| name | agentharness-audit |
| description | Audits any agent harness codebase against 13 meta-principles of quality, scores each 0 to 10 with code citations, generates a self-contained HTML report for operators, and writes an agent-readable AGENTS.md improvement directive with file-level architecture recommendations to reach 9.5 or better. Use when the user wants to audit or rate an agent harness, evaluate harness quality, score an agent framework, get a roadmap for improving a harness, run agentharness-audit, audit this harness, rate Hermes, audit Codex, or audit a Claude Code harness. Not for individual skill audits or general code review. |
| allowed-tools | Read, Write, Edit, Bash |
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
- Discover — map the codebase (system prompt, tool loading, memory, compaction, subagents, config)
- Score — evaluate each of the 13 principles with file:line citations (references/principles.md)
- 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.
- Directive — write
docs/agentharness-audit/<name>-AGENTS.md (references/agents-template.md)
- Print — output scores table to terminal
If no path is specified, audit the current working directory.
Hard rules
- Evidence before score. Every score requires a file:line citation or an explicit "not found" declaration.
- Both files always. HTML report and AGENTS.md are both produced on every run.
- Concrete recommendations only. AGENTS.md must name specific files and patterns — "implement Y at Z", never "improve X".
- Self-contained HTML path. The bundled template is the required source of truth.
html-artifact is provenance only; do not depend on loading it during an audit run.
- Not a harness → exit cleanly. If no agent loop, tool dispatch, or model calls are found, say so and stop.
Cross-references
Why this skill exists
Agent harnesses are evaluated on marketing claims, not the mechanisms that determine token
efficiency, self-improvement, and operator accessibility. This skill makes those mechanisms
legible and scoreable — enabling evidence-based harness selection and a concrete improvement
roadmap that the harness agent can execute autonomously.