AI_APICallError with GPT-5-codex in Zero Data Retention organizations

[khoi]

Description

Encountering an error when using GPT-5-codex in organizations with Zero Data Retention enabled.

Error Message

AI_APICallError: Item with id 'rs_06bf86e5bcb244a50168e1327555a481968c0242e88626fe66' not found. Items are not persisted for Zero Data Retention organizations. Remove this item from your input and try again.

Environment

• Model: GPT-5-codex
• Organization setting: Zero Data Retention enabled

[rekram1-node]

@khoi are you setting any reasoning effort settings for codex in your opencode.json?

[capdevc]

@rekram1-node

I'm having the same with a ZDR account.

What works for me is adding an explicit "store" = false

  "provider": {
    "openai": {
      "models": {
        "gpt-5-codex": {
          "options": {
            "store": false,
            "reasoningEffort": "high",
            "textVerbosity": "medium",
            "reasoningSummary": "auto",
            "include": ["reasoning.encrypted_content"]
          }
        }
    },

It now works whether I add the "reasoning.encrypted_content" inclusion or not.

My only guess, something in the reasoning summary handling is conditioned on "store" being there even though ZDR accounts treat all requests as store = false?

[rekram1-node]

OO good to know, yeah I think if you say store=false openai doesn't try to read the messages via id w/ responses api maybe?

[rekram1-node]

I will document this somewhere

[rekram1-node]

@capdevc does it work if you specify the option at top level? Allowing u to not have to specify it for each reasoning model? ex:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "openai": {
      "options": {
        "store": false
      },
      "models": {
        "gpt-5-codex": {
          "options": {
            "reasoningEffort": "high",
            "textVerbosity": "medium",
            "reasoningSummary": "auto",
            "include": ["reasoning.encrypted_content"]
          }
        }
      }
    }
  }
}

[capdevc]

@rekram1-node

No, sadly setting it at the provider did not work.

  "provider": {
    "openai": {
      "options": {
        "store": false
      },
      "models": {
        "gpt-5-codex": {
          "options": {
            "reasoningEffort": "high",
            "textVerbosity": "medium",
            "reasoningSummary": "auto",
            "include": ["reasoning.encrypted_content"]
          }
        },
...

And the problem reappears

[rekram1-node]

ah got it

[capdevc]

I think the ZDR issue has to do with "previous_response_id" handling... in a ZDR account this isn't enough since the reasoning tokens aren't stored at OpenAI, requiring you to receive the "encrypted_content" encrypted reasoning tokens and pass them back in if you want to keep the reasoning history in the model's context.

My guess is that somewhere the decision is being made on whether to pass along that encrypted reasoning content in the next message only if store is false instead of always attaching it. Maybe something upstream in the vercel ai package? I'm afraid I'm not super familiar with js/ts and the ecosystem.

apologies if you already know this, but: https://platform.openai.com/docs/guides/reasoning#keeping-reasoning-items-in-context

[rekram1-node]

i think you are right on the money yeah that is what I figured didnt know about store false option but yeah that makes sense ofc

ill document it

[khoi]

Using @capdevc config works.

[rekram1-node]

Do you have to set store to false for non reasoning models too? Like gpt-4.1?

[capdevc]

@rekram1-node

I just tried it and gpt-4.1 works out of the box, I have nothing set in my config for it, so no need for store = false

[rekram1-node]

ty!