Master enhancement orchestrator for plugins, agents, prompts, docs, hooks, and skills.
/enhance runs specialized analyzers over a repo's agent-facing files, verifies their findings, and reports them in one list ranked by certainty. The guidance targets current models: it flags dated prompt scaffolding (all-caps rule lists, "think step by step", step choreography, repeated rules) and never suggests adding emphasis.
/enhance # every enhancer whose content exists
/enhance --focus=agent # one enhancer
/enhance plugins/my-plugin --verbose
/enhance --apply # report, then apply HIGH certainty auto-fixes
/enhance --show-suppressed # findings hidden by learned suppressions
/enhance --no-learn # analyze without saving suppressions
/enhance --reset-learned # clear learned suppressions for this project
/enhance --export-learned # print learned suppressions as JSON
--focus values: plugin, agent, claudemd (alias claude-memory), docs, prompt, hooks, skills, cross-file.
Enhancers run in parallel, at most 4 at a time. In a harness without a subagent tool they run one after another inline.
| Enhancer | Checks | Model |
|---|---|---|
agent-enhancer |
Description as routing text, tool scope, model tier, prompt body | inherits session model |
prompt-enhancer |
Missing context, dated patterns, output contracts | inherits session model |
claudemd-enhancer |
CLAUDE.md / AGENTS.md accuracy, bloat, README duplication, cross-platform | inherits session model |
skills-enhancer |
Trigger quality, invocation control, tool scope, size | inherits session model |
docs-enhancer |
Broken links, structure, stale docs, retrieval readiness | sonnet |
hooks-enhancer |
Hook safety, exit codes, timeouts, matcher scope | sonnet |
plugin-enhancer |
Manifest, MCP tool schemas and descriptions, security | sonnet |
cross-file-enhancer |
Undeclared tools, missing agent references, duplicate or contradictory rules | sonnet |
Each enhancer runs its JavaScript analyzer from lib/enhance/, checks every finding against the file to drop false positives, and adds what a pattern analyzer cannot see.
| Level | Meaning | Auto-Fixable |
|---|---|---|
| HIGH | Definite issues | Some |
| MEDIUM | Likely improvements | No |
| LOW | Advisory suggestions (shown with --verbose) |
No |
Fixes apply only with --apply, only for HIGH certainty findings that have an auto-fix. Security and cross-file findings are never auto-fixed.
- An agentsys-supported harness (Claude Code, OpenCode, Codex, Cursor, Kiro)
- Node.js (for the analyzers)
MIT
