Codex functions.exec: Are shell_command and apply_patch result contracts officially documented?

ChatGPT Use cases and examples
1

Title:
Request for documented Codex functions.exec nested-tool result contract

Body:
I’m looking for official, normative documentation for the Codex app
functions.exec nested-tool return contract.

Could OpenAI clarify:

  1. Is a fulfilled shell_command result inside functions.exec identical to
    Responses API LocalShellCallOutput or ShellCallOutput?

  2. What are the exact schema and version for stdout, stderr, exit code,
    timeout, failure, and promise rejection?

  3. How can callers determine whether output was truncated?
    Are original length, complete length, or full-output digest available?

  4. Does status=“completed” normatively guarantee that the value delivered
    to functions.exec is complete and untruncated?

  5. What is the exact apply_patch result schema, including mutation outcome,
    affected paths, failure state, and completeness?

  6. Are nested-tool results guaranteed to be non-Proxy ordinary objects,
    structured-clone values, or host objects?

  7. Is this contract versioned and covered by compatibility guarantees?

Please link to official documentation where possible.

2

ChatGPT is a product not for you to program or think about. It is for you to consume.
The documentation is all about using it, not asking programming questions about it, as you don’t develop your own software to call ChatGPT, and the Codex SDK means you have given up writing software and instead committed to comprehending obtuse proprietary software. The Codex SDK uses an app server layer, abstracting the remote communication, but you do have the source if you really want to see.

Here’s a link to official documentation about as technically deep as you will get, how to set permissions on a sandbox for the equivalent of shell tool: Codex App Server | ChatGPT Learn

Here’s the SDK source, if you want to investigate ultimate calls to pay-per-use API (which does have moderate documentation), vs ChatGPT’s API and subscription account use:

3

Thanks—the App Server documentation is a useful lead. My question is narrower: does the Codex desktop functions.exec nested shell_command result map exactly to App Server command/exec or commandExecution?

If so, could you point me to the exact repository URL, commit/tag, source file, and generated type that define that mapping, including output truncation and completeness behavior?

Likewise, which App Server type maps to the nested apply_patch result, and what postcondition does a completed result guarantee?

I’m trying to distinguish an implementation-specific mapping from the public App Server protocol itself.

4

Hey yuta_yuuta, thanks for clarifying your question!

These two official OpenAI resources are more closely aligned with what you're looking for:

  1. Codex App Server Documentation — Covers command/exec, commandExecution, and fileChange result structures, including execution status and output fields.
  2. Codex Hooks — Tool Calls from Code Mode — Explains nested JavaScript tool-call behavior, including Promise resolution, rejection, and execution outcomes.

However, as _j highlighted, the exact implementation details remain unavailable in the public documentation, so these resources may not fully answer your questions about functions.exec result mappings and completeness guarantees.

You can also try reaching out to support@openai.com to see if they can provide further clarification on your request.

Hope this helps!

-Saif

5

I’m trying to identify the supported result contract for programmatic tool calls inside Codex—not the ChatGPT API or the Responses API.

The official App Server documentation shows that command/exec returns:

{
  "exitCode": 0,
  "stdout": "...",
  "stderr": ""
}

Reference:

In the Codex environment I’m using, functions.exec can invoke nested tools such as shell_command and apply_patch, but its declared return type is only Promise<unknown>.

Could someone clarify:

  1. Does the fulfilled value from nested shell_command map exactly to App Server command/exec, a commandExecution item, or another internal wrapper?
  2. If this mapping exists, what source file, generated type, version, or commit defines it?
  3. How can a client determine whether stdout and stderr are complete or truncated? Does a completed result guarantee that all output was delivered?
  4. What is the exact result schema and postcondition contract for nested apply_patch?
  5. Are returned JavaScript values guaranteed to be ordinary non-Proxy objects, structured-clone values, or another documented representation?
  6. Is this result contract versioned or covered by a compatibility guarantee?

If these are intentionally internal and unsupported contracts, confirmation of that would also answer the question.