ClearToolUsesEdit Does Not Persist Cleared State With Checkpointer, Causing Eviction Every Turn

The middleware applies ClearToolUsesEdit to a deep copy of request.messages and discards the copy after the model call. The cleared ToolMessages and their response_metadata.context_editing.cleared flags are not returned as state updates, so the checkpointer retains the original full-content messages. On the next turn, the raw state still contains all uncleared ToolMessages, the trigger remains true, and the same messages are re-cleared indefinitely.

Install langchain==1.3.1 and langchain-core==1.4.0. Create a persistent state list mimicking checkpointed messages. Instantiate ClearToolUsesEdit with trigger=10000, keep=5. Each turn append one AIMessage with tool call and one ToolMessage with 4000 characters. Run simulate_middleware_call which deepcopies state and calls edit.apply. Observe that edited messages are cleared but original state remains full; after turn 1, trigger always fires and cleared messages never accumulate flags in checkpoint state.

This fix relies on the agent runtime invoking after_model after the model call. If after_model is not called in a custom runtime or older LangChain version, the fix has no effect. RemoveMessage uses message IDs; ensure ToolMessage IDs are unique and stable across checkpoints. Concurrent turns or multi-threaded use may cause race conditions if the checkpointer is not transaction-safe. Rolling back to the default ContextEditingMiddleware will restore the original bug but does not lose data besides already-cleared tool contents (which can be re-fetched). Test with your specific checkpointer and agent topology before production.