niceeval 0.6.1 → 0.6.2

package/src/o11y/parsers/openclaw.ts

@@ -0,0 +1,310 @@
+// OpenClaw transcript / `agent --json` 解析器。OpenClaw 方言只住这里,不进 core
+// (契约见 docs/feature/adapters/sdk/openclaw/README.md:字段事实以真实 CLI 与 transcript
+// fixture 固定为准;fixture 未证明的形状这里只做防御式解析,认不出的行标 parseSuccess=false,
+// 绝不从最终文本猜测工具轨迹)。
+//
+// 行为轨两个来源(collection.md 的通道优先级):
+//   1. session transcript JSONL(~/.openclaw/agents/**/sessions/*.jsonl,pi-agent 系
+//      消息格式:role/content parts,toolCall/toolResult 按 call ID 配对)→ parseOpenClawTranscript;
+//   2. `openclaw agent --json` 的 stdout 结果封包(最终回复、session key、失败与 usage 摘要,
+//      无完整工具轨迹)→ parseOpenClawRunJson,只在 transcript 拿不到时兜底。
+
+import type { StreamEvent, Usage, ToolName, JsonValue } from "../../types.ts";
+import type { ParsedTranscript } from "./index.ts";
+import { GENERIC_VERB_ALIASES, normalizeToolName as normalizeShared } from "../tool-names.ts";
+
+/**
+ * OpenClaw 特有工具别名(CLI 自己的 transcript 词汇,不会撞用户域名,裸动词可 opt-in)
+ * + 通用裸动词基表。
+ */
+export const OPENCLAW_TOOL_ALIASES: Record<string, ToolName> = {
+  ...GENERIC_VERB_ALIASES,
+  read: "file_read",
+  write: "file_write",
+  edit: "file_edit",
+  process: "shell",
+  browser: "web_fetch",
+};
+
+function normalizeToolName(name: string): ToolName {
+  return normalizeShared(name, OPENCLAW_TOOL_ALIASES);
+}
+
+function get(obj: unknown, key: string): unknown {
+  return obj && typeof obj === "object" ? (obj as Record<string, unknown>)[key] : undefined;
+}
+
+function str(v: unknown): string | undefined {
+  return typeof v === "string" && v ? v : undefined;
+}
+
+function num(obj: unknown, ...keys: string[]): number {
+  for (const k of keys) {
+    const v = get(obj, k);
+    if (typeof v === "number" && Number.isFinite(v)) return v;
+  }
+  return 0;
+}
+
+function coerceArgs(value: unknown): JsonValue {
+  if (typeof value === "string") {
+    try {
+      return JSON.parse(value) as JsonValue;
+    } catch {
+      return value;
+    }
+  }
+  return (value ?? {}) as JsonValue;
+}
+
+/** content parts / 字符串 → 纯文本(text 类 part 拼接)。 */
+function extractText(content: unknown): string {
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    const parts: string[] = [];
+    for (const b of content) {
+      if (typeof b === "string") parts.push(b);
+      else {
+        const t = get(b, "text") ?? get(b, "content");
+        if (typeof t === "string" && t) parts.push(t);
+      }
+    }
+    return parts.join("\n");
+  }
+  return "";
+}
+
+/**
+ * OpenClaw session transcript(JSONL,pi-agent 系消息格式)→ 标准事件流 + 用量 + 压缩计数。
+ * assistant 消息的 text / thinking / toolCall parts → message / thinking / action.called,
+ * toolResult 消息按 toolCallId 配对成 action.result(isError → failed),消息级 usage 逐条
+ * 累加,compaction 条目计入压缩次数。认不出 / 解析失败的行标 `parseSuccess: false` 并跳过,
+ * 不猜测——transcript 不完整时负断言不可信(契约见 sdk/openclaw/README.md)。
+ */
+export function parseOpenClawTranscript(raw: string | undefined): ParsedTranscript {
+  const events: StreamEvent[] = [];
+  let inputTokens = 0;
+  let outputTokens = 0;
+  let cacheReadTokens = 0;
+  let cacheWriteTokens = 0;
+  let costUSD = 0;
+  let requests = 0;
+  let compactions = 0;
+  let parseSuccess = true;
+
+  if (!raw || !raw.trim()) {
+    return { events, usage: { inputTokens: 0, outputTokens: 0 }, compactions: 0, parseSuccess: true };
+  }
+
+  let synth = 0;
+  const nextSynthId = (): string => `oc_${++synth}`;
+
+  const addUsage = (usage: unknown): void => {
+    if (!usage || typeof usage !== "object") return;
+    // pi 系简写(input/output/cacheRead/cacheWrite/cost.total)与 snake_case 变体都认。
+    const input = num(usage, "input", "input_tokens", "inputTokens", "prompt_tokens");
+    const output = num(usage, "output", "output_tokens", "outputTokens", "completion_tokens");
+    const cacheRead = num(usage, "cacheRead", "cache_read_input_tokens", "cacheReadTokens");
+    const cacheWrite = num(usage, "cacheWrite", "cache_creation_input_tokens", "cacheWriteTokens");
+    const cost = num(usage, "cost") || num(get(usage, "cost"), "total");
+    if (input === 0 && output === 0 && cacheRead === 0 && cacheWrite === 0 && cost === 0) return;
+    inputTokens += input;
+    outputTokens += output;
+    cacheReadTokens += cacheRead;
+    cacheWriteTokens += cacheWrite;
+    costUSD += cost;
+    requests += 1;
+  };
+
+  for (const line of raw.split("\n")) {
+    const trimmed = line.trim();
+    if (!trimmed) continue;
+
+    let entry: unknown;
+    try {
+      entry = JSON.parse(trimmed);
+    } catch {
+      parseSuccess = false;
+      continue;
+    }
+
+    try {
+      // 会话条目可能带 { type: "message", message: {...} } 包装,也可能就是消息本体。
+      const wrappedMessage = get(entry, "message");
+      const entryType = get(entry, "type");
+      const msg =
+        wrappedMessage && typeof wrappedMessage === "object" ? wrappedMessage : entry;
+      const role = get(msg, "role");
+
+      if (entryType === "compaction" || entryType === "compact" || get(entry, "kind") === "compaction") {
+        compactions += 1;
+        events.push({ type: "compaction", reason: str(get(entry, "reason")) });
+        continue;
+      }
+      if (entryType === "error" && role === undefined) {
+        events.push({ type: "error", message: String(get(entry, "message") ?? get(entry, "error") ?? "error") });
+        continue;
+      }
+
+      if (role === "assistant") {
+        addUsage(get(msg, "usage"));
+        const content = get(msg, "content");
+        if (typeof content === "string") {
+          if (content) events.push({ type: "message", role: "assistant", text: content });
+          continue;
+        }
+        if (!Array.isArray(content)) continue;
+        for (const part of content) {
+          const partType = get(part, "type");
+          if (partType === "text") {
+            const text = str(get(part, "text"));
+            if (text) events.push({ type: "message", role: "assistant", text });
+          } else if (partType === "thinking" || partType === "reasoning") {
+            const text = str(get(part, "thinking")) ?? str(get(part, "reasoning")) ?? str(get(part, "text"));
+            if (text) events.push({ type: "thinking", text });
+          } else if (partType === "toolCall" || partType === "tool_call" || partType === "tool-call" || partType === "toolUse") {
+            const name = String(get(part, "name") ?? get(part, "toolName") ?? "unknown");
+            const callId = str(get(part, "id")) ?? str(get(part, "toolCallId")) ?? str(get(part, "tool_call_id")) ?? nextSynthId();
+            events.push({
+              type: "action.called",
+              callId,
+              name,
+              input: coerceArgs(get(part, "arguments") ?? get(part, "args") ?? get(part, "input")),
+              tool: normalizeToolName(name),
+            });
+          }
+        }
+        continue;
+      }
+
+      if (role === "user") {
+        const text = extractText(get(msg, "content") ?? get(msg, "text"));
+        if (text) events.push({ type: "message", role: "user", text });
+        continue;
+      }
+
+      if (role === "toolResult" || role === "tool") {
+        const callId =
+          str(get(msg, "toolCallId")) ?? str(get(msg, "tool_call_id")) ?? str(get(msg, "id")) ?? nextSynthId();
+        const isError = get(msg, "isError") === true || get(msg, "is_error") === true;
+        const output = get(msg, "output") ?? get(msg, "result") ?? get(msg, "content");
+        events.push({
+          type: "action.result",
+          callId,
+          output: (typeof output === "string" ? output : (extractText(output) || output)) as JsonValue,
+          status: isError ? "failed" : "completed",
+        });
+        continue;
+      }
+      // 其它条目(system prompt、订阅元数据等):无对应 StreamEvent。
+    } catch {
+      parseSuccess = false;
+    }
+  }
+
+  const usage: Usage = { inputTokens, outputTokens };
+  if (cacheReadTokens > 0) usage.cacheReadTokens = cacheReadTokens;
+  if (cacheWriteTokens > 0) usage.cacheWriteTokens = cacheWriteTokens;
+  if (requests > 0) usage.requests = requests;
+  if (costUSD > 0) usage.costUSD = costUSD;
+
+  return { events, usage, compactions, parseSuccess };
+}
+
+/** 便捷形态:只要 StreamEvent[]。 */
+export function parseOpenClaw(raw: string | undefined): StreamEvent[] {
+  return parseOpenClawTranscript(raw).events;
+}
+
+/** `openclaw agent --json` 结果封包里抠出的摘要(无完整工具轨迹,只兜底)。 */
+export interface OpenClawRunJson {
+  /** 最终回复文本(result.text / payloads[].text / reply 变体)。 */
+  text?: string;
+  /** 本次 run 的 session key(首轮 capture、后续 resume 用)。 */
+  sessionId?: string;
+  /** 封包报告的用量;没报就省略,不编造。 */
+  usage?: Usage;
+  /** 封包自报失败(error 字段 / status=error|failed / success=false)。 */
+  failed: boolean;
+}
+
+/**
+ * `openclaw agent --json` 的 stdout → 结果摘要。整段 stdout 先按单个 JSON 文档解析
+ * (含 pretty-print 多行),失败再逐行扫最后一个 JSON 对象(混了日志行的场景)。
+ * 完全解析不出时返回空摘要(`failed: false`)——失败判定交给调用方的 exitCode。
+ */
+export function parseOpenClawRunJson(stdout: string | undefined): OpenClawRunJson {
+  const doc = parseJsonDocument(stdout);
+  if (!doc) return { failed: false };
+
+  const result = get(doc, "result");
+  const textFromPayloads = (payloads: unknown): string | undefined => {
+    if (!Array.isArray(payloads)) return undefined;
+    const texts = payloads.map((p) => str(get(p, "text"))).filter((t): t is string => !!t);
+    return texts.length ? texts.join("\n") : undefined;
+  };
+  const text =
+    str(get(result, "text")) ??
+    textFromPayloads(get(result, "payloads")) ??
+    textFromPayloads(get(doc, "payloads")) ??
+    str(get(doc, "text")) ??
+    str(get(doc, "reply")) ??
+    str(get(doc, "message"));
+
+  const sessionId =
+    str(get(doc, "sessionId")) ??
+    str(get(doc, "session_id")) ??
+    str(get(doc, "sessionKey")) ??
+    str(get(result, "sessionId")) ??
+    str(get(result, "session_id"));
+
+  const rawUsage = get(doc, "usage") ?? get(result, "usage") ?? get(get(doc, "meta"), "usage");
+  let usage: Usage | undefined;
+  if (rawUsage && typeof rawUsage === "object") {
+    const input = num(rawUsage, "input", "input_tokens", "inputTokens", "prompt_tokens");
+    const output = num(rawUsage, "output", "output_tokens", "outputTokens", "completion_tokens");
+    if (input > 0 || output > 0) {
+      usage = { inputTokens: input, outputTokens: output };
+      const cacheRead = num(rawUsage, "cacheRead", "cache_read_input_tokens", "cacheReadTokens");
+      if (cacheRead > 0) usage.cacheReadTokens = cacheRead;
+      const cost = num(rawUsage, "cost") || num(get(rawUsage, "cost"), "total");
+      if (cost > 0) usage.costUSD = cost;
+    }
+  }
+
+  const status = get(doc, "status");
+  const failed =
+    get(doc, "error") != null ||
+    get(doc, "success") === false ||
+    (typeof status === "string" && /^(error|failed)$/i.test(status));
+
+  return {
+    ...(text !== undefined ? { text } : {}),
+    ...(sessionId !== undefined ? { sessionId } : {}),
+    ...(usage !== undefined ? { usage } : {}),
+    failed,
+  };
+}
+
+function parseJsonDocument(stdout: string | undefined): Record<string, unknown> | undefined {
+  if (!stdout || !stdout.trim()) return undefined;
+  const trimmed = stdout.trim();
+  try {
+    const doc = JSON.parse(trimmed) as unknown;
+    if (doc && typeof doc === "object" && !Array.isArray(doc)) return doc as Record<string, unknown>;
+  } catch {
+    // 混日志行:逐行找最后一个完整 JSON 对象
+  }
+  const lines = trimmed.split("\n").reverse();
+  for (const line of lines) {
+    const l = line.trim();
+    if (!l.startsWith("{") || !l.endsWith("}")) continue;
+    try {
+      const doc = JSON.parse(l) as unknown;
+      if (doc && typeof doc === "object") return doc as Record<string, unknown>;
+    } catch {
+      // 继续往上找
+    }
+  }
+  return undefined;
+}