H3 Shim

Python plugin for Hermes Core. Implements the H3 protocol on the Hermes side.

Install

The package is published to PyPI as of v0.1.0:

# PEP 668 distros (Ubuntu 24+, Debian 12+) refuse bare pip installs —
# always use a venv:
python3 -m venv .venv
source .venv/bin/activate   # run from the directory holding .venv
pip install hermes-h3-shim

From source, for any commit (including one that is not in a release yet):

pip install git+https://github.com/get-h3/shim

Release mechanics — build, twine upload, post-upload verification and the version-bump flow — are in docs/publishing.md.

This installs two CLI entry points:

• h3-test — the 46-test H3 compliance battery (h3-test --endpoint <url>)
• hermes-h3 — harness management (install, list, test, verify, scaffold, ...)

Quickstart

# 1. Stand up a demo harness first. The scaffold prints run instructions:
hermes-h3 scaffold --lang py
cd h3-harness-py
# One venv, explicitly activated. The scaffold's pyproject.toml depends on
# fastapi + uvicorn, so this venv is self-sufficient for running main.py.
cd h3-harness-py && python3 -m venv .venv && source .venv/bin/activate && pip install -e . && python main.py   # listens on :9191

# 2. In another terminal, run the test battery against it.
# The shim is NOT pre-deployed — install it into ITS OWN venv:
cd /tmp
python3 -m venv .venv-h3-test           # shim's own venv, separate from .venv
source /tmp/.venv-h3-test/bin/activate  # activate BY FULL PATH (not ./.venv)!
pip install hermes-h3-shim   # provides h3-test (or the git URL for an unreleased commit)
h3-test --endpoint http://localhost:9191

The scaffolded harness is a minimal H3 echo server — see hermes-h3 scaffold --help for --lang go|py|ts. Ready-made echo examples also live in the sibling SDK repos (get-h3/sdk-go/examples/echo, get-h3/sdk-python/src/h3_harness/examples/echo.py, get-h3/sdk-typescript/src/examples/echo.ts).

Exit codes

h3-test (and hermes h3 test / hermes-h3 test) uses three exit codes so CI can distinguish a real compliance failure from a wrong server:

See docs/integration.md for the full troubleshooting matrix.

Send a request

A harness is driven by POST /v1/process with a ProcessRequest body. Its four top-level fields — session_id, message, identity, context — are all required, and each nested object has its own required fields (see get-h3/protocol → schemas/v1/process-request.json and common.json). A valid, copyable payload:

{
  "session_id": "sess-7f3a9c",
  "message": {
    "role": "user",
    "content": "Book a flight to Medellín",
    "timestamp": "2026-08-20T14:05:42Z"
  },
  "identity": {
    "platform": "telegram",
    "chat_id": "-1001234567890",
    "user_name": "Alice",
    "user_id": "424242"
  },
  "context": {
    "history": [
      {"role": "user", "content": "Hi, I need help with travel"},
      {"role": "assistant", "content": "Sure — where are you flying from, and when?"}
    ],
    "tools": [
      {
        "name": "terminal",
        "description": "Execute shell commands on a Linux environment",
        "parameters": {
          "command": {"type": "string", "description": "Shell command to execute"}
        }
      }
    ],
    "models": [
      {
        "name": "deepseek-v4-pro",
        "provider": "deepseek",
        "cost_per_1k_input": 0.0011,
        "cost_per_1k_output": 0.0044,
        "context_window": 128000,
        "supports_vision": false,
        "supports_tool_calling": true
      }
    ],
    "config": {
      "max_iterations": 50,
      "timeout_seconds": 600
    },
    "session_state": {
      "turn_count": 1,
      "total_tool_calls": 0,
      "total_llm_calls": 1,
      "cost_so_far": 0.0003,
      "started_at": "2026-08-20T14:03:11Z"
    }
  }
}

Save it as process-request.json and POST it to the running harness:

curl -sS http://localhost:9191/v1/process \
  -H 'Content-Type: application/json' \
  -d @process-request.json

The harness answers with a Decision ({"decision": "tool_call", ...}, {"decision": "text", ...}, …) rather than a final answer — the shim loop executes that decision and POSTs the result back, so the harness never has to call a tool or a model itself. Optional fields (message.attachments, identity.thread_id, context.memory, context.skills, context.config.* beyond the two required keys) may be omitted.

Configuration

hermes-h3 keeps its harness/session state in one YAML file (~/.hermes/h3/config.yaml by default). The path is resolved in this order, highest first: --config <path> (accepted before or after a subcommand) → $HERMES_H3_CONFIG → ~/.hermes/h3/config.yaml. So HERMES_H3_CONFIG=/tmp/mine.yaml hermes-h3 list reads the override instead of your real config, and every command that writes the file (install, use, scaffold, uninstall) writes it there too. An unset or blank $HERMES_H3_CONFIG falls back to the default path. See docs/integration.md for details.

Components

• protocol.py — Pydantic models (generated from get-h3/protocol)
• client.py — REST client for harness communication
• loader.py — Harness discovery, health check, session routing
• shim_loop.py — Main H3ShimLoop
• native.py — Native Hermes loop wrapper
• test_battery.py — 46 compliance tests (THE GATE)
• cli.py — the hermes-h3 CLI (9 subcommands: install, list, pre-update-check, route, scaffold, test, uninstall, use, verify); the h3/ plugin in this repo exposes them as hermes h3 <cmd>

Development

Prerequisite: uv (make typecheck runs uv run --with mypy mypy src/; install via curl -LsSf https://astral.sh/uv/install.sh | sh or pip install uv). Everything else uses plain python3 + pip (see the Makefile).

make install   # create venv + install deps
make build     # verify imports
make test      # run tests
make lint      # ruff check
make fmt       # ruff format

Reference

• docs/integration.md — full install → register → route → verify guide
• Specs: get-h3/h3 → specs/05-Test-Battery.md, specs/06-Hermes-Core-Integration.md