GPT-5.6 Luna returns "Model not found" when User-Agent identity differs from the Responses Lite originator

[mulfyx]

Summary

Claude Code Proxy 0.1.15 returns HTTP 502 with Model not found: gpt-5.6-luna when it sends a Luna request through the Codex Responses Lite endpoint.

The same ChatGPT OAuth account and model work in Codex CLI 0.144.1 over both WebSocket and HTTP. A controlled request matrix points to the product identity in originator and User-Agent. CCP sends originator: codex_cli_rs for Responses Lite, while its default User-Agent is claude-code-proxy/0.1.15. The backend rejects that combination with HTTP 404.

Environment

• Claude Code Proxy 0.1.15
• Codex CLI 0.144.1
• Fedora 44, x86_64
• ChatGPT OAuth authentication
• Codex transport: HTTP
• Claude Code model value: gpt-5.6-luna[1m]
• Upstream model value: gpt-5.6-luna

The account's model cache reports gpt-5.6-luna with supported_in_api: true and use_responses_lite: true.

Reproduction

1. Authenticate CCP with ChatGPT OAuth.
2. Start CCP without overriding CCP_CODEX_USER_AGENT.
3. Send a request through its Anthropic-compatible endpoint:

curl --fail-with-body --no-buffer \
  'http://127.0.0.1:18765/v1/messages?beta=true' \
  -H 'content-type: application/json' \
  -H 'anthropic-version: 2023-06-01' \
  -H 'x-api-key: unused' \
  --data '{
    "model": "gpt-5.6-luna[1m]",
    "max_tokens": 64,
    "stream": false,
    "messages": [{"role": "user", "content": "Reply with luna-ok"}]
  }'

CCP returns HTTP 502. Its upstream error is Model not found: gpt-5.6-luna.

Control results

Using the same OAuth account:

• Codex CLI 0.144.1 over WebSocket succeeds.
• Codex CLI 0.144.1 forced to HTTP succeeds.
• A direct upstream request with CCP's default client identity returns HTTP 404.

I reproduced the issue with fresh CCP containers. Every case used the same image, OAuth files, HTTP transport, request body, and model. Only CCP_CODEX_USER_AGENT changed. CCP keeps the Responses Lite originator fixed at codex_cli_rs in all these cases.

originator	User-Agent	CCP result
codex_cli_rs	claude-code-proxy/0.1.15	HTTP 502, upstream Model not found
codex_cli_rs	codex_exec/0.144.1 ...	HTTP 502, upstream Model not found
codex_cli_rs	codex_cli_rs/0.144.1 ...	HTTP 200
codex_cli_rs	codex_cli_rs/0.144.1	HTTP 200
codex_cli_rs	codex_cli_rs/0.1.15	HTTP 200
codex_cli_rs	codex_cli_rs	HTTP 200

The full Codex CLI User-Agent and its version are not required in this environment. The bare codex_cli_rs value works. This narrows the trigger to the product identity rather than a specific Codex CLI version or platform suffix.

This test does not establish an official backend contract. It records the current routing behavior observed with this OAuth account and environment.

Relevant CCP behavior

src/providers/codex/auth/constants.rs defines the Responses Lite originator as codex_cli_rs:

pub const RESPONSES_LITE_ORIGINATOR: &str = "codex_cli_rs";

src/providers/codex/client.rs uses that value for Responses Lite requests, but builds the default User-Agent separately:

let originator = if use_responses_lite {
    RESPONSES_LITE_ORIGINATOR.to_string()
} else {
    config::codex_originator(ORIGINATOR)
};

let user_agent =
    config::codex_user_agent(&format!("claude-code-proxy/{}", env!("CARGO_PKG_VERSION")));

This produces the failing pair:

originator: codex_cli_rs
User-Agent: claude-code-proxy/0.1.15

Workaround

Setting a User-Agent whose product identity matches the Responses Lite originator fixes the request in this environment:

CCP_CODEX_USER_AGENT: "codex_cli_rs"

After restarting CCP, the same Luna request returns HTTP 200 and luna-ok.

Expected behavior

CCP should use a matching User-Agent product identity when it switches the Responses Lite originator to codex_cli_rs, or document the required CCP_CODEX_USER_AGENT override. A regression test should cover the originator and User-Agent pair used for Responses Lite requests.

Related reports

• 5.6 series support #23 reported the same Luna Model not found symptom and was later considered fixed.
• GPT-5.6 Luna resolves to a missing internal engine for non-Codex originators using ChatGPT OAuth openai/codex#31967 reports the same client identity and version-dependent routing behavior for third-party OAuth clients.

[raine]

Thanks, fixed in latest