Fix uncaught TypeError on string or non-dict message content

[adventurelands]

Closes #1064.

Small fix: assistant message with string content raises an uncaught TypeError

Thanks for open-sourcing this.

Minor one. parse_message handles string content for user messages but not
for assistant messages, so an assistant message whose content is a plain
string gets iterated character by character and raises a TypeError instead of
the documented MessageParseError. A non-dict block in the content list does
the same in both branches.

from claude_agent_sdk._internal.message_parser import parse_message
parse_message({"type": "user",      "message": {"content": "hi"}})               # ok
parse_message({"type": "assistant", "message": {"model": "m", "content": "hi"}})  # TypeError

The patch makes assistant mirror the user branch (wrap a string as one
TextBlock) and raises MessageParseError on a non-dict block. Added a test
that fails on main and passes with the patch. Nothing else touched.

Just figured I would pass it along. Thanks again.

Davis Brief, davisbrief@gmail.com

[adventurelands]

Filed #1064 with a minimal verified repro of the underlying crash (assistant content as a string and a non-dict block both raise an uncaught TypeError on latest main; the user branch already handles the string case). This PR closes it, with regression tests. Happy to rebase or reshape the string-to-TextBlock handling if you'd prefer a different approach. Thanks for taking a look.

[adventurelands]

@ashwin-ant @qing-ant — small fix here closing #1064 (verified repro: an assistant message with content as a string or a non-dict block raises an uncaught TypeError on latest main; the user branch already handles the string case). +126/−1 with regression tests. Mind kicking off CI / a quick look? Happy to reshape the string→TextBlock handling if you'd prefer a different approach. Thanks!

[ashwin-ant]

Thanks for the PR! The non-dict block guard is a real fix and I'd like to land that part. A couple of things to sort out first though:

The else raw_content arm actually makes things worse. If content is something weird like None or a bare dict, the ternary now passes it straight through into AssistantMessage.content (which is typed list[ContentBlock]). On main that case blows up immediately with a TypeError inside the parser; with this patch it returns a silently-broken object and the crash moves downstream to whoever iterates msg.content. That's the opposite of what we want here.

The string-content case isn't actually a thing for assistant messages. The CLI emits BetaMessage for assistant turns, and BetaMessage.content is always an array — the string shorthand only exists on the input side (MessageParam). That's why AssistantMessage.content is list[ContentBlock] while UserMessage.content is str | list[...]. So rather than normalizing a string into [TextBlock(...)], non-list assistant content should just raise MessageParseError like the other guards you added.

That also means we don't need the second AssistantMessage(...) return — duplicating those 8 field mappings is a drift hazard. Something like this keeps it to one constructor:

raw_content = data["message"]["content"]
if not isinstance(raw_content, list):
    raise MessageParseError(
        f"Invalid assistant content (expected list, got {type(raw_content).__name__})",
        data,
    )
content_blocks: list[ContentBlock] = []
for block in raw_content:
    if not isinstance(block, dict):
        raise MessageParseError(...)
    ...

Small test cleanup while you're in there:

• Drop the Place under "tests/"... line from the docstring
• test_normal_block_lists_unaffected and test_user_string_content_still_parses duplicate existing tests in test_message_parser.py — can go
• Swap the test_assistant_string_content_* tests for one that asserts MessageParseError on string content
• If you don't mind, fold what's left into class TestMessageParser in tests/test_message_parser.py so it lives with the other parser tests

The isinstance(block, dict) guards and the parametrized error test are good as-is. 👍