[BYOK] Reasoning effort never sent to OpenAI-compatible custom providers (spec id mismatch)

[soilderone]

Bug summary

When a custom (BYOK) provider is configured in ~/.commandcode/providers.json with reasoningEfforts on a model, selecting a reasoning effort (via /model, or --effort) never sends reasoning_effort in the request body. Upstream logs show no effort parameter. The same model + effort configured in opencode sends reasoning_effort correctly.

Environment

• command-code 1.51.3 (also reproduces in 1.50.1, 1.50.0, 1.49.1, 1.45.0, 1.38.2)
• macOS, OpenAI-compatible relays (e.g. https://api.qiuqiutoken.net/v1, https://sub2api.zehuamatrix.com/v1)

Reproduction

1. providers.json:

{
  "provider": {
    "effcap": {
      "baseURL": "http://127.0.0.1:18999/v1",
      "apiKey": false,
      "models": {
        "gpt-6-astra": {
          "contextWindow": 272000,
          "reasoningEfforts": ["low", "medium", "high", "xhigh", "max"]
        }
      }
    }
  }
}

2. command-code -p "hi" --model effcap/gpt-6-astra --effort high --local-only --max-turns 1 -t
3. Output prints Reasoning effort set to high for effcap/gpt-6-astra. but the captured request body does not contain reasoning_effort:

{"model": "gpt-6-astra", "max_tokens": 32000, "stream": true, ...}

Root cause (from dist/cli.mjs)

For BYOK providers the thinking hook is:

function thinkingHook(e){return({model,effort})=>{
  if(!effort||"off"===effort)return;
  const spec=e.specs.find(s=>s.id===model);
  return (spec?.reasoningEfforts??[]).includes(effort)
    ? "anthropic"===e.wire?{anthropic:{effort}}
      : "openai"===e.wire?{openai:{reasoningEffort}}
      : {[e.providerId]:{reasoning_effort:effort}}
    : void 0
}}

• toModelSpec builds specs with bare model ids (id: r.id, e.g. gpt-6-astra).
• The model passed to the hook is provider-qualified (effcap/gpt-6-astra / zehua/gpt-6-astra).
• specs.find(s => s.id === model) never matches → hook returns undefined → no providerOptions → reasoning_effort absent.

Workaround found

Adding model-level "options": {"reasoning_effort": "xhigh"} in providers.json works — it is merged into every request body. But it fixes the effort per model and ignores the session effort selection.

Expected

The declared reasoningEfforts + selected session effort should be transmitted (as reasoning_effort) for openai-compatible BYOK providers, e.g. by matching the provider-qualified model id (strip ${providerId}/ before the spec lookup).

[Coding-Dev-Tools]

Confirmed the missing field on 1.51.3, with one correction to the likely fix location: the current moduleFor() builds provider-qualified aliases and passes that expanded specs list into thinkingHook(). In the tested path, effcap/gpt-6-astra matches and the hook returns an option.

The loss occurs later: the OpenAI-compatible branch returns { [providerId]: { reasoning_effort: effort } }, while the pinned @ai-sdk/openai-compatible@2.0.74 expects reasoningEffort. Its request builder spreads extra options, then assigns reasoning_effort: compatibleOptions.reasoningEffort; that overwrites the snake-case value with undefined, which disappears in JSON.

Actual CLI → owned localhost captures, with identical dependencies:

Selection	Released CLI	Diagnostic: change only the hook key to reasoningEffort
--effort high	Field absent	reasoning_effort: "high"
--effort low	Field absent	reasoning_effort: "low"
Static option xhigh + selected high	"xhigh"	"high"

All six request-producing runs exited 0. Two additional --effort off controls were rejected before HTTP with “Unknown effort”; they do not establish in-band off behavior. These were independent launches on Windows x64 / Node 24.15.0, not interactive session-switch tests. No paid provider was called.

Suggested correction: use the adapter's camel-case provider option for the OpenAI-compatible hook, retaining the wire field as reasoning_effort. Regression coverage should include selected effort overriding a static default, unsupported efforts, qualified/namespaced IDs, provider switching and the separate native OpenAI/Anthropic paths. The diagnostic is evidence for this boundary, not a released fix or proof about older versions.

Bundle SHA256: d408a372d5d6c8a8851610e2a9840f806929595584d1cd8d909d501928c0037c. A sanitized runnable fixture, locked dependencies, source controls and exact observations are prepared.

Download the reproduction package (ZIP) — 58,688 bytes, SHA256 3da073024182e13bcd9c083346207fa41f8dffd64aecbe2588cef2665dbd07a5.

[ahmadbilaldev]

Taking a look.