[Bug]:

[vebaev]

Summary

OpenCodex proxy returns 404 for /v1/responses, and I canot see also nvidia models i put API for. I'm using latest codex-cli 0.144.4
``
`&ocx ensure
⚠️ Previous session (PID 139188) did not shut down cleanly. Codex state restored from journal.

• 120 models appended to Codex catalog (/home/vebaev/.codex/opencodex-catalog.json)
  Pointed Codex's built-in openai provider at the opencodex proxy (openai_base_url).
  Codex model catalog: /home/vebaev/.codex/opencodex-catalog.json
  Codex resume history: untouched (threads keep their native openai tag).
  All models now route through opencodex proxy (like OpenRouter).
  OpenAI models (gpt-5.5, etc.) are passed through to OpenAI.
  Custom models route to their configured providers.
  Fallback reference: /home/xxxx/.codex/opencodex.config.toml
  ✅ Proxy running on port 10100
  $ ocx sync
• 120 models appended to Codex catalog (/home/vebaev/.codex/opencodex-catalog.json)
  Pointed Codex's built-in openai provider at the opencodex proxy (openai_base_url).
  Codex model catalog: /home/vebaev/.codex/opencodex-catalog.json
  Codex resume history: untouched (threads keep their native openai tag).
  All models now route through opencodex proxy (like OpenRouter).
  OpenAI models (gpt-5.5, etc.) are passed through to OpenAI.
  Custom models route to their configured providers.
  Fallback reference: /home/xxxx/.codex/opencodex.config.toml`

Reproduction

Installed Opencodex
Start ocx init
Input nvidia api

Logs and screenshots

Area

CLI

Version

No response

OS

Linux

Config shape

Checks

• I searched existing issues and docs first.
• I removed secrets, tokens, and personal data from logs and screenshots.

[lidge-jun]

coudlyou capture screenshot on localhost:10100 models? or are you using vps??

[vebaev]

Can you advise what comands to run so i can capture you info?

[vebaev]

It seems ok, but i see in CLI and desktop and mobile codex app only gpt models?!

[lidge-jun]

and it didnt appears in your codex cli? I think there could be some cap of it? could you turn off and select a few?

[vebaev]

Inrestarted codex cli and now apeared I tried 3 free models from nvidia but gives error:

Or

Or just no response what so ever… just thinking

[lidge-jun]

Ive tried nvida models before alot I think it was so slow generating like 10s/tokens it seems time out error from it but Ill test tomorrow and Ill reply you

[vebaev]

Yes i tried it with opencode and same errors now seems overwhelmed. :) thanks i will try to test it!
So if i add new providers i have to restart cli codex in order models to appear…

[lidge-jun]

yes restart is needed but I think in mobile you can just load new session and I think it would appear

[lidge-jun]

Dug into this further with the exact error shapes from your screenshots — there are three separate things stacking here:

1. The 404 on nvidia/moonshotai/kimi-k2.6
NVIDIA's GET /v1/models (what opencodex uses to build the catalog — your "+120 models appended" matches NVIDIA's live list) is catalog metadata, not a "callable on your account" list. There are multiple reports of listed models returning 404 "function not found" for specific accounts (example, kimi-k2.5 case). I verified the proxy routing locally — the model id reaches NVIDIA unmangled — so this 404 is most likely NVIDIA-side for your account/model combination (I couldn't reproduce it with a keyed account). Worth trying moonshotai/kimi-k2-instruct or checking the model page on build.nvidia.com while logged in.

2. The 400s — two proxy-side incompatibilities found and patched on dev

• NIM's kimi models reject parallel_tool_calls: true with 400 This model only supports single tool-calls at once! (report) — opencodex sent it by default whenever tools were present. Now pinned to false for the nvidia provider.
• reasoning_effort isn't portable across NIM models (NVIDIA uses chat_template_kwargs natively); opencodex forwarded reasoning_effort: "medium" for kimi. Now suppressed for the NIM kimi family ids (kimi-k2.6, k2.5, k2-thinking, k2-instruct, k2-instruct-0905); gpt-oss on NIM keeps it since it's supported there.
• Also: bare errors like "Provider error 400" now surface the provider's actual JSON error text (when the upstream returns one), so the next report will show the real cause directly.

3. The "just thinking" hangs
Consistent with NIM free-tier behavior: limits are dynamic and often stricter than the nominal 40 RPM, and big models (kimi especially) are reported to queue for minutes (NVIDIA staff response). The proxy can't speed that up — trying a smaller model is the practical workaround.

The fixes are on dev and will go out in an upcoming release — not published yet, so nothing to do on your side except update when it lands (and restart ocx afterwards). I'll note it here when released. The NVIDIA-side account availability and capacity conditions described in #1 and #3 may still apply after the update.

[vebaev]

Thanks, and which provider and free models you reccomend with no such issues? I probably ill subscribe to opencode go, but want to have some free alternatives?

[Wibias]

@vebaev I am implementing some free providers into here atm. Just give me some time.

[vebaev]

@Wibias that will be great! It will be awsome if there is a way to refresh and include free models from openrouter and others somehow as the free models will vary in time and it is a thing when tou have to search them and turned on :) but dont know if it is possible :)

[Wibias]

For now i implemented OpenCode Free and MiiMoo Code Free. Just has to be merged. #128 #129