[Bug] Compaction summarization always fails with thinking models — default maxTokens=8192 is consumed by reasoning tokens

[knight-ni]

Environment

• dsh 0.1.0-rc.7 (reproduced on a live deployment; the relevant default is unchanged in 0.1.6-alpha.1)
• Model routed through an OpenAI-compatible endpoint, reasoning enabled (GLM-5.3; also applies to DeepSeek V4 with reasoning)

Symptom
Running /compact on a long session always fails with:

Compaction could not produce a useful summary. The conversation is unchanged; the attempt is recorded in the session log.

With auto-compaction this fails silently in a loop, so the session keeps growing until it hits the context window and outputs get starved.

Root cause
dsh-compaction-basic resolves the summarization budget as maxTokens ?? 8192. For reasoning models, reasoning_content counts against completion tokens: thinking alone exhausts the 8192 budget, leaving an empty or truncated summary, which the quality gate then rejects. Measured on our side: 331 reasoning tokens for a trivial one-line request; multi-minute analyses produce thousands more.

Relevant code (dsh-compaction-basic/lib/index.js, resolveConfig):

maxTokens: config.maxTokens ?? 8192,

Suggestions

1. Derive the default from the routed model's output capability (e.g. a fraction of the model's max output), or
2. Expose it via user config (settings.yaml / web UI).

Also worth documenting: the web layer disables compaction-basic (auto-compaction), while manual /compact still goes through the same default — so on web deployments there is currently no working path to compact long sessions with thinking models.

---

中文要点：思考型模型（reasoning 计入输出 token）执行压缩摘要时，默认 8192 的摘要输出预算被思考过程耗尽，摘要为空/截断后被质量门槛拒绝，导致 /compact 始终失败（自动压缩同样静默失败循环）。建议默认值随路由模型的输出能力自适应，或开放用户配置。

[wangzhanchao883]

同一个根因，这次在插件那边。第三方插件调模型出结构化内容（翻译、抽取、出题这类）的时候，
如果没显式关推理，推理那段也会把 maxTokens 吃掉。表现大概率是：请求正常返回、不抛异常，
你就是拿不到内容——静默失败，很容易查成是提示词的问题。

插件作者不用等官方改，代码里显式传 reasoningEffort: "off" 就行。最小复现和改法写在 #6857。

Same root cause on the plugin side — third-party plugins hit it too, and authors can work
around it today with reasoningEffort: "off" on structured-output calls. Repro + fix: #6857