fix(provider): prefer per-model temperature over agent override

[Alezander9]

Problem

ProviderTransform.temperature(model) in src/provider/transform.ts returns the temperature value a given model family requires (kimi-k2.* -> 1.0 or 0.6; gemini / glm-4.6 / glm-4.7 / minimax-m2 -> 1.0; qwen -> 0.55). Several of these are hard requirements: Moonshot rejects any other value for the kimi-k2 family with HTTP 400 invalid temperature: only 1 is allowed for this model.

The call site in src/session/llm.ts consulted that function in the wrong order:

input.agent.temperature ?? ProviderTransform.temperature(input.model)

The built-in title agent (src/agent/agent.ts:271) hard-codes temperature: 0.5. Because ?? short-circuits on the first non-nullish operand, agents 0.5 always won and a temperature of 0.5 was sent to kimi-k2.6, which then 400d.

I observed this in production telemetry across multiple users on kimi-k2.6 (the most common case) and the same code path will trip whenever an agent sets a temperature on any of the strict-required model families.

Fix

Swap the operands at the single call site:

ProviderTransform.temperature(input.model) ?? input.agent.temperature

The transform function exists precisely because the maintainer encoded model-aware knowledge of what each family expects; that knowledge should beat a generic agent default. For models not covered by the transform map (claude, openai, etc.), the function returns undefined and the agents preference still applies through normal ?? fallthrough — so the only behavior change is on the families the maintainer has already flagged as needing specific values.

This also aligns temperature with the adjacent topK (no agent override) and brings it closer to topP (still agent ?? transform, but topP has no known strict-required cases today).

Diff

 packages/opencode/src/session/llm.ts | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Test plan

• bun typecheck clean.
• Existing test/provider/transform.test.ts suite: 220 pass / 0 fail.
• Manual verification against Moonshot kimi-k2.6: title-generation no longer 400s.

[cubic-dev-ai]

No issues found across 3 files