@bolloon/bolloon-agent 0.4.33 → 0.5.1

package/dist/agents/pi-sdk.js

@@ -29,7 +29,10 @@ import { DeepThinkingEngine, AgentCoordinator } from '@bolloon/constraint-runtim
 import { WorkflowPivotLoop, createDefaultPivotConfig } from './workflow-pivot-loop.js';
 import { initDocumentReceiver } from './p2p-document-tools.js';
 import { startRun, recordStep, finishRun, readRun, budgetVerdict, recordDegradation, recordHarnessEvent, recordRecovery, setRunStatus, prepareResume, markRunRunning, buildResumeInstruction, argsDigestOf, repeatedFailureCount, classifyError as classifyRunError } from './run-store.js';
-import { createGoal, readGoal, attachRun, findActiveGoal, completeGoalIfEligible, setUnresolved, addEvidence, evaluateGoalCompletion } from './goal-store.js';
+import { createGoal, attachRun, findActiveGoal } from './goal-store.js';
+import { captureRunModelConfig } from '../llm/model-selection.js';
+// 2026-09-26: 工具名出网净化 (pi-ai.ts 唯一边界) + 回程派发还原 (原名 ↔ API 名)
+import { resolveApiToolName, expandKnownToolNames } from '../llm/tool-name.js';
 import { PiAgentHarness } from './pi-harness.js';
 import { DiscoveredAgentsManager, createSocialHeartbeat } from '../social/heartbeat.js';
 import { SkillRegistry } from '@bolloon/constraint-runtime';
@@ -645,6 +648,18 @@ export class PiAgentSession {
             return false;
         }
     }
+    /**
+     * 2026-09-26: 这个 Run 开始那一刻**真实生效**的模型配置快照, 由 startRun 落进 Run 记录。
+     * 拿不到 (配置读不出来) 就返回 undefined —— 宁可这个字段缺, 也不编一个假的进去。
+     */
+    async runModelSnapshot() {
+        try {
+            return await captureRunModelConfig();
+        }
+        catch {
+            return undefined;
+        }
+    }
     /** 2026-08-07: prompt 出口统一上报 token 用量到 ContextManager (fallback/pivot/react 全路径覆盖) —
      *   之前只有 runReActLoop 迭代内上报, chitchat/fallback/pivot 路径状态栏恒 0 */
     reportUsageToContextManager() {
@@ -749,6 +764,7 @@ export class PiAgentSession {
                     goalId: boundGoalId,
                     channelId: this.currentChannelId || undefined,
                     agentId: this.currentAgentId || undefined,
+                    modelConfig: await this.runModelSnapshot(),
                 });
                 this.lastRunId = rec.runId;
                 await attachRun(boundGoalId, rec.runId).catch(() => null);
@@ -757,6 +773,30 @@ export class PiAgentSession {
                     status: 'needs_human',
                     error: 'LLM 不可用: provider 未初始化或无 apiKey (fallback 不是执行结果, 需要配置或人工处理)',
                 });
+                // ★ M0: "工具/权限不可用"也是一条终止路径 → 同样进收尾漏斗 (规则 ④)。
+                const wiring = await import('./goal-flywheel-wiring.js');
+                const closed = await wiring.closeRunOnce({
+                    goalId: boundGoalId,
+                    runId: rec.runId,
+                    caller: 'runner',
+                    now: new Date().toISOString(),
+                    finalReview: `LLM 不可用 → fallback 收尾 (不是执行结果): ${response}`,
+                });
+                if (!wiring.isRefusal(closed) && closed.outcome) {
+                    const { reduceGoalState } = await import('./goal-state-reducer.js');
+                    await reduceGoalState({
+                        goalId: boundGoalId,
+                        intent: 'closure_outcome',
+                        now: new Date().toISOString(),
+                        by: 'runner',
+                        outcome: {
+                            goalStatus: wiring.goalStatusFromDecision(closed.outcome.decision),
+                            continuation: wiring.toGoalStoreContinuation(closed.outcome.continuation),
+                            reason: closed.outcome.decision.reason,
+                        },
+                        runId: rec.runId,
+                    });
+                }
                 console.log(`[PiAgent] LLM 不可用 → 本次仍落 Run ${rec.runId} (needs_human), 不伪装成正常执行`);
             }
             catch (err) {
@@ -1057,7 +1097,7 @@ export class PiAgentSession {
             const backoffMs = 1000 * Math.pow(2, attempt - 1); // 1s, 2s, 4s
             console.log(`[PiAgent] loop 自动重试 ${attempt}/${MAX_LOOP_RETRIES}, 等待 ${backoffMs}ms`);
             if (onStream) {
-                onStream({ type: 'status', content: `↻ 自动重试 loop ${attempt}/${MAX_LOOP_RETRIES} (${(backoffMs / 1000).toFixed(0)}s 后)`, tool: 'system' });
+                onStream({ type: 'status', internal: true, content: `↻ 自动重试 loop ${attempt}/${MAX_LOOP_RETRIES} (${(backoffMs / 1000).toFixed(0)}s 后)`, tool: 'system' });
             }
             // 中途 abort 也要响应
             await new Promise((resolve, reject) => {
@@ -1357,17 +1397,51 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
      * 2026-09-16 (M2): 从 checkpoint 恢复一次运行 —— **不是**重发原 prompt。
      * 语义: prepareResume (校验状态 + 读 checkpoint + 记 recovery + 落 recovering) → 用恢复指令
      * 驱动同一个 runId 继续 (历史保留), 非幂等动作由重放守卫挡住。
+     *
+     * 2026-09-26: 恢复前**反查一次模型配置**。Run 记录里存的 `modelConfig` 快照此前只被写进
+     * 快照和回显, 没人拿它核对现状 —— 于是"快照记的是 A, 现在盘上生效的是 B"这种情况会被
+     * 静默继续下去 (长任务的历史执行链就说不清了)。现在: 漂了就 `console.warn` 说清哪几个字段漂了,
+     * 并把结论放进返回值 (`modelDrift`) 让调用方能上报; 一致时也有一句明确结论。
+     * 判断失败 (读不到 Run 等) 不阻塞恢复。
      */
     async resumeRun(runId) {
+        let modelDrift;
+        let modelApplied;
+        try {
+            const { detectRunConfigDrift } = await import('../llm/model-selection.js');
+            modelDrift = (await detectRunConfigDrift(runId)) || undefined;
+            if (modelDrift?.drifted) {
+                console.warn('[pi-sdk] 恢复时模型配置已偏离 Run 快照:', modelDrift.message);
+            }
+            else if (modelDrift) {
+                console.log('[pi-sdk] 恢复前核对:', modelDrift.message);
+            }
+        }
+        catch { /* 核对本身失败不阻塞恢复 */ }
         const prep = await prepareResume(runId);
         if (!prep.ok || !prep.plan)
-            return { ok: false, reason: prep.reason };
+            return { ok: false, reason: prep.reason, ...(modelDrift ? { modelDrift } : {}) };
+        // 2026-09-26 (P6 × P7): 恢复一个在跑的 Run, 用的是**它自己那份快照**, 不是盘上现在的全局默认 ——
+        //   否则这次恢复就是"带着漂移继续跑", 之后再存快照会把漂移固化成历史。
+        //   只在"漂了"或"核对不了"时重装 (一致时运行时本来就是对的, 重装是白跑一趟)。
+        //   装配只有一处实现: `applyRunModelConfigToRuntime` → `installRuntime` (这里不自己 initMinimax)。
+        if (modelDrift && (modelDrift.drifted || !modelDrift.verified)) {
+            try {
+                const { applyRunModelConfigToRuntime } = await import('../llm/model-selection.js');
+                modelApplied = await applyRunModelConfigToRuntime(modelDrift.snapshot);
+                console.log(`[pi-sdk] 已按 Run 快照装配运行时: ${modelApplied.provider}/${modelApplied.model} (configHash ${String(modelApplied.configHash).slice(0, 12)})`);
+            }
+            catch (e) {
+                // 装配失败不阻塞恢复 (如实说; 调用方可以从返回值里看出没装成)
+                console.warn('[pi-sdk] 按 Run 快照装配运行时失败, 保持当前运行时:', String(e?.message || e).slice(0, 160));
+            }
+        }
         this.resumeRunId = runId;
         this.resumePlan = prep.plan;
         this.currentGoalId = prep.plan.goalId || this.currentGoalId;
         try {
             const reply = await this.prompt(buildResumeInstruction(prep.plan), {});
-            return { ok: true, reply };
+            return { ok: true, reply, ...(modelDrift ? { modelDrift } : {}), ...(modelApplied ? { modelApplied } : {}) };
         }
         finally {
             this.resumeRunId = '';
@@ -1437,7 +1511,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
         const loopActionLog = [];
         // 发送循环开始的事件
         if (onStream) {
-            onStream({ type: 'status', content: '🔄 开始 ReAct 循环...', tool: 'system' });
+            onStream({ type: 'status', internal: true, content: '🔄 开始 ReAct 循环...', tool: 'system' });
         }
         // React Harness: 循环开始 (重置 turn 计数 + 触发 harness sessionStart)
         // 失败静默 (fail-open), 不阻塞主循环
@@ -1469,7 +1543,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
             }
             const doneN = this.resumePlan?.completedSteps.length ?? 0;
             const guards = this.resumePlan?.replayGuards.length ?? 0;
-            onStream?.({ type: 'status', content: `♻️ 从 checkpoint 恢复运行 ${this.currentRunId} (已完成 ${doneN} 步, 非幂等重放守卫 ${guards} 条)`, tool: 'harness' });
+            onStream?.({ type: 'status', internal: true, content: `♻️ 从 checkpoint 恢复运行 ${this.currentRunId} (已完成 ${doneN} 步, 非幂等重放守卫 ${guards} 条)`, tool: 'harness' });
         }
         else {
             // 2026-09-16 (M2): 目标绑定 —— 有 goalId 就在该 Goal 下执行; 没有就建 Goal 再建 Run。
@@ -1512,6 +1586,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                     goalId: boundGoalId || undefined,
                     channelId: this.currentChannelId || undefined,
                     agentId: this.currentAgentId || undefined,
+                    modelConfig: await this.runModelSnapshot(),
                 });
                 this.currentRunId = rec.runId;
                 this.lastRunId = rec.runId;
@@ -1520,7 +1595,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                     // Run → Goal 反查链: runId → goalId → objective / success criteria
                     await attachRun(this.currentGoalId, rec.runId).catch((err) => console.warn('[PiAgent] attachRun 失败:', err?.message));
                 }
-                onStream?.({ type: 'status', content: `🧷 运行已登记 (run=${rec.runId}${this.currentGoalId ? `, goal=${this.currentGoalId}` : ''}, 预算 ${rec.budget.maxSteps} 步 / ${Math.round(rec.budget.deadlineMs / 60000)} 分钟)`, tool: 'harness' });
+                onStream?.({ type: 'status', internal: true, content: `🧷 运行已登记 (run=${rec.runId}${this.currentGoalId ? `, goal=${this.currentGoalId}` : ''}, 预算 ${rec.budget.maxSteps} 步 / ${Math.round(rec.budget.deadlineMs / 60000)} 分钟)`, tool: 'harness' });
             }
             catch (err) {
                 // 核心写失败: 不再 warn 后继续 —— 没有运行记录就不执行 (strict 模式默认如此)
@@ -1638,7 +1713,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
             if (shouldCompactBeforeIteration(estimatedTokensBefore, compactThreshold)) {
                 const tokensBeforeCompact = estimatedTokensBefore;
                 console.log(`[PiAgent] loop 入口 token ${tokensBeforeCompact} > ${compactThreshold}, 触发自动压缩`);
-                onStream?.({ type: 'status', content: `🗜️ loop 自动压缩 (token ${tokensBeforeCompact} > ${compactThreshold})`, tool: 'compactor' });
+                onStream?.({ type: 'status', internal: true, content: `🗜️ loop 自动压缩 (token ${tokensBeforeCompact} > ${compactThreshold})`, tool: 'compactor' });
                 try {
                     await this.maybeAutoCompact(onStream, signal);
                 }
@@ -1660,7 +1735,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
             // 调试日志：显示每次循环开始
             console.log(`[PiAgent] 循环 ${iteration}/${this.MAX_REACT_ITERATIONS} 开始`);
             if (onStream) {
-                onStream({ type: 'status', content: `🔄 循环 ${iteration}/${this.MAX_REACT_ITERATIONS}`, tool: 'loop' });
+                onStream({ type: 'status', internal: true, content: `🔄 循环 ${iteration}/${this.MAX_REACT_ITERATIONS}`, tool: 'loop' });
             }
             const context = this.buildContext();
             // M3.5 (2026-06-17): 也构造 messages 数组版本, 让 LLM 看到结构化 tool 角色
@@ -1849,7 +1924,10 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
             }
             // 路径 B: 文本解析 (parseAllToolCalls 收集全部)
             if (toolCalls.length === 0) {
-                const knownTools = new Set(Array.from(this.tools.keys()));
+                // 2026-09-26: 已知名集合 = 注册表原名 + 它们的 API 名 (净化后的).
+                //   出网时工具面给的是 API 名 (见 pi-ai.ts 的唯一净化边界), LLM 在文本里
+                //   回吐的也可能是 API 名 —— 只用原名集合过滤会把这类调用整个丢掉.
+                const knownTools = expandKnownToolNames(this.tools.keys());
                 toolCalls = parseAllToolCalls(reply, { tools: knownTools });
             }
             // 回退路径 C: 原生 parseToolCall (单个)
@@ -1858,6 +1936,12 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                 if (single)
                     toolCalls.push(single);
             }
+            // 2026-09-26: **回程派发的唯一还原点** —— LLM 回吐的是 API 名 (净化过的), 这里
+            //   还原成注册表真名再交给 this.tools.get(); 原名 (LLM 照 system prompt 抄的) 原样穿透.
+            for (const tc of toolCalls) {
+                if (typeof tc.name === 'string' && tc.name)
+                    tc.name = resolveApiToolName(tc.name);
+            }
             // 给每个 toolCall 分配稳定 id
             for (const tc of toolCalls) {
                 if (!tc.id) {
@@ -1882,7 +1966,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                     if (onStream) {
                         onStream({ type: 'tool', content: `🔧 调用工具 (${ti + 1}/${toolCalls.length}): ${toolCall.name}`, tool: toolCall.name });
                         if (toolCall.args && Object.keys(toolCall.args).length > 0) {
-                            onStream({ type: 'status', content: `📋 参数: ${JSON.stringify(toolCall.args)}`, tool: toolCall.name });
+                            onStream({ type: 'status', internal: true, content: `📋 参数: ${JSON.stringify(toolCall.args)}`, tool: toolCall.name });
                         }
                         onStream({
                             type: 'step_start',
@@ -2012,7 +2096,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                         const toolDurationMs = Date.now() - toolStart;
                         if (replaySkip) {
                             console.log(`[PiAgent] 恢复重放守卫: 跳过已完成的非幂等工具 ${toolCall.name}`);
-                            onStream?.({ type: 'status', content: `🛡️ 恢复保护: ${toolCall.name} 此前已成功执行, 本次不重复执行`, tool: toolCall.name });
+                            onStream?.({ type: 'status', internal: true, content: `🛡️ 恢复保护: ${toolCall.name} 此前已成功执行, 本次不重复执行`, tool: toolCall.name });
                         }
                         console.log(`[PiAgent] 工具 ${toolCall.name} 执行完成: success=${result.success} (${toolDurationMs}ms)`);
                         // 2026-09-16 (M3): 失败接线 —— 分类 + recovery 留痕 + 熔断 / 外部等待
@@ -2085,7 +2169,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                         }
                         if (onStream) {
                             if (result.success) {
-                                onStream({ type: 'status', content: `✅ ${toolCall.name} 执行成功`, tool: toolCall.name });
+                                onStream({ type: 'status', internal: true, content: `✅ ${toolCall.name} 执行成功`, tool: toolCall.name });
                                 if (result.output) {
                                     onStream({ type: 'tool', content: `📤 结果: ${result.output.substring(0, 200)}${result.output.length > 200 ? '...' : ''}`, tool: toolCall.name });
                                 }
@@ -2115,7 +2199,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                                 refineAttempts++;
                             }
                             if (onStream) {
-                                onStream({ type: 'status', content: `🔄 工具执行完成，继续循环...`, tool: 'loop' });
+                                onStream({ type: 'status', internal: true, content: `🔄 工具执行完成，继续循环...`, tool: 'loop' });
                             }
                         }
                         else {
@@ -2201,7 +2285,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                             content: `[dive-into stop condition] 你之前已成功执行了 ${unreported} 个工具, 但当前回复里没把它们的结果告诉用户. 请基于已有的工具结果 (在 history 里) 写一个完整总结回复给用户, 用 <final gen> 结尾. 不要再调工具.`
                         });
                         if (onStream) {
-                            onStream({ type: 'status', content: `🔄 还有 ${unreported} 个工具结果未汇报, 让 LLM 继续总结 (${unreportedRetries}/${MAX_UNREPORTED_RETRIES})`, tool: 'system' });
+                            onStream({ type: 'status', internal: true, content: `🔄 还有 ${unreported} 个工具结果未汇报, 让 LLM 继续总结 (${unreportedRetries}/${MAX_UNREPORTED_RETRIES})`, tool: 'system' });
                         }
                         continue;
                     }
@@ -2214,7 +2298,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                             content: `[dive-into stop condition] 已多次提示汇报工具结果仍未完成 (超过 ${MAX_UNREPORTED_RETRIES} 次). 现在直接基于你已知的信息写最终回复给用户, 用 <final gen> 结尾, 不要再调任何工具.`
                         });
                         if (onStream) {
-                            onStream({ type: 'status', content: `🔄 工具结果汇报超限, 强制收尾`, tool: 'system' });
+                            onStream({ type: 'status', internal: true, content: `🔄 工具结果汇报超限, 强制收尾`, tool: 'system' });
                         }
                     }
                     lastQualityScore = this.estimateResponseQuality(reply);
@@ -2244,7 +2328,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                         console.log(`[PiAgent] review ${loopReviewCount}/${DEFAULT_MAX_REVIEWS}: LLM 想 final 但先对齐需求深挖一次`);
                         this.messageHistory.push({ role: 'system', content: reviewDecision.hint });
                         if (onStream) {
-                            onStream({ type: 'status', content: `🔄 目标对齐 review ${loopReviewCount}/${DEFAULT_MAX_REVIEWS}: 深挖续跑`, tool: 'system' });
+                            onStream({ type: 'status', internal: true, content: `🔄 目标对齐 review ${loopReviewCount}/${DEFAULT_MAX_REVIEWS}: 深挖续跑`, tool: 'system' });
                         }
                         continue; // 让 LLM 看到 hint, 深挖或确认完成后再次 final
                     }
@@ -2268,14 +2352,14 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                 if (needsMoreWork && iteration < this.MAX_REACT_ITERATIONS) {
                     console.log(`[PiAgent] 继续循环处理 (${iteration}/${this.MAX_REACT_ITERATIONS}): needsMoreWork=${needsMoreWork}, hasError=${hasError}, containsToolCallIntent=${containsToolCallIntent}`);
                     if (onStream) {
-                        onStream({ type: 'status', content: `🔄 继续处理，循环 ${iteration}...`, tool: 'loop' });
+                        onStream({ type: 'status', internal: true, content: `🔄 继续处理，循环 ${iteration}...`, tool: 'loop' });
                     }
                     continue;
                 }
                 // 否则把这个当作可能的最终回答
                 finalResponse = reply;
                 if (onStream) {
-                    onStream({ type: 'status', content: `📝 提取最终回答，长度 ${reply.length}`, tool: 'system' });
+                    onStream({ type: 'status', internal: true, content: `📝 提取最终回答，长度 ${reply.length}`, tool: 'system' });
                 }
                 break;
             }
@@ -2293,7 +2377,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
         }
         // 通知前端循环完成
         if (onStream) {
-            onStream({ type: 'status', content: `✅ 处理完成，共 ${iteration - 1} 次循环`, tool: 'system' });
+            onStream({ type: 'status', internal: true, content: `✅ 处理完成，共 ${iteration - 1} 次循环`, tool: 'system' });
         }
         this.messageHistory.push({ role: 'assistant', content: finalResponse });
         // React Harness: 循环结束
@@ -2341,30 +2425,61 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
                 await recordDegradation({ kind: 'core', op: 'pi-sdk.finishRun', runId: this.currentRunId, message }).catch(() => { });
                 onStream?.({ type: 'error', content: `⚠️ ${message}`, tool: 'harness' });
             }
-            // 2026-09-16 (M4): Goal 侧完成门 —— Run 结束 ≠ Goal 完成; 只有判据全满足 + 有证据 + 无未解决项才算
+            // ★ 2026-09-25 (M0 接线冻结, 规则 ③ + ④): Run 结束**必过收尾漏斗**。
+            //
+            // 旧写法在这里另起了一套 Goal 侧收尾 (读证据 → `evaluateGoalCompletion` →
+            // `completeGoalIfEligible` / `setUnresolved`): 它**不收尾** —— 不写 Memory、不生成 Skill 候选、
+            // 不写权威 continuation、不留决策记录。于是同一条 Run 在 Supervisor 那里还会再收一次 (或反过来
+            // 这条路径根本不收), 全仓就有了两套"这一轮跑完意味着什么"。那一套已删除。
+            //
+            // 现在: `closeRunOnce` (幂等) → 9 步收尾 → 产物落盘; 再由 Goal reducer 把"下一步是什么"
+            // 落进 Goal (没有 Supervisor 宿主时 Runner 也要留下 continuation, 否则下一次还是从零开始)。
             if (this.currentGoalId) {
                 try {
-                    const full = await readRun(this.currentRunId).catch(() => null);
-                    const ev = (full?.evidence || []).slice(-10);
-                    if (ev.length)
-                        await addEvidence(this.currentGoalId, ev);
-                    const goal = await readGoal(this.currentGoalId);
-                    if (goal) {
-                        const verdict = evaluateGoalCompletion(goal);
-                        if (verdict.complete) {
-                            const done = await completeGoalIfEligible(this.currentGoalId);
-                            onStream?.({ type: 'status', content: `🎯 目标已达成: ${done.reason} (goal=${this.currentGoalId})`, tool: 'harness' });
-                        }
-                        else {
-                            // 不静默: 未达成的目标留在 active, 未解决项写清楚 (下次 prompt 会继续这个 Goal)
-                            const items = verdict.missing.length ? verdict.missing : [`run=${this.currentRunId} 结束但目标未达成: ${verdict.reason}`];
-                            await setUnresolved(this.currentGoalId, items);
-                            onStream?.({ type: 'status', content: `🎯 目标仍在进行 (未判完成): ${verdict.reason}`, tool: 'harness' });
-                        }
+                    const wiring = await import('./goal-flywheel-wiring.js');
+                    const { reduceGoalState } = await import('./goal-state-reducer.js');
+                    const nowIso = new Date().toISOString();
+                    const closed = await wiring.closeRunOnce({
+                        goalId: this.currentGoalId,
+                        runId: this.currentRunId,
+                        caller: 'runner',
+                        now: nowIso,
+                        finalReview: finalResponse ? String(finalResponse).slice(0, 2000) : '',
+                    });
+                    if (wiring.isRefusal(closed)) {
+                        onStream?.({ type: 'status', content: `⚠️ 收尾被拒 (不当作收过): ${closed.reason}`, tool: 'harness' });
+                    }
+                    else if (closed.outcome) {
+                        const view = closed.outcome;
+                        const applied = await reduceGoalState({
+                            goalId: this.currentGoalId,
+                            intent: 'closure_outcome',
+                            now: nowIso,
+                            by: 'runner',
+                            outcome: {
+                                goalStatus: wiring.goalStatusFromDecision(view.decision),
+                                continuation: wiring.toGoalStoreContinuation(view.continuation),
+                                reason: view.decision.reason,
+                            },
+                            runId: view.runId,
+                        });
+                        const line = view.decision.decision === 'complete'
+                            ? `🎯 目标已达成 (仍要过完成门): ${view.decision.reason}`
+                            : `🎯 目标仍在进行 (未判完成): ${view.decision.reason}`;
+                        onStream?.({
+                            type: 'status', internal: true,
+                            content: `${line}${applied.gateRejected ? ` [完成门拒绝: ${applied.gateRejected}]` : ''} `
+                                + `(收尾 ${view.steps} 步 · memory ${view.memories} · skill 候选 ${view.candidates} · goal=${this.currentGoalId})`,
+                            tool: 'harness',
+                        });
+                    }
+                    else {
+                        // 幂等: 别处 (或上一次) 已经收过尾 —— 事实读不回来就如实说, 不重收也不假装
+                        onStream?.({ type: 'status', content: `🎯 ${closed.reason}`, tool: 'harness' });
                     }
                 }
                 catch (err) {
-                    await recordDegradation({ kind: 'observational', op: 'pi-sdk.goalCompletion', runId: this.currentRunId, message: String(err?.message || err).slice(0, 160) }).catch(() => { });
+                    await recordDegradation({ kind: 'observational', op: 'pi-sdk.runClosure', runId: this.currentRunId, message: String(err?.message || err).slice(0, 160) }).catch(() => { });
                 }
             }
             this.currentRunId = '';
@@ -2542,7 +2657,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
         const estimated = this.estimateHistoryTokens();
         if (estimated > this.maxContextTokens() * 0.8) {
             console.warn(`[PiAgent] reactive compaction pre-check (${estimated} tokens > 80% threshold)`);
-            onStream?.({ type: 'status', content: '⚠️ reactive compaction 预检触发', tool: 'recovery' });
+            onStream?.({ type: 'status', internal: true, content: '⚠️ reactive compaction 预检触发', tool: 'recovery' });
             try {
                 const compacted = this.compressHistorySync(this.messageHistory);
                 this.messageHistory = compacted;
@@ -2696,7 +2811,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
     async maybeAutoCompact(onStream, signal) {
         if (this.messageHistory.length < 10)
             return; // 历史太短, 不值得压
-        onStream?.({ type: 'status', content: '🗜️ 评估是否需要压缩上下文...', tool: 'compactor' });
+        onStream?.({ type: 'status', internal: true, content: '🗜️ 评估是否需要压缩上下文...', tool: 'compactor' });
         // 注入 LLM (用 getMinimax().chat, 与 judgment 注入门 / ReAct 循环同一来源)
         // 给 Context Collapse (虚拟投影) 和 Auto-Compact (摘要) 共用
         const llm = getMinimax();
@@ -2724,7 +2839,7 @@ ${PiAgentSession.TOOL_SELECTION_GUIDE}
             const savedTokens = Math.max(0, beforeTokens - afterTokens);
             cm.markCompressStart(beforeTokens);
             onStream?.({
-                type: 'status',
+                type: 'status', internal: true,
                 content: `🗜️ 上下文压缩: ${stagesApplied || 'no-op'} | 节省 ${saved} 条 / ${savedTokens.toLocaleString()} tokens (剩余 ${result.history.length}, collapse=${isContextCollapseEnabled() ? 'on' : 'off'})`,
                 tool: 'compactor',
             });

package/dist/agents/run-store.js

@@ -24,14 +24,14 @@ export const RUN_TRANSITIONS = {
     queued: ['running', 'aborted', 'interrupted'],
     running: ['recovering', 'paused', 'awaiting_external', 'done', 'failed', 'aborted', 'interrupted', 'stalled', 'needs_human'],
     recovering: ['running', 'failed', 'aborted', 'needs_human', 'interrupted', 'stalled'],
-    paused: ['running', 'aborted', 'interrupted'],
-    awaiting_external: ['running', 'failed', 'aborted', 'interrupted', 'stalled'],
+    paused: ['running', 'aborted', 'interrupted', 'recovering'],
+    awaiting_external: ['running', 'failed', 'aborted', 'interrupted', 'stalled', 'recovering'],
     done: [],
     failed: [],
     aborted: [],
     interrupted: ['recovering', 'aborted'], // 允许"从 checkpoint 恢复"
     stalled: ['recovering', 'aborted', 'needs_human'],
-    needs_human: ['running', 'aborted'],
+    needs_human: ['running', 'aborted', 'recovering'],
 };
 export function canTransition(from, to) {
     if (from === to)
@@ -39,6 +39,8 @@ export function canTransition(from, to) {
     return (RUN_TRANSITIONS[from] || []).includes(to);
 }
 export const MAX_HARNESS_EVENTS = 50;
+/** 模型切换账本只留最近 N 条 (审计账本, 不是全量日志) */
+export const MAX_MODEL_SWITCHES = 20;
 const DEFAULT_HARNESS = {
     maxSteps: Number(process.env.BOLLOON_RUN_MAX_STEPS || 60),
     deadlineMs: Number(process.env.BOLLOON_RUN_DEADLINE_MS || 30 * 60_000),
@@ -343,6 +345,7 @@ export async function startRun(opts) {
         steps: [],
         budget: { maxSteps: cfg.maxSteps, deadlineMs: cfg.deadlineMs },
         recovery: [],
+        modelConfig: opts.modelConfig,
     };
     await coreWrite('startRun', rec.runId, () => writeRun(rec));
     return rec;
@@ -453,6 +456,56 @@ export async function recordHarnessEvent(runId, evt) {
         await recordDegradation({ kind: 'observational', op: 'recordHarnessEvent', runId, message: `${evt.event}/${evt.kind} 写入失败: ${String(err?.message || err).slice(0, 150)}` });
     }
 }
+/**
+ * 2026-09-26 (P7): 记一次模型切换决定 (Run 事件)。
+ *
+ * 与 `recordHarnessEvent` 同一纪律: 决策**已经**做出, 记账失败绝不反过来改决策 ——
+ * 所以这里只记降级 + 返回 `{ok:false}` 让调用方能如实上报, 不抛。
+ * 写两处事实: ① `modelSwitches` 账本 (机器可查的切换链); ② `harness` 事件 (人工审计的那本账)。
+ * 本次写入**不改** `modelConfig` —— Run 快照一旦定稿就不再改写。
+ */
+export async function recordModelSwitch(runId, evt) {
+    const full = {
+        ts: evt.ts || new Date().toISOString(),
+        ...evt,
+        runId: evt.runId || runId,
+    };
+    try {
+        const rec = await withRunLock(runId, async () => {
+            const cur = await readRun(runId);
+            if (!cur)
+                return null;
+            cur.modelSwitches = cur.modelSwitches || [];
+            cur.modelSwitches.push(full);
+            if (cur.modelSwitches.length > MAX_MODEL_SWITCHES) {
+                cur.modelSwitches = cur.modelSwitches.slice(-MAX_MODEL_SWITCHES);
+            }
+            // 人工审计账本: 同一件事在 harness 事件里也留一条 (kind 按结果映射, 不美化)
+            cur.harness = cur.harness || [];
+            cur.harness.push({
+                ts: full.ts,
+                event: 'model.switch',
+                kind: full.outcome === 'switched' ? 'note' : 'deny',
+                reason: `${full.outcome} mode=${full.mode} source=${full.source} → ${full.to.provider}/${full.to.model}#${full.to.configHash} · 守卫 ${full.guardsCarried} 条 · ${full.reason}`.slice(0, 400),
+                runId,
+                goalId: full.goalId,
+            });
+            if (cur.harness.length > MAX_HARNESS_EVENTS)
+                cur.harness = cur.harness.slice(-MAX_HARNESS_EVENTS);
+            cur.updatedAt = new Date().toISOString();
+            await writeRun(cur);
+            return cur;
+        });
+        if (!rec)
+            return { ok: false, reason: `run 不存在: ${runId}` };
+        return { ok: true, record: rec };
+    }
+    catch (err) {
+        const reason = `模型切换事件写入失败: ${String(err?.message || err).slice(0, 160)}`;
+        await recordDegradation({ kind: 'observational', op: 'recordModelSwitch', runId, message: reason });
+        return { ok: false, reason };
+    }
+}
 /** 防止"同一个工具 + 同一组参数"无限重试: 返回该指纹最近的连续失败次数 */
 export function repeatedFailureCount(rec, tool, argsDigest) {
     let n = 0;
@@ -572,6 +625,13 @@ function pidAlive(pid) {
         return e?.code === 'EPERM';
     }
 }
+let onRunTerminal = null;
+export function setOnRunTerminal(hook) {
+    onRunTerminal = hook;
+}
+export function getOnRunTerminal() {
+    return onRunTerminal;
+}
 /**
  * 孤儿对账: 启动时把 pid 已死的 running 记录改判 interrupted。
  * 不做这一层的话, 重载后 UI/CLI 会显示"还在跑"的幽灵运行 —— 那是最典型的假状态。
@@ -588,11 +648,14 @@ export async function reconcileOrphans() {
             continue;
         }
         try {
-            await finishRun(r.runId, {
+            const closed = await finishRun(r.runId, {
                 status: 'interrupted',
                 error: `进程 ${r.pid} 已不在 (刷新/重载/崩溃); 运行到此中断`,
             });
             interrupted.push(r.runId);
+            // 终止路径进唯一责任链 (拿不到 Goal 就不收, 不假装收过)
+            if (onRunTerminal)
+                await Promise.resolve(onRunTerminal(closed || r)).catch(() => null);
         }
         catch (err) {
             failed.push(r.runId);
@@ -615,11 +678,14 @@ export async function superviseRuns(now = Date.now()) {
             continue; // 交给 reconcileOrphans
         if (now - Date.parse(r.updatedAt) > cfg.staleMs) {
             try {
-                await finishRun(r.runId, {
+                const closed = await finishRun(r.runId, {
                     status: 'stalled',
                     error: `超过 ${Math.round(cfg.staleMs / 1000)}s 没有新进展 (可能是工具卡住或模型长时间无响应)`,
                 });
                 stalled.push(r.runId);
+                // 同上: 失速也是一条终止路径
+                if (onRunTerminal)
+                    await Promise.resolve(onRunTerminal(closed || r)).catch(() => null);
             }
             catch (err) {
                 failed.push(r.runId);

package/dist/agents/skill-readiness.js

@@ -12,7 +12,9 @@
  */
 import * as os from 'os';
 import { SkillsManager } from './skills-manager.js';
-import { readGoal, setContinuation, updateGoal, addEvidence, } from './goal-store.js';
+import { readGoal, updateGoal, addEvidence, } from './goal-store.js';
+// 2026-09-25 (M0 接线冻结, 规则 ②): Goal 状态与 continuation 变更只有一个漏斗
+import { reduceGoalState } from './goal-state-reducer.js';
 import { recordDegradation } from './run-store.js';
 export function parseSkillSpecs(list = []) {
     const required = [];
@@ -109,15 +111,45 @@ export async function ensureGoalSkillsReady(goal, opts = {}) {
 /** 门禁不过 → 写清事实并把 Goal 交给人 (不启动 Run, 不伪造失败) */
 export async function blockGoalOnSkills(goalId, res) {
     const reason = res.reason || '技能未就绪';
-    await setContinuation(goalId, {
-        wakeReason: 'needs_human', autoContinue: false, needsExternal: undefined,
-        skillReadiness: { ok: false, at: new Date().toISOString(), reason, missing: res.missing, drift: res.drift, degradations: res.degradations },
-    });
-    await updateGoal(goalId, { status: 'needs_human' }).catch(() => { });
-    await addEvidence(goalId, [`技能门禁拦截: ${reason}`]).catch(() => { });
+    // 规则 ②: 状态与 continuation 都经 Goal reducer 写 (单一漏斗)
+    await reduceGoalState({
+        goalId,
+        intent: 'skill_gate_block',
+        now: new Date().toISOString(),
+        by: 'skill-readiness',
+        reason,
+        skillReadiness: {
+            ok: false, at: new Date().toISOString(), reason,
+            missing: res.missing, drift: res.drift, degradations: res.degradations,
+        },
+    }).catch(() => null);
     for (const d of res.degradations)
         await recordDegradation({ kind: 'observational', op: 'skill-readiness', message: d }).catch(() => { });
 }
+/**
+ * 只**记事实**、不改状态: 把本地技能就绪检查的结论写进 `continuation.skillReadiness`。
+ *
+ * ★ 2026-09-25 (P5 验收修复): 飞轮裁决 = `delegate` 时本地技能门禁**不适用** (能力由子 Agent 凭合同
+ *   交付), 所以不能调 `blockGoalOnSkills` (那会把 Goal 打成 needs_human)。但"本地缺什么"必须仍然
+ *   可查 —— 原来只有 `blockGoalOnSkills` 写这个字段, 门禁一放行就没有任何记录了。
+ */
+export async function recordSkillReadiness(goalId, res) {
+    // 规则 ②: continuation 也经 Goal reducer 写 (只记事实, 不改状态)
+    await reduceGoalState({
+        goalId,
+        intent: 'skill_gate_record',
+        now: new Date().toISOString(),
+        by: 'skill-readiness',
+        skillReadiness: {
+            ok: res.ok,
+            at: new Date().toISOString(),
+            reason: res.reason,
+            missing: res.missing,
+            drift: res.drift,
+            degradations: res.degradations,
+        },
+    }).catch(() => null);
+}
 /** 人工批准技能升级: 重新冻结快照 (显式动作, 不隐式切换) */
 export async function approveSkillUpgrade(goalId, opts = {}) {
     const goal = await readGoal(goalId);
@@ -126,8 +158,18 @@ export async function approveSkillUpgrade(goalId, opts = {}) {
     const frozen = await freezeGoalSkills(goal, { ...opts, force: true });
     if (!frozen.ok)
         return { ok: false, reason: frozen.reason };
-    await setContinuation(goalId, { skillReadiness: { ok: true, at: new Date().toISOString(), reason: '人工批准技能升级' }, wakeReason: 'active', autoContinue: true });
-    await updateGoal(goalId, { status: 'active' }).catch(() => { });
-    await addEvidence(goalId, [`技能升级已被人工批准: ${(frozen.snapshot || []).map((s) => `${s.name}@${s.version}`).join(', ')}`]).catch(() => { });
+    // ★ 规则 ⑥ (Skill 不许绕过 验证+快照+可回退 通道): 批准是**人**的动作, 但必须留下通道证据 ——
+    //   快照 (skillSnapshot, 上面 freezeGoalSkills 已写) + 版本 + 变更原因 + 可回退 (fromVersion 仍在
+    //   Goal 上)。这里把"谁批的、为什么"连同新版本一起记进 Goal (走唯一漏斗, 规则 ②)。
+    const versions = (frozen.snapshot || []).map((s) => `${s.name}@${s.version}`);
+    const fromVersions = (goal.skillSnapshot || []).map((s) => `${s.name}@${s.version}`);
+    await reduceGoalState({
+        goalId,
+        intent: 'skill_upgrade_approved',
+        now: new Date().toISOString(),
+        by: 'human',
+        skillReadiness: { ok: true, at: new Date().toISOString(), reason: '人工批准技能升级' },
+        reason: `技能升级已被人工批准 (by=human; from=[${fromVersions.join(', ')}] → to=[${versions.join(', ')}]; 可回退: Goal.skillSnapshot 保留新版本快照, 内容哈希在 skills-manager 侧可核验)`,
+    }).catch(() => null);
     return { ok: true, snapshot: frozen.snapshot };
 }