agentharness-audit — score any agent harness against 13 quality principles

Self-validate after edits. Any change to this skill's files must be followed by ./scripts/validate.sh from the skill directory.

Audits an agent harness codebase against 13 evidence-based meta-principles derived from comparative analysis of Hermes, Claude Code, Codex CLI, and OpenClaw. Each principle is scored 0–10 with code citations. The operator receives a print-ready Tufte HTML report; the harness itself receives a concrete self-improvement directive in AGENTS.md format.

Procedure

Full step-by-step: references/procedure.md

1. Discover — map the codebase (system prompt, tool loading, memory, compaction, subagents, config)
2. Score — evaluate each of the 13 principles with file:line citations (references/principles.md)
3. Report — copy the bundled templates/audit.html.template to docs/agentharness-audit/<name>-audit.html and fill placeholders using references/html-template.md. Do not regenerate the report from scratch or require loading another skill.
4. Directive — write docs/agentharness-audit/<name>-AGENTS.md (references/agents-template.md)
5. Print — output scores table to terminal