mocode-ai 1.1.7 → 1.1.8

package/dist/session/compact.js

@@ -187,6 +187,39 @@ function flattenGroups(groups) {
     }
     return out;
 }
+/** 微压缩单个 group(原地):① tool 结果 content ② assistant tool_calls.arguments(provenance stub)
+ * ③ assistant 正文。只裁模型/工具产物,不动 user 原话;保 tool_call_id 配对。返回是否有改动。 */
+function microcompactGroup(g) {
+    let done = false;
+    for (const t of g.tools) {
+        const c = t.content;
+        if (typeof c === 'string' && c.length > MAX_OLD_TOOL_STUB) {
+            t.content = truncateMid(c, MAX_OLD_TOOL_STUB);
+            done = true;
+        }
+    }
+    const as = g.assistant;
+    if (as && as.role === 'assistant') {
+        if (typeof as.content === 'string' && as.content.length > MAX_OLD_TOOL_STUB) {
+            as.content = truncateMid(as.content, MAX_OLD_TOOL_STUB);
+            done = true;
+        }
+        // tool_calls 参数:整体超长才进,provenance stub 大字段(保 path + JSON 合法)。
+        if (Array.isArray(as.tool_calls)) {
+            for (const tc of as.tool_calls) {
+                const args = tc?.function?.arguments;
+                if (typeof args !== 'string' || args.length <= MAX_OLD_TOOL_STUB)
+                    continue;
+                const stubbed = stubToolCallArguments(args);
+                if (stubbed !== args) {
+                    tc.function.arguments = stubbed;
+                    done = true;
+                }
+            }
+        }
+    }
+    return done;
+}
 // ── 默认摘要器:复用 chat(),空 handlers 不打印 ──────────────────────────
 async function defaultSummarize(older, focus) {
     // 摘要前剥离多模态 user 消息里的图片(base64 会撑爆摘要 prompt;image 对摘要无信息量)。
@@ -256,7 +289,14 @@ export async function compactHistory(history, opts) {
         kept.unshift(g);
         keptTokens = nextTokens;
     }
-    const oldGroups = groups.slice(0, groups.length - kept.length);
+    let oldGroups = groups.slice(0, groups.length - kept.length);
+    // force(硬闸/手动强压):保护区不豁免——常规切分无旧区时只保最后一组,
+    // 其余全部进可压区(首轮/当前轮也一样)。仍按 group 边界切,不破坏 tool_call 配对。
+    if (oldGroups.length === 0 && opts.force && groups.length >= 2) {
+        kept.length = 0;
+        kept.push(groups[groups.length - 1]);
+        oldGroups = groups.slice(0, groups.length - 1);
+    }
     const noop = {
         compacted: false,
         summarized: false,
@@ -269,65 +309,27 @@ export async function compactHistory(history, opts) {
     if (oldGroups.length === 0) {
         // 没有旧区可压缩
         const protectedRatio = history.length > 0 ? kept.length / history.length : 0;
-        if (opts.force && history.length > 2) {
-            // 真·强制压:把降 keepBudget 当 old group 强行创建一组可压区
-            // 取最早的 user/assistant/tool 当 old,后一段当 kept
-            const mid = Math.max(1, Math.floor(history.length / 2));
-            const older = history.slice(1, mid);
-            const keptAfter = history.slice(mid);
-            // 走仅微压缩(LLM 可能不可用,不强行 summarize)
-            let microcompactDone2 = false;
-            for (const m of older) {
-                const c = m.content;
-                if (typeof c === 'string' && c.length > MAX_OLD_TOOL_STUB) {
-                    m.content = truncateMid(c, MAX_OLD_TOOL_STUB);
-                    microcompactDone2 = true;
-                }
-                const as = m;
-                if (as.role === 'assistant') {
-                    if (typeof as.content === 'string' && as.content.length > MAX_OLD_TOOL_STUB) {
-                        as.content = truncateMid(as.content, MAX_OLD_TOOL_STUB);
-                        microcompactDone2 = true;
-                    }
-                    if (Array.isArray(as.tool_calls)) {
-                        for (const tc of as.tool_calls) {
-                            const args = tc?.function?.arguments;
-                            if (typeof args !== 'string' || args.length <= MAX_OLD_TOOL_STUB)
-                                continue;
-                            const stubbed = stubToolCallArguments(args);
-                            if (stubbed !== args) {
-                                tc.function.arguments = stubbed;
-                                microcompactDone2 = true;
-                            }
-                        }
-                    }
-                }
+        if (opts.force && groups.length === 1) {
+            // force 但只剩单组(如首轮的单个 assistant+tools 组):原地微压缩、不动结构。
+            // 纯 user 原话组无可压项(不能毁掉用户请求),落到下方 noop 提示。
+            const single = groups[0];
+            const userOnly = single.tools.length === 0
+                && single.assistant?.role === 'user';
+            if (!userOnly && microcompactGroup(single)) {
+                const estimateAfter = estimatePromptTokens(history, activeTools, state.correction);
+                state.lastEstimate = estimateAfter;
+                state.lastUsage = undefined;
+                layout.contentWrite(`  ${ui.bold}${ui.accent}●${ui.reset} ${ui.accent}强制微压缩(单组)${ui.reset}  ${ui.dim}${estimateBefore} → ${estimateAfter} tokens${ui.reset}\n`);
+                return {
+                    compacted: true,
+                    summarized: false,
+                    estimateBefore,
+                    estimateAfter,
+                    reason: 'microcompact',
+                    protectedRatio,
+                    oldGroupCount: 0,
+                };
             }
-            const refs = collectArtifactRefs(older);
-            const summaryMsg = {
-                role: 'system',
-                content: older.length > 0
-                    ? `# 会话摘要(force)\n被跳过的早期对话 ${older.length} 条已微压缩(token 数减少)。${refs.length > 0 ? `\n[artifact refs: ${refs.join(', ')}]` : ''}`
-                    : `# 会话摘要(force)\n无内容。`,
-            };
-            const rebuilt = [history[0], summaryMsg, ...keptAfter];
-            history.length = 0;
-            history.push(...rebuilt);
-            pruneAfterCompaction(history);
-            const estimateAfter = estimatePromptTokens(history, activeTools, state.correction);
-            state.lastEstimate = estimateAfter;
-            state.lastUsage = undefined;
-            layout.contentWrite(`  ${ui.bold}${ui.accent}●${ui.reset} ${ui.accent}强制压缩(focus on early history)${ui.reset}  ${ui.dim}${estimateBefore} → ${estimateAfter} tokens${ui.reset}\n`);
-            return {
-                compacted: true,
-                summarized: false,
-                historyRebuilt: true,
-                estimateBefore,
-                estimateAfter,
-                reason: microcompactDone2 ? 'microcompact' : 'summarize',
-                protectedRatio,
-                oldGroupCount: 0,
-            };
         }
         // 细分 noop 类型,供 repl 文案
         const isEmpty = history.length <= 1; // 只有 system 提示
@@ -350,35 +352,8 @@ export async function compactHistory(history, opts) {
     //   ③ 旧 assistant 正文 content(模型长解释,回看价值低)。
     let microcompactDone = false;
     for (const g of oldGroups) {
-        for (const t of g.tools) {
-            const c = t.content;
-            if (typeof c === 'string' && c.length > MAX_OLD_TOOL_STUB) {
-                t.content = truncateMid(c, MAX_OLD_TOOL_STUB);
-                microcompactDone = true;
-            }
-        }
-        const as = g.assistant;
-        if (as && as.role === 'assistant') {
-            // ③ 旧正文 content
-            if (typeof as.content === 'string' && as.content.length > MAX_OLD_TOOL_STUB) {
-                as.content = truncateMid(as.content, MAX_OLD_TOOL_STUB);
-                microcompactDone = true;
-            }
-            // ② tool_calls 参数:整体超长才进,provenance stub 大字段(保 path + JSON 合法)。
-            //   stubToolCallArguments:大字符串字段 → "<N 字符,已省略>";path 永不省(/rollback 依赖)。
-            if (Array.isArray(as.tool_calls)) {
-                for (const tc of as.tool_calls) {
-                    const args = tc?.function?.arguments;
-                    if (typeof args !== 'string' || args.length <= MAX_OLD_TOOL_STUB)
-                        continue;
-                    const stubbed = stubToolCallArguments(args);
-                    if (stubbed !== args) {
-                        tc.function.arguments = stubbed;
-                        microcompactDone = true;
-                    }
-                }
-            }
-        }
+        if (microcompactGroup(g))
+            microcompactDone = true;
     }
     // 第二层:摘要——把旧区(微压缩后)压成一条 system 摘要
     const older = flattenGroups(oldGroups);
@@ -450,14 +425,14 @@ export async function compactHistory(history, opts) {
  * 自动压缩门槛:agent 每步调 chat() 前调用。
  * 用全量启发式估算(始终可用、安全侧、无 stale-usage 问题);超阈则压缩。
  *
- * 升级到 Budget Scheduler:可传 `report: BudgetReport`。传了就**按 ROI 调度**:
- *   - 只有当 `report.layers.history.overBudget` 或 `report.totalOver` 时才压;
- *   - 冷工具超(Cold Tool ROI 最低)→ 不压 history,留给调度器的 cold tools 路径处理。
- * 这样 cold tools 路径(L1 中截 / L2 relevance / L3 age stub)能先动,history
- * 摘要(最贵)只在 cold tools 解不开时才触发。
+ * The shared 80% pressure threshold is the only automatic trigger. Raw estimates
+ * participate so correction<1 cannot hide an oversized provider request.
+ *
+ * Budget Scheduler 模式在同一次 80% pressure 事件中先完成已启用的
+ * superseded / stale artifact / logs-search 清理，然后无条件继续压缩历史。
  *
- * 不传 report 时退化为原行为:仅看总占用是否超 `compactThreshold * window`——
- * 兼容老调用方(子 agent / 测试直接调时),零行为变化。
+ * 不传 report 时使用同一个共享阈值，供子 agent 和关闭 scheduler 时 fallback；
+ * 不再维护独立的自动压缩触发线。
  *
  * manual 选项(repl /compact 用):true 时旁路 autoCompact 开关与 ROI 阈,
  * 强制走 compactHistory(manual/force 参数透传)。返 CompactResult 给 caller 文案展示。
@@ -467,28 +442,37 @@ export async function maybeCompact(history, report, manualOpts, state = contextS
     const est = estimatePromptTokens(history, activeTools, state.correction);
     state.lastEstimate = est;
     const isManual = manualOpts?.manual === true;
+    // Hard guard: the scheduler has already attempted pressure cleanup; if raw
+    // content remains near the provider limit, compact even when correction < 1.
+    let hardCap = typeof report?.rawTotal === 'number'
+        && typeof report?.window === 'number'
+        && report.rawTotal >= DEFAULT_BUDGET_POLICY.pressureTriggerRatio * report.window;
     // 手动路径:旁路 autoCompact / report / 总阈三重门
     if (!isManual) {
-        if (!config.autoCompact)
+        if (!hardCap && !config.autoCompact)
             return;
         if (report) {
-            // 调度模式:只压 history 层超或兜底总超
-            const needsCompact = report.layers.history.overBudget || report.totalOver;
+            // Scheduler mode: the shared pressure report is the sole automatic trigger.
+            const needsCompact = hardCap || report.totalOver;
             if (!needsCompact)
                 return;
         }
         else {
-            // 老路径:总占用超阈才压
-            if (est < config.compactThreshold * config.contextWindowTokens)
+            // Fallback mode uses the same threshold; there is no second compact gate.
+            const raw = estimatePromptTokens(history, activeTools, 1);
+            hardCap = raw >= DEFAULT_BUDGET_POLICY.pressureTriggerRatio * config.contextWindowTokens;
+            const pressureLine = DEFAULT_BUDGET_POLICY.pressureTriggerRatio * config.contextWindowTokens;
+            if (!hardCap && est < pressureLine)
                 return;
         }
     }
     const r = await compactHistory(history, {
         window: config.contextWindowTokens,
-        threshold: config.compactThreshold,
+        threshold: DEFAULT_BUDGET_POLICY.pressureTriggerRatio,
         focus: manualOpts?.focus,
         manual: isManual,
-        force: manualOpts?.force,
+        // 硬闸触发时 force 直通保护区:首轮/当前轮不豁免。
+        force: manualOpts?.force === true || hardCap,
         tools: activeTools,
         contextState: state,
     });

package/dist/session/index.js

@@ -10,7 +10,6 @@ export { compactHistory, maybeCompact, capToolResultForHistory, truncateMid, con
 // 只执行可落地的 warn / compact_history；开关关闭时退化为 maybeCompact。
 // repl /compact 命令调 manualCompact(history, focus?):与自动路径共享决策，focus 透传摘要 prompt。
 export { runScheduler, manualCompact, createBudgetScheduler, } from './scheduler.js';
-export { dropContextFromHistory, formatDropResult, } from './drop.js';
 export { newSessionId, saveSession, loadSession, listSessions, sessionDir, } from './persist.js';
 export { appendCurrentSessionTrace, appendCurrentSessionTraceEvent, appendCurrentSessionRuntimeEvent, createTraceEvent, } from './trace.js';
 export { reduceTraceMetrics, readTraceEvents, readTraceMetrics } from './trace-metrics.js';

package/dist/session/scheduler.js

@@ -1,159 +1,155 @@
-// Context Budget Scheduler(执行层):评估五区预算并执行可落地的动作。
-//
-//   agent/core.ts (步前)              repl /compact 命令(手动)
-//     ↓ runScheduler(history, step)    ↓ manualCompact(history, focus?)
-//   session/scheduler.ts(本文件)
-//     ├─ evaluateBudget(history, window, step) → BudgetReport
-//     ├─ scheduleActions(report) → warn | compact_history
-//     └─ 执行 actions:
-//          - warn:            仅写日志
-//          - compact_history: 调 maybeCompact(history, report)
-//
-// push-time cap / relevance / lifecycle / age-aware sweep 在进入 scheduler 前独立完成，
-// scheduler 不再生成无法执行的 Cold/Hot tool action。每次决策写入 actionLog，供 /context 调试。
-// manualCompact 与自动路径共享 scheduleActions，但用户显式触发时强制追加 compact_history。
-//
-// 开关:
-//   - config.contextBudget !== false(默认 true):agent 调 runScheduler
-//   - 关时 agent 仍走原 maybeCompact(history)无 report 路径,零行为变化
-//   - 手动 /compact 走 manualCompact;关时退化直接调 compactHistory(history, { focus })
-import { evaluateBudget, scheduleActions, formatReport, } from '../context/budget.js';
+// Context Budget Scheduler: normal operation is metadata-only. History content is
+// rewritten only after real occupancy reaches the shared pressure threshold.
+import { evaluateBudget, scheduleActions, formatReport, DEFAULT_BUDGET_POLICY, } from '../context/budget.js';
 import { chatTools } from '../llm/index.js';
 import { config } from '../config/index.js';
 import { maybeCompact, contextState } from './compact.js';
 import * as layout from '../ui/layout.js';
 import { ui } from '../ui/theme.js';
 import { pruneStaleArtifacts, refreshArtifactFreshness } from '../context/artifacts.js';
-/** 创建 runAgentCore 闭包持有的 scheduler(每次 agent 启动一个新实例)。 */
+import { pruneSuperseded } from '../context/relevance.js';
+import { createAgeAwareEncodingState } from '../context/age-aware.js';
+function atPressure(report) {
+    return Math.max(report.rawTotal, report.total)
+        >= report.window * DEFAULT_BUDGET_POLICY.pressureTriggerRatio;
+}
+function emptyPressure(report) {
+    return {
+        triggered: false,
+        before: report.total,
+        after: report.total,
+        superseded: 0,
+        staleArtifacts: 0,
+        encodedLogsAndSearches: 0,
+    };
+}
+/** One scheduler instance is owned by one agent run. */
 export function createBudgetScheduler(state = contextState) {
-    const obs = {
+    const evaluate = (history, step, activeTools) => evaluateBudget(history, config.contextWindowTokens, step, state.correction, activeTools);
+    const scheduler = {
         lastRunLog: null,
         async runStep(history, step, activeTools = chatTools) {
-            // Re-check file-backed hashes first, then discard stale/rebuildable facts before budgeting.
+            // External file changes and mutations only update artifact metadata here.
             refreshArtifactFreshness(state, history);
-            pruneStaleArtifacts(state, history);
-            const report = evaluateBudget(history, config.contextWindowTokens, step, state.correction, activeTools);
+            const report = evaluate(history, step, activeTools);
+            const pressure = emptyPressure(report);
+            pressure.triggered = atPressure(report);
+            if (pressure.triggered) {
+                // A single 80% pressure event owns every history rewrite. Run all enabled
+                // low-cost cleanup first, then always compact; do not introduce per-stage
+                // thresholds or stop early when one stage happens to cross below 80%.
+                if (config.contextRelprune) {
+                    pressure.superseded = pruneSuperseded(history, report.hotBoundary);
+                }
+                pressure.staleArtifacts = pruneStaleArtifacts(state, history, report.hotBoundary);
+                if (config.contextOptimize) {
+                    const ageAware = createAgeAwareEncodingState(history);
+                    pressure.encodedLogsAndSearches = ageAware.sweepPressure(history, report.hotBoundary);
+                }
+                pressure.after = evaluate(history, step, activeTools).total;
+            }
+            // Use the trigger report intentionally: cleanup may reduce the current estimate,
+            // but crossing 80% commits this step to compacting for maximum token savings.
             const actions = scheduleActions(report);
             let compactHistoryCalled = false;
             let historyRebuilt = false;
-            for (const a of actions) {
-                if (a.kind === 'warn') {
-                    // system 超:写一行提示(配置漂移应由用户处理,不是调度器压)
-                    layout.contentWrite(`  ${ui.accent}●${ui.reset} ${ui.yellow}调度器警告 [${a.layer}] ${a.reason}${ui.reset}\n`);
-                }
-                else if (a.kind === 'compact_history') {
-                    // 路由到 maybeCompact；把结构重建信号传回 core，使 lifecycle 按新 index 恢复。
-                    const result = await maybeCompact(history, report, undefined, state, activeTools);
-                    compactHistoryCalled = true;
-                    historyRebuilt ||= result?.historyRebuilt === true;
+            for (const action of actions) {
+                if (action.kind === 'warn') {
+                    layout.contentWrite(`  ${ui.accent}●${ui.reset} ${ui.yellow}调度器警告 [${action.layer}] ${action.reason}${ui.reset}\n`);
+                    continue;
                 }
+                const result = await maybeCompact(history, report, undefined, state, activeTools);
+                compactHistoryCalled = true;
+                historyRebuilt ||= result?.historyRebuilt === true;
             }
             const log = {
                 step,
                 report,
                 actions,
+                pressure,
                 compactHistoryCalled,
                 ts: Date.now(),
             };
-            obs.lastRunLog = log;
-            // 暴露给 /context 共享读(repl / context 命令)
+            scheduler.lastRunLog = log;
             state.schedulerLog = log;
             return historyRebuilt;
         },
     };
-    return obs;
+    return scheduler;
 }
-/** 便捷:agent/core.ts 不需要每次 createBudgetScheduler,直接 runScheduler(history, step)。 */
 export async function runScheduler(history, step, state = contextState, activeTools = chatTools) {
-    const s = createBudgetScheduler(state);
-    return s.runStep(history, step, activeTools);
+    return createBudgetScheduler(state).runStep(history, step, activeTools);
 }
-/** 手动 /compact 入口(repl):与自动路径共享预算评估和可执行 action，但强制执行 history 摘要。
- *  即便 layers.history.overBudget=false 或 totalOver=false，manual 仍追加 compact_history，
- * 并把 focus 透传给 LLM 摘要 prompt。
- *
- *  关系:runScheduler 是「自动触发」,manualCompact 是「用户显式触发」,二者共享 scheduleActions。
- *
- *  force=true:即便 oldGroups 空(history 全在保护区)也强行把早期消息降级压一次。
- *  适合"history 太长,自动阈值从未触发,但用户想强制压"的场景。
- *
- *  退化:config.contextBudget === false 时直接调 compactHistory(history, { focus }),与改造前等价。
- *  返回 SchedulerRunLog + compactDetail 字段,供 repl 文案展示"为什么没压"。 */
+/** User-requested compaction bypasses automatic pressure gating. */
 export async function manualCompact(history, focus, opts) {
     if (config.contextBudget === false) {
-        // 退化:不经调度器,直压 history(等价于改造前,但带 manual/force 透传)
-        const r = await import('./compact.js').then(({ compactHistory }) => compactHistory(history, {
+        const result = await import('./compact.js').then(({ compactHistory }) => compactHistory(history, {
             window: config.contextWindowTokens,
-            threshold: config.compactThreshold,
+            threshold: DEFAULT_BUDGET_POLICY.pressureTriggerRatio,
             focus,
             manual: true,
             force: opts?.force,
         }));
-        return {
+        const report = evaluateBudget(history, config.contextWindowTokens, -1, contextState.correction);
+        const log = {
             step: -1,
-            report: evaluateBudget(history, config.contextWindowTokens, -1, contextState.correction),
+            report,
             actions: [{ kind: 'compact_history', focus }],
+            pressure: emptyPressure(report),
             compactHistoryCalled: true,
             ts: Date.now(),
             compactDetail: {
-                reason: r.reason,
-                estimateBefore: r.estimateBefore,
-                estimateAfter: r.estimateAfter,
-                protectedRatio: r.protectedRatio,
-                oldGroupCount: r.oldGroupCount,
+                reason: result.reason,
+                estimateBefore: result.estimateBefore,
+                estimateAfter: result.estimateAfter,
+                protectedRatio: result.protectedRatio,
+                oldGroupCount: result.oldGroupCount,
                 focus,
             },
         };
+        contextState.schedulerLog = log;
+        return log;
     }
     const report = evaluateBudget(history, config.contextWindowTokens, -1, contextState.correction);
     let actions = scheduleActions(report);
-    // 用户显式说「要压」:即使 report 不含 history 触发,仍追加 compact_history
-    const hasCompact = actions.some(a => a.kind === 'compact_history');
-    if (!hasCompact) {
+    if (!actions.some((action) => action.kind === 'compact_history')) {
         actions = [...actions, { kind: 'compact_history', focus }];
     }
     else if (focus) {
-        // 已有 compact_history(action 由 scheduleActions 产,无 focus)— 注入 focus
-        actions = actions.map(a => a.kind === 'compact_history' ? { ...a, focus } : a);
+        actions = actions.map((action) => action.kind === 'compact_history' ? { ...action, focus } : action);
     }
-    const s = createBudgetScheduler();
-    // 直接驱动 runStep 的执行逻辑(action list 我们已自己生成)
     let compactHistoryCalled = false;
     let compactDetail;
-    for (const a of actions) {
-        if (a.kind === 'warn') {
-            // 复用现有 contentWrite 路径(经 scheduler.runStep);这里略,只记录到 log
-        }
-        else if (a.kind === 'compact_history') {
-            const r = await maybeCompact(history, report, {
-                manual: true,
-                force: opts?.force,
-                focus: a.focus,
-            });
-            compactHistoryCalled = true;
-            if (r) {
-                compactDetail = {
-                    reason: r.reason,
-                    estimateBefore: r.estimateBefore,
-                    estimateAfter: r.estimateAfter,
-                    protectedRatio: r.protectedRatio,
-                    oldGroupCount: r.oldGroupCount,
-                    focus: a.focus,
-                };
-            }
+    for (const action of actions) {
+        if (action.kind !== 'compact_history')
+            continue;
+        const result = await maybeCompact(history, report, {
+            manual: true,
+            force: opts?.force,
+            focus: action.focus,
+        });
+        compactHistoryCalled = true;
+        if (result) {
+            compactDetail = {
+                reason: result.reason,
+                estimateBefore: result.estimateBefore,
+                estimateAfter: result.estimateAfter,
+                protectedRatio: result.protectedRatio,
+                oldGroupCount: result.oldGroupCount,
+                focus: action.focus,
+            };
         }
     }
     const log = {
         step: -1,
         report,
         actions,
+        pressure: emptyPressure(report),
         compactHistoryCalled,
         ts: Date.now(),
         compactDetail,
     };
-    s.lastRunLog = log;
     contextState.schedulerLog = log;
     return log;
 }
-// ── 调试导出:用于 scripts/check-budget.ts 把报告打到 stdout 调试 ────────
 export { evaluateBudget, scheduleActions, formatReport };

package/dist/session/trace-metrics.js

@@ -1,99 +1,16 @@
 import { readFileSync } from 'node:fs';
-/** 把 history 中所有 user/tool/assistant 消息的文本/参数汇总到一个大字符串,便于
- * 用单一正则扫描多个 marker。比逐条消息 join 更稳:LLM 流式 contentPart 数组
- * 与多模态混合时,joinStringField 内部把每个 part 展平成字符串。 */
-function joinHistoryForMarkerScan(history) {
-    if (!history)
-        return '';
-    const parts = [];
-    for (const message of history) {
-        if (typeof message.content === 'string') {
-            parts.push(message.content);
-        }
-        else if (Array.isArray(message.content)) {
-            for (const part of message.content) {
-                if (typeof part === 'string')
-                    parts.push(part);
-                else if (part && typeof part === 'object' && 'text' in part && typeof part.text === 'string') {
-                    parts.push(part.text);
-                }
-            }
-        }
-        if ('tool_call_id' in message && typeof message.tool_call_id === 'string') {
-            parts.push(message.tool_call_id);
-        }
-    }
-    return parts.join('\n');
-}
-/** QUAL-01 质量维度。
- *
- * 优先用硬事件(retry_reflection / checklist_triggered / ask_human_call),
- * 这是 core.ts 在接缝处显式 emit 的,可信度高。
- * 如果 events 里没有对应硬事件(老 trace / 旧 fixture 走 history 文本回放),fallback
- * 到 history 文本扫描 + tool_call_end(name === 'ask_human', status === 'success')。
- *
- * - `retry_reflection` → 反思重试注入次数(RETRY-01)
- * - `checklist_triggered` → PROMPT-02 触发次数
- * - `ask_human_call` (status === 'success') → ASK-01 触发次数
- *
- * 三类都给出 fallback 0,确保报告维度永远是 number,不会被 NaN 污染。 */
-function reduceQualityDimensions(events, history) {
-    // 1. 硬事件计数(优先级最高)。
-    let reflectionRounds = 0;
-    let checklistTriggered = 0;
-    let askHumanCount = 0;
-    let hasHardReflection = false;
-    let hasHardChecklist = false;
-    let hasHardAskHuman = false;
-    for (const event of events) {
-        if (event.type === 'retry_reflection') {
-            reflectionRounds += 1;
-            hasHardReflection = true;
-        }
-        else if (event.type === 'checklist_triggered') {
-            checklistTriggered += 1;
-            hasHardChecklist = true;
-        }
-        else if (event.type === 'ask_human_call' && event.data.status === 'success') {
-            askHumanCount += 1;
-            hasHardAskHuman = true;
-        }
-    }
-    // 2. fallback: 历史事件没有硬信号时,从 history 文本 + tool_call_end 推断。
-    //    保留旧 fixture / 回放 JSONL 的兼容能力。
-    if (!hasHardReflection || !hasHardChecklist || !hasHardAskHuman) {
-        if (!hasHardReflection) {
-            const historyText = joinHistoryForMarkerScan(history);
-            reflectionRounds += (historyText.match(/\[retry reflection:/g) ?? []).length;
-        }
-        if (!hasHardChecklist) {
-            const historyText = joinHistoryForMarkerScan(history);
-            checklistTriggered += (historyText.match(/\[checklist\]/g) ?? []).length;
-        }
-        if (!hasHardAskHuman) {
-            for (const event of events) {
-                if (event.type !== 'tool_call_end')
-                    continue;
-                if (event.data.name === 'ask_human' && event.data.status === 'success') {
-                    askHumanCount += 1;
-                }
-            }
-        }
-    }
-    return { reflectionRounds, askHumanCount, checklistTriggered };
+function countAskHumanCalls(events) {
+    return events.filter((event) => event.type === 'ask_human_call' && event.data.status === 'success').length;
 }
-export function reduceTraceMetrics(events, history) {
+export function reduceTraceMetrics(events) {
     const ends = events.filter((event) => event.type === 'tool_call_end');
     let recovered = false;
     let hadFailure = false;
     let successes = 0;
-    let toolRetries = 0;
     let tokens = 0;
     let hasTokens = false;
     for (const event of ends) {
         const status = String(event.data.status ?? 'error');
-        const retry = Number(event.data.retry ?? 0);
-        toolRetries += Number.isFinite(retry) ? retry : 0;
         if (status === 'success') {
             successes++;
             if (hadFailure)
@@ -113,21 +30,17 @@ export function reduceTraceMetrics(events, history) {
         }
     }
     const modelRetries = events.filter((event) => event.type === 'model_retry').length;
-    const firstValidation = events.find((event) => event.type === 'validation_end');
     const turnEnd = [...events].reverse().find((event) => event.type === 'turn_end');
-    const quality = reduceQualityDimensions(events, history);
     return {
         toolCalls: events.filter((event) => event.type === 'tool_call_start').length,
         toolFailures: ends.length - successes,
         toolRecovery: recovered,
         firstSuccessRate: ends.length ? successes / ends.length : 1,
         modelRetries,
-        toolRetries,
-        retries: modelRetries + toolRetries,
+        retries: modelRetries,
         tokens: hasTokens ? tokens : null,
         durationMs: Number(turnEnd?.data.durationMs ?? 0),
-        firstValidationPassed: firstValidation?.data.status === 'passed',
-        ...quality,
+        askHumanCount: countAskHumanCalls(events),
     };
 }
 /** Reads event JSONL; malformed/legacy summary lines are ignored. */

package/dist/session/trace.js

@@ -41,14 +41,5 @@ export function appendCurrentSessionTrace(trace) {
     const sessionId = getCurrentSessionId();
     if (!sessionId)
         return;
-    const validation = trace.validation
-        ? {
-            status: trace.validation.status,
-            level: trace.validation.level,
-            durationMs: trace.validation.durationMs,
-            verificationComplete: trace.validation.verificationComplete,
-            fingerprint: trace.validation.fingerprint,
-        }
-        : undefined;
-    appendTraceLine(sessionId, { ...trace, sessionId, validation });
+    appendTraceLine(sessionId, { ...trace, sessionId });
 }