Title-agent temperature 0.5 overrides model-required temperature, causing 400s on Moonshot kimi-k2.*

[Alezander9]

Description

In packages/opencode/src/session/llm.ts:172 the temperature is resolved as:

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

The built-in title agent hard-codes temperature: 0.5 (packages/opencode/src/agent/agent.ts:253), so ?? short-circuits and the agent default is used.

For model families that the provider strictly requires a specific temperature for (kimi-k2., glm-4.6/4.7, minimax-m2, gemini, qwen), ProviderTransform.temperature exists to encode that. Moonshot kimi-k2. returns HTTP 400 invalid temperature: only 1 is allowed for this model when 0.5 is sent.

Steps to reproduce

1. Configure Moonshot kimi-k2.6 as the provider/model.
2. Trigger a session that runs the title agent (any new session).
3. Moonshot rejects the request: 400 invalid temperature: only 1 is allowed for this model.

Suggested fix

Swap the operands so the model-aware transform wins, with agent preference as fallback:

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

For models without a transform entry (claude, openai, etc.) the function returns undefined and the agent preference still applies via ??.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Using Kimi K2.5 from Fireworks throws temperate related error #11134: Kimi K2.5 temperature-related error (closed) — same symptom of temperature being rejected, though from a different provider (Fireworks) and possibly a different root cause
• Venice not setting provider transform for Minimax or Kimi, or small options. #13552: Venice not setting provider transforms for Minimax/Kimi (closed) — related to ProviderTransform not being applied for Kimi models

If those issues didn't fully address the title-agent ?? operator precedence bug you've described, this report may still be a distinct issue.