DeepSeek Harness 플러그인 — 813 페이지

Reorder models in the DSH composer picker, store three per-million-token prices, and estimate per-turn and session cost with built-in DeepSeek peak/off-peak rates.

DeepSeek Harness 模块化模型选择器：根据模型来源（Provider 如 CodeBuddy、Qoder、DeepSeek 等）分块展示，支持分块下钻、全局搜索与推理强度调节。

Expose every model the DeepSeek Harness can already call as a loopback OpenAI/Anthropic-compatible gateway, so other local dev tools can use them. 把 DSH 能调用的所有模型开放成本机 OpenAI / Anthropic 兼容网关，供其他开发软件使用。

DSH web plugin: per-model thinking-intensity tiers, provider default reasoning, modalities and gateway compat switches for custom llm-pi-ai providers, edited inside the official Models settings card. 模型能力：思考强度档位 / 默认思考强度 / 支持模态 / 兼容开关。

DSH Settings section to edit per-model input modalities (text/image) AND reasoningEfforts (incl. custom wire values), for all configured providers, models and modelOverrides at once — a companion to the lighter input-only WJZ-P plugin. Works on DSH Web and DSH Desktop.

Model selection as a three-level cascading menu (Provider -> Series -> Full model) for the composer model seat.

给 DeepSeek Harness 的模型列表加上两级级联选择（先选线路、再按 DeepSeek/GPT/其他挑模型）、搜索框，以及一条可拖动的推理强度能量条。Two-level model cascade (provider → family-grouped model), search, and a draggable reasoning-effort bar for DeepSeek Harness.

Model Check：DeepSeek Harness 桌面 / Web 通用的模型检测插件，支持 ModelTrace 与糖果题联合检测、响应路由及行为指纹。

DSH web plugin: collapse the model picker by provider, with a pinned quick bar for expand-all / collapse-all / focus-current / filter / reset.

DSH compatibility guard: repairs compaction limits, fail-safe sandbox escalation, and missing tool descriptions for third-party models; capability lookups are cached and the deprecated prompt-injection path is off by default.

DeepSeek Harness Web GUI 插件：在设置面板为 llm-pi-ai 提供方按模型配置思考开关/思考等级（reasoningEfforts），为官方 DeepSeek 配置思考开关（thinking）与思考等级（reasoningEffort），并为各提供方模型配置输入类型（llm-pi-ai 写 input、官方 DeepSeek 写 inputModalities）。

Unified model control for DeepSeek Harness: model picker (search/favorites/fold/chips/facts) + model management (channels, per-model capabilities) + reasoning-effort advisory (knowledge base, per-model defaults), in one seat and one settings page.

Semantic model routing plugin for DeepSeek Harness

Local or remote AI model deployment benchmarking for real hardware targets

DeepSeek Harness plugin: LLM model selection & deployment analysis (deployability, VRAM, TTFT, latency, throughput, power for 38 models × 20 GPUs/NPUs)

Publishes the model this DeepSeek Harness profile is using as an OpenAI-compatible HTTP endpoint, bound to a configurable LAN or public address.

Modeling abstraction and automated review for designs and code, including a refactoring harness: UML diagrams, executable verification of machines and protocols with before/after regression, data-model and migration checks, concurrency claim routing, export to external checkers, dependency-graph review, and report diffing. Native plugin for DeepSeek Harness (dsh).

在 dsh 模型目录的每一行直接设置该模型的思考强度范围与支持的输入模态 (Per-model reasoning-effort range and accepted input modalities, right in the dsh Models catalog)

DeepSeek Harness plugin: extended per-model LLM configuration (reasoning efforts, input modalities, compat flags) as a companion settings page

Model failover for DeepSeek Harness with a Web UI seat beside the composer model picker: pick one or more providers, pool = all models inside them; on request failure (RATE_LIMIT/QUOTA/TIMEOUT/...) auto-retry with the next untried model. Runtime tool model_failover (status/set/clear/enable/disable) plus an in-page panel.

Model fallback chains for DeepSeek Harness: an ordered per-model failover list, editable in Settings with drag-to-reorder, applied at the kernel's agent/request seam.

模型故障转移队列：P1 限额/限流/不可用时自动切到队列下一个模型继续，全部图形界面配置 | Model fallback queue for DeepSeek Harness: auto-switch to the next model when the current one hits quota or rate limits

Model filter plugin: adds search box to model selection with fuzzy search and highlighting

DSH 模型流式缺陷修复：opencode 聚合端点上 muse-spark-1.2 不发送 finish_reason/[DONE]，导致 pi-ai 每次对话结尾报 "Stream ended without finish_reason"（TRANSPORT）并触发重试。本插件在 llm/stream waterfall 上把「已输出内容但结尾被截断」的 finish 改写为正常 stop，直连即可使用，无需代理。