[BUG] LiteLLM reports wrong output token count (max_tokens vs max_output_tokens)

[fabb]

Problem (one or two sentences)

When trying to use Sonnet 4.5 via LiteLLM via Google Vertex, this error shows up:

LiteLLM streaming error: 400 litellm.BadRequestError: VertexAIException BadRequestError - b'{"type":"error","error":{"type":"invalid_request_error","message":"max_tokens: 200000 > 64000, which is the maximum allowed number of output tokens for claude-sonnet-4-5-20250929"},"request_id":"req_vrtx_011CTeGWyomNL2s6LacBN6w5"}'. Received Model Group=claude-sonnet-4-5

The issue seems to be a confusion of max_tokens and max_output_tokens:

The problem is most likely in this line:

Roo-Code/src/api/providers/fetchers/litellm.ts

Line 59 in 13534cc

maxTokens: modelInfo.max_tokens || 8192,

I think in this line, max_output_tokens should be used if available, and max_tokens only as fallback.

Context (who is affected and when)

LiteLLM with Sonnet 4.5 on Google Vertex

Reproduction steps

1. Add Sonnet 4.5 via LiteLLM via Google Vertex

Expected result

Prompts should work

Actual result

Prompts fail, because the requests ask for 200k output tokens, where the maximum is 64k

Variations tried (optional)

No response

App Version

3.28.14

API Provider (optional)

LiteLLM

Model Used (optional)

Sonnet 4.5 via Google Vertex

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I'm exploring a potential fix for this issue where LiteLLM incorrectly uses max_tokens instead of max_output_tokens for Claude Sonnet 4.5 via Google Vertex. I'll investigate the suggested fix in litellm.ts and open a PR to share progress.

[roomote]

I've opened PR #8455 with a fix for this issue: #8455

The fix modifies the LiteLLM fetcher to prefer max_output_tokens when available, falling back to max_tokens for backward compatibility. This should resolve the issue where Claude Sonnet 4.5 via Google Vertex was incorrectly using the context window size (200k) instead of the output token limit (64k).

The implementation includes comprehensive test coverage to ensure the correct behavior in all scenarios. All unit tests and code quality checks are passing.

[davidwindell]

Also seeing the same in non-vertex usage of Sonnet via LiteLLM here.

[anengineerdude]

Is this a roo-code issue or something that should be fixed in the LiteLLM proxy as well? Encountering the same issue in Kilo with LiteLLM.

[fabb]

Is this a roo-code issue or something that should be fixed in the LiteLLM proxy as well? Encountering the same issue in Kilo with LiteLLM.

Not sure: https://github.com/BerriAI/litellm/blob/12dfd14b5281bed673e284b300d148eea35c9d65/model_prices_and_context_window.json#L3