Identity neutralization misses current Codex CLI wording

[byongshintv]

Client or integration

Codex CLI

Area

Provider adapter

Summary

neutralizeIdentity() only replaces one exact Codex identity sentence. Codex CLI 0.145.0 sends a different current sentence, so routed non-OpenAI models receive the GPT-5/OpenAI identity and may identify themselves as Codex instead of the selected model.

Expected: known Codex identity variants are replaced with NEUTRAL_IDENTITY_LINE before a request is forwarded to a routed provider.

Reproduction

1. Run OpenCodex 2.7.41 with Codex CLI 0.145.0 and an openai-chat routed non-OpenAI model.
2. Ask the selected model to identify itself.
3. Observe that the forwarded Codex instruction contains:

You are Codex, an agent based on GPT-5.

4. Inspect src/adapters/identity.ts. The exact string currently neutralized is:

You are Codex, a coding agent based on GPT-5.

5. Because the strings differ, String.replace() leaves the current identity instruction intact. In the observed run, the routed non-OpenAI model explicitly identified itself as Codex/GPT-5 rather than the selected model.

Version

2.7.41

Operating system

Ubuntu 22.04

Provider and model

Custom OpenAI-compatible provider using the openai-chat adapter / routed non-OpenAI model

Logs or error output

Expected neutralized input:
You are a coding agent. Do not claim to be GPT-5 or to be made by OpenAI.

Actual identity line left in the routed request:
You are Codex, an agent based on GPT-5.

Screenshots and supporting files

None. The minimal source-level reproduction above shows the exact-string mismatch.

Redacted configuration

{
  "providers": {
    "example": {
      "adapter": "openai-chat",
      "baseUrl": "https://redacted.example/v1",
      "liveModels": true
    }
  }
}

Suggested fix

Handle a small allowlist of known exact variants, or use a narrowly scoped expression such as:

/You are Codex, (?:a coding agent|an agent) based on GPT-5(?:\.[0-9]+)?\./g

A broad You are Codex.* expression should be avoided because it could rewrite unrelated content. Tests should cover both known variants and unrelated text remaining unchanged.

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.

[github-actions]

_{Automated translation bookkeeping — detected language: English.}