pi-provider-sicau-pmt 0.1.0 → 0.2.0

package/README.md

@@ -1,6 +1,6 @@
 # pi-provider-sicau-pmt
 
-[Pi](https://pi.dev) provider for a self-hosted [qoder2api](https://github.com/shuishuipingan/qoder2api-hub) gateway. Adds the gateway's models to Pi with correct thinking-effort support, and asks for the API key through `/login` instead of hand-editing config files.
+[Pi](https://pi.dev) provider for a self-hosted `oh-pmt-api` gateway (OpenAI-compatible). Adds the gateway's models to Pi with correct context windows, output limits and thinking-effort tables, and asks for the API key through `/login` instead of hand-editing config files.
 
 ## Install
 
@@ -24,24 +24,32 @@ pi -e npm:pi-provider-sicau-pmt
 
 ## What it registers
 
-| Model | Context | Thinking effort |
-|---|---|---|
-| DeepSeek-Flash | 1M | low / high / max, off |
-| GLM-5.3-Flash | 1M | high / max |
-| Qwen-3.8-Flash | 180K | low / medium / xhigh, off |
+The model list comes from the gateway's `/v1/models` at startup. Models added to the gateway show up in Pi without editing this package; the built-in table below is only a fallback for offline startup and the first run after install.
 
-The gateway's `/v1/models` publishes no capability metadata, so context windows and effort tables are declared statically from Qoder's official catalog.
+| Model | Context | Output | Thinking effort | Image input |
+|---|---|---|---|---|
+| Qwen-3.8-Flash | 180K | 131072 | low / medium / xhigh, off | no |
+| DeepSeek-Flash | 1M | 131072 | low / high / max, off | no |
+| GLM-5.3-Flash | 1M | 131072 | low / high / max | yes |
+| MiniMax-M3.1-Flash | 1M | 128000 | low / medium / high / xhigh / max | yes |
+| DeepSeek-V4.1-Flash | 1M | 128000 | low / high / max, off | yes |
+| Space-Bunny | 1M | 128000 | low / medium / high / xhigh / max | yes |
+| Step-5-Preview | 1M | 64000 | low / medium / high | yes |
 
-`compat.supportsDeveloperRole` is disabled on every model. Pi switches the system prompt to the `developer` role on any model with `reasoning: true`, and the upstream Qoder API validates message roles against a serde enum that accepts only `system`, `assistant`, `user` and `tool` — a `developer` message rejects the whole request.
+The gateway's `/v1/models` publishes no capability metadata, so capabilities live in `src/models.ts`, keyed by upstream key. Context windows and output limits come from Qoder's official catalog and WorkBuddy's product config; effort tables come from the vendors' declared effort lists.
+
+Duplicate display ids are collapsed: the gateway lists `GLM-5.3-Flash` and `MiniMax-M3.1-Flash` twice, once per upstream, and the id-form request always resolves to the same backend, so one entry is enough.
+
+`compat.supportsDeveloperRole` is disabled on every model. Pi switches the system prompt to the `developer` role on any model with `reasoning: true`, and the Qoder upstream validates message roles against a serde enum that accepts only `system`, `assistant`, `user` and `tool` — a `developer` message rejects the whole request.
 
 ## Point it at your own gateway
 
-`BASE_URL` in `src/index.ts` is pinned to `https://api.hhy.homes/v1`. To use a different qoder2api deployment, fork this package, change `BASE_URL` and the model table, and install your fork.
+`BASE_URL` in `src/index.ts` is pinned to `https://api.hhy.homes/v1`. To use a different deployment, fork this package, change `BASE_URL`, and rewrite the capability table in `src/models.ts` for your models.
 
 ## Requirements
 
-- Pi 1.4.0 or newer.
-- A qoder2api gateway that forwards `reasoning_effort` and reads `max_completion_tokens`. Gateways without that plumbing still work, but the thinking selector has no effect and the output limit falls back to the server default.
+- Pi 1.4.0 or newer — the model list is refreshed through `ProviderConfig.refreshModels`.
+- A gateway that forwards `reasoning_effort` and reads `max_completion_tokens`. Gateways without that plumbing still work, but the thinking selector has no effect and the output limit falls back to the server default.
 
 ## License
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "pi-provider-sicau-pmt",
-  "version": "0.1.0",
-  "description": "Pi provider for a self-hosted qoder2api gateway (OpenAI-compatible), with thinking-effort support and a /login API-key prompt",
+  "version": "0.2.0",
+  "description": "Pi provider for a self-hosted oh-pmt-api gateway (OpenAI-compatible), with dynamic model discovery, thinking-effort support and a /login API-key prompt",
   "type": "module",
   "license": "MIT",
   "publishConfig": {

package/src/index.ts

@@ -11,89 +11,63 @@
  * 没给 apiKey 时自动补一个默认的 API key 登录，/login 里弹 secret 输入框，
  * 存成 {"type":"api_key","key":...}，正是我们要的。
  *
- * 上游能力（思考档位、上下文窗口）取自 Qoder 官方目录，中转站的 /v1/models
- * 不公布这些信息，所以写死在下面的表里。
+ * 模型列表走 refreshModels 动态拉取：中转站的 /v1/models 决定有哪几个模型，
+ * 能力（上下文窗口、输出上限、思考档位）由 models.ts 里的表按 upstream_key 补。
  */
 import type { ExtensionAPI, ProviderModelConfig } from "@earendil-works/pi-coding-agent";
+import { SEED_MODELS, modelsFromCatalog, type GatewayModel } from "./models.ts";
 
 const PROVIDER_ID = "sicau-pmt";
 const BASE_URL = "https://api.hhy.homes/v1";
 
-/**
- * 每个模型官方的思考档位，取自 Qoder 目录的 thinking_config：
- *   Qwen3.8-Flash     low / medium / xhigh，默认 medium，可关闭
- *   GLM-5.3-Flash     high / max，默认 max，不能关闭
- *   DeepSeek-Flash    low / high / max，默认 high，可关闭
- * 关不掉的模型 off 留 null，pi 就不显示关闭选项。
- *
- * supportsDeveloperRole 必须关掉：pi 在 reasoning 模型上会把系统提示词的
- * 角色换成 developer，而上游只认 system，整条请求会被拒。
- */
-const MODELS: ProviderModelConfig[] = [
-  {
-    id: "Qwen-3.8-Flash",
-    name: "Qwen 3.8 Flash",
-    reasoning: true,
-    thinkingLevelMap: {
-      off: "disabled",
-      minimal: null,
-      low: "low",
-      medium: "medium",
-      high: null,
-      xhigh: "xhigh",
-      max: null,
-    },
-    input: ["text"],
-    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-    contextWindow: 180000,
-    maxTokens: 8192,
-    compat: { supportsDeveloperRole: false },
-  },
-  {
-    id: "GLM-5.3-Flash",
-    name: "GLM 5.3 Flash",
-    reasoning: true,
-    thinkingLevelMap: {
-      off: null,
-      minimal: null,
-      low: null,
-      medium: null,
-      high: "high",
-      xhigh: null,
-      max: "max",
-    },
-    input: ["text"],
-    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-    contextWindow: 1000000,
-    maxTokens: 8192,
-    compat: { supportsDeveloperRole: false },
-  },
-  {
-    id: "DeepSeek-Flash",
-    name: "DeepSeek Flash",
-    reasoning: true,
-    thinkingLevelMap: {
-      off: "disabled",
-      minimal: null,
-      low: "low",
-      medium: null,
-      high: "high",
-      xhigh: null,
-      max: "max",
-    },
-    input: ["text"],
-    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-    contextWindow: 1000000,
-    maxTokens: 8192,
-    compat: { supportsDeveloperRole: false },
-  },
-];
+/** 目录缓存时间。同一个会话里 pi 会多次触发刷新，没必要每次都打网关。 */
+const CATALOG_TTL_MS = 5 * 60 * 1000;
+/** 网关的 /v1/models 只是个静态列表，不需要等太久。 */
+const CATALOG_TIMEOUT_MS = 8 * 1000;
+
+let cached: ProviderModelConfig[] = SEED_MODELS;
+let cachedAt = 0;
+
+/** 从网关拉一次目录。失败返回 null，由调用方回落到缓存或内置表。 */
+async function fetchCatalog(apiKey: string, signal: AbortSignal): Promise<ProviderModelConfig[] | null> {
+  const timeout = AbortSignal.timeout(CATALOG_TIMEOUT_MS);
+  try {
+    const response = await fetch(`${BASE_URL}/models`, {
+      headers: apiKey ? { authorization: `Bearer ${apiKey}` } : {},
+      signal: AbortSignal.any([signal, timeout]),
+    });
+    if (!response.ok) return null;
+    const payload = (await response.json()) as { data?: GatewayModel[] };
+    const entries = Array.isArray(payload?.data) ? payload.data : [];
+    const models = modelsFromCatalog(entries);
+    return models.length > 0 ? models : null;
+  } catch {
+    return null;
+  }
+}
 
 export default function sicauPmt(pi: ExtensionAPI) {
   pi.registerProvider(PROVIDER_ID, {
     name: "SICAU-PMT",
     baseUrl: BASE_URL,
     api: "openai-completions",
-    models: MODELS,
+    models: SEED_MODELS,
+    /**
+     * pi 在启动时先用 allowNetwork=false 调一次，从缓存恢复；随后在交互模式下用
+     * allowNetwork=true 再调一次，这时才真的访问网关。返回值会替换掉上面注册的
+     * 模型列表，所以任何路径都必须返回非空列表。
+     */
+    async refreshModels(context) {
+      if (!context.allowNetwork) return cached;
+      if (!context.force && Date.now() - cachedAt < CATALOG_TTL_MS) return cached;
+
+      const key = context.credential?.type === "api_key" ? (context.credential.key ?? "") : "";
+      const models = await fetchCatalog(key, context.signal);
+      if (!models) return cached;
+
+      cached = models;
+      cachedAt = Date.now();
+      return models;
+    },
   });
 }

package/src/models.ts

@@ -0,0 +1,224 @@
+/**
+ * 模型能力表 + 目录合并规则。
+ *
+ * 中转站的 /v1/models 只给出 id、upstream_key 和 owned_by，没有上下文窗口、输出
+ * 上限、思考档位这些能力数据。所以能力写在这张表里，按 upstream_key 索引——
+ * upstream_key 在上游唯一，而展示用的 id 会撞车（GLM-5.3-Flash 和
+ * MiniMax-M3.1-Flash 在 Qoder 与 WorkBuddy 各有一条，id 相同）。
+ *
+ * 合并规则：/v1/models 决定有哪几个模型，这张表只补能力。表里没有的模型用保守
+ * 默认值，仍然可用，网关将来新增模型也不会从列表里消失。
+ */
+import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
+
+/**
+ * 档位表类型。ProviderChatModelConfig 没有从包入口导出，所以从联合类型里按
+ * 「有 reasoning 字段」把 chat 分支取出来。
+ */
+type LevelMap = NonNullable<Extract<ProviderModelConfig, { reasoning: boolean }>["thinkingLevelMap"]>;
+
+/** 一条模型的能力描述，合并时按 upstream_key 查。 */
+interface Capability {
+  name: string;
+  contextWindow: number;
+  maxTokens: number;
+  /** 是否按思考模型注册。false 时 pi 只给「关」，也不下发 reasoning_effort。 */
+  reasoning: boolean;
+  thinkingLevelMap?: LevelMap;
+  input: ("text" | "image")[];
+}
+
+/** 上游把 none / off / disabled / disable / false / 0 / no 都当关闭，这里统一用 disabled。 */
+const OFF = "disabled";
+
+/**
+ * 由厂商声明的档位列表生成 thinkingLevelMap。
+ *
+ * pi 的规则：到 high 为止的标准档位，没写等于「用上游默认值」；xhigh 与 max 必须
+ * 显式写非 null 才可选。所以没声明的档位一律写 null，避免选了 max 却静默退回默认档。
+ * off 传 null 表示该模型关不掉思考，pi 就不显示关闭选项。
+ */
+function levelMap(efforts: string[], off: string | null): LevelMap {
+  const supported = new Set(efforts);
+  const map: LevelMap = { off };
+  for (const level of ["minimal", "low", "medium", "high", "xhigh", "max"] as const) {
+    map[level] = supported.has(level) ? level : null;
+  }
+  return map;
+}
+
+/**
+ * 该模型关不掉思考：off 写 null，pi 就不显示关闭选项，也不下发 reasoning_effort。
+ *
+ * 档位仍按厂商声明的写。网关这条链路眼下不会把 reasoning_effort 转给上游，选了也
+ * 只是回到默认档；但 pi 需要至少一个可选档位（全是 null 会被 clampThinkingLevel
+ * 退成 off，反而把思考模型当成关闭思考用）。等网关补上转发，这张表不用改。
+ */
+function alwaysThinking(efforts: string[]): LevelMap {
+  return levelMap(efforts, null);
+}
+
+/**
+ * 能力表。上下文窗口与输出上限取自 Qoder 官方目录
+ * （~/.pi/agent/qoder*-models-cache.json）和 WorkBuddy 产品配置
+ * （~/.workbuddy/cache/acc-product-config-v3.json）。
+ *
+ * key 是网关实际路由到的上游 key，不是 owned_by。网关的 owned_by 是
+ * `qoder-<region>` 硬编码，看不出真实厂商：请求进来时 resolve_model 先按上游 key
+ * 精确匹配、再按去掉区域后缀的小写形式匹配，所以 id 为 GLM-5.3-Flash 的请求会命中
+ * WorkBuddy 的 glm-5.3-flash，而不是 Qoder 的 gfmodel（后者已因缺订阅返回 403）。
+ */
+const CAPABILITIES: Record<string, Capability> = {
+  // ---- 走 Qoder 适配器：网关会把 reasoning_effort 传下去，档位真实生效 ----
+  qfmodel: {
+    name: "Qwen 3.8 Flash",
+    // Qoder 目录里 max_input_tokens 是 180000，另有 200K/400K/1M 可选档，默认 200K
+    contextWindow: 180000,
+    maxTokens: 131072,
+    reasoning: true,
+    thinkingLevelMap: levelMap(["low", "medium", "xhigh"], OFF),
+    input: ["text"],
+  },
+  dfmodel: {
+    name: "DeepSeek Flash",
+    contextWindow: 1000000,
+    maxTokens: 131072,
+    reasoning: true,
+    thinkingLevelMap: levelMap(["low", "high", "max"], OFF),
+    input: ["text"],
+  },
+
+  // ---- 走 WorkBuddy 适配器 ----
+  // 该适配器只转发温度、top_p 这类采样字段，reasoning_effort 不在白名单里，所以
+  // 这几个模型眼下的档位调不动（实测任何档位都没有变化）。档位仍按厂商声明的写，
+  // 界面上跟厂商本身一致；等网关补上转发就自然生效。
+  "glm-5.3-flash": {
+    name: "GLM 5.3 Flash",
+    contextWindow: 1000000,
+    maxTokens: 131072,
+    reasoning: true,
+    thinkingLevelMap: alwaysThinking(["low", "high", "max"]),
+    input: ["text", "image"],
+  },
+  "minimax-m3.1-flash": {
+    name: "MiniMax M3.1 Flash",
+    contextWindow: 1000000,
+    maxTokens: 128000,
+    reasoning: true,
+    thinkingLevelMap: alwaysThinking(["low", "medium", "high", "xhigh", "max"]),
+    input: ["text", "image"],
+  },
+  "deepseek-v4.1-flash": {
+    name: "DeepSeek V4.1 Flash",
+    contextWindow: 1000000,
+    maxTokens: 128000,
+    reasoning: true,
+    thinkingLevelMap: alwaysThinking(["low", "high", "max"]),
+    input: ["text", "image"],
+  },
+  "space-bunny": {
+    name: "Space Bunny",
+    contextWindow: 1000000,
+    maxTokens: 128000,
+    reasoning: true,
+    thinkingLevelMap: alwaysThinking(["low", "medium", "high", "xhigh", "max"]),
+    input: ["text", "image"],
+  },
+  "step-5-preview": {
+    name: "Step 5 Preview",
+    contextWindow: 1000000,
+    maxTokens: 64000,
+    reasoning: true,
+    thinkingLevelMap: alwaysThinking(["low", "medium", "high"]),
+    input: ["text", "image"],
+  },
+};
+
+/**
+ * 表里查不到时用的默认值。中转站不校验 max_tokens，设大了不会报错，但真按 1M
+ * 输出截断会很难排查，所以给一个大多数模型都吃得下的值。
+ */
+const FALLBACK: Capability = {
+  name: "",
+  contextWindow: 200000,
+  maxTokens: 32000,
+  reasoning: false,
+  input: ["text"],
+};
+
+/**
+ * supportsDeveloperRole 保持关闭：pi 在 reasoning 模型上会把系统提示词的角色换成
+ * developer，而 Qoder 那条链路用 serde 枚举校验 role，只认 system / assistant /
+ * user / tool，多一个 developer 整条请求会被拒。网关现在会把 developer 转成 system，
+ * 但关掉这个开关没有代价，留着当保险。
+ */
+const COMPAT = { supportsDeveloperRole: false } as const;
+
+/** /v1/models 里我们用到的那几个字段。 */
+export interface GatewayModel {
+  id: string;
+  upstream_key?: string;
+  owned_by?: string;
+}
+
+/**
+ * 查出这条目录项真实对应的能力。
+ *
+ * 先按 upstream_key 查，查不到再按小写 id 查：网关的 resolve_model 是「先按上游 key
+ * 精确匹配，再按去掉区域后缀的小写名字匹配」，所以 id 为 GLM-5.3-Flash 的请求实际
+ * 落到 WorkBuddy 的 glm-5.3-flash，而不是列表里排在前面的 Qoder 那条 gfmodel
+ * （它已经因为缺订阅返回 403）。按同样的顺序查表，才能跟路由结果对上。
+ */
+function capabilityFor(entry: GatewayModel): Capability {
+  const byKey = CAPABILITIES[entry.upstream_key || ""];
+  if (byKey) return byKey;
+  return CAPABILITIES[entry.id.toLowerCase()] ?? FALLBACK;
+}
+
+/** 把网关返回的一条目录项补成 pi 的模型定义。 */
+export function toProviderModel(entry: GatewayModel): ProviderModelConfig {
+  const cap = capabilityFor(entry);
+  return {
+    id: entry.id,
+    name: cap.name || entry.id,
+    reasoning: cap.reasoning,
+    ...(cap.thinkingLevelMap ? { thinkingLevelMap: cap.thinkingLevelMap } : {}),
+    input: cap.input,
+    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+    contextWindow: cap.contextWindow,
+    maxTokens: cap.maxTokens,
+    compat: COMPAT,
+  };
+}
+
+/**
+ * 把网关目录去重成 pi 模型列表。
+ *
+ * 同一展示 id 可能出现两次（GLM-5.3-Flash、MiniMax-M3.1-Flash 在 Qoder 与 WorkBuddy
+ * 各有一条）。id 形式的请求总是走 resolve_model 的同一条匹配路径、落到同一个上游，
+ * 所以按 id 保留第一条即可。
+ */
+export function modelsFromCatalog(entries: GatewayModel[]): ProviderModelConfig[] {
+  const seen = new Set<string>();
+  const models: ProviderModelConfig[] = [];
+  for (const entry of entries) {
+    if (!entry.id || seen.has(entry.id)) continue;
+    seen.add(entry.id);
+    models.push(toProviderModel(entry));
+  }
+  return models;
+}
+
+/**
+ * 内置目录，服务启动、离线、或网关暂时不可用时先用它，保证模型始终可选。
+ * 与 2026-10-25 的 /v1/models 一致（重复 id 已按网关路由结果归并）。
+ */
+export const SEED_MODELS: ProviderModelConfig[] = modelsFromCatalog([
+  { id: "Qwen-3.8-Flash", upstream_key: "qfmodel" },
+  { id: "GLM-5.3-Flash", upstream_key: "glm-5.3-flash" },
+  { id: "DeepSeek-Flash", upstream_key: "dfmodel" },
+  { id: "MiniMax-M3.1-Flash", upstream_key: "minimax-m3.1-flash" },
+  { id: "DeepSeek-V4.1-Flash", upstream_key: "deepseek-v4.1-flash" },
+  { id: "Space-Bunny", upstream_key: "space-bunny" },
+  { id: "Step-5-Preview", upstream_key: "step-5-preview" },
+]);