@bolloon/bolloon-agent 0.1.35 → 0.1.36

package/src/context-compaction/auto-compact.ts

@@ -0,0 +1,153 @@
+/**
+ * Auto-Compact — 第 5 层: LLM 摘要 (最后手段)
+ *
+ * 论文:
+ *   - 兜底层, budget > 100% 时才触发
+ *   - 用 LLM 生成完整摘要, 替代前 K 对 user/assistant
+ *   - 摘要 prompt 明确要求"必须保留出现过的 tool names"
+ *   - 缓存到 ~/.bolloon/sessions/<channel>/compaction-cache.json
+ *   - LLM 失败 → fallback 保留原文
+ *
+ * 关键不变量: 这是**破坏性**的 (折叠 N 对成 1 条), 调用方接受新 history
+ *
+ * 失败静默: LLM 失败 → 返回原 history (与现状一致)
+ */
+
+import * as crypto from 'crypto';
+import * as fs from 'fs/promises';
+import * as os from 'os';
+import * as path from 'path';
+import type { Message, StageResult, StageOptions } from './types.js';
+
+const DEFAULT_COLLAPSE_PAIRS = 5;
+const CACHE_DIR = () => path.join(process.env.HOME || os.homedir() || '/tmp', '.bolloon', 'sessions', 'compaction-cache');
+
+function cacheKey(history: Message[], collapsePairs: number): string {
+  const slice = history.slice(0, collapsePairs * 2);
+  const joined = slice.map((m) => `${m.role}:${m.content}`).join('|');
+  return crypto.createHash('sha1').update(joined).digest('hex').slice(0, 12);
+}
+
+async function readCache(scope: string, key: string): Promise<string | null> {
+  try {
+    const file = path.join(CACHE_DIR(), `${scope}_${key}.json`);
+    const raw = await fs.readFile(file, 'utf-8');
+    const obj = JSON.parse(raw);
+    if (typeof obj.summary === 'string' && Date.now() - obj.ts < 24 * 60 * 60 * 1000) {
+      return obj.summary;
+    }
+  } catch { /* cache miss or parse err */ }
+  return null;
+}
+
+async function writeCache(scope: string, key: string, summary: string): Promise<void> {
+  try {
+    await fs.mkdir(CACHE_DIR(), { recursive: true });
+    const file = path.join(CACHE_DIR(), `${scope}_${key}.json`);
+    await fs.writeFile(file, JSON.stringify({ ts: Date.now(), summary, key }), 'utf-8');
+  } catch (err) {
+    console.warn('[compactor] cache write failed (silent):', err);
+  }
+}
+
+const SUMMARIZE_SYSTEM = `你是一个对话历史压缩助手.
+你的任务: 把给定的对话历史压缩成 1-2 句话的摘要, 保留所有出现过的 tool names.
+要求:
+- 保留所有用户的关键需求
+- 保留所有调用过的 tool names (例如 read_file, shell_exec, use_skill 等)
+- 保留关键决策和发现
+- 不要添加原对话中没有的信息
+- 输出纯文本, 不要 JSON, 不要 markdown 标题`;
+
+export async function autoCompact(history: Message[], opts: StageOptions = {}): Promise<StageResult> {
+  try {
+    if (opts.skip) return { history, applied: false, detail: 'skipped' };
+    const collapsePairs = opts.autoCompactCollapsePairs ?? DEFAULT_COLLAPSE_PAIRS;
+
+    // 找前 K 对的边界
+    let pairsFound = 0;
+    let cutTo = -1;
+    for (let i = 0; i < history.length; i++) {
+      if (history[i].role === 'user') {
+        pairsFound++;
+        if (pairsPairsOver(pairsFound, collapsePairs)) {
+          cutTo = i;
+          break;
+        }
+      }
+    }
+    if (cutTo <= 0) {
+      return { history, applied: false, detail: 'history too short' };
+    }
+
+    const toCollapse = history.slice(0, cutTo);
+    const remaining = history.slice(cutTo);
+    const scope = opts.cacheScope || 'default';
+
+    // 1. 查缓存 (key 基于实际被切的消息, 不用理论 collapsePairs)
+    const key = crypto.createHash('sha1').update(
+      toCollapse.map((m) => `${m.role}:${m.content}`).join('|')
+    ).digest('hex').slice(0, 12);
+    let summary = await readCache(scope, key);
+
+    // 2. 没缓存就调 LLM
+    if (!summary) {
+      if (!opts.llmChat) {
+        // 没有 LLM 就 fallback 保留原文
+        console.warn('[compactor] autoCompact: no llmChat injected, returning original');
+        return { history, applied: false, detail: 'no llmChat available' };
+      }
+      try {
+        const userPrompt = `以下是 ${cutTo} 条较早的对话历史, 请压缩为 1-2 句摘要:\n\n${formatForSummary(toCollapse)}`;
+        summary = await opts.llmChat(SUMMARIZE_SYSTEM, userPrompt);
+        if (!summary || summary.trim().length === 0) {
+          return { history, applied: false, detail: 'llm returned empty' };
+        }
+        // 写缓存 (静默, 但 await 确保测试可重现)
+        try { await writeCache(scope, key, summary); } catch { /* ignored */ }
+      } catch (err) {
+        console.warn('[compactor] autoCompact LLM call failed (silent, fallback):', err);
+        return { history, applied: false, detail: 'llm call failed' };
+      }
+    }
+
+    // 3. 折叠: 1 条 summary 消息替代 N 条原文
+    const summaryMsg: Message = {
+      role: 'system',
+      content: `[Auto-Compact Summary] ${summary}`,
+    } as any;
+    return {
+      history: [summaryMsg, ...remaining],
+      applied: true,
+      detail: `collapsed ${cutTo} messages into 1 summary (cache key: ${key})`,
+    };
+  } catch (err) {
+    console.warn('[compactor] autoCompact failed (silent, returning original):', err);
+    return { history, applied: false, detail: 'error' };
+  }
+}
+
+function pairsPairsOver(found: number, target: number): boolean {
+  return found > target;
+}
+
+function formatForSummary(messages: Message[]): string {
+  return messages.map((m) => {
+    if (m.role === 'user') return `用户: ${m.content}`;
+    if (m.role === 'assistant') return `助手: ${m.content.substring(0, 500)}`;
+    if (m.role === 'tool') {
+      const r = (m as any).toolResult;
+      return `工具: ${typeof r === 'string' ? r : JSON.stringify(r).substring(0, 300)}`;
+    }
+    return m.content;
+  }).join('\n');
+}
+
+// ============================================================
+// 测试钩子
+// ============================================================
+
+export function _resetAutoCompactCacheForTest(): void {
+  // 实际缓存由 readCache/writeCache 管理, 这里只暴露接口
+  // 测试可以走 fs.rm CACHE_DIR() 清理
+}

package/src/context-compaction/budget-gate.ts

@@ -0,0 +1,32 @@
+/**
+ * Budget Gate — 预算判定 (不修改消息)
+ *
+ * 论文: 这是判定阶段, 不是压缩阶段. 触发条件:
+ *   - ratio < 0.8   → fit, 不压缩
+ *   - 0.8 ≤ ratio ≤ 1.0 → triggerNextLayer (跑下一层)
+ *   - ratio > 1.0   → triggerNextLayer, 全部层都跑
+ *
+ * 失败静默: 任何异常 → 视为 fit (与现状一致, 不冒险压缩)
+ */
+
+import type { Message } from './types.js';
+import { estimateTokens } from './token-estimator.js';
+import type { BudgetReport } from './types.js';
+
+const FIT_THRESHOLD = 0.8;
+
+export function budgetGate(history: Message[], maxTokens: number = 8000): BudgetReport {
+  try {
+    const estimatedTokens = estimateTokens(history as any);
+    const ratio = maxTokens > 0 ? estimatedTokens / maxTokens : 0;
+    return {
+      fit: ratio < FIT_THRESHOLD,
+      estimatedTokens,
+      ratio,
+      triggerNextLayer: ratio >= FIT_THRESHOLD,
+    };
+  } catch (err) {
+    console.warn('[compactor] budgetGate failed (silent, treating as fit):', err);
+    return { fit: true, estimatedTokens: 0, ratio: 0, triggerNextLayer: false };
+  }
+}

package/src/context-compaction/budget-reduce.ts

@@ -0,0 +1,37 @@
+/**
+ * Budget Reduction — 第 1 层: 单条消息大小限制
+ *
+ * 论文: 总是跑. 截断 > 阈值的单条消息, 加 ... [truncated, original=N chars]
+ *
+ * 关键: 不破坏 messageHistory 内存结构, 只改 buildContext 输出
+ * 但因为 buildContext 需要最终字符串, 这里实际返回新 history
+ * (内容相同的引用 = 内存未真正变更)
+ *
+ * 失败静默: 异常 → 返回原 history
+ */
+
+import type { Message, StageResult, StageOptions } from './types.js';
+
+const DEFAULT_MAX_CHARS = 4000;
+
+export function budgetReduce(history: Message[], opts: StageOptions = {}): StageResult {
+  try {
+    if (opts.skip) return { history, applied: false, detail: 'skipped' };
+    const maxChars = opts.budgetReduceMaxChars ?? DEFAULT_MAX_CHARS;
+    let changed = false;
+    const out: Message[] = history.map((m) => {
+      if (m.content && m.content.length > maxChars) {
+        changed = true;
+        return {
+          ...m,
+          content: m.content.substring(0, maxChars) + `\n... [truncated, original=${m.content.length} chars]`,
+        };
+      }
+      return m;
+    });
+    return { history: out, applied: changed, detail: changed ? `truncated >${maxChars} chars` : 'no oversized messages' };
+  } catch (err) {
+    console.warn('[compactor] budgetReduce failed (silent, returning original):', err);
+    return { history, applied: false, detail: 'error' };
+  }
+}

package/src/context-compaction/context-collapse.ts

@@ -0,0 +1,72 @@
+/**
+ * Context Collapse — 第 4 层: 读时虚拟投影 (非破坏)
+ *
+ * 论文:
+ *   - 不修改 messageHistory
+ *   - 在 buildContext() 输出阶段, 前 K 对 user/assistant 替换为 [collapsed: <summary>]
+ *   - feature flag BOLLOON_CONTEXT_COLLAPSE 默认关闭
+ *
+ * 实现策略:
+ *   - 由于 buildContext 阶段才输出最终字符串, 这里"虚拟投影"实际意味着:
+ *     返回一个新 history, 前 K 对用一个虚拟的 "collapsed" summary 消息替代
+ *   - 但 memory 层的 messageHistory 不变 (由调用方负责回滚或仅在 buildContext 内用)
+ *
+ * 实践: 本 stage 返回一个标记 collapsed=true 的 history, 调用方可以选择:
+ *   (a) 仅在 buildContext 字符串拼接时替换 (理想)
+ *   (b) 临时替换 (本次 LLM call 结束后回滚)
+ *   (c) 永久替换 (但这违背"非破坏"原则)
+ *
+ * 本模块选 (b): 返回新 history, 标记 collapsedMeta 字段, 文档说明调用方应仅用于构建 LLM 输入
+ *
+ * 失败静默: 异常 → 返回原 history
+ */
+
+import type { Message, StageResult, StageOptions } from './types.js';
+
+const DEFAULT_COLLAPSE_PAIRS = 5;
+const DEFAULT_SUMMARY = '[collapsed: N earlier rounds — summary not yet generated]';
+
+export function isContextCollapseEnabled(): boolean {
+  return process.env.BOLLOON_CONTEXT_COLLAPSE === '1';
+}
+
+export function contextCollapse(history: Message[], opts: StageOptions = {}): StageResult {
+  try {
+    if (opts.skip || !isContextCollapseEnabled()) {
+      return { history, applied: false, detail: isContextCollapseEnabled() ? 'skipped' : 'feature flag off' };
+    }
+    const collapsePairs = opts.contextCollapseCollapsePairs ?? DEFAULT_COLLAPSE_PAIRS;
+
+    // 找前 K 对 user/assistant 的边界
+    let pairsFound = 0;
+    let cutTo = -1;
+    for (let i = 0; i < history.length; i++) {
+      if (history[i].role === 'user') {
+        pairsFound++;
+        if (pairsFound > collapsePairs) {
+          cutTo = i;
+          break;
+        }
+      }
+    }
+    if (cutTo <= 0) {
+      return { history, applied: false, detail: 'history too short' };
+    }
+
+    // 把前 cutTo 条压成 1 条 virtual summary message
+    const collapsedCount = cutTo;
+    const summaryMsg: Message = {
+      role: 'system',
+      content: `[collapsed: ${collapsedCount} earlier messages (${collapsePairs} pairs) — virtual projection, original history preserved]`,
+    } as any;
+    const out: Message[] = [summaryMsg, ...history.slice(cutTo)];
+    return {
+      history: out,
+      applied: true,
+      detail: `virtual-collapsed first ${collapsedCount} messages`,
+    };
+  } catch (err) {
+    console.warn('[compactor] contextCollapse failed (silent, returning original):', err);
+    return { history, applied: false, detail: 'error' };
+  }
+}

package/src/context-compaction/index.ts

@@ -0,0 +1,24 @@
+/**
+ * Context Compaction — 公共导出 (barrel)
+ *
+ * 设计: 严格对齐 Claude Code 论文 5 层压缩流水线
+ *   1. Budget Reduction   (单条消息大小限制)
+ *   2. Snip               (裁掉老历史, feature flag 关闭)
+ *   3. Microcompact       (cache-aware 细粒度压缩)
+ *   4. Context Collapse   (读时虚拟投影, feature flag 关闭)
+ *   5. Auto-Compact       (LLM 摘要, 兜底)
+ *
+ * 主要入口: `compactPipeline(history, opts)`
+ * 失败静默: 任何 stage 抛错 → pipeline 返回原 history
+ */
+
+export type { Message, StageResult, StageOptions, PipelineResult, BudgetReport, CompactionContext } from './types.js';
+
+export { compactPipeline, _resetPipelineForTest } from './pipeline.js';
+export { budgetGate } from './budget-gate.js';
+export { budgetReduce } from './budget-reduce.js';
+export { snip, isSnipEnabled } from './snip.js';
+export { microcompact } from './microcompact.js';
+export { contextCollapse, isContextCollapseEnabled } from './context-collapse.js';
+export { autoCompact, _resetAutoCompactCacheForTest } from './auto-compact.js';
+export { estimateTokens, estimateStringTokens } from './token-estimator.js';

package/src/context-compaction/microcompact.ts

@@ -0,0 +1,54 @@
+/**
+ * Microcompact — 第 3 层: cache-aware 细粒度压缩
+ *
+ * 论文: 总是跑 (时间触发), 关键作用是保留 prompt cache 命中
+ * 实现 (简化版): 把老的 toolResult 折叠为 "(tool result: <name>, N chars)"
+ *   - 保留最近 N 条 tool_result 完整 (默认 3)
+ *   - 之前的折叠
+ *
+ * 关键: 不破坏 messageHistory 内存结构
+ *
+ * 失败静默: 异常 → 返回原 history
+ */
+
+import type { Message, StageResult, StageOptions } from './types.js';
+
+const DEFAULT_KEEP_RECENT = 3;
+
+export function microcompact(history: Message[], opts: StageOptions = {}): StageResult {
+  try {
+    if (opts.skip) return { history, applied: false, detail: 'skipped' };
+    const keepRecent = opts.microcompactKeepRecent ?? DEFAULT_KEEP_RECENT;
+
+    // 找所有 tool 消息的索引
+    const toolIndices: number[] = [];
+    for (let i = 0; i < history.length; i++) {
+      if (history[i].role === 'tool') toolIndices.push(i);
+    }
+    if (toolIndices.length <= keepRecent) {
+      return { history, applied: false, detail: 'no old tool_results to fold' };
+    }
+
+    // 要折叠的 = 前 (total - keepRecent) 条
+    const toFold = new Set(toolIndices.slice(0, toolIndices.length - keepRecent));
+    let changed = false;
+    const out: Message[] = history.map((m, i) => {
+      if (toFold.has(i) && m.role === 'tool') {
+        changed = true;
+        // 尝试从相邻的 assistant 消息拿 tool name
+        const name = (m as any).toolName || 'unknown';
+        const originalLen = m.content?.length || 0;
+        return {
+          ...m,
+          content: `(tool result: ${name}, ${originalLen} chars — folded by microcompact)`,
+          toolResult: undefined,
+        };
+      }
+      return m;
+    });
+    return { history: out, applied: changed, detail: `folded ${toFold.size} old tool_results, kept last ${keepRecent}` };
+  } catch (err) {
+    console.warn('[compactor] microcompact failed (silent, returning original):', err);
+    return { history, applied: false, detail: 'error' };
+  }
+}

package/src/context-compaction/pipeline.ts

@@ -0,0 +1,133 @@
+/**
+ * Context Compaction Pipeline — 5 层编排
+ *
+ * 严格对齐 Claude Code 论文: 每次 LLM call 前顺序执行 5 个 shaper, 由便宜到重, 任一层 fit 就短路.
+ *
+ *   1. Budget Reduction   (总是, 截断 > 阈值的单条消息)
+ *   2. Snip               (feature flag BOLLOON_SNIP_ENABLED)
+ *   3. Microcompact       (总是, 折叠老 tool_result)
+ *   4. Context Collapse   (feature flag BOLLOON_CONTEXT_COLLAPSE, 读时虚拟投影)
+ *   5. Auto-Compact       (LLM 摘要, 兜底)
+ *
+ * 短路: 任一层后 budget fit (< 0.8) → 返回该层结果, 不跑后续.
+ *
+ * 关键不变量:
+ *   - 任何异常 → 整 pipeline 返回原 history (与现状一致)
+ *   - 第 1-3 层不破坏 messageHistory 内存引用 (内容可改)
+ *   - 第 4 层读时投影, memory 不变
+ *   - 第 5 层破坏性, 折叠 N 对成 1 条
+ *
+ * 失败静默: 整 pipeline 用 try/catch 包, 任何 stage 失败不影响整体
+ */
+
+import type { Message, StageResult, StageOptions, PipelineResult, CompactionContext } from './types.js';
+import { budgetGate } from './budget-gate.js';
+import { budgetReduce } from './budget-reduce.js';
+import { snip, isSnipEnabled } from './snip.js';
+import { microcompact } from './microcompact.js';
+import { contextCollapse, isContextCollapseEnabled } from './context-collapse.js';
+import { autoCompact } from './auto-compact.js';
+
+const DEFAULT_MAX_TOKENS = 8000;
+
+export async function compactPipeline(
+  history: Message[],
+  opts: StageOptions = {}
+): Promise<PipelineResult> {
+  const maxTokens = opts.maxTokens ?? DEFAULT_MAX_TOKENS;
+  const ctx: CompactionContext = {
+    inputHistory: history,
+    history,
+    estimatedTokens: 0,
+    opts,
+    stages: [],
+  };
+
+  try {
+    // Stage 1: Budget Reduction
+    let result = budgetReduce(history, opts);
+    ctx.history = result.history;
+    ctx.stages.push(stageRecord('budgetReduce', result, history));
+    let gate = budgetGate(ctx.history, maxTokens);
+    ctx.estimatedTokens = gate.estimatedTokens;
+    if (gate.fit) return finalize(ctx, result.applied);  // 短路但可能 budgetReduce 实际改了内容
+
+    // Stage 2: Snip (feature flag)
+    if (isSnipEnabled()) {
+      const before = ctx.history;
+      result = snip(ctx.history, opts);
+      ctx.history = result.history;
+      ctx.stages.push(stageRecord('snip', result, before));
+      gate = budgetGate(ctx.history, maxTokens);
+      ctx.estimatedTokens = gate.estimatedTokens;
+      if (gate.fit) return finalize(ctx, true);
+    } else {
+      ctx.stages.push({ stage: 'snip', applied: false, before: ctx.history.length, after: ctx.history.length, detail: 'feature flag off' });
+    }
+
+    // Stage 3: Microcompact
+    const before3 = ctx.history;
+    result = microcompact(ctx.history, opts);
+    ctx.history = result.history;
+    ctx.stages.push(stageRecord('microcompact', result, before3));
+    gate = budgetGate(ctx.history, maxTokens);
+    ctx.estimatedTokens = gate.estimatedTokens;
+    if (gate.fit) return finalize(ctx, result.applied || true);
+
+    // Stage 4: Context Collapse (feature flag)
+    if (isContextCollapseEnabled()) {
+      const before4 = ctx.history;
+      result = contextCollapse(ctx.history, opts);
+      ctx.history = result.history;
+      ctx.stages.push(stageRecord('contextCollapse', result, before4));
+      gate = budgetGate(ctx.history, maxTokens);
+      ctx.estimatedTokens = gate.estimatedTokens;
+      if (gate.fit) return finalize(ctx, true);
+    } else {
+      ctx.stages.push({ stage: 'contextCollapse', applied: false, before: ctx.history.length, after: ctx.history.length, detail: 'feature flag off' });
+    }
+
+    // Stage 5: Auto-Compact (兜底)
+    const before5 = ctx.history;
+    result = await autoCompact(ctx.history, opts);
+    ctx.history = result.history;
+    ctx.stages.push(stageRecord('autoCompact', result, before5));
+
+    return finalize(ctx, true);
+  } catch (err) {
+    console.warn('[compactor] pipeline failed (silent, returning original):', err);
+    return {
+      history,
+      estimatedTokens: 0,
+      stages: ctx.stages,
+      compacted: false,
+    };
+  }
+}
+
+function stageRecord(name: string, result: StageResult, before: Message[]) {
+  return {
+    stage: name,
+    applied: result.applied,
+    before: before.length,
+    after: result.history.length,
+    detail: result.detail,
+  };
+}
+
+function finalize(ctx: CompactionContext, compacted: boolean): PipelineResult {
+  return {
+    history: ctx.history,
+    estimatedTokens: ctx.estimatedTokens,
+    stages: ctx.stages,
+    compacted,
+  };
+}
+
+// ============================================================
+// 测试钩子
+// ============================================================
+
+export function _resetPipelineForTest(): void {
+  // pipeline 是 stateless, 但保留 API 一致性
+}

package/src/context-compaction/snip.ts

@@ -0,0 +1,51 @@
+/**
+ * Snip — 第 2 层: 裁掉老 history
+ *
+ * 论文: feature flag 关闭 (默认), 用户开启后才生效
+ * 开启环境变量: BOLLOON_SNIP_ENABLED=1
+ *
+ * 行为: 只保留最近 N 对 user/assistant (默认 20 对)
+ * 工具结果消息按其所属 round 一起保留
+ *
+ * 失败静默: 异常 → 返回原 history
+ */
+
+import type { Message, StageResult, StageOptions } from './types.js';
+
+const DEFAULT_KEEP_PAIRS = 20;
+
+export function isSnipEnabled(): boolean {
+  return process.env.BOLLOON_SNIP_ENABLED === '1';
+}
+
+export function snip(history: Message[], opts: StageOptions = {}): StageResult {
+  try {
+    if (opts.skip || !isSnipEnabled()) {
+      return { history, applied: false, detail: isSnipEnabled() ? 'skipped' : 'feature flag off' };
+    }
+    const keepPairs = opts.snipKeepPairs ?? DEFAULT_KEEP_PAIRS;
+
+    // 从后往前数 N 对 user/assistant 边界
+    let pairsFound = 0;
+    let cutFrom = -1;
+    for (let i = history.length - 1; i >= 0; i--) {
+      const m = history[i];
+      if (m.role === 'user') {
+        pairsFound++;
+        if (pairsFound >= keepPairs) {
+          cutFrom = i;
+          break;
+        }
+      }
+    }
+
+    if (cutFrom <= 0) {
+      return { history, applied: false, detail: 'history too short' };
+    }
+    const out = history.slice(cutFrom);
+    return { history: out, applied: true, detail: `kept last ${keepPairs} pairs, dropped ${cutFrom} older messages` };
+  } catch (err) {
+    console.warn('[compactor] snip failed (silent, returning original):', err);
+    return { history, applied: false, detail: 'error' };
+  }
+}

package/src/context-compaction/token-estimator.ts

@@ -0,0 +1,36 @@
+/**
+ * Token Estimator — 4 字符 ≈ 1 token 的粗估
+ *
+ * 为什么不引 tiktoken 依赖:
+ * - bolloon 默认 npm 用户, 引依赖增加 install 体积
+ * - Claude Code 论文也指出: 精确 token 计数不是关键, 渐进压缩才是
+ * - 4 字符/token 在中文/英文混合场景下误差约 ±20%, 对预算判定足够
+ *
+ * 失败静默: 任何异常 → 返回估算值 (不抛错)
+ */
+
+const CHARS_PER_TOKEN = 4;
+
+export function estimateTokens(messages: Array<{ content: string; toolResult?: unknown; toolCall?: unknown }>): number {
+  let total = 0;
+  for (const m of messages) {
+    total += Math.ceil((m.content?.length || 0) / CHARS_PER_TOKEN);
+    if (m.toolResult) {
+      try {
+        const s = typeof m.toolResult === 'string' ? m.toolResult : JSON.stringify(m.toolResult);
+        total += Math.ceil(s.length / CHARS_PER_TOKEN);
+      } catch { /* circular ref, ignore */ }
+    }
+    if (m.toolCall) {
+      try {
+        const s = typeof m.toolCall === 'string' ? m.toolCall : JSON.stringify(m.toolCall);
+        total += Math.ceil(s.length / CHARS_PER_TOKEN);
+      } catch { /* ignore */ }
+    }
+  }
+  return total;
+}
+
+export function estimateStringTokens(s: string): number {
+  return Math.ceil(s.length / CHARS_PER_TOKEN);
+}