Handle Anthropic tool_use retries on ValidationError

[kelvin-tran]

Auto-retry when is not working when using Anthropic Claude tool use because of this error:

Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.2: `tool_use` ids were found without `tool_result` blocks immediately after: toolu_01KELz18M6qYvnwPEYi9wZKx. Each `tool_use` block must have a corresponding `tool_result` block in the next message.'}, 'request_id': 'req_011CT5ysHdXtzNDMm2WwNgKc'}

This is when first attempt resulted in ValidationError, eg,:

<generation number="1">
<exception>
    1 validation error
...

To Reproduce
Something like the below when the client is instantiated with mode=instructor.Mode.ANTHROPIC_TOOLS or the equivalent:

client.messages.create(
    model=model_id,
    messages=instruct_msgs,
    system=system_prompt,
    response_model=response_model,
)

Expected behavior and fix
That retried requests will contain the tool call result and successful generate responses (whether valid or not)

Root cause
In instructor/core/retry.py line 599, the handle_reask_kwargs() function calls:

exception = InstructorError.from_exception(exception, failed_attempts=failed_attempts)

This converts pydantic ValidationError exceptions to instructor ValidationError exceptions, so by the time the exception reaches reask_anthropic_tools(), it's an instructor ValidationError, not a pydantic one.

Fix Applied

File: instructor/providers/anthropic/utils.py
Change:

        if (
            content.type == "tool_use"
            and isinstance(exception, ValidationError)
            and content.name == exception.title
        ):

to

        if (
            content.type == "tool_use"
            and isinstance(exception, ValidationError | InstructorValidationError)   # should really just be InstructorValidationError, but being safe
            and content.name == exception.title
        ):

Seems only the providers/anthropic/utils.py has the problematic pattern as it is the only provider that:

• uses isinstance(exception, ValidationError) checks
• accesses the exception.title attribute
• imports ValidationError from pydantic

So OpenAI, Mistral, Gemini, Cohere, Cerebras, Fireworks, Writer, XAI, Bedrock, Perplexity, VertexAI seem unaffected.

---

Important

Fixes auto-retry issue in reask_anthropic_tools() by handling both ValidationError and InstructorValidationError for Anthropic provider.

• Behavior:
  • Fixes auto-retry issue in reask_anthropic_tools() in utils.py by handling both ValidationError and InstructorValidationError.
  • Ensures tool call results are included in retried requests for Anthropic Claude tool.
• Exceptions:
  • Imports InstructorValidationError from core.exceptions in utils.py.
  • Updates isinstance check to include InstructorValidationError in reask_anthropic_tools().
• Scope:
  • Only affects Anthropic provider; other providers like OpenAI, Mistral, etc., are unaffected.

^{This description was created by for b113c93. You can customize this summary. It will automatically update as commits are pushed.}

[ellipsis-dev]

Important

Looks good to me! 👍

Reviewed everything up to b113c93 in 29 seconds. Click for details.

• Reviewed 21 lines of code in 1 files
• Skipped 0 files when reviewing.
• Skipped posting 1 draft comments. View those below.
• Modify your settings and rules to customize what types of comments Ellipsis leaves. And don't forget to react with 👍 or 👎 to teach Ellipsis.

1. instructor/providers/anthropic/utils.py:161

• Draft comment:
  Using the union operator (ValidationError | InstructorValidationError) is concise, but consider using a tuple (i.e., isinstance(exception, (ValidationError, InstructorValidationError))) for better compatibility with Python versions < 3.10 if needed.
• Reason this comment was not posted:
  Confidence changes required: 80% <= threshold 85% None

Workflow ID: wflow_kWOx7kI2oe9JZ86u

^{You can customize by changing your verbosity settings, reacting with 👍 or 👎, replying to comments, or adding code review rules.}

[kelvin-tran]

This is caused by changes in #1804. The InstructorError.from_exception() method that enriches exceptions with failed attempts data effectively creates a different class of ValidationError, such that conditionals (eg, exception handlers/if branches) that use isinstance(e, ValidationError, where ValidationError is the pydantic-type err, no longer work as previously