Skip to content

Copilot Tools (unitorch-copilot-cli)¤

Copilot is the agent-facing integration layer for unitorch. It adapts registered copilot tools into Python functions, CLI commands, remote clients, and skill metadata so agents can discover and invoke ML capabilities with stable parameter contracts.

Copilot tools work alongside the lifecycle CLIs. Use the workflow commands directly for training, inference, evaluation, and FastAPI serving, and use unitorch-copilot-cli for component discovery, tool invocation, remote client access, and skill-oriented automation.

Install UniTorch before running python3 -m commands or npm skill scripts. For local development, use python3 -m pip install -e ., then run commands without PYTHONPATH=src.

unitorch-copilot-cli <tool-name> [--key value ...]

Use workflow commands directly:

unitorch-train
unitorch-infer
unitorch-eval
unitorch-fastapi

Discovery¤

# List copilot tool names
unitorch-copilot-cli

# Install, export, validate, or uninstall generated skill markdown files
python3 -m unitorch.cli.copilots.skills install all --folder ./skills --force true
python3 -m unitorch.cli.copilots.skills export all --folder ./agent-skills
python3 -m unitorch.cli.copilots.skills validate --folder ./skills
python3 -m unitorch.cli.copilots.skills uninstall all --folder ./skills

Generated Skills¤

UniTorch copilot skill files are generated under the root skills/unitorch-copilot-tools/ package. That generated package lives beside the hand-written root skills such as skills/unitorch-config-ini and skills/unitorch-train-model. Each generated SKILL.md includes Hermes/OpenClaw-friendly frontmatter, CLI and Python invocation examples, parameter tables, FastAPI adapter details when available, and verification notes.

Use npm scripts from this repository:

npm run generate-skills
npm run validate-skills

Use the npm-distributed skills installer when installing UniTorch repository skills into a local agent skill folder:

npx skills add fuliucansheng/unitorch
npx skills add fuliucansheng/unitorch --folder ./agent-skills

The command above is provided by the external open-agent skills npm package. UniTorch only provides the root skills/ layout for that ecosystem to copy; this repository does not publish or alias a local installer bin. A default install can report:

{
  "repo": "fuliucansheng/unitorch",
  "folder": "/home/decu/.hermes/skills",
  "copied": [
    "unitorch-config-ini",
    "unitorch-copilot-tools",
    "unitorch-infer-model",
    "unitorch-replace-decorator",
    "unitorch-serve-fastapi",
    "unitorch-train-model"
  ]
}

The external installer copies the published root skills/ package; it does not generate, export, or validate skill markdown.

Use the repository npm or python3 -m commands above when updating the checked-in skills/ tree. Use npx skills add fuliucansheng/unitorch when you only need to install the packaged repository skills into an agent environment.

Python Adapter¤

from unitorch.cli.copilots import classify_image

result = classify_image(
    image_path="a.jpg",
    config="config.ini",
    top_k=3,
)

For local execution, pass either an existing pipeline instance or a config path that can build the underlying unitorch pipeline.

Copilot adapters register with the same CLI registry style used by other unitorch components:

from unitorch.cli import register_copilot_tool

@register_copilot_tool(name="core/copilot/classify_image")
def classify_image(...):
    ...

Remote Client Adapter¤

from unitorch.cli.copilots import CopilotClient

client = CopilotClient(endpoint="http://127.0.0.1:5000")
result = client.classify_image(image_path="a.jpg")

CopilotClient can also start a local unitorch-fastapi subprocess from a FastAPI config file:

from unitorch.cli.copilots import CopilotClient

with CopilotClient(config="fastapi.ini", host="127.0.0.1", port=5000) as client:
    result = client.classify_image(image_path="a.jpg")

If port is omitted, the client chooses an available local port and passes it to the subprocess.

The remote client talks to services started with unitorch-fastapi. Remote calls use the same media contract as unitorch-fastapi: parameters are sent as query params, images are uploaded as JPEG multipart files, videos are uploaded as video multipart files, and responses can be JSON, image, or video bytes.

core/copilot/pkg_infos¤

Lists all registered components in the current unitorch installation.

# List every registered type
unitorch-copilot-cli core/copilot/pkg_infos

# Filter by type
unitorch-copilot-cli core/copilot/pkg_infos --name model
unitorch-copilot-cli core/copilot/pkg_infos --name process
unitorch-copilot-cli core/copilot/pkg_infos --name fastapi

Available --name values: process, copilot_tool, model, fastapi, score, dataset, loss, optimizer, scheduler, task, writer.