fix: (1) strip whitespace when parsing action_stop_tokens (2) avoid truncation could split multimodal placeholder token sequences

[zengxingchen]

Problem1

When configuring action_stop_tokens with comma-separated values like:action_stop_tokens='</code>, </search>'
The parser was not stripping whitespace, resulting in:
Token 1: </code> ✅
Token 2: </search> ❌ (with leading space)
This caused the model to fail to stop at , leading to incorrect behavior.

Solution1

Modified the token parsing logic in verltool_agent_loop.py (line 173) to strip whitespace:

# Before
cls.action_stop_tokens = [x for x in f.read().split(',') if x]

# After  
cls.action_stop_tokens = [x.strip() for x in f.read().split(',') if x.strip()]

Problem1

Similar to bugs identified in verl-project/verl#4050.
When truncating agent responses to fit within response_length limits, the truncation could split multimodal placeholder token sequences (e.g., <|vision_start|><|image_pad|><|vision_end|>) in the middle. This causes lots of issues.

Solution2

When truncating responses by response_length, walk backwards to avoid cutting multimodal placeholder tokens (e.g., <|vision_start|>, <|image_pad|>, <|vision_end|>, <|audio_bos|>, etc.) in half.

This ensures that multimodal token sequences are either kept completely or discarded as complete blocks, preventing token corruption that could cause model errors during training.