This PR aims to fix the issue claude code exiting due to parsing failure of a single json message over multiple stream reads Issue 6 Issue 15
The default maximum size of asyncio stream is 64KB. This limit can easily be breached when claude code invokes a tool to read a file of over 1000 lines.
Since a single message has been split over multiple stream reads, attempting to decode it as json leads to JSONDecodeError.

To demonstrate this, let's assume the asyncio stream limit is 5 characters, then the json message {"a": "bc"} would be split into 3 consecutive stream reads of {"a": , "bc" , }
Attempting to parse each of these individually will lead to JSONDecodeError error.

This PR includes a test case in test_subprocess_buffering.py to demonstrate the issue.
The test case fails without the fix.

To fix the issue, a buffer (accumulator variable) is initialised as an empty string. If there is a JSONDecodeError, the stream read is appended to the buffer. This cumulative buffer will be parsed. This will continue until the buffer has the complete JSON and is parsed successfully.

To avoid the buffer increasing infinitely, an arbitrary limit of 1MB (for now) has been put on the buffer, post which it'll result in an OverflowError.