A saga cannot rewind remote history. Each compensation is a new, fallible business operation with its own identity, policy, evidence, and possible manual outcome.
A successful agent run is only part of a CI review. Track the checked revision, keep test results independent, and handle incomplete output before publishing findings.
Adding context: fork to a skill runs it in an isolated subagent instead of inline. The skill body becomes the subagent's task, the noisy work stays contained, and your main conversation gets back only the useful result.
Background subagents run concurrently, but they can't stop to ask you questions or negotiate permissions mid-run. Here is when concurrency saves time and when it just fails more quietly.
Most skill failures are not caused by bad instructions in the body — they are caused by a weak description. That is the line Claude uses to decide whether a skill is relevant enough to load. Here is how to fix it.