What is the most robust way of downcasting an `anyhow::Error`?

I am using `anyhow` extensively in my application. At the outer layer of an event loop, I am then downcasting the error to check how it needs it be handled and which log-level is appropriate.

So far, I've used `root_cause` for this but recently, that has failed on my because I was using `bail!` further down but without applying any additional `.context`. Hence, the resulting `anyhow::Error` did in fact not have a source and the check never succeeded, causing the error handling to be buggy.

I then discovered the interactions of `context` and the `is` and `downcast` methods.

My question now is: What is the most robust way of checking whether an `anyhow::Error` is / contains a particular type? It seems like to cover all bases, I need to both check `is` and `root_cause`.

Type-erasure is great for ergonomics but less so if it can easily be misused or broken by a refactoring (like the removal / addition of `context`).

Should `is` be extended to also walk all sources of the original error prior to being turned into an `anyhow::Error`?