Does skills support scripts in sdk?

[Smile0524]

Does skills support scripts in sdk?

[grandocean (173787247)]

Plugin / skill («Does skills support scripts in sdk?»)

Desktop/asar apps often symlink packages into an archive; third-party plugins that fs.realpath the path will crash. CLI + real node_modules is the escape hatch.

Skills are optional markdown, not a workflow engine — if tools already describe the steps, the model may never open SKILL.md. Put mandatory order in code or a single wrapper tool.

Reload/restart the host after install; stale caches are common.

[Gaolenyue (glyahh)]

Yes—the Python SDK can use scripts bundled with a skill, but skill discovery and script execution are separate steps:

1. SkillsMiddleware discovers SKILL.md metadata and exposes it to the model.
2. The model reads the instructions through read_file.
3. It can run the helper through execute if the configured backend supports command execution (SandboxBackendProtocol). Merely loading a skill does not execute its scripts.

For a trusted local development example:

from deepagents import create_deep_agent
from deepagents.backends import LocalShellBackend

backend = LocalShellBackend(root_dir="/work/project")
agent = create_deep_agent(
    model=model,  # your configured chat model
    backend=backend,
    skills=["/skills/"],
)

Directory layout:

/work/project/skills/demo/
  SKILL.md
  scripts/hello.py

In SKILL.md, tell the agent to run the interpreter and script at their execution-environment paths, for example python /work/project/skills/demo/scripts/hello.py. With the backend's virtual filesystem mode, /skills/demo/SKILL.md is a file-tool path; do not assume /skills/demo/scripts/hello.py is also a real host-shell path. The interpreter and the script's dependencies must exist in the execution environment.

The default StateBackend is not a shell executor. For hosted/untrusted workloads use an isolated execution backend: LocalShellBackend runs directly on the host, and root_dir is a working directory, not shell isolation.

I checked this locally with deepagents==0.7.13, langchain==1.4.0, and langgraph==1.2.11: a deterministic model driver caused skill discovery, read_file, then execute; an actual Python helper printed the expected marker. This verifies the SDK/tool path, not whether a live model will choose the skill reliably. Prepared with Codex assistance and locally verified.

Relevant implementation: SkillsMiddleware and LocalShellBackend.