INPUT_VALUE overwritten in _process_input_data leading to loss of multi-message context

[rite7sh]

Summary

SpanAttributes.INPUT_VALUE is overwritten multiple times in _process_input_data, resulting in loss of context for multi-message inputs.

In the current implementation, the final assignment reduces the input to only the first message, which breaks trace correctness for conversational inputs.

---

Affected Code

In _span_io_handler.py:

if eval_input:
    eval_input_str = " \n ".join(map(str, eval_input))
    span.set_attribute(SpanAttributes.INPUT_VALUE, eval_input_str)

if eval_input and len(eval_input) > 0:
    span.set_attribute(SpanAttributes.INPUT_VALUE, eval_input[0])

[pmady]

ran into this while reviewing the codebase for #164. the issue is in _span_io_handler.py lines 55-63 -- INPUT_VALUE gets set three times in sequence:

1. line 55: full json of input_content (correct, has all messages)
2. line 61: joined eval_input strings (loses structure but keeps all messages)
3. line 63: eval_input[0] only (drops everything except the first message)

line 63 always wins and you lose multi-turn context. looks like lines 62-63 are redundant with the block at 59-61 anyway (both check if eval_input). removing 62-63 should fix it without changing behavior for the eval path.

happy to open a PR if that approach makes sense.

[gh-raju]

Thanks @pmady for the precise diagnosis agreed. In _span_io_handler.py, the eval_input[0] assignment (around line 63) overwrites the full multi-message value set earlier at line 55, so everything after the first message is dropped, and lines 62–63 are redundant with the if eval_input block just above.

If you are not already working on it, I would like to pick this up.

Plan:

• Remove the redundant eval_input[0] assignment so the structured/full-context value set earlier is preserved (matches your analysis), keeping the eval path behavior unchanged.
• Add a regression test that asserts INPUT_VALUE retains all messages for a multi-turn input — the exact case that's silently lost today.

Could a maintainer assign this to me? Happy to step back if @pmady wants to carry their own fix.

[pmady]

Sure go for it, @gh-raju , happy that the diagnosis was helpful. This is precisely what I had in mind, remove the unnecessary assignment of eval_input[0] and include the multiple message regression test. Looking forward to reviewing the pull request.

[gh-raju]

Working on it, @pmady I will open a PR against dev shortly with the fix and a multi message regression test.