Declarative Skill Configuration

A read-only S7 value bundling a prompt extension, original ellmer tools, and declared package and provider requirements. Load a Skill into an Agent with agent$load_skill(skill).

Usage

Skill(
  name,
  version = "0.0.0",
  description = NULL,
  prompt = NULL,
  tools = list(),
  requires = list(),
  path = NULL
)

Arguments

name: Non-empty skill name, retained as supplied.
version: Non-empty version string. Defaults to "0.0.0"; skill_create() retains its "1.0.0" default.
description: Optional description.
prompt: Optional system prompt extension.
tools: List of tools created with ellmer::tool().
requires: List with optional packages and providers entries. Each entry contains strings, or is NULL or an empty sequence. Other declarative metadata is retained but not evaluated.
path: Optional path to the skill directory. Construction does not inspect or load this path.

Value

A read-only Skill S7 value.

Details

Construct skills with Skill(...), skill_create(), or skill_load(). Read properties with $, @, or S7::prop(). To revise configuration, edit S7::props(skill) and pass that list to do.call(Skill, fields). Individual and bulk property replacement are rejected, including initially NULL fields. Requirements are declarative package/provider sequences.

Executable tools are composed without cloning. Their closures, clients, services, and caller-owned environments retain their original semantics; read-only Skill configuration does not freeze state inside those tools. Neither construction nor skill_check_requirements() executes tool code. skill_load() remains the explicit boundary that can source declared tools.

RDS can preserve the value and serializable R object graphs after Deputy loads in the receiving process. It is not a portable transport for live services or connections. AgentDefinition YAML uses explicit host registries to attach skills; it does not embed executable Skill objects.

Examples

concise <- Skill("concise", prompt = "Answer in one short paragraph.")
fields <- S7::props(concise)
fields$prompt <- "Answer in one sentence."
shorter <- do.call(Skill, fields)
shorter$prompt
#> [1] "Answer in one sentence."
concise$prompt
#> [1] "Answer in one short paragraph."
skill_check_requirements(concise)$ok
#> [1] TRUE