Runs an activated skill as a forked sub-conversation in the parent agent's process.

A skill activation is not a new agent and does not spawn a new supervision tree. Instead, it runs a synchronous LLM loop in-process using a forked copy of the parent's messages and an expanded tool list.

The sub-loop:

• Uses the parent's system prompt with the rendered skill body appended.
• Inherits the parent's :tools and adds the skill's underlying tool module (deduplicated by module).
• Does NOT expose activate_skill or subagent-delegation tools — nested skill activation is not supported.
• Loops LLM.stream/2 plus in-process tool execution until the LLM returns a final assistant message with no tool calls.
• Returns that final text as the activate_skill tool's result.

Events streamed by the sub-loop (%Delta{}, tool call events, tool results, token usage) are forwarded to the parent's caller pid tagged with a synthetic agent name "<parent>/skill:<skill_name>", so chat printers that match by agent-name prefix surface them alongside the parent's own events.

Wrapped in Hooks.call(:skill_activation) so [:skill_kit, :skill_activation, :start/:stop] telemetry fires and :pre_skill_activation / :post_skill_activation hooks can observe or deny the activation.

Implementation

The LLM + tool-dispatch loop is the shared SkillKit.Agent.SubLoop primitive. This module's responsibility is preparing the skill-specific scope (tool list with skill's tool appended, fork the parent messages to drop the trailing activate_skill tool use, compose the skill-prefixed sub-agent name) and wrapping the run in the :skill_activation hook.