[DSH adapter] Keep slow or parameter-limited host LLM routes from causing false failures #2484

Pre-submission checklist | 提交前检查

• I have searched existing issues and this hasn't been mentioned before | 我已搜索现有问题，确认此问题尚未被提及
• I have read the project documentation and confirmed this issue doesn't already exist | 我已阅读项目文档并确认此问题尚未存在
• This issue is specific to MemOS and not a general software issue | 该问题是针对 MemOS 的，而不是一般软件问题

Bug Description | 问题描述

Problem

In a DSH profile using @memtensor/memos-local-plugin 2.0.20 with a host-routed reasoning model, optional auxiliary LLM work can exceed the adapter's short foreground recall budget. Some routes also reject the optional temperature parameter. In this situation automatic recall or memos_search can fall back unnecessarily, and an optional timeout may make the shared host-LLM health indicator look disconnected even though normal DSH chat works.

Expected behavior

Optional features should degrade independently: retry once without temperature when the provider explicitly rejects that parameter; keep foreground recall within its SLA by allowing the optional LLM reranker to be skipped in favor of deterministic ranking; and ensure summary/classification timeouts do not mark a healthy host route as disconnected. Summary timeout should be passed through to the host completion call or be configurable for reasoning models.

Local workaround

My profile currently carries a pnpm patch that (1) skips the optional LLM reranker for DSH automatic recall and the explicit search tool, (2) retries unsupported temperature without that parameter and caches the route capability until invalidation, (3) increases the summarizer timeout and forwards timeoutMs, and (4) uses deterministic session classifiers for host-routed models. These are workarounds, not claims that this exact policy is right for every adapter.

Request

Could these behaviors be made first-class/configurable in the DSH adapter? I can provide the local diff or split this into focused PRs if the direction is useful. No credentials or secret values are included here.

How to Reproduce | 如何重现

1. Run DSH Desktop with @memtensor/memos-local-plugin 2.0.20 and configure MemOS to delegate auxiliary LLM calls to the active DSH host route.
2. Use a reasoning model/route that takes longer than the 3-second foreground recall bound or rejects the optional temperature parameter.
3. Trigger automatic recall or call memos_search, then observe the fallback/timeout and shared LLM health signal.
4. Separately trigger background summarization and observe the default timeout behavior.

Environment | 环境信息

• DSH Desktop: 0.2.0-rc.2 (Windows x64)
• Plugin: @memtensor/memos-local-plugin 2.0.20
• LLM: DSH host-routed reasoning model; provider/model deliberately omitted because it is profile-specific

Additional Context | 其他信息

I checked the existing DSH adapter documentation and searched open issues for skipLlmFilter, unsupportedTemperatureRoutes, and credentialRef; I did not find a duplicate focused issue. The current upstream adapter describes a 3-second recall bound and deterministic safeCutoff fallback. My patch keeps that bounded behavior while avoiding the optional reranker on foreground calls.

Willingness to Implement | 实现意愿

• I'm willing to implement this myself | 我愿意自己解决
• I would like someone else to implement this | 我希望其他人来解决