Proposal: workflow-preview skill (execution contract before run)

[mvanhorn]

Motivation

The collection's value is that users keep control instead of handing the whole process to a framework. Yet a user often cannot see a workflow's concrete cost and side effects until it has started. #894 and #826 report unexpected token cost. #1012 shows hidden skill dependencies. #591 asks for deterministic routing guidance, but a static guide cannot answer what this invocation will load, write, delegate, or ask.

Last-30-days community traffic around installs and host collisions reinforces the same theme: people need clearer contracts before committing a session, not after a long run has already started.

Proposed solution

Add an explicit workflow-preview skill. Given a target skill and optional task description, it reads only the target SKILL.md, directly referenced local files, setup pointers, and current repository conventions. It returns:

• skills and prerequisites likely to be invoked
• blocking questions and where user input is required
• files/artifacts that may be created or changed
• external mutations that the workflow may request
• worker fanout and bounded/parallel stages
• stop, approval, and resume points
• a qualitative cost band with the evidence behind it
• unresolved branches it cannot predict

The preview must not run the target, launch workers, create files, or claim a branch will occur when it depends on future evidence.

Out of scope: a static workflow guide only, auto-running the target, inventing token estimates without evidence from the skill text.

Happy to open a draft PR if this preview boundary looks right. Looking for maintainer signal before writing code.