Skills Management

Skills are server-side slash commands (/skill-name) that expand into prompt snippets before the request reaches the LLM. They are managed under Admin → Tools → Skills.

Skill Sources (Three Tabs)

Meine Skills (Local)

Admin-created or imported skills stored in /app/skills/. These are immediately available for invocation once enabled.

• Create: Click New Skill, enter name (lowercase, hyphens only), description, and body
• Edit / Delete: Via the tile actions
• Enable / Disable: Toggle button on each tile

Upstream (anthropics/skills)

Official Anthropic skills from github.com/anthropics/skills.

• Git Pull / Clone: The Git Pull button clones the repository on a fresh installation (no .git present) or performs git pull --ff-only on subsequent calls
• Audit required: Every upstream skill must pass an internal security audit before the Import button becomes active
• Audit: Click the Audit button on a tile; the LLM security check runs and the tile updates with a verdict badge (Safe / Warning / Blocked)
• Import: Copies the audited skill to /app/skills/ — blocked skills cannot be imported

Community (skills.sh)

305+ community skills from alirezarezvani/claude-skills.

• Pull Community Skills: Clones the repository and copies all valid SKILL.md files to /app/skills/community/
• Activate: Copies an audited community skill to /app/skills/ — requires a passed audit

Audit System

Internal LLM Audit

Applies to both upstream and community skills. An LLM (default: phi4:14b) checks for:

1. Shell command execution (subprocess, os.system, exec, eval)
2. Network access (HTTP, sockets, curl, wget)
3. File system writes outside the working directory
4. Prompt injection (system prompt overrides, role confusion)
5. Data exfiltration (sending data to external URLs)
6. Credential access (environment variables, config files)
7. Privilege escalation (sudo, chmod, chown)

Verdicts:

Badge	Meaning
Audited: Safe (green)	No findings — can be imported/activated
Audited: Warning (yellow)	Minor findings — can still be imported, use judgment
Blocked (red)	Critical findings — import/activation is blocked
Unaudited (grey)	No audit run yet — import/activation not possible

Audit reports are saved as {name}.audit.json alongside each skill file.

External Audit (skills.sh)

Community skill tiles additionally show ratings scraped from skills.sh/audits:

Badge	Source	Meaning
Gen: Safe / unsafe	Gen Agent Trust Hub	Independent safety classification
Socket: N	Socket.dev	Number of security alerts detected
Snyk: Low/Med/High/Critical	Snyk	Dependency vulnerability risk level

External audit data is cached for 24 hours. Use "Ext. Audits aktualisieren" to force a refresh.

The internal LLM audit is always required regardless of the external rating. External badges are informational only.

Hard-Lock Execution Gate

Skills must also be approved in the skill_registry database table before they can be invoked. Every invocation attempt is logged to skill_audit_log with outcome (executed / blocked / error). A skill that passes the file-level audit but is not registry-approved remains blocked at invocation time.