dsh-provider-sync

中文 | English below

让 OpenRouter 路由的模型清单自己保持最新,并给支持推理的模型写好思考强度档位。

它解决什么

~/.dsh/settings.yaml 里 llm-pi-ai.providers..models 是一次性拉下来的快照:之后发布的新模型不会出现;而且条目没有 reasoningEfforts,dsh 的推理等级选择器就会回答 "does not support reasoning effort",尽管 OpenRouter 的 API 本身接受 reasoning.effort。

它做什么

• 启动后(等设置服务就绪,且距上次同步超过间隔时)以及每 intervalHours(默认 24)拉一次公开的 GET https://openrouter.ai/api/v1/models(无需密钥),通过 dsh 的设置服务合并进每一个 OpenRouter 路由(与「模型」页同一条写入路径,由 core 校验)。
• 合并规则:新 id 追加到清单末尾(你原有的顺序不动);已有 id 的 name / contextWindow / maxTokens 只在「仍是本插件上次写的值(或目录值)」时刷新——你手改过的值原样保留;OpenRouter 路由下 anthropic/ 开头的模型名后附「【claude系模型不可走openrouter原生api搜索,建议自己配置api】」,模型选择器里自然可见(联网搜索插件对 Claude 保留原模型走工具检索);supported_parameters 含 reasoning 且没有手写 reasoningEfforts 的模型写入 {off: null, minimal, low, medium, high, xhigh, max}(off 表示什么都不发,其余对应 reasoning.effort);:online 变体按其基础模型重推;手改过的字段与已从目录消失的 id 一律保留。
• 设置页卡片「模型清单同步」:上次同步、间隔、立即同步、每条路由的计数。

路由

GET /dsh-provider-sync/status · POST /dsh-provider-sync/sync · POST /dsh-provider-sync/settings {intervalHours?, anthropicNote?}(套件统一的本机 + 同源防线;类型不对回 400,不会悄悄重置)。配置文件 ~/.dsh/provider-sync.json(含本插件写过的字段记录 written、自动写入的档位 autoEfforts、失败计数与上次尝试时间——失败后按 5 → 15 → 60 分钟退避重试)。凭据中心页/区的「刷新模型清单」按钮调用的就是 /sync。

Claude 提示后缀:OpenRouter 路由下所有 anthropic/ 与手写的 ~anthropic/…-latest 别名条目,名字后面都会附「【claude系模型不可走openrouter原生api搜索,建议自己配置api】」(用户指定的原句,任何界面语言都一样)——这是唯一会碰手写名字的情况,且只追加。设置卡片和启动器凭据页上的复选框对应 anthropicNote,关掉后下一次同步把提示从名字里去掉。

依据

• openrouter.ai/docs/api-reference/list-available-models — 字段 id / name / context_length / top_provider.max_completion_tokens / supported_parameters
• openrouter.ai/docs/use-cases/reasoning-tokens — reasoning.effort: max / xhigh / high / medium / low / minimal / none
• core llm-pi-ai 目录规则 — reasoningEfforts 里 off 可为 null(不发送),其余档位写线上的值;OpenRouter 路由由 pi-ai 以 reasoning: { effort } 发送。

---

dsh-provider-sync (English)

Keeps every OpenRouter route's model list in DeepSeek Harness current, and declares reasoning-effort levels for the models that support them.

The list under llm-pi-ai.providers..models is a one-time snapshot: models released later never appear, and no entry carries reasoningEfforts, so the reasoning picker answers "does not support reasoning effort" although OpenRouter accepts reasoning.effort. This plugin pulls the public catalog on boot and every intervalHours (24 by default), merges it through the settings service (the same validated write the Models page makes), adds new ids, refreshes name / context window / output limit, writes {off: null, minimal, low, medium, high, xhigh, max} for reasoning-capable models that have no hand-written map, re-derives :online variants from their base, and never overwrites what you edited by hand. A Settings card shows the last sync, the interval, a sync-now button and per-route counts.

Claude note and settings (English)

• Every Anthropic entry on an OpenRouter route — catalog ids and the hand-written ~anthropic/…-latest aliases alike — gets the fixed suffix "【claude系模型不可走openrouter原生api搜索,建议自己配置api】" appended to its name (the owner's own wording, the same in every UI language). It marks that OpenRouter's native web search never answers a Claude request carrying the harness's tool schemas; dsh-web-search-plus keeps those models on the tool search. This is the one case where a hand-written name is touched, and only by appending. The checkbox on the Settings card (and on the launcher's Credentials page) is anthropicNote; turning it off strips the suffix at the next sync, which the toggle runs at once.
• POST /dsh-provider-sync/settings accepts { intervalHours?: whole number 1–720, anthropicNote?: boolean }; anything else is a 400, never a silent reset.
• A failed sync is retried after 5, then 15, then 60 minutes; status reports failures, lastAttemptAt, and this run's per-route record (models / added / updated / noted / reasoning / at).
• Hand edits stay: name / contextWindow / maxTokens are refreshed only while the field still holds what this plugin wrote last (or the catalog value); new ids are appended, the owner's order is kept.