skills_executor

Skill examples

A Skill is an immutable, revisioned context package: SKILL.md guidance plus indexed resources. SkillLibrary owns installation and package truth; AgentExecution owns task-scoped selection and binding; TaskContext and ContextReader own progressive disclosure.

Agently.skills_executor remains a thin management/compatibility facade. It can configure the library, install/list/inspect/read Skills, and build compatibility context packs. Its legacy TaskDAG helper is not an executor-ready integration until host code adapts the real TaskDAGContext; do not present direct registration as a working example. The facade does not choose execution routes, run Skill-local strategies, actionize scripts, or grant capabilities.

Recommended execution:

execution = (
    agent
    .use_skills([skill_id], mode="required")
    .input(task)
    .output({"summary": (str, "...")})
)
result = await execution.async_get_data()

Released convenience calls remain adapters over the same path:

compat = await agent.async_run_skills_task(
    task,
    skills=[skill_id],
    mode="required",
    output={"summary": (str, "...")},
)
result = compat.output

Current runnable examples:

python examples/skills_executor/01_basic_declarative_skills.py
python examples/skills_executor/07_agently_skills_availability_check.py
python examples/skills_executor/08_architecture_diagram_skill.py
python examples/skills_executor/09_skill_script_exec.py
python examples/skills_executor/10_model_pool_key_pool_resolution.py
python examples/skills_executor/11_conditional_resource_read.py
python examples/skills_executor/12_complete_task_selection.py

09_skill_script_exec.py uses two requests in one Session. The first does not need the offered Skill; the second creates a fresh AgentExecution, selects the Skill from the current request, then host code enables the stable restricted Python Action only in that execution. Scripts remain resources; the framework does not generate an Action per script or per request.

11_conditional_resource_read.py uses a real model to choose supporting files from the conditions in the already-read root Skill. Configure AGENTLY_BASE_URL, AGENTLY_API_KEY, and AGENTLY_MODEL explicitly. It compares outline and handoff phases through the existing execution context reader, without mounting Actions or generating a final business answer.

12_complete_task_selection.py uses a real model to select optional Skills from all declared goals, criteria and original task facts. Configure the same AGENTLY_* variables; AGENTLY_REQUEST_OPTIONS accepts provider-specific JSON options. It inspects preparation and binding reuse without executing scripts.