Repository navigation
[Bug]: #126
Description
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
- 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.
Activity
lidge-jun commentedon Jul 14, 2026
coudlyou capture screenshot on localhost:10100 models? or are you using vps??
vebaev commentedon Jul 14, 2026
Can you advise what comands to run so i can capture you info?
vebaev commentedon Jul 14, 2026
It seems ok, but i see in CLI and desktop and mobile codex app only gpt models?!
lidge-jun commentedon Jul 14, 2026
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 commentedon Jul 14, 2026
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 commentedon Jul 14, 2026
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 commentedon Jul 14, 2026
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 commentedon Jul 14, 2026
yes restart is needed but I think in mobile you can just load new session and I think it would appear
lidge-jun commentedon Jul 14, 2026
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-k2.6NVIDIA'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 tryingmoonshotai/kimi-k2-instructor checking the model page on build.nvidia.com while logged in.2. The 400s — two proxy-side incompatibilities found and patched on
devparallel_tool_calls: truewith 400This model only supports single tool-calls at once!(report) — opencodex sent it by default whenever tools were present. Now pinned tofalsefor the nvidia provider.reasoning_effortisn't portable across NIM models (NVIDIA useschat_template_kwargsnatively); opencodex forwardedreasoning_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."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
devand will go out in an upcoming release — not published yet, so nothing to do on your side except update when it lands (and restartocxafterwards). 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.fix(nvidia): harden NIM provider for kimi family + surface openai-cha…
vebaev commentedon Jul 14, 2026
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 commentedon Jul 14, 2026
@vebaev I am implementing some free providers into here atm. Just give me some time.
vebaev commentedon Jul 14, 2026
@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 commentedon Jul 15, 2026
For now i implemented OpenCode Free and MiiMoo Code Free. Just has to be merged. #128 #129
lidge-jun commentedon Jul 15, 2026
Closing this out — everything in this thread has now shipped:
nvidia/moonshotai/kimi-k2.6404: NIM provider hardening for the kimi family went out in v2.7.18 (a818004). Note that NVIDIA's catalog can still list models that return 404 for specific accounts — that part is upstream account-level availability.Upgrade with
npm i -g @bitkyc08/opencodex, restartocx(and restart Codex CLI / open a new session on mobile so the new models appear). Both providers work keyless out of the box from the Providers page.If the kimi 404 or anything else here still reproduces on 2.7.19, feel free to reopen or file a fresh issue.
fix(nvidia): harden NIM provider for kimi family + surface openai-cha…