Class SkillToolset

constructor

• new SkillToolset(
      skills: Record<string, Skill> | Skill[],
      options?: {
          additionalTools?: (BaseTool | BaseToolset)[];
          allowInlineScripts?: boolean;
          codeExecutor?: BaseCodeExecutor;
          registry?: SkillRegistry;
          scriptOutputDir?: string;
      },
  ): SkillToolset

  Parameters

  • skills: Record<string, Skill> | Skill[]
  • options: {
        additionalTools?: (BaseTool | BaseToolset)[];
        allowInlineScripts?: boolean;
        codeExecutor?: BaseCodeExecutor;
        registry?: SkillRegistry;
        scriptOutputDir?: string;
    } = {}

    • OptionaladditionalTools?: (BaseTool | BaseToolset)[]

    • OptionalallowInlineScripts?: boolean

      Whether to expose the run_skill_inline_script tool, which executes model-provided script content in the configured code executor. This is disabled by default because arbitrary inline-script execution is a sensitive capability; opt in explicitly by setting this to true. Execution additionally remains gated behind a human-in-the-loop confirmation.

    • OptionalcodeExecutor?: BaseCodeExecutor

    • Optionalregistry?: SkillRegistry

    • OptionalscriptOutputDir?: string

      Directory that files produced by run_skill_script / run_skill_inline_script are written into. Output file names come from the executed script, so this must be a directory dedicated to throwaway script output — never the application's working directory or a source tree.

      Defaults to a private, randomly named directory created under the OS temp dir on first use. Nothing removes that directory: close() runs once per invocation on a toolset instance that concurrent invocations share (see Runner), so it cannot tell whose output is still in use. Set this to own the location and the lifetime of the output.

  Returns SkillToolset

additionalTools

OptionalcodeExecutor

Optionalregistry

skills

close

• close(): Promise<void>

  Closes the toolset.

  NOTE: This method is invoked, for example, at the end of an agent server's lifecycle or when the toolset is no longer needed. Implementations should ensure that any open connections, files, or other managed resources are properly released to prevent leaks.

  Returns Promise<void>

  A Promise that resolves when the toolset is closed.

getOrFetchSkill

• getOrFetchSkill(name: string, invocationId?: string): Promise<Skill | undefined>

  Parameters

  • name: string
  • OptionalinvocationId: string

  Returns Promise<Skill | undefined>

getScriptOutputDir

• getScriptOutputDir(): Promise<string>

  Resolves the directory that script output files are materialized into.

  Script output file names are attacker-influenced (a prompt-injected skill controls what its script writes), so they must never be resolved against the host application's working directory. When no scriptOutputDir was configured this hands back a private temp directory instead, created once and reused for the lifetime of the toolset.

  Returns Promise<string>

getSkill

• getSkill(name: string): Skill | undefined

  Parameters

  • name: string

  Returns Skill | undefined

getTools

• getTools(context?: ReadonlyContext): Promise<BaseTool[]>

  Returns the tools that should be exposed to LLM.

  Parameters

  • Optionalcontext: ReadonlyContext

    Context used to filter tools available to the agent. If not defined, all tools in the toolset are returned.

  Returns Promise<BaseTool[]>

  A Promise that resolves to the list of tools.

processLlmRequest

• processLlmRequest(toolContext: Context, llmRequest: LlmRequest): Promise<void>

  Processes the outgoing LLM request for this toolset. This method will be called before each tool processes the llm request.

  Use cases:

  • Instead of let each tool process the llm request, we can let the toolset process the llm request. e.g. ComputerUseToolset can add computer use tool to the llm request.

  Parameters

  • toolContext: Context
  • llmRequest: LlmRequest

  Returns Promise<void>