deepagent-hermes 0.1.2

Closed-loop reflection / skill creation agent on LangGraph + deepagents. Faithful reproduction of Nous Research's Hermes Agent.

deepagent-hermes

A faithful reproduction of Nous Research's Hermes Agent on top of LangGraph + deepagents + langgraph-stream-parser.

Status: v0.1.0 live on PyPI. Spec at SPEC.md. Release notes in CHANGELOG.md. The runtime is verified end-to-end against a real Anthropic model — both the memory loop and the skill-creation loop close autonomously; see examples/dogfood.py and examples/dogfood_procedural.py for the traces.

What it is

A deepagents-built agent with a closed reflection→skill-creation loop:

• After ~10 tool-using iterations, a review subagent runs in the background, writes/patches a SKILL.md capturing the pattern it just exercised, and ships it to a skill library.
• Next session, the agent reads the library at startup, sees the new skill's description in its system prompt, and can skill_view(name) to load the full body on demand (progressive disclosure per the agentskills.io spec).
• A weekly curator consolidates skills into umbrellas and archives stale ones.
• A frozen-snapshot memory (MEMORY.md + USER.md) preserves prefix-cache hits for the entire session.
• FTS5 session search indexes every past conversation in a local SQLite DB.
• Bundled MarkdownProvider that keyword-searches <HERMES_HOME>/memories/notes/*.md — drop hand-authored long-form context there and the agent surfaces relevant sections on demand. Zero external dependencies.

Designed to be loaded into the existing deepagent-* host family (deepagent-code, deepagent-lab, cowork-dash, deepagent-vscode) without UI changes — set DEEPAGENT_AGENT_SPEC=deepagent_hermes.agent:graph in any of them.

Installation

pip install deepagent-hermes

Or with uv (recommended):

uv venv .venv
. .venv/Scripts/activate      # Windows
. .venv/bin/activate          # macOS / Linux
uv pip install deepagent-hermes

Optional extras

pip install "deepagent-hermes[openai]"     # OpenAI / OpenRouter / any OpenAI-wire provider
pip install "deepagent-hermes[daytona]"    # Daytona sandbox terminal backend
pip install "deepagent-hermes[modal]"      # Modal sandbox terminal backend
pip install "deepagent-hermes[ssh]"        # paramiko-backed SSH terminal backend
pip install "deepagent-hermes[dev]"        # tests + lint (contributors only)

Picking a model

By default the agent uses anthropic:claude-sonnet-4-5-20250929 and needs ANTHROPIC_API_KEY set. Swap the model via --model on the CLI or model.default in deepagent-hermes.toml — any init_chat_model string works.

OpenAI / OpenRouter

pip install "deepagent-hermes[openai]"
export OPENAI_API_KEY=sk-…                   # or: OPENROUTER_API_KEY=sk-or-v1-…
export OPENAI_BASE_URL=https://openrouter.ai/api/v1   # only for OpenRouter
deepagent-hermes chat --model openai:openai/gpt-4o-mini

For OpenRouter specifically you usually also want:

export DEEPAGENT_HERMES_MODEL_DEFAULT="openai:openai/gpt-4o-mini"
export DEEPAGENT_HERMES_MODEL_AUX="openai:openai/gpt-4o-mini"

so the reflection subagent uses the same cheap model.

Verify your setup

deepagent-hermes verify

does one live round-trip against the configured model and confirms the prompts, bundled skills, and FTS5 store all wire up correctly. Run this first on any fresh install — if it passes, chat will work.

Quick start

# show resolved config + sources
deepagent-hermes --show-config

# interactive chat
deepagent-hermes chat

# from inside chat:
#   /skills            list available skills
#   /model anthropic:claude-haiku-4-5-20251001    switch models
#   /memory            dump current memory snapshot
#   /compress          force context compression
#   /quit

Load into an existing host

Any deepagent-* host with langgraph-stream-parser>=0.2 host conventions can run this agent:

# deepagent-code
DEEPAGENT_AGENT_SPEC="deepagent_hermes.agent:graph" deepagent-code

# deepagent-lab — set the same in deepagents.toml under [agent]
echo 'spec = "deepagent_hermes.agent:graph"' >> deepagents.toml
deepagent-lab

Configuration

deepagent-hermes.toml (project) or ~/.deepagent-hermes/config.toml (global). Layered resolution: defaults < TOML < DEEPAGENT_HERMES_* env < CLI overrides. See SPEC §2 for every field; deepagent-hermes --show-config prints the resolved value + source of each.

Architecture

See SPEC.md for the full 21-section requirements doc. Top-level layout:

• src/deepagent_hermes/agent.py — the compiled graph (entry point for hosts)
• src/deepagent_hermes/config.py — HermesConfig(HostConfig) resolver
• src/deepagent_hermes/state.py — HermesState (extends AgentState)
• src/deepagent_hermes/reflection.py — closed-loop middleware + review subagent
• src/deepagent_hermes/skills/ — SkillLibrary, loader, tools
• src/deepagent_hermes/memory/ — frozen-snapshot memory + provider ABC
• src/deepagent_hermes/store/sqlite_fts.py — BaseStore with FTS5
• src/deepagent_hermes/search/session_search.py — session_search tool
• src/deepagent_hermes/compression.py — HermesCompressionMiddleware
• src/deepagent_hermes/caching.py — AnthropicCachingS3Middleware
• src/deepagent_hermes/budget.py — IterationBudgetMiddleware
• src/deepagent_hermes/tools/ — registry + 33 toolsets + 6 terminal envs
• src/deepagent_hermes/cron/ — daemon + cronjob tool
• src/deepagent_hermes/plugins/ — discovery + lifecycle hooks
• src/deepagent_hermes/cli.py — deepagent-hermes entry point
• prompts/ — verbatim/paraphrased system-prompt building blocks

Status by subsystem

Subsystem	Status
Config + state + agent factory	✅ working
Reflection loop (10-iter / 10-turn triggers, subagent review)	✅ working — verified live
Skill library + agentskills.io validator	✅ working
Skill loader (system-prompt injection + progressive disclosure)	✅ working
skill_view / skill_manage / skills_list tools	✅ working
Frozen-snapshot memory (MEMORY.md / USER.md)	✅ working — verified live (702 bytes written autonomously)
SQLite FTS5 store + session_search (3 modes)	✅ working
MarkdownProvider (bundled, default)	✅ keyword search over <HERMES_HOME>/memories/notes/*.md — zero deps
Iteration budget middleware	✅ working
Compression middleware (13-section template)	✅ working
Anthropic system_and_3 caching strategy	✅ working
Tool registry + 33-toolset enum	✅ working
LocalEnvironment terminal backend	✅ working (Git Bash on Windows)
DockerEnvironment	✅ working (gated on docker info reachability)
SshEnvironment	✅ working (paramiko-backed, behind [ssh] extra)
SingularityEnvironment	✅ working (auto-detects singularity / apptainer)
DaytonaEnvironment / ModalEnvironment	✅ lazy SDK with defensive attribute probing (extras-gated)
Cron daemon + cronjob tool	✅ working (deliverers: local, stdout, agentmail)
Plugin loader (4 discovery sources)	✅ working (13 of 17 lifecycle hooks wired)
CLI + v1-essentials slash commands	✅ working
Curator (skill lifecycle)	✅ basic
Bundled skills	✅ 26 from nousresearch/hermes-agent (MIT, attributed)
Self-evolution integration	📄 docs only (separate offline repo)

License

MIT. See LICENSE. This project is a faithful reproduction of the design ideas in Nous Research's Hermes Agent — see NOTICE for attribution.