ValueError: Found AIMessages with tool_calls that do not have a corresponding ToolMessage

TL;DR

Somewhere your message history lost the tool results while keeping the AI message that requested them. Providers reject that shape. The fix is to never split the pair: when you trim or filter history, drop the orphaned AI message too, or rebuild the list so every tool call has its ToolMessage.

ValueError: Found AIMessages with tool_calls that do not have a corresponding ToolMessage

Fix it

1. Find where history gets modified: trimming, filtering, summarization, or a custom get_session_history. Expected: you locate the spot that drops ToolMessages.
2. Keep pairs together. When you slice history, cut at message boundaries so an AIMessage with tool_calls is never separated from its ToolMessage replies. Expected: re-running the same conversation no longer raises.
3. If you filter messages (for example removing old tool output to save tokens), also remove the AI message whose tool_call_ids you deleted. Expected: the provider sees a clean alternating history.
4. In LangGraph, let ToolNode append tool results instead of hand-rolling message surgery, and use the add_messages reducer so updates merge by id. Expected: pairing is maintained by the framework.

When this applies

• You trim, summarize, or filter chat history and then call the model again.
• You rebuild history on resume (from a checkpointer or your own store) and some ToolMessages are missing.

When this does NOT apply

• The error names a different missing piece (a ToolMessage without a matching AI call is the reverse problem).
• You never use tools; then the history shape issue is something else entirely.

Compatibility

• langchain-core 0.1+, LangGraph 0.1+. Applies to any provider; OpenAI and Anthropic both validate the pairing.

Root cause

Tool calling is a two-message contract: the AI message says "call this tool with this id" and a later ToolMessage says "here is the result for that id". Providers validate the contract server-side. Any history manipulation that deletes one side but keeps the other produces a conversation no provider will accept.

Edge cases

• Summarization that compresses tool exchanges into prose must remove both messages, not just the tool output.
• Parallel tool calls produce several ToolMessages for one AI message; keep all of them or drop the whole group.
• Streaming chunks reassembled by hand can lose tool_call ids; prefer the framework's chunk-merging helpers.