2026 Hermes Agent Skills: продвинутый гайд — GEPA, Skill Bundles и Tap publish

~18 мин чтения · MACCOME

Аудитория: инженеры, уже развернувшие Hermes (install runbook), но упирающиеся в routing Skills, Bundles или GEPA pipeline.Результат: полная спецификация SKILL.md, Progressive Disclosure (Level 0–2), Conditional Activation, Tap publish, GEPA+DSPy evolution, plugin namespace и authoring patterns.Структура: шесть bottlenecks, матрица Skills/Memory/Prompts, bundles, hub, 8 шагов Tap, GEPA, case study, FAQ, ресурсы.

Шесть bottlenecks: почему Hermes не «учится» без Skills-архитектуры

В начале 2026 Hermes Agent превысил 160 000 GitHub stars. Тезис the agent that grows with you реализован через Skills — стандартизированный procedural memory layer, а не через размер модели. Механизмы загрузки и персистентности определяют, сработает ли compounding.

  1. Смешение Skills и Memory: MEMORY.md хранит факты (declarative); Skills — процедуры с on-demand load и нулевым token cost до activation.
  2. description как summary: Level-0 router получает только name+description (~3K tokens суммарно) — неточный routing = skill never loads.
  3. Без Bundles: три отдельных /skill вместо atomic /backend-dev workflow load.
  4. Conditional Activation отключена: primary и fallback tools одновременно в prompt — лишние tokens без functional gain.
  5. Нет Git на ~/.hermes/skills/: GEPA PR и team Tap без audit trail; rollback невозможен.
  6. Ephemeral host: container restart стирает SQLite trajectories и learned skills — compounding (~38% token reduction, 30-day report) обнуляется.

Skills vs Memory vs Prompts: архитектурная матрица

Измерение Prompt Memory Skill
Персистентность текущая session cross-session cross-session
Load trigger always in context inject each session on-demand (Progressive Disclosure)
Token до activation full cost small, stable zero (только metadata L0)
Тип данных произвольная инструкция факты, preferences procedure (step-by-step)
Sharing сложно private publishable as Tap

Memory architecture (3 layers): MEMORY.md + SQLite FTS5 + Provider.

SKILL.md: agentskills.io и Progressive Disclosure

Стандарт agentskills.io обеспечивает portability между Hermes, Claude Code, Cursor. Путь: ~/.hermes/skills/<category>/<name>/SKILL.md.

Три уровня загрузки (token budget control)

Level Payload Trigger Tokens
0 name + description session init ~3000 total
1 full SKILL.md body /skill or LLM match file size
2 references/, scripts/ runtime execution per file; script stdout only in context
markdown
---
name: github-code-review
description: |
  Use when reviewing PRs, checking security or style.
  Do NOT use for writing new code.
metadata:
  hermes:
    requires_toolsets: [terminal]
    fallback_for_tools: [web_search]
---

# GitHub Code Review

## Procedure
1. `gh pr diff`
2. Security checklist
3. Structured comment

## Common Pitfalls
- 403 rate limit: `--rate-limit 100`
- Large diff: load references/chunking.md

Skill Bundles: atomic workflow load

YAML в ~/.hermes/skill-bundles/<slug>.yaml. Команда /mlops-deploy загружает vllm, llama-cpp, github-pr-workflow, systematic-debugging одновременно.

Priority rules: bundle > single skill с тем же name; missing skills skip silently с warning; bundle не модифицирует system prompt — Prompt Cache остаётся valid.

yaml
name: mlops-deploy
description: Model deployment pipeline with monitoring.
skills:
  - vllm
  - llama-cpp
  - github-pr-workflow
  - systematic-debugging
instruction: |
  Run inference benchmarks before and after deploy.
  Document quantization in PR description.

Conditional Activation: toolset-aware routing

Четыре поля в metadata.hermes управляют visibility skill в runtime tool registry:

  • requires_toolsets / requires_tools: hide если tool missing.
  • fallback_for_toolsets / fallback_for_tools: hide если primary tool present — паттерн DuckDuckGo vs paid web_search.

При наличии FIRECRAWL_KEY fallback skill исчезает из Level-0 metadata — measurable token savings без manual toggle. TUI hermes skills позволяет per-platform enable (CLI, Telegram, Discord).

Skills Hub, Tap и open-source repos

bash
hermes skills install official/research/arxiv
hermes skills install github:openai/skills/k8s
hermes skills tap add github:your-org/your-skills-tap
hermes skills tap update
hermes skills tap list
Repo Focus Notes
hermes-agent Official skills Source of truth
hermeshub Community registry Prompt injection scan
awesome-hermes-skills Production skills MLOps, Deep Research
ai-agent-skills 191 skills, 28 categories Hermes/Claude/Cursor

8 шагов: publish Skill Tap для команды

  1. GitHub repo с категориями (mlops/, research/).
  2. SKILL.md per skill; validate: skills-ref validate ./my-skill.
  3. skills.sh.json для Hub groupings (optional).
  4. README + license (MIT recommended).
  5. Register tap: hermes skills tap add github:org/repo; private: --token $GH_TOKEN.
  6. Team onboarding: один command per developer.
  7. Git ~/.hermes/skills/ для version control и rollback.
  8. Update cycle: hermes skills tap update; после changes — /reset или --now (cache invalidation).

GEPA + DSPy: evolution без fine-tuning weights

GEPA (Genetic-Pareto Prompt Evolution) — ICLR 2026 Oral. Optimizer анализирует execution trajectories в SQLite, генерирует SKILL.md variants, оценивает Pareto frontier (success rate × token efficiency × latency). Repo: hermes-agent-self-evolution. Cost: $2–10 per run (API only, no GPU).

5 stages: trajectory collection → reflective failure analysis → targeted mutation (10–20 variants) → multi-objective Pareto eval → human-reviewed PR.

bash
export HERMES_AGENT_PATH=~/.hermes
python -m evolution.skills.evolve_skill \
    --skill github-code-review \
    --iterations 10 \
    --eval-source sessiondb

# Cross-agent traces (experimental):
# --trace-dirs ~/.claude/traces,~/.hermes/sessions

4 guardrails: pytest 100%, skill ≤15KB, prompt cache compatibility, semantic drift check. Roadmap: Phase 1 Skills (live), Phases 2–5 tool descriptions, system prompts, tool code, full automation (planned).

Plugin skills и authoring patterns

Namespace plugin:skill — opt-in load via skill_view("superpowers:writing-plans"); не pollute default skills_list.

  • description: «Use when…» + explicit exclusions — routing precision.
  • Common Pitfalls: concrete failure modes + fixes — GEPA training signal.
  • Size limits: SKILL.md <500 lines; >15KB blocks GEPA; split to references/.
  • skill_manage: programmatic patch/create; agent_writes_require_approval: true в config.yaml.

Case study: blog-workflow bundle

yaml
name: blog-workflow
skills:
  - seo-keyword-research
  - outline-generator
  - code-example-validator
  - bilingual-checker
instruction: |
  Research SEO keywords before writing.
  All code examples must be runnable.

Три hard metrics (июнь 2026)

  • 160 000+ GitHub stars Hermes Agent — growth driver = Skills system, not model size.
  • ~3000 tokens Level-0 metadata overhead — justification for precise description fields.
  • 3→19 skills, ~38% token reduction on repeat tasks (30-day field report) — requires 24/7 host with persistent ~/.hermes/ on Apple Silicon UMA.

Resources

Заключение: procedures на диске, uptime на выделенном Mac

SKILL.md + Bundles + Conditional Activation + GEPA формируют closed learning loop без weight fine-tuning. Limits альтернатив: (a) ephemeral prompts; (b) shared VPS без isolated ~/.hermes/; (c) laptop sleep прерывает Gateway и trajectory accumulation.

Для GEPA runs, team Taps и launchd Gateway на Apple Silicon UMA с FileVault-at-rest — выделенный Mac Mini M4 через MACCOME. Цены: тарифы аренды; поддержка: центр помощи.

FAQ

Skills vs MCP в Hermes?

Skills — procedural docs; MCP — tool interfaces. Skill описывает когда и как вызывать MCP tools — complementary layers.

Почему старая версия Skill?

/reset или install --now после edit. Production: Git tag + documented rollout window.

GEPA-safe для production?

4 guardrails + manual PR review. SessionDB с PII требует purge policy перед GEPA run.

24/7 hosting для Skills?

Compounding и Tap sync требуют persistent disk. Сравните регионы на странице тарифов; FAQ: центр помощи.