Encrypted Reasoning Content Filtering

{
  "error": {
    "message": "The encrypted content gAAA...xrM= could not be verified. Reason: Encrypted content organization_id did not match the target organization.",
    "type": "invalid_request_error",
    "code": "invalid_encrypted_content"
  }
}

​

When do you run into this problem?

• You use a tool such as Codex / OpenCode to call the Responses endpoint
• The tool automatically adds include: ["reasoning.encrypted_content"] to the request, and in the next request sends back the reasoning.encrypted_content (or the type:"reasoning" item) returned last time
• Meanwhile, Poixe, as an aggregation gateway, schedules across multiple Keys / multiple organizations (org/project) for capacity and stability This causes the “encrypted fragment generated in the previous turn” to mismatch the “organization/Key actually hit in the next turn”, triggering a verification failure (the organization_id did not match you see)

​

How Poixe handles it: filtering encrypted-reasoning fields

1. Removing include.reasoning.encrypted_content from the request (so the upstream does not return this field)
2. Cleaning up reasoning fragments in the historical items of the request body
   • Deleting items with type: "reasoning", as well as all encrypted_content fields

​

Impact and notes

• Filtering does not “delete your conversation content”; your input/messages are still forwarded normally
• However, in stateless multi-turn flows, without the “reasoning continuation boost” from reasoning.encrypted_content, some agent tasks may show:
  • Reduced multi-turn consistency
  • Repeated planning and a stronger sense of lost context (this is a side effect of the inherent conflict between the upstream mechanism design and the proxy architecture)

​

Need help?

​

References

​

Official documentation (recommended reading first)

• Responses API: the include parameter and related fields
  https://platform.openai.com/docs/api-reference/responses/create#responses_create-include
• Reasoning guide: reasoning outputs and the purpose of encrypted_content
  https://platform.openai.com/docs/guides/reasoning#encrypted-reasoning-items
• Migrating to Responses: when to use stateful vs. stateless, and notes on encrypted_content
  https://platform.openai.com/docs/guides/migrate-to-responses#4-decide-when-to-use-statefulness
• Cookbook: reasoning items example (Responses API)
  https://developers.openai.com/cookbook/examples/responses_api/reasoning_items

​

Azure reference (corroboration of the same concept)

• Azure OpenAI Responses: stateless multi-turn and related fields
  https://learn.microsoft.com/zh-cn/azure/ai-foundry/openai/how-to/responses?view=foundry-classic&utm_source=chatgpt.com&tabs=python-key#encrypted-reasoning-items

​

Community discussions / Issues (troubleshooting references)

• Codex official repo discussion thread (consolidated discussion of related problems)
  https://github.com/openai/codex/issues/8120
• OpenCode community feedback: second-turn / resumed-chat problems (corroboration)
  https://github.com/anomalyco/opencode/issues/5104?utm_source=chatgpt.com
• new-api project: PR proposing removal of reasoning (reference for engineering-side handling)
  https://github.com/QuantumNous/new-api/pull/2662