v0.11.9 - Fix hybrid-thinking models on DashScope

Highlights

Thinking mode for DeepSeek V4 and Kimi K2.6 now works correctly when routed through Aliyun DashScope (百炼) — previously it silently degraded to non-thinking mode.

What changed

The two model families are served via two incompatible OpenAI-compatible APIs that disagree on the field name for thinking control:

Backend	Field
DeepSeek 官方 / Moonshot 官方	thinking: {"type": "enabled" | "disabled"}
Aliyun DashScope / 百炼	top-level enable_thinking: <bool>

v0.11.8 only sent the official-API form, which DashScope quietly ignored, so any DashScope-backed kimi-k2.6 / deepseek-v4-pro / deepseek-v4-flash agent ran without thinking even with enable_thinking = true configured.

v0.11.9 sends both fields for both model families. Whichever backend doesn't recognize a given field ignores it, so a single agent config now works across DashScope, Moonshot official, DeepSeek official, and LiteLLM-proxied routes.

Affected models

• kimi-k2.6 (and provider-prefixed slugs like moonshotai/kimi-k2.6)
• deepseek-v4-pro / deepseek-v4-flash (and dashscope/deepseek-v4-*)

reasoning_effort mapping (high / max) was already correct for both providers and is unchanged.

How to verify

After upgrading, send a reasoning-heavy prompt to your DashScope-backed Kimi K2.6 or DeepSeek V4 agent — a thinking block should now appear above the final answer in the TUI.

---

Full Changelog: v0.11.8...v0.11.9