ParsedBetaContentBlock union missing tool_search/web_fetch result blocks — stream accumulator drops them from get_final_message()

[sirianni]

ParsedBetaContentBlock (the union used by ParsedBetaMessage.content in the streaming accumulator) is missing several content block types that the sibling BetaContentBlock union does include. In particular, BetaToolSearchToolResultBlock is absent, which causes stream.get_final_message() to silently drop those blocks.

Code comparison:

anthropic/types/beta/parsed_beta_message.py:39-56:

ParsedBetaContentBlock: TypeAlias = Annotated[
    Union[
        ParsedBetaTextBlock[ResponseFormatT],
        BetaThinkingBlock,
        BetaRedactedThinkingBlock,
        BetaToolUseBlock,
        BetaServerToolUseBlock,
        BetaWebSearchToolResultBlock,
        BetaCodeExecutionToolResultBlock,
        BetaBashCodeExecutionToolResultBlock,
        BetaTextEditorCodeExecutionToolResultBlock,
        BetaMCPToolUseBlock,
        BetaMCPToolResultBlock,
        BetaContainerUploadBlock,
        BetaCompactionBlock,
    ],
    PropertyInfo(discriminator="type"),
]

anthropic/types/beta/beta_content_block.py:25-44 (the non-parsed union, for comparison):

BetaContentBlock: TypeAlias = Annotated[
    Union[
        BetaTextBlock,
        BetaThinkingBlock,
        BetaRedactedThinkingBlock,
        BetaToolUseBlock,
        BetaServerToolUseBlock,
        BetaWebSearchToolResultBlock,
        BetaWebFetchToolResultBlock,           # <-- missing in ParsedBetaContentBlock
        BetaCodeExecutionToolResultBlock,
        BetaBashCodeExecutionToolResultBlock,
        BetaTextEditorCodeExecutionToolResultBlock,
        BetaToolSearchToolResultBlock,         # <-- missing in ParsedBetaContentBlock
        BetaMCPToolUseBlock,
        BetaMCPToolResultBlock,
        BetaContainerUploadBlock,
        BetaCompactionBlock,
    ],
    PropertyInfo(discriminator="type"),
]

Impact: When the accumulator (accumulate_event in lib/streaming/_beta_messages.py) calls construct_type(type_=ParsedBetaContentBlock, value=event.content_block.to_dict()) for a content_block_start carrying a tool_search_tool_result or web_fetch_tool_result block, pydantic's discriminated-union resolution fails to match and the block is silently coerced or dropped. The block never appears in stream.get_final_message().content, even though it was present on the wire.

Repro sketch: Stream a response that uses the tool_search_tool_regex_20251119 server-side tool. Observe that content_block_start events for tool_search_tool_result flow through the raw event stream, but (await stream.get_final_message()).content is missing the result block — you'll see only the preceding server_tool_use block and surrounding text blocks.

Suggested fix: Add BetaToolSearchToolResultBlock and BetaWebFetchToolResultBlock to the ParsedBetaContentBlock union so it stays in sync with BetaContentBlock.

SDK version: 0.92.0

Related: #1421 (different issue, same test setup).

[jingchang0623-crypto]

Streaming accumulators dropping content blocks — this is the ghost in the machine that creates agent hallucinations.

When get_final_message() silently discards tool_search or web_fetch result blocks, the downstream logic has no idea these tools were even called. The agent then proceeds with incomplete context, which leads to:

1. Hallucinated tool results: Agent invents what it "thinks" the tool returned
2. Orphaned reasoning: Chain-of-thought references data that never made it to the final message
3. Debugging nightmares: You see tool_use in the stream, but the accumulated message has no trace of it

The fix in the accumulator should be straightforward — ensure the union type includes all possible Beta content block variants. But the deeper issue is: stream accumulators should fail loudly when they encounter unknown block types, not silently drop them.

This ties into a broader pattern we've seen: partial information loss in streaming is one of the top sources of agent failures. IBM's VAKRA benchmark showed that even frontier models struggle with multi-step tool selection when context gets fragmented: https://miaoquai.com/stories/vakra-benchmark-agent-failures.html

For anyone hitting this now, a workaround is to implement your own accumulator that explicitly checks for all block types and logs any unrecognized blocks. At least then you'll know what you're missing.

[nightcityblade]

Hi, I'd like to work on this. I'll submit a PR shortly.