Refactor anthropic message format to support tool messages & Allows Anthropic Bedrock clients

[lazyhope]

Instead of concatenating strings with "\n\n", we use the following format to merge consecutive Anthropic messages of the same roles:

{
  "role": "user", 
  "content": [
    {"type": "text", "text": "Hello, Claude"},
    {"type": "text", "text": "How is it going?"}
  ]
}

Which also allows other message types like image objects, tool invocation (for assistant), tool results, etc. As a comparsion, this version of tool support is able to pass the anthropic tests which currently fails.

---

	🚀 This PR description was created by Ellipsis for commit 9705a6f.

Summary:

This PR refactors the anthropic message format to support tool messages, updates the reask_messages and merge_consecutive_messages functions, and modifies the corresponding tests.

Key points:

• Refactored reask_messages function in /instructor/retry.py to include the original response and format error messages differently.
• Updated merge_consecutive_messages function in /instructor/utils.py to handle the new message format.
• Updated tests in /tests/test_utils.py to reflect these changes.

---

Generated with ❤️ by ellipsis.dev

[ellipsis-dev]

👍 Looks good to me!

• Reviewed the entire pull request up to 9705a6f
• Looked at 117 lines of code in 3 files
• Took 46 seconds to review

More info

• Skipped 0 files when reviewing.
• Skipped posting 1 additional comments because they didn't meet confidence threshold of 85%.

1. instructor/retry.py:45:

• Assessed confidence : 50%
• Comment:
  The error message in the assertion could be more descriptive. Instead of 'Tool use ID not found in the response', it could be 'Expected tool use ID in the response but none found'.

assert tool_use_id is not None, "Expected tool use ID in the response but none found"

• Reasoning:
  The changes in the PR seem to be logically correct and follow the best practices. The code is not overly complicated and the variable names are descriptive. The function naming follows a consistent pattern. The changes in the library code are reflected in the tests. The assertions have well-formatted error messages. However, I noticed that the error message in the assertion at line 45 in retry.py could be more descriptive. Instead of 'Tool use ID not found in the response', it could be 'Expected tool use ID in the response but none found'.

Workflow ID: wflow_rMlGfFzdJ9S22wtA

---

Not what you expected? You can customize the content of the reviews using rules. Learn more here.

[jxnl]

1. skip anthropic bedrock tests
2. add some logic to avoid doing the merge for openai keys

[jxnl]

is this good to merge?

[lazyhope]

is this good to merge?

Yeah, I was waiting for it to be merged these days 😂