Executor prompt context
This reference explains the internal executor-prompt dogfood skill under
plugins/harness-prompts/skills/executor-prompt/. It is not an eval oracle.
The runtime uses the packaged references/system.md text as the fixed
instruction prefix and appends the selected skill's SKILL.md at the marked
call site in src/skillroll/runtime/execution.py.
The important boundary is that a selected skill can request one generic
world_action tool, while the simulated World and semantic judge remain
separate model actors. Eval files under this skill test observable use of the
boundary; they are excluded from the model-readable skill bundle and never
become executor instructions.