Composer-based dynamic skill curation for Hermes Agent.
Walks a YAML capability tree once to build a flat index (~50KB, ~400 skills), then uses a single LLM call to curate a query-specific bundle of 3-20 skills — complete with load levels (★/▸/·), confidence scores, and reasoning.
Old approach: recursive LLM tree descent (5 levels × branching 3 = ~243 calls/query). Unusable for real-time.
New approach: flat index pre-filter → single LLM curation → 1 call/query, sub-second latency.
User Query → flat index pre-filter (top 50) → LLM picks best 3-20 → inject hints
Each skill gets:
| Field | Meaning |
|---|---|
name |
Skill name (call skill_view(name) to load) |
load_as |
must ★ / should ▸ / consider · |
confidence |
high / medium / low |
reason |
Why this skill fits the query |
[Skill Capability Chain]
These skills are curated for this query.
Call skill_view('<name>') to load each one.
★ cloudflare-tunnel — Tunnel deployment + credential management
▸ infrastructure-as-code — Terraform for Cloudflare tunnels
· devops — Broader deployment workflows
pip install skill-retriever
skill-retriever install # optional: install bundled community skillsNo plugin development needed — the Hermes plugin is registered automatically.
# Rebuild the flat index (after adding new skills)
skill-retriever rebuild
# Compose a bundle for a query
skill-retriever compose "deploy a cloudflare tunnel"
# Show index info
skill-retriever info| Point | What | When |
|---|---|---|
pre_llm_call hook |
Inject curated bundle | Every turn |
| Subagent dispatch | Inject per-task bundle | delegate_task() |
| Deepthink planning | Phase-sectioned coverage | Planning phase |
| Skills logger | JSONL usage tracking | Every compose |
See ARCHITECTURE.md for flat-index details, plugin wiring, and the reasoning model handling (LongCat-2.0 reasoning_content field).
- Hermes Agent v0.18+
- Python 3.10+
- ~50MB for flat index
- OpenAI-compatible LLM endpoint (LongCat, OpenAI, etc.)
| Env Variable | Default | Description |
|---|---|---|
SKILL_RETRIEVER_DISABLE |
— | Set 1 to disable |
SKILL_RETRIEVER_LLM_MODEL |
from Hermes config | LLM model override |
SKILL_RETRIEVER_LLM_API_KEY |
from Hermes config | API key |
SKILL_RETRIEVER_LLM_BASE_URL |
from Hermes config | Base URL |
SKILL_RETRIEVER_TEMPERATURE |
0.3 |
LLM temperature |
skill-retriever/
├── plugin/ # Hermes plugin (pre_llm_call hook)
├── src/
│ ├── skill_retriever/ # Core engine
│ │ ├── cli_compose.py # CLI (rebuild, compose, info)
│ │ ├── compose.py # Composer: skill curation (single LLM call)
│ │ ├── config.py # LLM discovery (borrow Hermes config)
│ │ ├── build_flat_index.py
│ │ ├── search/ # Tree search (CLI only, not for real-time)
│ │ ├── subagent_binding.py # Per-task bundle injection
│ │ ├── planning_with_skills.py # Phase-sectioned bundles
│ │ └── skill_usage_logger.py # JSONL tracking
│ ├── scanner.py # Hermes skills scanner
├── data/ # Skill corpus (gitignored)
├── tests/ # pytest suite
└── ARCHITECTURE.md
- Composer prompts for confidence per skill
- Bundle is capped to LLM token budget (max ~50 skills in prompt)
- Reasoning models (LongCat-2.0, DeepSeek R1) return responses in
reasoning_content— handled correctly
MIT. Community skills may have separate licenses.