HeadersDict is a subclass of CaseInsensitiveDict with two new features:
1. Preserve the case of the header key from the first time it was set.
This means that later munging won't modify the key case. (You can
force picking up the new case with `del` followed by setting.)
2. If the value is a list or tuple, unpack it and store the first
element. This is the same as how `Request.add_header()` used to work.
For backward compatibility this commit preserves `Request.add_header()` but
marks it deprecated.
There is a weird quirk in HTTP. You can send the same header twice.
For this reason, headers are represented by a dict, with lists as the
values. However, it appears that HTTPlib is completely incapable of
sending the same header twice. This puts me in a weird position: I want
to be able to accurately represent HTTP headers in cassettes, but I
don't want the extra step of always having to do [0] in the general
case, i.e. request.headers['key'][0]
In addition, some servers sometimes send the same header more than once,
and httplib *can* deal with this situation.
Futhermore, I wanted to keep the request and response cassette format as
similar as possible.
For this reason, in cassettes I keep a dict with lists as keys, but once
deserialized into VCR, I keep them as plain, naked dicts.
So the stubs were getting out of hand, and while trying to add support for the
putrequest and putheader methods, I had an idea for a cleaner way to handle
the stubs using the VCRHTTPConnection more as a proxy object. So
VCRHTTPConnection and VCRHTTPSConnection no longer inherit from HTTPConnection
and HTTPSConnection. This allowed me to get rid of quite a bit of
copy-and-pasted stdlib code.
This is a backwards-incompatible change that will store headers
as a list rather than a dictionary. The reason being that you can
have multiple values for a single header, and concatenating them
together with commas can create an unparseable string (sometimes
the header values can have commas in them)
`responses_of` replaces `response_of`, since each request can have
several matching responses now.
This breaks backwards compatibility if you are using the
response_of method, so a version bump will be required.
