pi-models-discovery 1.6.0 → 1.6.1

package/README.md

@@ -77,21 +77,28 @@ Forces a rediscovery of every discovery provider and updates the local cache, no
 - **Offline / fetch failure**: handwritten `models` in models.json (if any) are kept as a fallback, and an explicit warning is surfaced via in-session notify — never a silent degradation. One provider failing does not affect the others.
 - **apiKey resolution** (discovery request only): supports literals and `$ENV_VAR` / `${ENV_VAR}` interpolation; `!command` values skip discovery with an explicit warning (chat requests are still resolved by pi itself and are unaffected).
 - Default parameters for discovered models: `reasoning: true`, `input: ["text", "image"]`, zero cost, `contextWindow` 1M, `maxTokens` 64K, `compat.supportsDeveloperRole: false`. Provider-level `compat` is merged into every discovered model.
-- Discovered models declare `thinkingLevelMap: { xhigh: "xhigh", max: "max" }`, so `xhigh` and `max` appear in `/thinking`; the standard levels stay omitted and keep pi's provider default mapping. `/models` cannot report per-model thinking capability, so this is a blanket default for every discovered model: when a provider rejects those levels, override it per model through `modelOverrides` in models.json (set `xhigh` / `max` to `null` to hide them), or declare the model yourself in a handwritten `models` entry.
+- Discovered models default `thinkingLevelMap.xhigh` / `max` to `null`; extra tiers are no longer assumed to work on every backend. Standard tiers keep Pi's default mapping. Opt in with provider-level `compat.thinkingLevelMap` when the service supports extra tiers, or set any unsupported tier to `null`.
+- For non-reasoning models, set provider-level `compat.reasoning: false` so `/thinking` offers only `off`. `reasoning` and `thinkingLevelMap` are extracted to the model's top level; other `compat` fields (such as `supportsReasoningEffort` and `thinkingFormat`) are forwarded normally. Mixed-model gateways can use `modelOverrides` by model.id. Existing caches apply the current thinking configuration at startup without a network refresh.
+
+Example for a local non-reasoning service. To enable verified extra tiers instead, use `"thinkingLevelMap": { "xhigh": "xhigh", "max": "max" }` and remove `"reasoning": false`.
 
 ```json
 {
   "providers": {
-    "llm-proxy": {
+    "local": {
+      "baseUrl": "http://127.0.0.1:8000/v1",
+      "api": "openai-completions",
       "discoverModels": true,
-      "modelOverrides": {
-        "some-model": { "thinkingLevelMap": { "xhigh": null, "max": null } }
+      "compat": {
+        "reasoning": false,
+        "supportsReasoningEffort": false,
+        "thinkingLevelMap": { "xhigh": null, "max": null }
       }
     }
   }
 }
 ```
-- Model metadata may carry `name` / `context_window` (or `contextWindow`) / `max_tokens` (or `maxTokens`); defaults are used when absent.
+- Model metadata may carry `name` / `context_window` (or `contextWindow`, or vLLM's `max_model_len`) / `max_tokens` (or `maxTokens`); defaults are used when absent.
 
 The old `/model-discovery`, `/model-discovery-refresh`, `/pi-model-discovery`, and `/pi-model-discovery-refresh` names remain available as compatibility aliases.
 

package/README.zh-CN.md

@@ -77,21 +77,28 @@ pi install npm:pi-models-discovery
 - **离线 / 拉取失败**：保留 models.json 里手写的 `models`（如有，作为回退），并通过会话内 notify 显式警告，不静默降级；单个 provider 失败不影响其他 provider。
 - **apiKey 解析**（仅发现请求）：支持字面量与 `$ENV_VAR` / `${ENV_VAR}` 插值；`!command` 形式跳过发现并显式警告（聊天请求仍由 pi 自身解析执行，不受影响）。
 - 发现的模型默认参数：`reasoning: true`、`input: ["text", "image"]`、cost 全 0、`contextWindow` 1M、`maxTokens` 64K、`compat.supportsDeveloperRole: false`；provider 级 `compat` 会合并进每个发现的模型。
-- 发现的模型声明 `thinkingLevelMap: { xhigh: "xhigh", max: "max" }`，让 `xhigh` / `max` 出现在 `/thinking`；标准档位保持缺省，仍由 pi 的 provider 默认映射决定。`/models` 无法逐个模型返回思考能力，所以这是对所有发现模型一刀切的默认值：某个 provider 拒绝这些档位时，在 models.json 的 `modelOverrides` 里按 model.id 覆盖（把 `xhigh` / `max` 置 `null` 可隐藏），或改用手写 `models` 条目自行声明。
+- 发现的模型默认将 `thinkingLevelMap.xhigh` / `max` 设为 `null`，不再假定所有后端支持这些额外档位。标准档位仍沿用 Pi 默认映射。若服务支持额外档位，可在 provider 的 `compat.thinkingLevelMap` 中显式开启，也可将任一不支持的档位置为 `null`。
+- 非推理模型可设置 provider 的 `compat.reasoning: false`，让 `/thinking` 只提供 `off`。`reasoning` 和 `thinkingLevelMap` 会提取到模型顶层，其余 `compat` 字段（如 `supportsReasoningEffort`、`thinkingFormat`）正常转发。混合模型网关可用 `modelOverrides` 按 model.id 覆盖。旧缓存会在启动时应用当前推理配置，无需联网刷新。
+
+下面是本地非推理服务的配置；若要开启已确认支持的额外档位，改用 `"thinkingLevelMap": { "xhigh": "xhigh", "max": "max" }` 并移除 `"reasoning": false`。
 
 ```json
 {
   "providers": {
-    "llm-proxy": {
+    "local": {
+      "baseUrl": "http://127.0.0.1:8000/v1",
+      "api": "openai-completions",
       "discoverModels": true,
-      "modelOverrides": {
-        "some-model": { "thinkingLevelMap": { "xhigh": null, "max": null } }
+      "compat": {
+        "reasoning": false,
+        "supportsReasoningEffort": false,
+        "thinkingLevelMap": { "xhigh": null, "max": null }
       }
     }
   }
 }
 ```
-- 模型元数据可携带 `name` / `context_window`（或 `contextWindow`）/ `max_tokens`（或 `maxTokens`），缺失时用默认值。
+- 模型元数据可携带 `name` / `context_window`（或 `contextWindow`，或 vLLM 的 `max_model_len`）/ `max_tokens`（或 `maxTokens`），缺失时用默认值。
 
 旧的 `/model-discovery`、`/model-discovery-refresh`、`/pi-model-discovery` 和 `/pi-model-discovery-refresh` 名称仍作为兼容别名保留。
 

package/SKILL.md

@@ -23,7 +23,7 @@ description: "配置与排查 models.json 的动态模型发现、缓存刷新
 
 ## 思考等级
 
-发现的模型默认声明 `thinkingLevelMap: { xhigh: "xhigh", max: "max" }`（对所有模型一刀切）。若某个模型不认 `xhigh` / `max`，用 models.json 的 `modelOverrides` 按 model.id 把它们置 `null`。
+发现的模型默认隐藏 `xhigh` / `max`（映射为 `null`），标准档位沿用 Pi 默认映射。provider 的 `compat.thinkingLevelMap` 可显式开启支持的档位或将不支持的档位置 `null`；非推理服务用 `compat.reasoning: false` 关闭推理，必要时配合 `compat.supportsReasoningEffort: false`。这两个扩展配置字段会提取到模型顶层，其余 compat 正常转发。逐模型差异用 `modelOverrides` 按 model.id 覆盖。旧缓存启动时会应用当前配置，无需联网刷新。
 
 ## 验证
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-models-discovery",
-  "version": "1.6.0",
+  "version": "1.6.1",
   "description": "Pi extension: discover models from {baseUrl}/models for providers marked with discoverModels in models.json",
   "type": "module",
   "main": "./index.ts",

package/src/index.ts

@@ -35,9 +35,8 @@
  * - 注册 refreshModels：打开 /model 触发在线刷新时重新发现；
  *   离线初始化（allowNetwork=false）返回上次成功列表（含缓存），尚无成功记录时抛错，
  *   以免空列表清掉 models.json 手写 models。
- * - 发现模型默认声明 thinkingLevelMap { xhigh: "xhigh", max: "max" }，让 xhigh/max 出现在
- *   /thinking；标准档位缺省沿用 pi 的 provider 默认映射。这是对所有发现模型一刀切的默认值，
- *   上游不认这些值时用 models.json 的 provider.modelOverrides 按 model.id 覆盖。
+ * - 发现模型默认隐藏 xhigh/max；provider.compat.thinkingLevelMap 可显式开启或覆盖档位，
+ *   provider.compat.reasoning 可关闭推理；逐模型差异用 provider.modelOverrides 覆盖。
  * - 发现请求的 apiKey 解析仅支持字面量与 $ENV_VAR/${ENV_VAR} 插值；
  *   "!command" 形式的 apiKey 跳过发现（显式警告），pi 发起聊天请求时仍由 pi 自身解析。
  * - /config:model-discovery 命令对 models.json 的修改立即生效（registerProvider 运行时可直接调用）；
@@ -96,6 +95,8 @@ interface ModelsResponse {
 		name?: string;
 		context_window?: number;
 		contextWindow?: number;
+		/** vLLM reports the served context window here. */
+		max_model_len?: number;
 		max_tokens?: number;
 		maxTokens?: number;
 	}>;
@@ -298,6 +299,21 @@ function resolveEnvValue(raw: string): { value: string | null; error: string | n
 	return { value, error: null };
 }
 
+/** 扩展专用的推理默认值从 compat 提取，不作为请求兼容字段转发。 */
+function modelThinking(providerCompat: Record<string, unknown> | undefined) {
+	const { reasoning, thinkingLevelMap, ...compat } = providerCompat ?? {};
+	return {
+		reasoning: typeof reasoning === "boolean" ? reasoning : true,
+		thinkingLevelMap: {
+			xhigh: null,
+			max: null,
+			...(thinkingLevelMap && typeof thinkingLevelMap === "object" && !Array.isArray(thinkingLevelMap)
+				? thinkingLevelMap as ProviderModelConfig["thinkingLevelMap"] : {}),
+		},
+		compat: { supportsDeveloperRole: false, ...compat },
+	};
+}
+
 export function buildModel(
 	id: string,
 	name: string | undefined,
@@ -308,18 +324,11 @@ export function buildModel(
 	return {
 		id,
 		name: name ?? id,
-		reasoning: true,
-		// xhigh/max 必须给非 null 值才会出现在 /thinking（缺省键等同不支持）；
-		// 标准档位保持缺省，仍走 pi 的 provider 默认映射，语义不变。
-		thinkingLevelMap: { xhigh: "xhigh", max: "max" },
+		...modelThinking(providerCompat),
 		input: ["text", "image"],
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 		contextWindow: contextWindow ?? 1_000_000,
 		maxTokens: maxTokens ?? 65_536,
-		compat: {
-			supportsDeveloperRole: false,
-			...providerCompat,
-		},
 	};
 }
 
@@ -375,7 +384,7 @@ async function fetchModels(
 			buildModel(
 				m.id,
 				m.name,
-				m.context_window ?? m.contextWindow,
+				m.context_window ?? m.contextWindow ?? m.max_model_len,
 				m.max_tokens ?? m.maxTokens,
 				entry.compat,
 			),
@@ -486,6 +495,8 @@ function registerFromCache(
 	notices: Notice[],
 ): void {
 	if (!validateEntry(entry, notices)) return;
+	// 旧缓存含硬编码的 xhigh/max，启动时按当前 provider 配置重建，保持离线可用。
+	models = models.map((model) => ({ ...model, ...modelThinking(entry.compat) }));
 	const state: LastModelsState = { lastModels: models };
 	const refreshModels = createRefreshModels(entry, notices, state);
 	pi.registerProvider(entry.id, {