ppxans-harness 2.4.0

package/config/ppx.json

@@ -0,0 +1,151 @@
+{
+    "agent":  {
+                  "name":  "皮皮虾",
+                  "yuan":  "ppx",
+                  "max_tool_rounds":  8,
+                  "tool_result_budget":  4000,
+                  "max_tool_error_retry":  2,
+                  "proactive":  {
+                                    "enabled":  true,
+                                    "interval_ms":  3600000
+                                },
+                  "evolve":  {
+                                 "enabled":  true,
+                                 "every_calls":  20,
+                                 "min_interval_ms":  30000
+                             }
+              },
+    "user":  {
+                 "name":  "兄弟"
+             },
+    "providers":  [
+                      {
+                          "id":  "dsh",
+                          "backend":  "deepseek",
+                          "dsh_root":  ".deps/deepseek-harness",
+                          "timeout_ms":  180000
+                      },
+                      {
+                          "id":  "openai",
+                          "base_url":  "https://api.openai.com/v1",
+                          "api_key_env":  "OPENAI_API_KEY",
+                          "model":  "gpt-4o-mini",
+                          "timeout_ms":  180000,
+                          "context_window":  128000
+                      },
+                      {
+                          "id":  "deepseek",
+                          "base_url":  "https://api.deepseek.com/v1",
+                          "api_key_env":  "DEEPSEEK_API_KEY",
+                          "model":  "deepseek-chat",
+                          "timeout_ms":  180000,
+                          "context_window":  65536
+                      },
+                      {
+                          "id":  "dashscope",
+                          "base_url":  "https://dashscope.aliyuncs.com/compatible-mode/v1",
+                          "api_key_env":  "DASHSCOPE_API_KEY",
+                          "model":  "qwen-turbo",
+                          "timeout_ms":  180000,
+                          "context_window":  131072
+                      },
+                      {
+                          "id":  "qwen-vl",
+                          "base_url":  "https://dashscope.aliyuncs.com/compatible-mode/v1",
+                          "api_key_env":  "DASHSCOPE_API_KEY",
+                          "model":  "qwen-vl-max",
+                          "vision":  true,
+                          "timeout_ms":  180000,
+                          "context_window":  32768
+                      },
+                      {
+                          "id":  "lmstudio",
+                          "base_url":  "http://127.0.0.1:1234/v1",
+                          "api_key":  "lm-studio",
+                          "model":  "gemma-4-e2b-uncensored-hauhaucs-aggressive-q8_k_p",
+                          "vision":  false,
+                          "timeout_ms":  180000
+                      },
+                      {
+                          "id":  "zhipu",
+                          "base_url":  "https://open.bigmodel.cn/api/paas/v4",
+                          "api_key_env":  "ZHIPU_API_KEY",
+                          "model":  "glm-5v-turbo",
+                          "vision":  true,
+                          "timeout_ms":  180000,
+                          "context_window":  65536
+                      }
+                  ],
+    "_optional_engines":  {
+                              "_note":  "可选引擎。火山引擎 volcengine 需先填真实 endpoint(model 不能留 REPLACE_WITH_YOUR_ENDPOINT 占位), 把它加回 providers 数组并设 VOLCENGINE_API_KEY 即启用；本地引擎(openclaw/dsh)默认不启用, 要用时把对象移入 providers 并设对应路径环境变量。",
+                              "openclaw":  {
+                                               "id":  "openclaw",
+                                               "backend":  "openclaw",
+                                               "mjs":  "",
+                                               "session_key":  "ppx:main",
+                                               "timeout_ms":  180000
+                                           }
+                          },
+    "memory":  {
+                   "enabled":  true,
+                   "token_budget":  2500,
+                   "decay_per_day":  0.02,
+                   "hit_bonus":  5,
+                   "base_importance":  10,
+                   "compile_threshold":  4.5,
+                   "forget_speed":  1,
+                   "memory_ttl_days":  90
+               },
+    "audit":  {
+                  "enabled":  true,
+                  "_note":  "工具调用审计哈希链 (吸收自 ppx-v2)。append-only + SHA-256 链式防篡改, 落 data/logs/audit.ndjson。设 false 可关闭 (性能敏感场景)。校验: ppx 对话里调 audit_verify 工具, 或 npm run audit:verify。"
+              },
+    "embedding":  {
+                      "base_url":  "http://127.0.0.1:1234/v1",
+                      "api_key":  "lm-studio",
+                      "model":  "text-embedding-nomic-embed-text-v1.5"
+                  },
+    "experience":  {
+                       "enabled":  true
+                   },
+    "selfheal":  {
+                     "enabled":  true,
+                     "check_interval_ms":  60000
+                 },
+    "tools":  {
+                  "enabled":  true
+              },
+    "channels":  {
+                     "http":  {
+                                  "enabled":  true,
+                                  "port":  8899,
+                                  "auth_token":  ""
+                              },
+                     "log":  {
+                                 "enabled":  true,
+                                 "target":  "console"
+                             },
+                     "feishu":  {
+                                    "enabled":  false,
+                                    "appId":  "",
+                                    "appSecret":  "",
+                                    "verifyToken":  ""
+                                },
+                     "wechat":  {
+                                    "enabled":  false,
+                                    "path":  "/wechat/webhook",
+                                    "token":  "",
+                                    "encodingAESKey":  "",
+                                    "corpId":  "",
+                                    "corpSecret":  "",
+                                    "agentId":  ""
+                                }
+                 },
+    "security":  {
+                     "allow_all":  false,
+                     "command_timeout_ms":  30000,
+                     "deny":  [
+
+                              ]
+                 }
+}

package/package.json

@@ -0,0 +1,69 @@
+{
+  "name": "ppxans-harness",
+  "version": "2.4.0",
+  "description": "PPXANS-Harness - 皮皮虾神经系 (ANS) + Harness 一体化智能体内核. 零运行时依赖纯Node. 自愈 + 自学习 + 五层记忆 + SHA-256 审计哈希链 + 多 Agent 军团 + 治理内核(deny-wins/熔断/seam) + 进化内核(Playbook/故障记忆) + 符号画布 + supervisor. Agent Nervous System harness in pure Node, zero runtime dependencies.",
+  "type": "module",
+  "license": "Apache-2.0",
+  "author": "chen6896qqwee <chen6896qqwee@users.noreply.github.com>",
+  "engines": {
+    "node": ">=20"
+  },
+  "main": "src/agent/index.js",
+  "bin": {
+    "ppx": "bin/ppx.js",
+    "ppxans": "bin/ppx.js",
+    "ppx-serve": "bin/ppx-serve.js",
+    "ppx-channels": "bin/ppx-channels.js"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/chen6896qqwee/PPXANS-Harness.git"
+  },
+  "homepage": "https://github.com/chen6896qqwee/PPXANS-Harness",
+  "bugs": {
+    "url": "https://github.com/chen6896qqwee/PPXANS-Harness/issues"
+  },
+  "scripts": {
+    "start": "node src/agent/index.js",
+    "chat": "node src/cli.js",
+    "serve": "node src/server.js",
+    "web": "node scripts/start-web.js",
+    "web:build": "npm run build --prefix web",
+    "release": "node scripts/release.js",
+    "eval": "node scripts/eval.js",
+    "dsh": "cd .deps/deepseek-harness && pnpm dsh",
+    "dsh:install": "cd .deps/deepseek-harness && pnpm install",
+    "dsh:build": "cd .deps/deepseek-harness && pnpm run build",
+    "selfheal": "node scripts/selfheal-bench.js",
+    "selfheal:run": "node src/selfheal/run.js",
+    "audit:verify": "node scripts/audit-verify.js",
+    "bench": "node scripts/bench.js",
+    "test": "node --test --test-force-exit test/*.test.js",
+    "prepublishOnly": "npm run selfheal && npm test"
+  },
+  "keywords": [
+    "agent",
+    "ai",
+    "llm",
+    "memory",
+    "autonomous",
+    "self-healing",
+    "zero-dependency",
+    "audit-chain",
+    "memory-governance",
+    "agent-nervous-system"
+  ],
+  "files": [
+    "bin/",
+    "src/",
+    "config/",
+    "README.md",
+    "LICENSE"
+  ],
+  "exports": {
+    ".": "./src/agent/index.js",
+    "./server": "./src/server.js",
+    "./cli": "./src/cli.js",
+    "./audit": "./src/audit/audit-chain.js"
+  }
+}

package/src/agent/context.js

@@ -0,0 +1,179 @@
+// src/agent/context.js - Agent 历史/上下文管理 (从 index.js 拆分, mixin 挂回 prototype)
+// 重构第三刀后续 (2026-09-15): 历史裁剪/token 预算/会话压缩从 PPXAgent 类中抽出,
+// 方法以 mixin 方式挂回 prototype, 实例行为与调用方完全不变 (测试走 agent._xxx 不受影响)。
+
+import { estimateTokens } from "../utils/text.js";
+import { transcriptToText, buildCompactionMessages } from "../memory/compaction.js";
+
+// 辅助 LLM 调用短超时 (压缩/提炼等非主对话调用):
+// 模型不可用/网络不通时快速失败降级, 避免阻塞主对话
+const AUX_LLM_TIMEOUT_MS = 10000;
+// 上下文窗口感知: 未知窗口的保守默认 (绝不放大历史) + 历史占用窗口的安全比例上限
+const DEFAULT_CONTEXT_WINDOW = 8192;
+const DEFAULT_CONTEXT_RATIO = 0.6;
+
+export const contextMethods = {
+  // ---- 多轮会话历史 (吸收 dsh "会话即事实源") ----
+  // 历史从事件日志投影, 再按预算裁剪 (裁剪只发生在投影层, 日志本身不可变)
+  // v0.6.6 优化: 信息量感知裁剪 (学自 Claude Code Microcompact 思路)
+  //   旧版: 纯按条数硬截 + 尾部 token 预算, 可能裁掉关键决策/工具结果轮次
+  //   新版: 优先保留"含关键信息"的轮次(指令/数字/路径/结论/工具结果), 纯寒暄让位
+  _historyPriority(m) {
+    const s = String(m?.content || "");
+    if (!s) return 0;
+    let p = 0;
+    // 长消息(含工具结果/详细决策)权重高
+    if (s.length > 120) p += 2;
+    // 含指令/结论/数字/路径/文件等关键信号
+    if (/[查|算|计算|写|建|改|创建|删除|修复|总结|分析|设置|配置|执行|运行|启动|停止|提交|部署|安装|生成|编译|测试]/.test(s)) p += 2;
+    if (/[0-9]{2,}|[%.¥$元%]|[:：][0-9]/.test(s)) p += 1;
+    if (/[A-Za-z]:[\\\/]|\.(js|py|md|json|txt|ts|go|rs|cpp|java|log)\b/.test(s)) p += 2;
+    if (/失败|错误|报错|异常|成功|完成|结果|结论|决定|方案|建议/.test(s)) p += 2;
+    // 纯寒暄/简短确认权重低
+    if (/^(你好|hi|hello|在吗|谢谢|好的|ok|嗯|是的|对|收到|知道|了解|再见|拜拜)/i.test(s.trim())) p -= 3;
+    return p;
+  },
+
+  // provider 上下文窗口 -> 历史 token 预算硬上限:
+  // 用 llm.context_window(未配置回退 config.memory.context_window) * 安全比例, 与显式预算取小。
+  // 目的: 本地小上下文模型即使没配 history_token_budget, 也不会把历史塞爆窗口。
+  _histTokenCap() {
+    const window = Number(this.llm?.context_window) || Number(this.config?.memory?.context_window) || DEFAULT_CONTEXT_WINDOW;
+    const ratio = Number(this.config?.memory?.context_window_ratio) || DEFAULT_CONTEXT_RATIO;
+    return Math.max(200, Math.floor(window * ratio));
+  },
+
+  // 会话历史裁剪 (中心函数): 条数上限 + token 预算, 信息量感知, 必保最近一条。
+  // opts.budget / opts.maxItems 可覆盖 (溢出降档重试时传更紧预算)。
+  // token 预算 = min(显式 history_token_budget, 窗口硬上限), 两者都收紧, 取小者。
+  _trimHistory(hist, opts = {}) {
+    let h = [...hist];
+    const maxItems = opts.maxItems != null ? opts.maxItems : (Number(this.config.memory?.max_history_items) || 40);
+    const cfgBudget = Number(this.config.memory?.history_token_budget) || 4000;
+    const tokenBudget = opts.budget != null ? opts.budget : Math.min(cfgBudget, this._histTokenCap());
+    // 1) 条数上限: 超限时按信息量淘汰 (低信息量优先, 从旧到新)
+    if (h.length > maxItems) {
+      const scored = h.map((m, i) => ({ m, i, p: this._historyPriority(m) }));
+      scored.sort((a, b) => (a.p - b.p) || (a.i - b.i));
+      const drop = scored.length - maxItems;
+      const dropped = new Set(scored.slice(0, drop).map((x) => x.i));
+      scored.sort((a, b) => a.i - b.i);
+      h = scored.filter((x) => !dropped.has(x.i)).map((x) => x.m);
+    }
+    // 2) token 预算: 信息量感知裁剪 (替代旧版"丢最旧前缀")
+    //    必保最近一条, 其余按信息量从高到低补足, 低信息量轮次让位
+    let total = h.reduce((a, m) => a + estimateTokens(m.content), 0);
+    if (total > tokenBudget) {
+      const keep = new Set();
+      let used = 0;
+      const lastIdx = h.length - 1;
+      keep.add(lastIdx); used += estimateTokens(h[lastIdx].content);
+      const rest = h.slice(0, lastIdx)
+        .map((m, i) => ({ m, i, p: this._historyPriority(m) }))
+        .sort((a, b) => (b.p - a.p) || (a.i - b.i));
+      for (const { m, i } of rest) {
+        const t = estimateTokens(m.content);
+        if (used + t > tokenBudget) continue;
+        keep.add(i); used += t;
+      }
+      h = h.filter((_, i) => keep.has(i));
+    }
+    return h;
+  },
+
+  // 绝对硬裁剪兜底 (第九轮 review P1: 不依赖 LLM 也能保证历史放得下):
+  // 在 _trimHistory 基础上再加一道绝对底线 — 超条数则保留最近 N 轮,
+  // 超 token 则从最新向前贪心保留到预算内 (最近信息优先, 必保最后一条)。
+  // 即便 config 异常 (预算极大/极小的模型), 注入的历史也不会超过窗口安全比例。
+  _ensureContextFit(hist, { budget = this._histTokenCap(), maxItems } = {}) {
+    let h = [...hist];
+    const itemCap = maxItems != null ? maxItems : (Number(this.config.memory?.max_history_items) || 40);
+    // 环节 1: 条数绝对兜底 — 超上限只留最近 itemCap 条
+    if (h.length > itemCap) h = h.slice(-itemCap);
+    // 环节 2: token 绝对兜底 — 最近优先贪心直到塞满预算 (必保最后一条)
+    let total = h.reduce((a, m) => a + estimateTokens(m.content), 0);
+    if (total > budget && h.length) {
+      const keep = [];
+      let used = 0;
+      for (let i = h.length - 1; i >= 0; i--) {
+        const t = estimateTokens(h[i].content);
+        // 至少保留最后一条 (当前对话), 其余超预算跳过
+        if (keep.length === 0 || used + t <= budget) {
+          keep.unshift(h[i]); used += t;
+        }
+      }
+      h = keep;
+    }
+    return h;
+  },
+
+  // 溢出降档: 把已组好的消息数组按「更紧历史预算」重建 (消息完整性安全版)。
+  //  - 保留全部 system (角色/记忆/经验)
+  //  - 自最后一条 user 起的一切消息原样保留 (含 in-flight 的 assistant tool_calls + tool 配对,
+  //    绝不剪切成"孤立的 tool 消息"导致 API 400)
+  //  - 只对最后一条 user 之前的旧历史做最近优先硬裁剪到 budget 内
+  _shrinkMessagesForOverflow(messages, budget) {
+    let i = 0;
+    while (i < messages.length && messages[i] && messages[i].role === "system") i++;
+    const systems = messages.slice(0, i);
+    const rest = messages.slice(i);
+    if (!rest.length) return messages; // 只有 system, 无裁剪空间, 原样返回
+    // 从后向前找最后一条 user 作为「进行中单元」起点 (含其后的 tool 配对)
+    let lastUser = rest.length - 1;
+    while (lastUser > 0 && rest[lastUser].role !== "user") lastUser--;
+    const tail = rest.slice(lastUser);          // 完整保留 (结束于 user 或 in-flight 工具单元)
+    const mid = rest.slice(0, lastUser);        // 仅剪这里的历史
+    const itemCap = Math.max(2, Number(this.config.memory?.max_history_items) || 40);
+    const trimmed = this._ensureContextFit(mid, { budget, maxItems: itemCap });
+    return [...systems, ...trimmed, ...tail];
+  },
+
+  _getSession(sessionKey) {
+    // 先信息量感知裁剪, 再 + 绝对硬兜底: 即便 config 异常/压缩失败, 历史也放得下
+    const raw = this.sessionStore.deriveCompacted(sessionKey || "default");
+    return this._ensureContextFit(this._trimHistory(raw));
+  },
+
+  // 追加一轮对话为不可变事件 (append-only, 永不重写日志)
+  // v1.1.1: user+assistant 一次批量落盘 (skipFlush), 一轮对话只写一次磁盘而非两次
+  _pushTurn(sessionKey, userMsg, assistant) {
+    const k = sessionKey || "default";
+    this.sessionStore.append(k, "user/message", { content: String(userMsg) }, Date.now(), { skipFlush: true });
+    if (assistant) this.sessionStore.append(k, "assistant/message", { content: String(assistant) }, Date.now(), { skipFlush: true });
+    this.sessionStore.flush(k);
+  },
+
+  // 加载历史: 先尝试结构化压缩(超阈值), 再按预算裁剪
+  async _loadHistory(sessionKey) {
+    const k = sessionKey || "default";
+    await this._maybeCompact(k);
+    return this._getSession(k).map((m) => ({ ...m }));
+  },
+
+  // 会话压缩: 未压缩部分超 token 阈值时, 把最旧一半压成结构化摘要并持久化到日志
+  // (吸收 OpenClaw compaction: 摘要替换被压缩区间, 日志本身不可变)
+  async _maybeCompact(sessionKey) {
+    if (!this.llm) return;
+    if (!(await this._auxLlmReady())) return; // 模型不可用时跳过压缩 (交给 _trimHistory 硬裁剪)
+    const events = this.sessionStore.replay(sessionKey);
+    let upToSeq = 0;
+    for (const e of events) if (e.type === "compaction/summary") upToSeq = e.data?.upToSeq || 0;
+    const tail = events.filter((e) => e.seq > upToSeq && (e.type === "user/message" || e.type === "assistant/message"));
+    if (!tail.length) return;
+    const tokenBudget = Number(this.config.memory?.history_token_budget) || 4000;
+    const total = tail.reduce((a, e) => a + estimateTokens(e.data?.content), 0);
+    if (total <= tokenBudget * 1.5) return; // 未超阈值不压缩
+    const split = Math.floor(tail.length / 2);
+    const old = tail.slice(0, split);
+    if (old.length < 2) return; // 太少不值得压
+    const lastSeq = old[old.length - 1].seq;
+    const transcript = transcriptToText(old.map((e) => ({ role: e.type === "user/message" ? "user" : "assistant", content: e.data?.content })));
+    try {
+      const r = await this.llm.chat(buildCompactionMessages(transcript), { timeoutMs: AUX_LLM_TIMEOUT_MS, retryMax: 0 });
+      const summary = r?.content;
+      if (summary) this.sessionStore.append(sessionKey, "compaction/summary", { summary, upToSeq: lastSeq });
+    } catch {
+      // 压缩失败静默降级, 交给 _trimHistory 硬裁剪
+    }
+  }
+};