Repository navigation
INPUT_VALUE overwritten in _process_input_data leading to loss of multi-message context #151
Description
Activity
ran into this while reviewing the codebase for #164. the issue is in
_span_io_handler.pylines 55-63 --INPUT_VALUEgets set three times in sequence:- line 55: full json of
input_content(correct, has all messages) - line 61: joined
eval_inputstrings (loses structure but keeps all messages) - 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.
- line 55: full json of
Thanks @pmady for the precise diagnosis agreed. In
_span_io_handler.py, theeval_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 theif eval_inputblock 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.
- Remove the redundant
Summary
SpanAttributes.INPUT_VALUEis 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: