@bolloon/bolloon-agent 0.2.15 → 0.3.1

package/dist/llm/system-prompt/layers/tool/goal_handoff.md

@@ -0,0 +1,77 @@
+---
+added_at: 2026-07-10
+last_reviewed_at: 2026-07-10
+ttl_days: 270
+author: yuanjie
+---
+
+<!-- tool.goal_handoff@1.0.0 -->
+# Goal Handoff 工具 (park_goal / resume_goal / continue_goal_background)
+
+**目标接力 3 件套** — 切 channel / 切用户 / 切对端时, 保持"目标不中断".
+
+## 何时用 (决策树)
+
+```
+当前 task 还在跑, 但要切换上下文
+  ├─ 切到另一个 channel (用户主动 / UI 切)
+  │    → park_goal(reason='channel_switch')
+  │    → 切完调 resume_goal
+  │
+  ├─ 用户几小时没回来 (你被 hook 启动做后台)
+  │    → park_goal(reason='user_away') — 如果之前还没 park
+  │    → resume_goal({ newSession: true }) 在新 channel 续
+  │
+  ├─ 等对端 peer 回应 (>1min 没音讯)
+  │    → park_goal(reason='awaiting_external')
+  │    → 不主动 resume, 等对端 ack
+  │
+  └─ 主动把目标推到对端 (任务大 / 想分工)
+       → continue_goal_background(peer_did, p2pSendMessage)
+       (内部已含 park, 推完不返回)
+```
+
+## park_goal 必传参数
+
+```typescript
+{
+  goalRef: {
+    goalId: string,        // 已存在 (从 session metadata 读) 或新生成
+    targetId: string,      // ⚠️ 用户视角的稳定描述, 不允许 "the task"
+    createdBy: 'user' | 'agent' | 'peer',
+    createdAt: ISO 字符串,
+    originChannel: string, // 当前 session id
+  },
+  reason: 'channel_switch' | 'user_away' | 'awaiting_external' | 'peer_handoff',
+}
+```
+
+返回 GoalHandle: `{ goalId, targetId, state: 'parked', taskId?, error? }`.
+`error` 字段 = 不抛错, 静默记录到 `goal-parked.jsonl`, 允许 LLM 继续响应.
+
+## resume_goal 必传参数
+
+```typescript
+resumeGoal(goalId: string, {
+  newSession?: boolean,   // true = 在新 session key 下续, 旧 session 保留
+  channelId?: string,     // 指定 channelId 恢复 (默认 = originChannel)
+})
+```
+
+恢复过程: 加载末 30 条消息 → 写回 session → 关联 task status 改 'running' →
+触发 `onGoalResumed` 写 `goal-resumed.jsonl`.
+
+## continue_goal_background 注意
+
+推给对端时**内部已过滤**隐私 (judgment / persona 不发), LLM 不必再过滤.
+但 LLM **应该**确认:
+
+- 对端节点**在线** (先 `list_peers` 看)
+- 对端**有足够 context** (同 `core.identity` layer, 共享人格)
+- 任务**可分解** (不要推一坨未拆解的大任务)
+
+## 失败兜底
+
+- park 失败 → 不阻塞切换, 静默 warn
+- resume 找不到 goalId → 返回 `{ error: 'goal X 未找到' }`, LLM 应该给用户解释
+- 跨机器 continue 失败 (P2P outbox 满) → 自动入 outbox 重试, 标 `state: 'continued_background'`

package/dist/llm/system-prompt/layers/tool/p2p_request.md

@@ -0,0 +1,61 @@
+---
+added_at: 2026-07-10
+last_reviewed_at: 2026-07-10
+ttl_days: 270
+author: yuanjie
+---
+
+<!-- tool.p2p_request@1.0.0 -->
+# P2P 工具集 (list_peers / send_message / broadcast / send_to_channel / check_inbox / agent_call)
+
+P2P 工具的**语义区别** — 选错了会污染对端或浪费 token.
+
+## 决策树 (按场景选)
+
+```
+想 "知道有哪些节点在线"
+  → list_peers
+
+想 "给某节点发一条短消息 (不建 channel)"
+  → send_message(peer_id, message)
+    e.g. 问对方是否接任务 / 通知进展 / 简单寒暄
+
+想 "广播给所有节点"
+  → broadcast_message(message)
+    e.g. 广播"我刚发布了新版本 v0.2.16"
+    ⚠️ 慎用 — 每次广播每节点都收一条, token 消耗 = 节点数 × 消息长度
+
+想 "建一个长期 channel 与某节点多轮协作"
+  → send_to_channel(channel_id='', message, peer_did)
+    channel_id 留空 = 自动建; peer_did 绑定 = 后续消息通过 P2P 自动同步到对端
+    用 channel 的场景: 跨多轮的复杂协作 / 需要保留上下文 / 切换后还能找到
+
+想 "查看我收到的所有消息 (本地 + 远程)"
+  → check_inbox(max=50)
+    返回按时间倒序; 触发 onMessage hook 的消息都进 _inboxMessages
+
+想 "调对端 agent 执行一个完整任务 (含 LLM 推理)"
+  → agent_call(peer_did, task, options)
+    ⚠️ 对端会启动独立 LLM 轮次, 消耗对端 token; 你拿回的是结构化结果
+```
+
+## 离线和连接
+
+所有 send_* 工具**失败不报错给用户** — 自动入 outbox (`~/.bolloon/outbox/`),
+等连接恢复自动 flush. 工具返回 `{ success: false, error: "queued" }` 视为成功.
+
+如果你需要确认"对端真的收到了", 用 `check_inbox` 看是否有 ack 类型消息.
+
+## 隐私过滤 (LLM 必做)
+
+`send_message` / `send_to_channel` / `broadcast_message` 之前:
+
+- ❌ 不发 judgment 库内容 (调 `humanValueStore.list` 看 privacy 标签)
+- ❌ 不发 API key / 凭证 / 路径里的私密信息
+- ✅ 可发: targetId / 任务描述 / 公开文档摘要 / 代码片段
+- ⚠️ 摘要后发: 用户偏好 (改写为通用描述, 不带具体值)
+
+## 接收端注意
+
+`check_inbox` 拿到的不一定都是人类消息 — 也可能是对端 agent 调 `agent_call` 推过来的任务.
+判断: 消息 metadata 里的 `fromDid` / `peerName` 字段; 是 DID 形式 = agent, 人类名 = 人.

package/dist/llm/system-prompt/registry.js

@@ -91,7 +91,6 @@ const STATIC_LAYERS = [
     { id: 'core.tone', version: '1.0.0', priority: 110, appliesTo: ['all'], source: 'static-md', maxChars: 500, meta: DEFAULT_META(TTL_KNOWLEDGE) },
     { id: 'core.wellbeing', version: '1.0.0', priority: 120, appliesTo: ['all'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_KNOWLEDGE) },
     { id: 'core.evenhandedness', version: '1.0.0', priority: 130, appliesTo: ['all'], source: 'static-md', maxChars: 300, meta: DEFAULT_META(TTL_KNOWLEDGE) },
-    { id: 'core.memory_system', version: '1.0.0', priority: 140, appliesTo: ['all'], source: 'static-md', maxChars: 200, meta: DEFAULT_META(TTL_KNOWLEDGE) },
     { id: 'core.artifacts_storage', version: '1.0.0', priority: 145, appliesTo: ['never'], source: 'static-md', maxChars: 0, meta: DEFAULT_META(TTL_KNOWLEDGE) },
     { id: 'core.network_filesystem', version: '1.0.0', priority: 148, appliesTo: ['never'], source: 'static-md', maxChars: 0, meta: DEFAULT_META(TTL_KNOWLEDGE) },
     // ── role/ ── 阶段 0 只用 expert, 其他 3 个停用 (节省 + 阶段 0 不分 role)
@@ -103,6 +102,17 @@ const STATIC_LAYERS = [
     { id: 'channel.local', version: '1.0.0', priority: 150, appliesTo: ['local'], source: 'static-md', maxChars: 500, meta: DEFAULT_META(TTL_CHANNEL) },
     { id: 'channel.p2p-visitor', version: '1.0.0', priority: 150, appliesTo: ['p2p-visitor'], source: 'static-md', maxChars: 700 },
     { id: 'channel.p2p-agent', version: '1.0.0', priority: 150, appliesTo: ['p2p-agent'], source: 'static-md', maxChars: 700 },
+    // 2026-07-10 双栖 agent 网络新增 (按 plan 阶段 3):
+    { id: 'channel.p2p-peer-sync', version: '1.0.0', priority: 150, appliesTo: ['local', 'p2p-agent'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_CHANNEL) },
+    { id: 'channel.p2p-proactive', version: '1.0.0', priority: 150, appliesTo: ['p2p-agent'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_CHANNEL) },
+    { id: 'channel.human-async', version: '1.0.0', priority: 150, appliesTo: ['local'], source: 'static-md', maxChars: 500, meta: DEFAULT_META(TTL_CHANNEL) },
+    { id: 'channel.session-handoff', version: '1.0.0', priority: 150, appliesTo: ['local', 'p2p-visitor', 'p2p-agent'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_CHANNEL) },
+    // ── core/ ──
+    { id: 'core.memory_system', version: '1.0.0', priority: 140, appliesTo: ['all'], source: 'static-md', maxChars: 200, meta: DEFAULT_META(TTL_KNOWLEDGE) },
+    // 2026-07-10 双栖 agent 网络新增 (按 plan 阶段 4):
+    // priority 90 (在 channel 150 之后) — 保证 channel 装完后再装 core, 避免 channel 预算被 core 吃掉
+    // maxChars 700 → 400: 7 个新 layer 总预算 3900 chars, 必须让出空间给 tool.p2p_request (700) + tool.goal_handoff (600)
+    { id: 'core.external-engagement', version: '1.0.0', priority: 90, appliesTo: ['all'], source: 'static-md', maxChars: 400, meta: DEFAULT_META(TTL_KNOWLEDGE) },
     // ── tool/ (按工具调用嵌对应 layer) ──
     { id: 'tool.bash', version: '1.0.0', priority: 250, appliesTo: ['tool:bash'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_TOOL) },
     { id: 'tool.web_search', version: '1.0.0', priority: 250, appliesTo: ['tool:web_search'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_TOOL) },
@@ -111,6 +121,9 @@ const STATIC_LAYERS = [
     { id: 'tool.image_search', version: '1.0.0', priority: 250, appliesTo: ['never'], source: 'static-md', maxChars: 0, meta: DEFAULT_META(TTL_TOOL) },
     { id: 'tool.artifacts', version: '1.0.0', priority: 250, appliesTo: ['never'], source: 'static-md', maxChars: 0, meta: DEFAULT_META(TTL_TOOL) },
     { id: 'tool.manifest', version: '1.0.0', priority: 250, appliesTo: ['tool:manifest'], source: 'static-md', maxChars: 500, meta: DEFAULT_META(TTL_TOOL) },
+    // 2026-07-10 双栖 agent 网络新增 (按 plan 阶段 5):
+    { id: 'tool.p2p_request', version: '1.0.0', priority: 250, appliesTo: ['tool:send_message', 'tool:send_to_channel', 'tool:check_inbox', 'tool:list_peers', 'tool:agent_call', 'tool:broadcast_message'], source: 'static-md', maxChars: 700, meta: DEFAULT_META(TTL_TOOL) },
+    { id: 'tool.goal_handoff', version: '1.0.0', priority: 250, appliesTo: ['tool:park_goal', 'tool:resume_goal', 'tool:continue_goal_background'], source: 'static-md', maxChars: 600, meta: DEFAULT_META(TTL_TOOL) },
 ];
 /**
  * 动态 layer (运行时计算, 例如 project context, judgment 注入)
@@ -141,8 +154,13 @@ const DYNAMIC_LAYERS = [
  * P-Action 4 (2026-06-15): 总字符上限 15000 → 4500.
  * 阶段 0 单次 system prompt 控制在 ≤ 4.5KB (≈ 1125 tokens).
  * 配合单 layer maxChars 收紧 + 6 layer 停用, 每轮 chat 节省 ≈ 2625 tokens.
+ *
+ * 2026-07-10 改造: 4500 → 6500 → 8000 (≈ 2000 tokens).
+ *   原因: 7 个新 layer (4 channel + 1 core + 2 tool) 合计 ≈ 3500 chars.
+ *   实际测算: 现有 11 个 layer 装配后 ≈ 6930 chars; 必须松绑到 8000 才能保证 tool.* 也装入.
+ *   每次 chat 多 ~875 tokens input (7% 增), 可接受 — 双栖 agent 网络的"目标接力"必须让 LLM 看到.
  */
-const TOTAL_BUDGET = 4500;
+const TOTAL_BUDGET = 8000;
 export async function assembleSystemPrompt(ctx) {
     // 1. 收集所有 layer
     const allLayers = [];

package/dist/web/client.js

@@ -125,6 +125,13 @@
           node.appendChild(marker);
           node.appendChild(label2);
           node.appendChild(args);
+          const errContainer = document.createElement("div");
+          errContainer.className = "step-timeline-error-wrap";
+          errContainer.style.display = "none";
+          const errEl = document.createElement("span");
+          errEl.className = "step-timeline-error";
+          errContainer.appendChild(errEl);
+          node.appendChild(errContainer);
           if (existing[htmlIdx] && existing[htmlIdx] !== node) {
             listEl.replaceChild(node, existing[htmlIdx]);
           } else {
@@ -142,6 +149,16 @@
           argsEl.textContent = argStr;
           argsEl.style.display = argStr ? "" : "none";
         }
+        const errWrap = node.querySelector(".step-timeline-error-wrap");
+        const errEl2 = node.querySelector(".step-timeline-error");
+        if (step.status === "error" && step.error && errWrap && errEl2) {
+          errEl2.textContent = String(step.error);
+          errEl2.setAttribute("title", String(step.error));
+          errWrap.style.display = "";
+        } else if (errWrap && errEl2) {
+          errWrap.style.display = "none";
+          errEl2.textContent = "";
+        }
         htmlIdx++;
       }
       while (listEl.children.length > htmlIdx) {
@@ -356,6 +373,31 @@
   function parseToolCall(content, ctx) {
     if (!content) return null;
     const strippedContent = content.replace(/<think[\s\S]*?<\/think/g, "");
+    try {
+      const result = /* @__PURE__ */ function _diagProbe() {
+        return null;
+      }();
+      const probe = function _doParse() {
+        const m1 = strippedContent.match(/<invoke\s+name=["']([\w]+)["']/);
+        if (m1) {
+          return { name: m1[1], args: { __probe_invoketag: "1" } };
+        }
+        const m2 = strippedContent.match(/<function_calls>[\s\S]*?<invoke\s+name=["']([\w]+)["']/);
+        if (m2) {
+          return { name: m2[1], args: { __probe_function_calls: "1" } };
+        }
+        const m3 = strippedContent.match(/\{[\s\S]*?"name"\s*:\s*["']([\w]+)["']/);
+        if (m3) {
+          return { name: m3[1], args: { __probe_json_name: "1" } };
+        }
+        return null;
+      }();
+      console.warn(
+        "[parseToolCall diag] rawLen=" + content.length + " strippedLen=" + strippedContent.length + " probeName=" + (probe?.name ?? "null") + " rawHead=" + JSON.stringify(content.slice(0, 500))
+      );
+    } catch (diagErr) {
+      console.warn("[parseToolCall diag] diag-self-failed:", String(diagErr));
+    }
     const jsonPatterns = [
       // markdown json code block + OpenAI  块, 同时匹配 arguments/input 字段
       /(?:```(?:json|json5)?\s*\n?)?\{[\s\S]*?"name"\s*:\s*["'](\w+)["']\s*,\s*["']?(?:arguments|input)["']?\s*:\s*(\{[\s\S]*?\})\s*\}/

package/dist/web/style.css

@@ -4740,3 +4740,30 @@ body:has(> .api-config-page) {
 }
 .step-timeline-more:hover { text-decoration: underline; }
 
+
+/* 2026-07-12: 工具原始 error 展示 — 之前 step.error 数据存了但没渲染, 用户只看到
+   LLM 改写的"X 必填"误导调试. 现在展示原始 error (mono 字体 + 红色 + 缩进).
+   默认 hidden, error 状态时 step-timeline.ts 显示出来. */
+.step-timeline-error-wrap {
+  flex-basis: 100%;
+  margin: 2px 0 0 0;
+  padding-left: 16px;
+}
+.step-timeline-error {
+  display: inline-block;
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 10.5px;
+  color: var(--warning);
+  background: rgba(255, 165, 0, 0.08);
+  border-left: 2px solid var(--warning);
+  padding: 2px 6px;
+  border-radius: 0 3px 3px 0;
+  word-break: break-word;
+  white-space: pre-wrap;
+  max-width: 100%;
+}
+.step-timeline-node[data-status="error"] .step-timeline-error {
+  /* error 状态下加粗, 用户视线引导 */
+  font-weight: 500;
+}
+

package/dist/web/ui/step-timeline.js

@@ -98,6 +98,13 @@ function render(timelineEl) {
         node.appendChild(marker);
         node.appendChild(label2);
         node.appendChild(args);
+        const errContainer = document.createElement("div");
+        errContainer.className = "step-timeline-error-wrap";
+        errContainer.style.display = "none";
+        const errEl = document.createElement("span");
+        errEl.className = "step-timeline-error";
+        errContainer.appendChild(errEl);
+        node.appendChild(errContainer);
         if (existing[htmlIdx] && existing[htmlIdx] !== node) {
           listEl.replaceChild(node, existing[htmlIdx]);
         } else {
@@ -115,6 +122,16 @@ function render(timelineEl) {
         argsEl.textContent = argStr;
         argsEl.style.display = argStr ? "" : "none";
       }
+      const errWrap = node.querySelector(".step-timeline-error-wrap");
+      const errEl2 = node.querySelector(".step-timeline-error");
+      if (step.status === "error" && step.error && errWrap && errEl2) {
+        errEl2.textContent = String(step.error);
+        errEl2.setAttribute("title", String(step.error));
+        errWrap.style.display = "";
+      } else if (errWrap && errEl2) {
+        errWrap.style.display = "none";
+        errEl2.textContent = "";
+      }
       htmlIdx++;
     }
     while (listEl.children.length > htmlIdx) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bolloon/bolloon-agent",
-  "version": "0.2.15",
+  "version": "0.3.1",
   "type": "module",
   "description": "P2P AI Document Agent - 全局安装后执行 `bolloon` 启动产品",
   "main": "dist/cli-entry.js",
@@ -48,7 +48,7 @@
     "src/constraint-runtime"
   ],
   "dependencies": {
-    "@bolloon/bolloon-agent": "^0.2.15",
+    "@bolloon/bolloon-agent": "^0.3.1",
     "@bolloon/constraint-runtime": "0.1.0",
     "@capacitor/core": "^8.4.1",
     "@capacitor/ios": "^8.4.1",