugin disappears upstream, a ceremony is abandoned — but the steering text still asserts the old world. Nothing fails loudly, because the harness is instructions, not code: no test turns red when a rule lies. - **They stop working while still being true.** Every word is accurate, but the skills never trigger, the rules are too long to be followed, and the loop closes on the agent's own say-so instead of a real check. A perfectly accurate harness that nothing obeys is still a broken harness. - **They keep working while being dangerous.** Every instruction is accurate, live, and cheap — and one of the