fix(anthropic): coerce max_tokens to int to fix validation error (#1519)

[carlosjarenom]

Changes

• backend/app/agent/agent_model.py: Added coercion of max_tokens to integer for Anthropic models.

Root Cause

max_tokens can arrive as a string from:

• Environment variables (all env vars are strings)
• JSON config from frontend
• Legacy extra_params

Anthropic API requires max_tokens to be an integer, causing 400 error:

Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'max_tokens: Input should be a valid integer'}}

Fix

Added type coercion with fallback to default (128000) on invalid values:

if "max_tokens" in model_config and isinstance(model_config["max_tokens"], str):
    try:
        model_config["max_tokens"] = int(model_config["max_tokens"])
    except (ValueError, TypeError):
        logger.warning(...)
        model_config["max_tokens"] = 128000

Fixes

• [BUG] Anthropic API Key cannot configured in MacOS desktop app #1519: Anthropic API Key cannot be configured in MacOS desktop app (and Linux/Win)

[4pmtong]

Thanks for the contribution and for looking into these issues. @carlosjarenom

I found a few blocking points:

• The Groq change is outside this PR’s description. Also, [BUG] Groq API incompatibility with Browser Agent tool schema #1511 uses Groq through the OpenAI-compatible provider, so the model_platform == "groq" check will not run. On the current CAMEL version, the browser_type schema is already strict. Please remove this change or submit it separately with a failing regression test.
• The Anthropic conversion only runs when creating the runtime agent. The model settings flow calls /model/validate first, where a string max_tokens can still fail. Please use shared normalization for both runtime and validation.
• Please do not silently replace an invalid value with 128000, since token limits differ between Anthropic models. Validation should report invalid values clearly, and any runtime fallback should be conservative.
• Please add regression tests for both the validation and runtime paths.

We’ll keep this PR open for now. If you have time to update it, we’ll be happy to review it again.