Metadata stripping mutates the original computer-task conversation

[sylvesterkaczmarek]

Summary

strip_all_metadata() is intended to create a cleaned conversation for recorder output, but it uses Pydantic's shallow model_copy() and then assigns filtered metadata onto each copied message.

Because the messages are shared objects in a shallow copy, those assignments also mutate the message metadata in the original conversation.

Current behavior

convo = convo.model_copy(
    update={"metadata": ...}
)
for msg in convo.messages:
    msg.metadata = {...}

model_copy() is shallow unless deep=True, so convo.messages[i] still refers to the same message object as the source conversation.

Impact

Calling the recorder/logging path can remove metadata from the solver's live/final conversation rather than only from the serialized copy. Logging therefore has an unintended side effect on result data and any later consumers of the conversation.

Proposed resolution

Deep-copy the conversation before rewriting message metadata, while keeping the existing metadata allow-list behavior.

Add a regression that strips metadata from a copied conversation and verifies the returned copy is filtered while the original conversation and message metadata remain unchanged.