[adivekar-utexas]

The bug

Command Code's validator advertises low|medium|high|xhigh|max — xhigh is a native and distinct tier. But open-sse/executors/base/reasoningEffort.ts had Command Code in its isMaxTierTarget list, which rewrote xhigh → max unconditionally:

if (isMaxTierTarget && effortStr === "xhigh") {
  return writeEffortValue(b, "max", c);
}

writeEffortValue writes to every carrier, so both top-level reasoning_effort and the nested reasoning.effort were silently changed. Observed on the wire: a payload setting "reasoning": {"effort": "xhigh"} went upstream as "reasoning": {"effort": "max"}.

Why it is wrong

1. It spends a different reasoning level than the caller asked for.
2. The premise is false for Command Code. The same file quotes Command Code's own 400 message — expected one of "low"|"medium"|"high"|"xhigh"|"max" — which is the proof that xhigh is accepted.
3. It preempted the correct mechanism that already exists further down: if (supportsXHigh) return body; // model accepts xhigh natively. supportsXHighEffort() defaults to true for unlisted models, so the fix is to stop short-circuiting before it.

The change

Command Code is removed from isMaxTierTarget. Nothing else moves.

Case	Before	After
command-code + xhigh, unlisted model	max	xhigh
command-code + max	max	max
GLM 5.1+ / DeepSeek V4+ / Kimi K3+ + xhigh, any provider	max	max
ollama-cloud or opencode-go + xhigh	max	max
openrouter + xhigh	xhigh	xhigh

The model-level ceilings are deliberately kept, because those families genuinely top out at max in their own native API.

Tests

Two existing tests encoded the bug and are inverted to pin the fix:

• command-code maps normalized xhigh back to max becomes command-code preserves native xhigh on every carrier, asserting both reasoning_effort and nested reasoning.effort survive.
• The Qwen case asserted xhigh → max on cmd. It now asserts max on opencode-go (unchanged behaviour) and xhigh on command-code.

Three added cases bound the change so it cannot be over-applied: literal max still passes through, max-tier model families still collapse on command-code, and ollama-cloud is unchanged.

Sweep: every unit test mentioning xhigh was run — 123 suites, 1938 assertions passing. chatcore-translation-paths has 3 failures (tests 2/64/72, none about effort) that are pre-existing — confirmed identical with this patch stashed. ESLint clean, open-sse-typecheck 0 errors, check:file-size OK, complexity ratchets OK (2931/3218 and 1328/1437).

open-sse/executors/base/reasoningEffort.ts is not in the Stryker mutate set, so no tap.testFiles registration is needed.