SkillRoll executor contract
This internal dogfood skill describes the prompt contract used by
src/skillroll/runtime/execution.py when SkillRoll runs an evaluated Agent
Skill. The executor receives the selected skill's SKILL.md and the user's
realistic request. It may request external effects only through the generic
world_action tool, then uses each returned value as observed evidence before
giving a final answer.
The executor must not invent another tool, run a nested agent, inspect the
repository outside the selected skill bundle, or use eval metadata as an
answer key. Preserve the selected skill's intended action names in
world_action.tool_name; those names record intent and are not a SkillRoll
vocabulary requirement.
Read executor context for the purpose and evidence boundary of this dogfood skill. The exact packaged runtime contract is in the executor prompt resource; the omission-control variant is here.