skill-and-command-dispatch

To run slash commands in an isolated forked subagent, set the context configuration to `fork` in the SKILL.md frontmatter, which executes the prompt in a separate subagent and returns only the final result to prevent multi-step reasoning from cluttering the main session.

Inline execution shares the current conversation history and runs directly in the main session, while fork execution runs a subagent in isolation and returns only the final result, preventing intermediate reasoning steps from cluttering the main conversation.

You configure allowedTools and agent selection by defining them in the SKILL.md frontmatter, which specifies the prompt templates, tool access permissions, and execution environment required for getPromptForCommand integration.

Use PromptCommand for LLM-driven templates defined in SKILL.md files, and use LocalCommand for deterministic local actions. PromptCommand supports multi-step research workflows, while LocalCommand handles predictable, non-LLM tasks.

Yes, you can prevent multi-step research workflows from cluttering your main conversation by configuring the command context to `fork`, which isolates the multi-step reasoning in a subagent and returns only the final result.

Yes, execution isolation applies to both SKILL.md-based PromptCommand skills and built-in PromptCommand or LocalCommand scenarios, allowing you to control inline or forked execution for automating prompts and local actions.