Bug reasoning omni + opencode

[iHanatos]

I’m using ChatGPT 5-5 xHigh, but OmniRoute logs show “medium.” Why?

ChatGPT is connected to OpenCode through environment variables.

[diegosouzapw]

Hey @iHanatos! Two pieces here:

1. OpenCode-go sends reasoning_effort on the wire but the value depends on how the model is configured in your OpenCode models.json. If the model entry lacks an explicit reasoning_effort, OpenCode-go defaults to medium. The "xHigh" label in the OpenCode UI is a preset that maps to high only for certain provider templates -- when the model is routed via OmniRoute (custom OpenAI-compat base URL), OpenCode falls back to the default.

2. Fix: edit your OpenCode model config (or env) and pin reasoning_effort: "high" explicitly for the model entry you point at chatgpt-5-5. Then OmniRoute will pass it through verbatim (the translator maps it to ChatGPT's reasoningEffort field).

If you can paste the OpenCode model JSON for that entry, I can tell you exactly which key is missing.

[iHanatos]

Hey @iHanatos! Two pieces here:

1. OpenCode-go sends reasoning_effort on the wire but the value depends on how the model is configured in your OpenCode models.json. If the model entry lacks an explicit reasoning_effort, OpenCode-go defaults to medium. The "xHigh" label in the OpenCode UI is a preset that maps to high only for certain provider templates -- when the model is routed via OmniRoute (custom OpenAI-compat base URL), OpenCode falls back to the default.
2. Fix: edit your OpenCode model config (or env) and pin reasoning_effort: "high" explicitly for the model entry you point at chatgpt-5-5. Then OmniRoute will pass it through verbatim (the translator maps it to ChatGPT's reasoningEffort field).

If you can paste the OpenCode model JSON for that entry, I can tell you exactly which key is missing.

[iHanatos]

[iHanatos]

My OmniRoute connection to OpenCode is configured through Windows environment variables.

[iHanatos]

I’m not using a custom connection in OpenCode through JSON.

[diegosouzapw]

Got it @iHanatos -- the env-var setup changes the fix shape. Since you can't edit OpenCode's models.json, do this instead:

Use a model alias that encodes the effort: in your OpenCode environment, point the model name at gpt-5.5-xhigh (or whichever suffix you want -- high, xhigh). OmniRoute parses the suffix at open-sse/handlers/chatCore.ts:3235 and propagates it as the wire-level reasoning effort, no JSON edit needed.

Concretely: wherever OpenCode lets you type the model name (env var OPENCODE_MODEL or the model selector in the UI), use gpt-5.5-xhigh instead of gpt-5.5. OmniRoute strips the suffix before forwarding to Codex/upstream but keeps the effort.

If the OpenCode UI doesn't accept arbitrary model names (preset list only), the alternative is to create a model alias in OmniRoute: /dashboard/models -> New Alias -> name: gpt-5.5, target: codex/gpt-5.5-xhigh. Then OpenCode keeps requesting plain gpt-5.5 and OmniRoute rewrites it to xhigh on the way out.

Verify in /dashboard/logs -- the executed model should read gpt-5.5-xhigh. If it still shows medium, paste the log line + your omniroute --version and I'll trace it.

[diegosouzapw]

Hey @iHanatos -- circling back. Did the model-alias approach sort out the medium vs xhigh issue? Since you connect OpenCode via Windows env vars (no models.json to edit), the fix was to point the model name at the effort-encoded alias, e.g. set gpt-5.5-xhigh (instead of gpt-5.5) wherever OpenCode takes the model -- OmniRoute parses the -xhigh suffix and propagates it as the wire-level reasoning_effort.

If that worked, great -- I'll close this out. If you still see medium in the logs after switching to the suffixed name, drop the OmniRoute log line showing the model + effort and I'll trace it. Otherwise I'll close in ~7 days for inactivity.

[diegosouzapw]

Closing for inactivity -- the model-alias approach (omniroute/claude-sonnet-4-5-thinking or similar) should route the right effort token without needing to edit OpenCode's models.json. If the issue persists on the latest version with Windows env-var setup, please reopen with the exact model alias you are using and the error. Thanks!