bug: Claude thinking path forces temperature and extended-thinking, breaking Opus 4.7+ / Opus 5 / Sonnet 5 / Fable 5

[JNK234]

Problem

ClaudeProvider.createProviderRequest unconditionally sets temperature = 1.0 and thinking: {type: "enabled"} whenever thinking is on (ClaudeProvider.scala:92, :98):

if (isThinking) {
  ...
  // Anthropic requires temperature=1.0 when thinking is enabled
  baseRequest("temperature") = 1.0
  ...
  baseRequest("thinking") = ujson.Obj(
    "type" -> "enabled",
    "budget_tokens" -> budget
  )
}

Both assumptions are now wrong for the current Claude generation (verified against Anthropic's model overview, Jul 2026):

1. temperature is deprecated on Claude Opus 4.7 and later and returns HTTP 400 on any non-default value. Sending temperature = 1.0 to Opus 5, Sonnet 5, or Fable 5 makes the request fail outright.
2. Extended thinking (thinking.type: "enabled") is not supported on Fable 5, Opus 5, or Sonnet 5, and is deprecated on Opus 4.6 / Sonnet 4.6. Those models use adaptive thinking instead, configured with an effort parameter (Fable 5 is always-on; Opus 5 / Sonnet 5 / Opus 4.6-4.8 accept effort, default high).

So the thinking path works only for the older family (claude-haiku-4-5, claude-sonnet-4-5, claude-opus-4-5, claude-opus-4-1) and hard-fails on everything newer.

Note Haiku 4.5 is the inverse of the newer models: extended thinking supported, adaptive not.

What needs to be done

1. Branch on model generation in ClaudeProvider: extended-thinking models keep the current thinking block and temperature = 1.0; adaptive-thinking models get effort and no temperature key at all.
2. Suppress temperature entirely for Opus 4.7+ even on the non-thinking path — the deprecation is not limited to thinking requests.
3. Map our existing reasoning_effort config values (none|low|medium|high|xhigh) onto Anthropic's effort parameter.
4. Decide where the generation split lives. ReasoningModelDetector already string-matches Claude versions, but adding another hardcoded family check there compounds the inconsistency noted in cleanup: remove or wire currently unused provider and registry helpers #31 — a ProviderDescriptor field or a models.yaml capability flag would be cleaner.
5. Tests covering: extended-thinking model, adaptive-thinking model, and a non-thinking request to an Opus 4.7+ model (asserting no temperature key is sent).

Related

• bug: three provider default models are missing from the bundled model registry #62 — our Anthropic default claude-3-5-haiku-latest is a retired model (underlying claude-3-5-haiku-20241022 retired 19 Feb 2026), so the default config cannot reach Anthropic at all.
• claude-opus-4-1-20250805 retires 5 Aug 2026.
• cleanup: remove or wire currently unused provider and registry helpers #31 — hardcoded model-family string matching in ReasoningModelDetector.