Claude API 400 invalid_request_error: 'messages.N: `tool_use` ids were found without `tool_result` blocks immediately after: toolu_...' — how do I fix my message history?

Cause: when Claude's response contains tool_use blocks, the very next message must be a user message that includes a tool_result block for every one of those tool_use ids. The error means at least one tool call in the history has no matching result right after it. Common reasons: - A tool threw an exception, timed out or was cancelled, and the agent loop skipped adding its result. - Several tools were called in parallel and only some results were sent back. - History was trimmed, summarized or reloaded from storage, and the cut separated an assistant tool_use message from its results. - The user message with the results put text before the tool_result blocks, or results were split across several messages, or a framework converting between OpenAI and Anthropic formats reordered blocks. - The user interrupted mid-turn and a new user message was appended after an unanswered tool_use. How to check: log the messages array right before the failing request, find message N from the error, and confirm the next message has a tool_result for every tool_use id in it. Fix: 1. Always return one tool_result per tool_use id, in a single user message immediately after the assistant message. If the tool failed, still return a tool_result with is_error: true and a short error text. 2. Put the tool_result blocks first in that user message's content; any extra text goes after them. 3. When trimming or summarizing history, drop an assistant tool_use message together with its tool_result message, never one without the other. 4. If the user interrupts, add tool_results (for example is_error: true, 'cancelled by user') for the pending calls before the new user text. 5. Persist whole content blocks (type, id, tool_use_id), not just text. 6. Server tools such as web search or web fetch keep their result in the assistant turn; don't strip or move those blocks. After: add a pre-request check that pairs every tool_use id with a tool_result and repairs gaps with error results, so one failed tool doesn't break the whole conversation. Sources: https://platform.claude.com/docs/en/agents-and-tools/tool-use/implement-tool-use https://github.com/openai/openai-agents-python/issues/1797 https://feedback.boltai.com/p/webfetch-servertooluse-block-sent-without-its-toolresult-anthropic-400-on