skill-creator: the eval loop assumes it may spawn subagents; in harnesses that forbid unrequested spawning, the main path cannot be followed

What happened

While creating a new skill in Claude Code, "Running and evaluating test cases" opens by spawning a with-skill and a baseline subagent per test case. That harness's instructions forbid spawning agents unless the user explicitly asks. So the documented main path couldn't be followed, and testing fell back to an inline sanity check. The skill describes that fallback only for Claude.ai, not as a general "no subagents here" path.

Suggested fix

Before the eval loop, add an environment check: can subagents be spawned here without the user asking? If not, take the inline path, and offer the parallel with-skill/baseline loop to the user as an opt-in ("I can run N paired test agents; want me to?"). The inline path then becomes the general fallback, instead of being tied to one product surface.

Related: #518 (eval-loop rough edges) and #1467 (the PyYAML dependency, the other environment assumption hit in the same session).

Principle

Before a skill's workflow depends on an environment capability (subagents, a package, a browser, a CLI), it should check that the capability is there and state a fallback. An unchecked assumption fails in the middle of the procedure, where recovering costs the most.