Skills.Runtime (fnord v0.9.29)

View Source

Runtime helpers for executing skills.

This module owns the glue between skill definitions (TOML) and the runtime components used to execute them:

  • model preset parsing (skill model string -> AI.Model.t())
  • tool tag mapping (skill tools -> AI.Tools.toolbox())
  • response_format validation

Keeping these helpers in one place avoids duplicating the execution rules between the agent (AI.Agent.Skill) and the tool entry points.

Summary

Types

model_error()
response_format_error()
tool_tag()
toolbox_error()

Functions

allowed_model_presets()

Return the list of allowed model presets.

allowed_toolboxes()

Return the list of allowed toolboxes.

model_from_string(model)

Resolve a model preset string (from skill TOML) into an AI.Model struct.

reasoning_preamble()

Returns the reasoning preamble that is prepended to every skill's system prompt.

toolbox_from_tags(tags)

Build a toolbox from skill tool tags.

validate_response_format(map)

Validate a response_format value from a skill.

Types

model_error()

@type model_error() :: {:unknown_model_preset, String.t()}

response_format_error()

@type response_format_error() ::
  {:invalid_response_format, term()} | {:missing_response_format_type, map()}

tool_tag()

@type tool_tag() :: String.t()

toolbox_error()

@type toolbox_error() ::
  {:unknown_tool_tag, tool_tag()} | {:missing_basic_tool_tag, [tool_tag()]}

Functions

allowed_model_presets()

@spec allowed_model_presets() :: [String.t()]

Return the list of allowed model presets.

This list is intended for interactive selection.

allowed_toolboxes()

@spec allowed_toolboxes() :: [tool_tag()]

Return the list of allowed toolboxes.

Toolboxes are the skill's tool tags; they select tool groups.

model_from_string(model)

@spec model_from_string(String.t()) :: {:ok, AI.Model.t()} | {:error, model_error()}

Resolve a model preset string (from skill TOML) into an AI.Model struct.

Supported values:

  • smart
  • balanced
  • fast
  • web
  • large_context
  • large_context:<speed> where speed is smart|balanced|fast

The plain large_context form preserves the default behavior by calling AI.Model.large_context/0.

reasoning_preamble()

@spec reasoning_preamble() :: String.t()

Returns the reasoning preamble that is prepended to every skill's system prompt.

This establishes baseline reasoning discipline for all skill agents.

toolbox_from_tags(tags)

@spec toolbox_from_tags([tool_tag()]) ::
  {:ok, AI.Tools.toolbox()} | {:error, toolbox_error()}

Build a toolbox from skill tool tags.

Tags are mapped to AI.Tools.with_* groupers. Toolbox construction is deterministic and ignores input order.

The basic tag is required; it acts as the toolbox entrypoint.

validate_response_format(map)

@spec validate_response_format(nil | map()) ::
  {:ok, nil | map()} | {:error, response_format_error()}

Validate a response_format value from a skill.

nil is allowed and means default text responses.

When present, the response format must be a map and should include a type key.