feat: 新增 LLM checker 支持大模型服务应用层拨测

基于 AI SDK v6 实现 openai/openai-responses/anthropic 三类 provider 的 http/stream 模式调用
支持 output/finishReason/usage/stream 等完整 expect 断言链路
新增 9 个源文件和 5 个测试文件共 78 个测试
更新 README/DEVELOPMENT/probes.example.yaml 和 probe-config.schema.json

src/server/checker/runner/llm/observation.ts

@@ -0,0 +1,131 @@
+import type { APICallError } from "ai";
+
+import type {
+  LlmCheckObservation,
+  LlmHttpMetadata,
+  LlmMode,
+  LlmProvider,
+  LlmStreamObservation,
+  LlmUsageObservation,
+} from "./types";
+
+export function buildObservationFromApiCallError(
+  error: APICallError,
+  provider: LlmProvider,
+  model: string,
+  mode: LlmMode,
+): LlmCheckObservation {
+  const http: LlmHttpMetadata | null =
+    error.statusCode !== undefined
+      ? {
+          headers: error.responseHeaders ?? {},
+          status: error.statusCode,
+          statusText: "",
+        }
+      : null;
+
+  return {
+    finishReason: null,
+    http,
+    mode,
+    model,
+    outputText: null,
+    provider,
+    rawFinishReason: null,
+    stream: null,
+    usage: null,
+    warnings: [],
+  };
+}
+
+export function buildObservationFromGenerateText(
+  provider: LlmProvider,
+  model: string,
+  mode: LlmMode,
+  result: {
+    finishReason: string;
+    rawFinishReason: null | string | undefined;
+    text: string;
+    usage: { inputTokens: number; outputTokens: number; totalTokens?: number | undefined };
+    warnings?: string[];
+  },
+  http: LlmHttpMetadata | null,
+): LlmCheckObservation {
+  return {
+    finishReason: result.finishReason,
+    http,
+    mode,
+    model,
+    outputText: result.text,
+    provider,
+    rawFinishReason: result.rawFinishReason ?? null,
+    stream: null,
+    usage: {
+      inputTokens: result.usage.inputTokens ?? 0,
+      outputTokens: result.usage.outputTokens ?? 0,
+      totalTokens: result.usage.totalTokens ?? (result.usage.inputTokens ?? 0) + (result.usage.outputTokens ?? 0),
+    },
+    warnings: result.warnings ?? [],
+  };
+}
+
+export async function buildObservationFromStreamText(
+  provider: LlmProvider,
+  model: string,
+  mode: LlmMode,
+  fullStream: AsyncIterable<unknown>,
+  http: LlmHttpMetadata | null,
+  startMs: number,
+): Promise<LlmCheckObservation> {
+  let outputText = "";
+  let firstTokenMs: null | number = null;
+  let completed = false;
+  let finishReason: null | string = null;
+  let rawFinishReason: null | string = null;
+  let usage: LlmUsageObservation | null = null;
+  const warnings: string[] = [];
+
+  for await (const part of fullStream) {
+    const p = part as Record<string, unknown>;
+    const type = p["type"] as string;
+
+    if (type === "text-delta") {
+      const delta = p["textDelta"] as string;
+      if (delta !== "") {
+        firstTokenMs ??= Math.round(performance.now() - startMs);
+        outputText += delta;
+      }
+    } else if (type === "finish") {
+      completed = true;
+      finishReason = (p["finishReason"] as string) ?? null;
+      rawFinishReason = (p["rawFinishReason"] as string | undefined) ?? null;
+
+      const totalUsage = p["totalUsage"] as
+        | undefined
+        | { inputTokens: number; outputTokens: number; totalTokens: number };
+      const partUsage = p["usage"] as undefined | { inputTokens: number; outputTokens: number };
+
+      usage = {
+        inputTokens: totalUsage?.inputTokens ?? partUsage?.inputTokens ?? 0,
+        outputTokens: totalUsage?.outputTokens ?? partUsage?.outputTokens ?? 0,
+        totalTokens: totalUsage?.totalTokens ?? (partUsage?.inputTokens ?? 0) + (partUsage?.outputTokens ?? 0),
+      };
+    } else if (type === "error") {
+      const err = p["error"] as Error | undefined;
+      warnings.push(err?.message ?? "stream error");
+    }
+  }
+
+  return {
+    finishReason,
+    http,
+    mode,
+    model,
+    outputText: outputText || null,
+    provider,
+    rawFinishReason,
+    stream: { completed, firstTokenMs } satisfies LlmStreamObservation,
+    usage,
+    warnings,
+  };
+}