Skills Management and Security

Relevant source files

• agent/background_review.py
• agent/curator.py
• agent/skill_commands.py
• agent/skill_utils.py
• gateway/platforms/yuanbao_media.py
• hermes_cli/curator.py
• hermes_cli/skills_config.py
• hermes_cli/skills_hub.py
• scripts/build_skills_index.py
• skills/creative/songwriting-and-ai-music/SKILL.md
• skills/software-development/simplify-code/SKILL.md
• tests/agent/test_curator.py
• tests/agent/test_curator_classification.py
• tests/agent/test_curator_reports.py
• tests/agent/test_skill_commands.py
• tests/agent/test_skill_invocation_description.py
• tests/agent/test_skill_utils.py
• tests/cron/test_rewrite_skill_refs.py
• tests/gateway/test_yuanbao_media_ssrf.py
• tests/hermes_cli/test_curator_archive_prune.py
• tests/hermes_cli/test_curator_recent_run_notice.py
• tests/hermes_cli/test_curator_status.py
• tests/hermes_cli/test_skills_config.py
• tests/hermes_cli/test_skills_hub.py
• tests/run_agent/test_background_review.py
• tests/run_agent/test_background_review_cache_parity.py
• tests/run_agent/test_background_review_cost_controls.py
• tests/run_agent/test_background_review_toolset_restriction.py
• tests/scripts/test_build_skills_index_health.py
• tests/tools/test_force_dangerous_override.py
• tests/tools/test_hub_lock_non_utf8_68053.py
• tests/tools/test_skill_manager_tool.py
• tests/tools/test_skill_usage.py
• tests/tools/test_skills_guard.py
• tests/tools/test_skills_hub.py
• tests/tools/test_skills_hub_clawhub.py
• tests/tools/test_skills_sync.py
• tests/tools/test_skills_tool.py
• tests/website/__init__.py
• tests/website/test_extract_skills.py
• tests/website/test_generate_skill_docs.py
• tools/skill_manager_tool.py
• tools/skill_usage.py
• tools/skills_guard.py
• tools/skills_hub.py
• tools/skills_sync.py
• tools/skills_tool.py
• website/docs/reference/optional-skills-catalog.md
• website/docs/reference/skills-catalog.md
• website/docs/user-guide/features/curator.md
• website/docs/user-guide/skills/bundled/creative/creative-songwriting-and-ai-music.md
• website/docs/user-guide/skills/bundled/mlops/mlops-evaluation-evaluating-llms-harness.md
• website/docs/user-guide/skills/bundled/mlops/mlops-inference-serving-llms-vllm.md
• website/docs/user-guide/skills/bundled/software-development/software-development-simplify-code.md
• website/docs/user-guide/skills/optional/creative/creative-audiocraft-audio-generation.md
• website/docs/user-guide/skills/optional/creative/creative-heartmula.md
• website/docs/user-guide/skills/optional/mlops/mlops-models-segment-anything-model.md
• website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/reference/optional-skills-catalog.md
• website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/reference/skills-catalog.md
• website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/features/skills.md
• website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/skills/bundled/creative/creative-songwriting-and-ai-music.md
• website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/skills/optional/creative/creative-heartmula.md
• website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/skills/optional/productivity/productivity-shop.md
• website/scripts/extract-skills.py
• website/scripts/generate-skill-docs.py
• website/scripts/prebuild.mjs
• website/sidebars.ts
• website/src/pages/skills/index.tsx
• website/src/pages/skills/styles.module.css

This page documents the skills management system, which allows the agent to create, modify, and delete skills—reusable procedural knowledge for recurring task types. Skills are the agent's "how-to" memory: they capture proven approaches, workflows, and solutions that can be applied to similar problems in the future.

For information about discovering and installing community skills from the hub, see 8.2 Skills Hub. This page focuses on programmatic management, security scanning, and the underlying synchronization mechanisms.

---

Purpose and Scope

The skills system provides:

• Programmatic skill creation: The agent can create new skills based on successful task completions via the skill_manage tool tools/skill_manager_tool.py3-12
• Modification operations: Create, edit, patch, or delete skills via the skill_manage tool tools/skill_manager_tool.py14-20
• Security scanning: Automatic regex-based static analysis of all skill content before installation or update via Skills Guard tools/skills_guard.py3-9
• Supporting files: References, templates, scripts, and assets alongside SKILL.md tools/skills_tool.py14-26
• Validation: Strict name, frontmatter, and platform compatibility constraints tools/skill_manager_tool.py176-177 agent/skill_utils.py175-197
• Synchronization: Automatic seeding and updating of bundled skills from the repository to the user directory tools/skills_sync.py3-7

All skills reside in a single location (~/.hermes/skills/) and follow a standardized structure with YAML frontmatter tools/skills_tool.py28-46

Sources: tools/skill_manager_tool.py3-20 tools/skills_tool.py14-26 tools/skills_sync.py3-7 tools/skills_guard.py3-9 agent/skill_utils.py175-197 tools/skills_tool.py28-46

---

Skills Directory Structure

All skills are stored in SKILLS_DIR (~/.hermes/skills/) with optional category subdirectories tools/skills_tool.py139-143 This directory is the single source of truth for agent edits, hub installs, and bundled skills tools/skills_tool.py139-141

Sources: tools/skills_tool.py14-26 tools/skills_tool.py139-143 tools/skills_sync.py57 tools/skill_manager_tool.py28 tools/skills_hub.py72-75

SKILL.md Format

Every skill must have a SKILL.md file with YAML frontmatter followed by markdown content tools/skills_tool.py28-46 The skill_manage tool and internal utilities validate these fields tools/skill_manager_tool.py176-177

Field	Requirement	Description
name	Required	Max 64 chars, lowercase alphanumeric + hyphens tools/skill_manager_tool.py176 tools/skill_manager_tool.py114-124
description	Required	Max 1024 chars tools/skills_tool.py163 tools/skill_manager_tool.py177
platforms	Optional	List of macos, linux, windows tools/skills_tool.py34-36 Maps to sys.platform agent/skill_utils.py21-25
prerequisites	Optional	Env var names and command checks tools/skills_tool.py37-41

Sources: tools/skills_tool.py28-46 tools/skill_manager_tool.py114-124 tools/skills_tool.py163 tools/skill_manager_tool.py177 agent/skill_utils.py21-25

---

Skill Management Tools

The primary programmatic interfaces for skills are skills_list and skill_view, which implement a "progressive disclosure" architecture to save context tokens tools/skills_tool.py9-13

Tool Implementation and Flow

The following diagram bridges the agent's intent to the underlying code entities that perform file operations and security checks.

Sources: tools/skills_tool.py60-66 tools/skill_manager_tool.py21 tools/skill_manager_tool.py127 agent/skill_commands.py138 agent/skill_utils.py123 agent/skill_utils.py200 tools/skill_manager_tool.py180

---

Security and Prompt Injection Scanning

Hermes implements layers of security scanning for skill content and memory:

1. Memory Scanning

The memory tool, which manages persistent notes in MEMORY.md and USER.md, scans all content for injection/exfiltration patterns before storage tools/memory_tool.py88-91

• Scope: Uses the "strict" scope (broadest pattern set) because memory entries are injected as a frozen snapshot into the system prompt tools/memory_tool.py78-83
• Implementation: Delegates to _first_threat_message from tools.threat_patterns tools/memory_tool.py85-91

Sources: tools/memory_tool.py78-91

2. Skills Guard (Static Analysis)

The Skills Guard is a security scanner for externally-sourced skills tools/skills_guard.py3-9

• Trust Levels: Defines builtin, trusted (e.g., openai/skills, NVIDIA/skills), and community levels tools/skills_guard.py11-53
• Policy Enforcement: community skills with any caution or dangerous findings are blocked tools/skills_guard.py55-65
• Threat Patterns: Detects environment exfiltration (e.g., curl with API_KEY), credential store access (e.g., ~/.ssh), and programmatic env dumps tools/skills_guard.py101-168

Sources: tools/skills_guard.py3-168

3. Skill Manager Guard

The skill_manage tool can optionally scan agent-created skills if skills.guard_agent_created is enabled tools/skill_manager_tool.py108-115

• Blocking Logic: If a scan returns a "dangerous" verdict, the tool returns a failure message to the agent, allowing it to retry with safer content tools/skill_manager_tool.py139-148

Sources: tools/skill_manager_tool.py108-151

---

Synchronization and Bundled Skills

Bundled skills are seeded into the user's ~/.hermes/skills/ directory on install and updated via a manifest-based sync mechanism tools/skills_sync.py3-7

Manifest Logic (v2)

The sync process uses ~/.hermes/skills/.bundled_manifest to track skill versions via MD5 hashes tools/skills_sync.py8-57

Scenario	Logic
New Skill	Copied to user directory, hash recorded tools/skills_sync.py204
Unmodified Skill	If user copy matches manifest hash, it is updated if the repo version changed tools/skills_sync.py210-212
Modified Skill	If user copy differs from origin hash, sync is skipped to protect customizations tools/skills_sync.py214-215
Deleted Skill	If entry exists in manifest but file is missing in user dir, it is not re-added tools/skills_sync.py217-218

Sources: tools/skills_sync.py3-22 tools/skills_sync.py69-75 tools/skills_sync.py204-218

---

Character and Token Limits

To maintain efficiency and prevent context overflow, the system enforces limits on skill metadata and memory content:

• Skill Name: Max 64 characters (MAX_NAME_LENGTH) tools/skills_tool.py162 tools/skill_manager_tool.py176
• Skill Description: Max 1024 characters (MAX_DESCRIPTION_LENGTH) tools/skills_tool.py163 tools/skill_manager_tool.py177
• Memory Limits: MEMORY.md is limited to 2200 characters, and USER.md to 1375 characters by default tools/memory_tool.py167

Sources: tools/skills_tool.py162-163 tools/skill_manager_tool.py176-177 tools/memory_tool.py167