@capybara-hq/plugin-model-fake 0.1.0-rc.3 → 0.1.0-rc.5

package/.capybara-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "schemaVersion": 1,
   "name": "model-fake",
-  "version": "0.1.0-rc.3",
+  "version": "0.1.0-rc.5",
   "publisher": "Capybara",
   "description": {
     "en": "Deterministic scripted model provider (model-provider@2) for runtime tests, e2e tests and keyless demos",
@@ -16,7 +16,7 @@
     "entry": "./dist/index.mjs",
     "manifest": {
       "name": "model-fake",
-      "version": "0.1.0-rc.3",
+      "version": "0.1.0-rc.5",
       "sdk": "^0.1.0-rc.0",
       "provides": [
         "model-provider@2"
@@ -98,12 +98,69 @@
                     }
                   },
                   "title": "Document input"
+                },
+                "kind": {
+                  "type": "string",
+                  "enum": [
+                    "chat",
+                    "embedding",
+                    "rerank",
+                    "image_generation",
+                    "video_generation",
+                    "speech_to_text",
+                    "text_to_speech"
+                  ],
+                  "x-ui": {
+                    "label": {
+                      "zh": "类型",
+                      "en": "Kind"
+                    }
+                  },
+                  "title": "Kind"
+                },
+                "publisher": {
+                  "type": "string",
+                  "x-ui": {
+                    "label": {
+                      "zh": "发布者",
+                      "en": "Publisher"
+                    }
+                  },
+                  "title": "Publisher"
+                },
+                "dimensions": {
+                  "type": "array",
+                  "items": {
+                    "type": "integer",
+                    "exclusiveMinimum": 0,
+                    "maximum": 9007199254740991
+                  },
+                  "x-ui": {
+                    "label": {
+                      "zh": "向量维度",
+                      "en": "Embedding dimensions"
+                    }
+                  },
+                  "title": "Embedding dimensions"
+                },
+                "maxBatch": {
+                  "type": "integer",
+                  "exclusiveMinimum": 0,
+                  "maximum": 9007199254740991,
+                  "x-ui": {
+                    "label": {
+                      "zh": "单次最大输入数",
+                      "en": "Max inputs per call"
+                    }
+                  },
+                  "title": "Max inputs per call"
                 }
               },
               "required": [
                 "id"
               ]
             },
+            "x-org": "override",
             "x-ui": {
               "label": {
                 "zh": "模型",

package/dist/index.mjs

@@ -8,7 +8,9 @@ const __capybara_dirname = __capybaraPathDirname(__capybara_filename);
 // src/index.ts
 import {
   defineExtension,
-  MODEL_PROVIDER_EXTENSION_POINT_V2
+  MODEL_PROVIDER_EXTENSION_POINT_V2,
+  ModelError as ModelError2,
+  orgAccess
 } from "@capybara-hq/plugin-sdk";
 import { z } from "zod";
 
@@ -107,6 +109,7 @@ function mediaSummary(req) {
   }
   return "none";
 }
+var SENDER_LINE = /^From [^\n]+$/;
 function lastUser(req) {
   for (let i = req.messages.length - 1; i >= 0; i--) {
     const m = req.messages[i];
@@ -114,7 +117,9 @@ function lastUser(req) {
     let text = "";
     let result = "";
     let hasToolResult = false;
-    for (const b of m.content) {
+    const first = m.content[0];
+    const blocks = m.content.length > 1 && first?.type === "text" && SENDER_LINE.test(first.text) ? m.content.slice(1) : m.content;
+    for (const b of blocks) {
       if (b.type === "text") text += b.text;
       else if (b.type === "tool_result") {
         hasToolResult = true;
@@ -177,9 +182,9 @@ function scripted(req) {
   if (multi) {
     try {
       const calls = JSON.parse(multi[1]);
-      const [first2, ...more] = calls;
-      if (first2 === void 0) throw new Error("empty");
-      return { kind: "tool", name: first2.name, input: first2.input, more };
+      const [first, ...more] = calls;
+      if (first === void 0) throw new Error("empty");
+      return { kind: "tool", name: first.name, input: first.input, more };
     } catch {
       return {
         kind: "fail",
@@ -204,29 +209,72 @@ function scripted(req) {
     return { kind: "text", text: words.join(" "), chunk: 64, delayMs: 0 };
   }
   if (hasToolResult) return std(`Tool result: ${result}`);
-  const first = req.tools[0];
-  if (first !== void 0) return { kind: "tool", name: first.name, input: {} };
+  const only = req.tools.length === 1 ? req.tools[0] : void 0;
+  if (only !== void 0) return { kind: "tool", name: only.name, input: {} };
   return std(`You said: ${text}`);
 }
+function embedVector(text, dim) {
+  const out = [];
+  for (let block = 0; out.length < dim; block++) {
+    const h = createHash("sha256").update(`${block}:${text}`).digest();
+    for (let i = 0; i + 1 < h.length && out.length < dim; i += 2) {
+      out.push(h.readUInt16BE(i) / 65535 * 2 - 1);
+    }
+  }
+  return out;
+}
 function createFakeProvider(opts) {
   const sleep = opts.sleep ?? defaultSleep;
-  const configured = opts.models !== void 0 && opts.models.length > 0 ? opts.models : [{ id: "fake-1" }];
-  const infos = configured.map((m) => ({
-    id: m.id,
-    displayName: m.id,
-    contextWindow: m.contextWindow ?? 32e3,
-    maxOutputTokens: 4096,
-    capabilities: {
-      tools: true,
-      vision: m.vision ?? false,
-      documents: m.documents ?? false,
-      reasoning: false,
-      caching: true,
-      midConversationSystem: true
-    },
-    ...m.streamIdleTimeoutMs !== void 0 ? { streamIdleTimeoutMs: m.streamIdleTimeoutMs } : {},
-    parameters: parameters()
-  }));
+  const ERROR_KINDS = [
+    "rate_limited",
+    "overloaded",
+    "server",
+    "timeout",
+    "network",
+    "auth",
+    "invalid_request",
+    "context_overflow",
+    "aborted"
+  ];
+  function infosOf(models) {
+    const configured = models !== void 0 && models.length > 0 ? models : [{ id: "fake-1" }];
+    return configured.map((m) => {
+      const embedding = m.kind === "embedding";
+      const dims = m.dimensions ?? [8, 16];
+      return {
+        id: m.id,
+        displayName: m.id,
+        ...m.kind !== void 0 ? { kind: m.kind } : {},
+        ...m.publisher !== void 0 ? { publisher: m.publisher } : {},
+        ...embedding ? {
+          embedding: {
+            dimensions: dims,
+            defaultDimensions: m.dimensions?.[0] ?? 8,
+            maxInputTokens: 512,
+            maxBatch: m.maxBatch ?? 4
+          }
+        } : {},
+        contextWindow: m.contextWindow ?? 32e3,
+        maxOutputTokens: 4096,
+        capabilities: {
+          tools: true,
+          vision: m.vision ?? false,
+          documents: m.documents ?? false,
+          reasoning: false,
+          caching: true,
+          midConversationSystem: true
+        },
+        ...m.streamIdleTimeoutMs !== void 0 ? { streamIdleTimeoutMs: m.streamIdleTimeoutMs } : {},
+        parameters: parameters()
+      };
+    });
+  }
+  const setupInfos = infosOf(opts.models);
+  function infosFor(call) {
+    if (call?.config === void 0 || opts.modelsOf === void 0) return setupInfos;
+    const models = opts.modelsOf(call.config);
+    return models === void 0 ? setupInfos : infosOf(models);
+  }
   async function cacheUsage(req) {
     const cached = req.system.filter((s) => s.cache).map((s) => s.text);
     const prefix = cached.join("\n");
@@ -240,11 +288,34 @@ function createFakeProvider(opts) {
     return { read: 0, write: tokens(cachedChars), cachedChars };
   }
   return {
-    async listModels() {
-      return infos.map((i) => structuredClone(i));
+    async listModels(call) {
+      return infosFor(call).map((i) => structuredClone(i));
+    },
+    async embed(req, call) {
+      if (call.signal.aborted) throw abortError();
+      const info = infosFor(call).find((i) => i.id === req.model && i.kind === "embedding");
+      if (info?.embedding === void 0) {
+        throw new ModelError("invalid_request", `Unknown embedding model "${req.model}"`);
+      }
+      const failure = /^fail:(\w+)/.exec(req.input[0] ?? "");
+      if (failure && ERROR_KINDS.includes(failure[1])) {
+        throw new ModelError(failure[1], `scripted failure: ${failure[1]}`);
+      }
+      const dim = req.dimensions ?? info.embedding.defaultDimensions;
+      if (!info.embedding.dimensions.includes(dim)) {
+        throw new ModelError("invalid_request", `Unsupported dimensions ${dim}`);
+      }
+      if (req.input.length > info.embedding.maxBatch) {
+        throw new ModelError("invalid_request", "Too many inputs in one embed call");
+      }
+      return {
+        vectors: req.input.map((t) => embedVector(t, dim)),
+        usage: { inputTokens: req.input.reduce((n, t) => n + Math.ceil(t.length / 4), 0) }
+      };
     },
-    async *stream(req, { signal }) {
-      if (!infos.some((i) => i.id === req.model)) {
+    async *stream(req, call) {
+      const { signal } = call;
+      if (!infosFor(call).some((i) => i.id === req.model && (i.kind ?? "chat") === "chat")) {
         throw new ModelError("invalid_request", `Unknown model "${req.model}"`);
       }
       if (signal.aborted) throw abortError();
@@ -307,20 +378,35 @@ function createFakeProvider(opts) {
 // src/index.ts
 var label = (zh, en) => ({ ui: { label: { zh, en } } });
 var DeploymentConfig = z.object({
-  models: z.array(
-    z.object({
-      id: z.string().min(1).meta(label("模型 ID", "Model ID")),
-      contextWindow: z.number().int().positive().optional().meta(label("上下文窗口", "Context window")),
-      streamIdleTimeoutMs: z.number().int().positive().optional().meta(label("流空闲超时（毫秒）", "Stream idle timeout (ms)")),
-      /** Declares image / document input support (default false). */
-      vision: z.boolean().optional().meta(label("图片输入", "Image input")),
-      documents: z.boolean().optional().meta(label("文档输入", "Document input"))
-    })
-  ).optional().meta(label("模型", "Models"))
+  models: orgAccess(
+    z.array(
+      z.object({
+        id: z.string().min(1).meta(label("模型 ID", "Model ID")),
+        contextWindow: z.number().int().positive().optional().meta(label("上下文窗口", "Context window")),
+        streamIdleTimeoutMs: z.number().int().positive().optional().meta(label("流空闲超时（毫秒）", "Stream idle timeout (ms)")),
+        /** Declares image / document input support (default false). */
+        vision: z.boolean().optional().meta(label("图片输入", "Image input")),
+        documents: z.boolean().optional().meta(label("文档输入", "Document input")),
+        kind: z.enum([
+          "chat",
+          "embedding",
+          "rerank",
+          "image_generation",
+          "video_generation",
+          "speech_to_text",
+          "text_to_speech"
+        ]).optional().meta(label("类型", "Kind")),
+        publisher: z.string().optional().meta(label("发布者", "Publisher")),
+        dimensions: z.array(z.number().int().positive()).optional().meta(label("向量维度", "Embedding dimensions")),
+        maxBatch: z.number().int().positive().optional().meta(label("单次最大输入数", "Max inputs per call"))
+      })
+    ).optional().meta(label("模型", "Models")),
+    "override"
+  )
 });
 var index_default = defineExtension({
   name: "model-fake",
-  version: "0.1.0-rc.3",
+  version: "0.1.0-rc.5",
   sdk: "^0.1.0-rc.0",
   provides: [MODEL_PROVIDER_EXTENSION_POINT_V2],
   config: { deployment: DeploymentConfig },
@@ -328,7 +414,20 @@ var index_default = defineExtension({
     const { models } = DeploymentConfig.parse(ctx.config ?? {});
     ctx.provide(
       MODEL_PROVIDER_EXTENSION_POINT_V2,
-      createFakeProvider({ kv: ctx.kv, ...models !== void 0 ? { models } : {} })
+      createFakeProvider({
+        kv: ctx.kv,
+        ...models !== void 0 ? { models } : {},
+        modelsOf: (config) => {
+          const parsed = DeploymentConfig.safeParse(config ?? {});
+          if (!parsed.success) {
+            throw new ModelError2(
+              "invalid_request",
+              `Invalid model-fake config: ${parsed.error.issues.map((i) => `${i.path.join(".")}: ${i.message}`).join("; ")}`
+            );
+          }
+          return parsed.data.models;
+        }
+      })
     );
   }
 });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@capybara-hq/plugin-model-fake",
-  "version": "0.1.0-rc.3",
+  "version": "0.1.0-rc.5",
   "type": "module",
   "license": "Apache-2.0",
   "description": "Deterministic scripted model provider (model-provider@2) for runtime tests, e2e tests and keyless demos",