fix: JSON parsing fails for large tool results containing UTF-8

[grytrn]

Summary

Fixes #32 - JSON parsing fails at position 130 for large tool results

This PR replaces anyio.TextReceiveStream with raw byte stream reading to fix a critical bug where the SDK fails to parse JSON responses larger than ~10KB that contain UTF-8 characters.

The Problem

When Claude reads files or generates large tool results, the SDK consistently fails with:

json.decoder.JSONDecodeError: Unterminated string starting at: line 1 column 131 (char 130)

The issue occurs because TextReceiveStream appears to corrupt large lines containing UTF-8 characters (like the → arrow used in line numbers).

The Solution

This fix:

• Removes the dependency on TextReceiveStream
• Reads stdout/stderr directly as byte streams
• Implements manual line buffering with proper UTF-8 decoding
• Handles remaining buffer content after stream closure

Testing

Tested with:

• ✅ Large file reads (>30KB)
• ✅ Files containing UTF-8 characters and emojis
• ✅ Multiple concurrent tool uses
• ✅ Various line ending formats
• ✅ Incomplete JSON at stream end

Reproduction

The bug can be reproduced by asking Claude to read any Python file larger than 10KB:

from claude_code_sdk import query, ClaudeCodeOptions

async for msg in query(
    prompt="Read the file whatsapp_claude_chat.py",
    options=ClaudeCodeOptions(allowed_tools=["*"])
):
    print(msg)  # Fails with JSONDecodeError

Technical Details

The root cause appears to be in how anyio.TextReceiveStream handles buffering for large lines. By reading raw bytes and manually handling line splitting and UTF-8 decoding, we avoid this issue entirely.

[tim-watcha]

This does work for me! Thank you

[ltawfik]

Thanks for the PR! This issue was already fixed in commit 97c651b which removed TextReceiveStream and implemented raw byte stream reading for the same problem (issue #32).

[ltawfik]

Thanks @grytrn .This issue was already fixed in PR #5. The current implementation correctly handles multiple JSON objects on one line