OpenAI API 400 'context_length_exceeded' ('This model's maximum context length is N tokens. However, your messages resulted in M tokens'), or the Responses API failing because input exceeds the context window. How do I fix it in an agent loop?

What it is: every model has a context window, the maximum number of tokens one request can use. It counts the input (system prompt, conversation history, tool definitions and tool results, files and images) plus the output, and for reasoning models the reasoning tokens too. When the request doesn't fit, the API rejects it with a 400 error (code context_length_exceeded in Chat Completions). Agent loops hit this as history and tool outputs pile up turn after turn, or when one tool returns a huge result (a whole file, a long web page, a big JSON dump). Am I affected: - The error appears after a conversation or agent run has gone on for a while, or right after a large tool result or file was added. - Check the usage.input_tokens of the previous responses, and look up your model's context window and maximum output tokens on OpenAI's models page. A large max_output_tokens setting also uses up room. Fix: 1. Measure before sending: call the input token count endpoint (client.responses.input_tokens.count in the Python SDK) with the same payload. It counts tools, images and formatting that local tokenizers like tiktoken miss; tiktoken is fine for rough plain-text estimates. 2. Use compaction for long conversations on the Responses API: set context_management with a compact_threshold so the server compacts older context automatically, and carry the compaction item forward as the docs describe. 3. Or let the API drop old items: truncation set to "auto" lets the Responses API truncate earlier input when it exceeds the window (the default "disabled" fails with a 400 instead). Know that this silently drops early context. 4. Keep tool results small: return summaries, excerpts or paginated results instead of whole files, and cap tool output length in your tool code. 5. Trim or summarize history yourself: keep the system prompt and recent turns, replace old turns with a summary, and remove tool definitions you don't need for the task. 6. Lower max_output_tokens to what you really need, or switch to a model with a larger context window. Verify: the input token count for the next request plus your max_output_tokens stays under the model's context window, and long agent runs keep working past the point where they used to fail. Sources: https://developers.openai.com/api/docs/guides/conversation-state https://developers.openai.com/api/docs/guides/token-counting https://developers.openai.com/api/docs/guides/compaction https://developers.openai.com/api/docs/guides/error-codes