Message with `finish_reason = "max_tokens"` can contain `toolUse`, but the tool use isn't returned

[azriel-healthpoint]

Describe the bug

Sometimes a tool call in a message isn't returned, and instead a message is logged:

Unknown tag in message content toolUse

This was hit when using Haiku 4.5 as an Evalite / autoevals Factuality scorer.

Please complete the following information:

• Which API you used: [e.g. /chat/completions]
• Which model you used: au.anthropic.claude-haiku-4-5-20251001-v1:0

I think the API is /chat/completions -- it's this call; the issue stems from here:

bedrock-access-gateway/src/api/models/bedrock.py

Line 876 in 7e03ab0

if finish_reason == "tool_use":

For the finish_reason, I was receiving "max_tokens", and content was:

[{
  "toolUse": {
    "toolUseId": "tooluse_abcdef123456_abcd1234",
    "name": "select_choice",
    "input": { "choice": "D" }
  }
}]

Expected behavior

I expected the tool use to be returned

Additional context

I'll open a PR (#198) with some code changes that got this to work -- just check for "toolUse" even if the finish_reason isn't "tool_use". I'm not sure if it's the right solution for all cases -- it worked for my case.

If anyone else hits the same issue, it's because maxTokens is hardcoded to 512 in autoevals, and presumably Haiku 4.5's tokenization is different and exceeds that.