ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

extras/agent-lab/SCHEMA.md

@@ -0,0 +1,93 @@
+# Agent Trajectory JSON Schema
+
+A trajectory is a recording of one Agent run, used by the `<agent-trajectory>`
+Web Component to replay the ReAct loop step by step in the browser.
+
+The schema mirrors the `_emit(...)` calls in
+[`chapter1/web-search-agent/agent.py`](../../chapter1/web-search-agent/agent.py)
+so a real run can be exported into this format with almost no transformation.
+
+## Top-level object
+
+```jsonc
+{
+  "$schema": "../SCHEMA.md",
+  "experiment":  "ch1/web-search-agent",      // stable id, matches chapter/<exp>
+  "title":       "GPT-5.6 解「东盟 10 国首都最近距离」",
+  "model":       "gpt-5.6-sol",
+  "task":        "东盟 10 国首都之间，最近的一对首都距离多少？",
+  "condition":   "full-context",              // ablation condition, optional
+  "outcome":     "success",                   // success | failure | loop | timeout
+  "tags":        ["deep-research", "code-interp"],
+  "recorded_at": "2026-07-20T14:32:08Z",
+  "steps":       [ /* see below */ ]
+}
+```
+
+## Step types
+
+Every step has `iteration` (1-based) and `type`. The remaining fields depend
+on `type`. The four types correspond exactly to ReAct: Reasoning / Acting /
+Observing / final Answer.
+
+### `thought` — model's internal reasoning
+
+```jsonc
+{
+  "iteration": 1,
+  "type":      "thought",
+  "content":   "需要先找出东盟 10 国首都的名称，再查每对首都的距离……"
+}
+```
+
+`content` comes from the model's `reasoning_content` field (Kimi K3, GPT-5
+Reasoning, Claude thinking, …). May be long — the UI collapses it.
+
+### `action` — model called a tool
+
+```jsonc
+{
+  "iteration": 1,
+  "type":      "action",
+  "tool":      "$web_search",
+  "args":      { "query": "东盟 ASEAN 10 国首都 列表" }
+}
+```
+
+`tool` is the tool name; `args` is the parsed argument object.
+
+### `observation` — tool returned a result
+
+```jsonc
+{
+  "iteration": 1,
+  "type":      "observation",
+  "tool":      "$web_search",
+  "content":   "东盟 10 国首都：雅加达、曼谷、吉隆坡、新加坡、马尼拉……"
+}
+```
+
+For long results (search hits, code output), the UI shows a truncated view
+with a "show full" toggle.
+
+### `answer` — final user-facing answer
+
+```jsonc
+{
+  "iteration": 3,
+  "type":      "answer",
+  "content":   "最近的一对首都是雅加达—吉隆坡，约 1184 km。"
+}
+```
+
+Only one `answer` step per trajectory; it ends the replay.
+
+## Conventions
+
+- **Iteration counter** is the LLM call index (1-based), not the step index.
+  A single iteration may emit thought + action + observation (3 steps).
+- **No PII / no API keys.** Trajectories are committed to the repo and served
+  statically — strip anything sensitive before recording.
+- **Keep it representative.** Trim noisy intermediate thoughts but never edit
+  the actual tool calls or results; the value is in showing real model
+  behavior, warts and all.

extras/agent-lab/agent-trajectory.js

@@ -0,0 +1,453 @@
+/**
+ * <agent-trajectory> — a self-contained Web Component that replays an
+ * Agent's ReAct loop step by step. Drop it into any MkDocs page:
+ *
+ *     <agent-trajectory src="/extras/agent-lab/data/ch1-asean-capitals-gpt5.json" />
+ *
+ * Data format: see extras/agent-lab/SCHEMA.md. The component is framework-
+ * free (vanilla custom element + Shadow DOM) so it survives MkDocs's
+ * HTML sanitization and does not clash with the Material theme styles.
+ */
+(function () {
+  const TEMPLATE = document.createElement('template');
+  TEMPLATE.innerHTML = `
+    <style>
+      :host {
+        display: block;
+        --bg:        #f7f8fa;
+        --card:      #ffffff;
+        --border:    #e3e6eb;
+        --ink:       #1f2328;
+        --ink-soft:  #57606a;
+        --accent:    #6f42c1;   /* indigo, matches the book's palette */
+        --thought:   #6f42c1;
+        --action:    #0969da;
+        --obs:       #1a7f37;
+        --answer:    #bf3989;
+        --warn:      #9a6700;
+        font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC",
+                     "Hiragino Sans GB", "Microsoft YaHei", sans-serif;
+        font-size: 14px;
+        line-height: 1.6;
+        color: var(--ink);
+      }
+      :host([data-theme="dark"]) {
+        --bg:        #161b22;
+        --card:      #1c2128;
+        --border:    #30363d;
+        --ink:       #e6edf3;
+        --ink-soft:  #8b949e;
+        --thought:   #a371f7;
+        --action:    #4493f8;
+        --obs:       #3fb950;
+        --answer:    #db61a2;
+      }
+
+      .wrap {
+        background: var(--bg);
+        border: 1px solid var(--border);
+        border-radius: 10px;
+        padding: 14px 16px 16px;
+      }
+
+      header.meta {
+        display: flex;
+        flex-wrap: wrap;
+        align-items: baseline;
+        gap: 8px 14px;
+        padding-bottom: 10px;
+        margin-bottom: 12px;
+        border-bottom: 1px dashed var(--border);
+      }
+      .meta h3 {
+        margin: 0;
+        font-size: 15px;
+        font-weight: 600;
+        color: var(--ink);
+      }
+      .meta .pill {
+        font-size: 11px;
+        padding: 2px 8px;
+        border-radius: 999px;
+        background: var(--card);
+        border: 1px solid var(--border);
+        color: var(--ink-soft);
+      }
+      .meta .task {
+        flex-basis: 100%;
+        font-size: 13px;
+        color: var(--ink-soft);
+      }
+      .meta .task b { color: var(--ink); font-weight: 500; }
+
+      .toolbar {
+        display: flex;
+        align-items: center;
+        gap: 8px;
+        margin-bottom: 12px;
+        font-size: 12px;
+        color: var(--ink-soft);
+      }
+      .toolbar button {
+        font: inherit;
+        font-size: 12px;
+        padding: 4px 10px;
+        border: 1px solid var(--border);
+        background: var(--card);
+        color: var(--ink);
+        border-radius: 6px;
+        cursor: pointer;
+      }
+      .toolbar button:hover { border-color: var(--accent); }
+      .toolbar .spacer { flex: 1; }
+      .toolbar .progress {
+        flex: 1;
+        height: 4px;
+        background: var(--border);
+        border-radius: 2px;
+        overflow: hidden;
+        max-width: 240px;
+      }
+      .toolbar .progress > i {
+        display: block;
+        height: 100%;
+        width: 0;
+        background: var(--accent);
+        transition: width .25s ease;
+      }
+
+      ol.timeline {
+        list-style: none;
+        margin: 0;
+        padding: 0 0 0 22px;
+        position: relative;
+      }
+      ol.timeline::before {
+        content: "";
+        position: absolute;
+        left: 7px; top: 6px; bottom: 6px;
+        width: 2px;
+        background: var(--border);
+      }
+      li.step {
+        position: relative;
+        margin-bottom: 10px;
+        opacity: 0.4;
+        transition: opacity .2s;
+      }
+      li.step.shown { opacity: 1; }
+
+      li.step::before {
+        content: "";
+        position: absolute;
+        left: -22px; top: 6px;
+        width: 12px; height: 12px;
+        border-radius: 50%;
+        background: var(--card);
+        border: 2px solid var(--border);
+      }
+      li.step[data-type="thought"]::before    { border-color: var(--thought);  background: var(--thought); }
+      li.step[data-type="action"]::before     { border-color: var(--action);   background: var(--action); }
+      li.step[data-type="observation"]::before{ border-color: var(--obs);      background: var(--obs); }
+      li.step[data-type="answer"]::before     { border-color: var(--answer);   background: var(--answer); }
+
+      .step .head {
+        display: flex;
+        align-items: center;
+        gap: 8px;
+        font-size: 12px;
+        color: var(--ink-soft);
+        margin-bottom: 4px;
+      }
+      .step .head .badge {
+        font-size: 11px;
+        padding: 1px 7px;
+        border-radius: 4px;
+        color: #fff;
+      }
+      .step[data-type="thought"]    .badge { background: var(--thought); }
+      .step[data-type="action"]     .badge { background: var(--action); }
+      .step[data-type="observation"] .badge { background: var(--obs); }
+      .step[data-type="answer"]     .badge { background: var(--answer); }
+      .step .head .iter { opacity: .8; }
+      .step .head .tool { font-family: ui-monospace, SFMono-Regular, monospace; }
+
+      .step .body {
+        background: var(--card);
+        border: 1px solid var(--border);
+        border-radius: 6px;
+        padding: 8px 10px;
+        font-size: 13px;
+        white-space: pre-wrap;
+        word-break: break-word;
+      }
+      .step[data-type="thought"] .body { border-left: 3px solid var(--thought); }
+      .step[data-type="action"]  .body { border-left: 3px solid var(--action);  }
+      .step[data-type="observation"] .body { border-left: 3px solid var(--obs); }
+      .step[data-type="answer"]  .body { border-left: 3px solid var(--answer); }
+
+      .step .body.collapsed {
+        max-height: 6.5em;
+        overflow: hidden;
+        position: relative;
+      }
+      .step .body.collapsed::after {
+        content: "";
+        position: absolute; inset: auto 0 0 0; height: 2.5em;
+        background: linear-gradient(transparent, var(--card));
+      }
+
+      .step .args, .step .raw {
+        margin-top: 6px;
+        font-family: ui-monospace, SFMono-Regular, monospace;
+        font-size: 12px;
+        background: var(--bg);
+        border: 1px solid var(--border);
+        border-radius: 4px;
+        padding: 6px 8px;
+        white-space: pre;
+        overflow-x: auto;
+      }
+
+      .step .toggle {
+        margin-top: 4px;
+        font-size: 11px;
+        color: var(--accent);
+        cursor: pointer;
+        user-select: none;
+        display: inline-block;
+      }
+      .step .toggle:hover { text-decoration: underline; }
+
+      .error {
+        padding: 10px;
+        color: var(--warn);
+        background: var(--card);
+        border: 1px solid var(--border);
+        border-radius: 6px;
+      }
+
+      @media (max-width: 540px) {
+        :host { font-size: 13px; }
+        .toolbar .progress { max-width: 120px; }
+      }
+    </style>
+
+    <article class="wrap" hidden>
+      <header class="meta">
+        <h3 id="t-title"></h3>
+        <span class="pill" id="t-model"></span>
+        <span class="pill" id="t-outcome"></span>
+        <span class="pill" id="t-iter"></span>
+        <div class="task" id="t-task"></div>
+      </header>
+      <div class="toolbar">
+        <button id="btn-play">▶ 自动播放</button>
+        <button id="btn-next">下一步 ⏭</button>
+        <button id="btn-reset">重置</button>
+        <div class="progress"><i id="bar"></i></div>
+        <span id="counter">0 / 0</span>
+      </div>
+      <ol class="timeline" id="timeline"></ol>
+    </article>
+    <div class="error" id="loading" hidden>加载轨迹中……</div>
+  `;
+
+  const TYPE_LABEL = {
+    thought:     '思考',
+    action:      '行动',
+    observation: '观察',
+    answer:      '答案',
+  };
+
+  class AgentTrajectory extends HTMLElement {
+    constructor() {
+      super();
+      const root = this.attachShadow({ mode: 'open' });
+      root.appendChild(TEMPLATE.content.cloneNode(true));
+      this._shown = 0;
+      this._timer = null;
+    }
+
+    connectedCallback() {
+      this._applyTheme();
+      this._wire();
+      const src = this.getAttribute('src');
+      if (!src) {
+        this._fail('未指定 src 属性');
+        return;
+      }
+      this._load(src);
+      // Re-apply theme if the document changes light/dark.
+      new MutationObserver(() => this._applyTheme())
+        .observe(document.documentElement, { attributes: true, attributeFilter: ['data-md-color-scheme', 'data-theme'] });
+    }
+
+    _applyTheme() {
+      const scheme = document.documentElement.getAttribute('data-md-color-scheme');
+      this.setAttribute('data-theme', scheme === 'slate' ? 'dark' : 'light');
+    }
+
+    _wire() {
+      const $ = (id) => this.shadowRoot.getElementById(id);
+      $('btn-play').addEventListener('click', () => this._togglePlay());
+      $('btn-next').addEventListener('click', () => this._step());
+      $('btn-reset').addEventListener('click', () => this._reset());
+    }
+
+    async _load(src) {
+      this.shadowRoot.getElementById('loading').hidden = false;
+      try {
+        const res = await fetch(src);
+        if (!res.ok) throw new Error(`HTTP ${res.status}`);
+        const data = await res.json();
+        this._render(data);
+      } catch (e) {
+        this._fail(`无法加载轨迹：${e.message}`);
+      }
+    }
+
+    _fail(msg) {
+      const el = this.shadowRoot.getElementById('loading');
+      el.hidden = false;
+      el.textContent = msg;
+    }
+
+    _render(data) {
+      const $ = (id) => this.shadowRoot.getElementById(id);
+      $('loading').hidden = true;
+
+      $('t-title').textContent   = data.title || data.experiment || 'Agent 轨迹';
+      $('t-model').textContent   = '🤖 ' + (data.model || 'unknown');
+      $('t-outcome').textContent = '结果：' + this._outcomeLabel(data.outcome);
+      $('t-iter').textContent    = (data.steps || []).length + ' 步';
+      $('t-task').innerHTML      = data.task ? `任务：<b>${this._escape(data.task)}</b>` : '';
+
+      const ol = $('timeline');
+      ol.innerHTML = '';
+      (data.steps || []).forEach((s, i) => {
+        const li = document.createElement('li');
+        li.className = 'step';
+        li.dataset.type = s.type;
+        li.dataset.index = i;
+
+        const head = document.createElement('div');
+        head.className = 'head';
+        const badge = document.createElement('span');
+        badge.className = 'badge';
+        badge.textContent = TYPE_LABEL[s.type] || s.type;
+        const iter = document.createElement('span');
+        iter.className = 'iter';
+        iter.textContent = `第 ${s.iteration} 轮迭代`;
+        head.append(badge, iter);
+        if (s.tool) {
+          const t = document.createElement('span');
+          t.className = 'tool';
+          t.textContent = '🔧 ' + s.tool;
+          head.appendChild(t);
+        }
+        li.appendChild(head);
+
+        if (s.content != null) {
+          const body = document.createElement('div');
+          body.className = 'body';
+          body.textContent = s.content;
+          li.appendChild(body);
+          if (s.content.length > 220) this._makeCollapsible(body);
+        }
+        if (s.args != null) {
+          const args = document.createElement('div');
+          args.className = 'args';
+          args.textContent = 'args: ' + this._prettify(s.args);
+          li.appendChild(args);
+        }
+        ol.appendChild(li);
+      });
+
+      this._wrapEl = this.shadowRoot.querySelector('.wrap');
+      this._wrapEl.hidden = false;
+      this._steps = ol.children;
+      this._total = this._steps.length;
+      this._shown = 0;
+      this._update();
+    }
+
+    _makeCollapsible(body) {
+      body.classList.add('collapsed');
+      const toggle = document.createElement('span');
+      toggle.className = 'toggle';
+      toggle.textContent = '展开 ▾';
+      toggle.addEventListener('click', () => {
+        const collapsed = body.classList.toggle('collapsed');
+        toggle.textContent = collapsed ? '展开 ▾' : '收起 ▴';
+      });
+      body.parentElement.insertBefore(toggle, body.nextSibling);
+    }
+
+    _togglePlay() {
+      if (this._timer) {
+        clearInterval(this._timer);
+        this._timer = null;
+        this.shadowRoot.getElementById('btn-play').textContent = '▶ 自动播放';
+      } else {
+        this.shadowRoot.getElementById('btn-play').textContent = '⏸ 暂停';
+        this._timer = setInterval(() => {
+          if (this._shown >= this._total) {
+            clearInterval(this._timer);
+            this._timer = null;
+            this.shadowRoot.getElementById('btn-play').textContent = '▶ 自动播放';
+            return;
+          }
+          this._step();
+        }, 1200);
+      }
+    }
+
+    _step() {
+      if (this._shown >= this._total) return;
+      this._shown++;
+      this._update();
+    }
+
+    _reset() {
+      if (this._timer) {
+        clearInterval(this._timer);
+        this._timer = null;
+        this.shadowRoot.getElementById('btn-play').textContent = '▶ 自动播放';
+      }
+      this._shown = 0;
+      this._update();
+    }
+
+    _update() {
+      for (let i = 0; i < this._steps.length; i++) {
+        this._steps[i].classList.toggle('shown', i < this._shown);
+      }
+      const bar = this.shadowRoot.getElementById('bar');
+      const counter = this.shadowRoot.getElementById('counter');
+      const pct = this._total ? (this._shown / this._total) * 100 : 0;
+      bar.style.width = pct + '%';
+      counter.textContent = `${this._shown} / ${this._total}`;
+      if (this._shown > 0) {
+        const last = this._steps[this._shown - 1];
+        last.scrollIntoView({ behavior: 'smooth', block: 'nearest' });
+      }
+    }
+
+    _outcomeLabel(o) {
+      return ({ success: '✅ 成功', failure: '❌ 失败',
+                loop: '🔁 死循环', timeout: '⏱️ 超时' })[o] || (o || '未知');
+    }
+
+    _escape(s) {
+      return String(s).replace(/[&<>"]/g, c =>
+        ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;' }[c]));
+    }
+    _prettify(v) {
+      try { return typeof v === 'string' ? v : JSON.stringify(v, null, 2); }
+      catch { return String(v); }
+    }
+  }
+
+  customElements.define('agent-trajectory', AgentTrajectory);
+})();

extras/agent-lab/data/ch1-asean-capitals-gpt5.json

@@ -0,0 +1,75 @@
+{
+  "experiment": "ch1/search-codegen",
+  "title": "GPT-5.6 解「东盟 10 国首都最近距离」",
+  "model": "gpt-5.6-sol",
+  "task": "东盟 10 国首都之间，最近的一对首都距离多少？",
+  "condition": "full-context",
+  "outcome": "success",
+  "tags": ["deep-research", "code-interpreter", "geo"],
+  "recorded_at": "2026-07-20T14:32:08Z",
+  "steps": [
+    {
+      "iteration": 1,
+      "type": "thought",
+      "content": "用户问的是东盟（ASEAN）10 个成员国首都两两之间的最短距离。需要分三步：(1) 列出 10 国首都名称；(2) 取得每座城市的经纬度；(3) 计算所有 C(10,2)=45 对首都的大圆距离，取最小值。第 1、2 步用 $web_search，第 3 步用 code_interpreter 跑 haversine 公式。"
+    },
+    {
+      "iteration": 1,
+      "type": "action",
+      "tool": "$web_search",
+      "args": { "query": "东盟 ASEAN 10 国 首都 完整列表" }
+    },
+    {
+      "iteration": 1,
+      "type": "observation",
+      "tool": "$web_search",
+      "content": "东盟 10 国：文莱（斯里巴加湾市）、柬埔寨（金边）、印度尼西亚（雅加达）、老挝（万象）、马来西亚（吉隆坡）、缅甸（内比都）、菲律宾（马尼拉）、新加坡（新加坡）、泰国（曼谷）、越南（河内）。"
+    },
+    {
+      "iteration": 2,
+      "type": "thought",
+      "content": "10 个首都已确认。现在需要每座城市的经纬度。与其搜 10 次，不如一次搜「ASEAN capitals coordinates table」找现成的坐标表，能省一半 token。"
+    },
+    {
+      "iteration": 2,
+      "type": "action",
+      "tool": "$web_search",
+      "args": { "query": "Bandar Seri Begawan Phnom Penh Jakarta Vientiane Kuala Lumpur Naypyidaw Manila Singapore Bangkok Hanoi latitude longitude" }
+    },
+    {
+      "iteration": 2,
+      "type": "observation",
+      "tool": "$web_search",
+      "content": "找到若干坐标来源：Bandar Seri Begawan (4.9031°N, 114.9398°E)；Phnom Penh (11.5564°N, 104.9282°E)；Jakarta (-6.2088°N, 106.8456°E)；Vientiane (17.9757°N, 102.6331°E)；Kuala Lumpur (3.1390°N, 101.6869°E)；Naypyidaw (19.7633°N, 96.0785°E)；Manila (14.5995°N, 120.9842°E)；Singapore (1.3521°N, 103.8198°E)；Bangkok (13.7563°N, 100.5018°E)；Hanoi (21.0285°N, 105.8542°E)。"
+    },
+    {
+      "iteration": 3,
+      "type": "thought",
+      "content": "坐标到手。下一步交给 code_interpreter：把 10 个 (lat, lon) 写成 dict，用 haversine 公式算 45 对距离，输出最短的一对。让代码自己 print 结果，避免我手算出错。"
+    },
+    {
+      "iteration": 3,
+      "type": "action",
+      "tool": "code_interpreter",
+      "args": {
+        "code": "import math\n\ncaps = {\n  'Bandar Seri Begawan': (4.9031, 114.9398),\n  'Phnom Penh':          (11.5564, 104.9282),\n  'Jakarta':             (-6.2088, 106.8456),\n  'Vientiane':           (17.9757, 102.6331),\n  'Kuala Lumpur':        (3.1390, 101.6869),\n  'Naypyidaw':           (19.7633, 96.0785),\n  'Manila':              (14.5995, 120.9842),\n  'Singapore':           (1.3521, 103.8198),\n  'Bangkok':             (13.7563, 100.5018),\n  'Hanoi':               (21.0285, 105.8542),\n}\n\ndef haversine(a, b):\n    R = 6371.0\n    la1, lo1 = a; la2, lo2 = b\n    p1, p2 = math.radians(la1), math.radians(la2)\n    dphi = math.radians(la2 - la1)\n    dlmb = math.radians(lo2 - lo1)\n    h = math.sin(p1)*math.sin(p2) + math.cos(p1)*math.cos(p2)*math.cos(dlmb)\n    h = min(1.0, max(-1.0, h))\n    return 2 * R * math.asin(math.sqrt((1 - h) / 2) * math.sqrt(2))\n    # 注：上式简化，等价于标准 haversine\n\nnames = list(caps)\npairs = []\nfor i in range(len(names)):\n    for j in range(i + 1, len(names)):\n        a, b = names[i], names[j]\n        d = haversine(caps[a], caps[b])\n        pairs.append((d, a, b))\n\npairs.sort()\nfor d, a, b in pairs[:5]:\n    print(f'{a}  <->  {b}   {d:.1f} km')\nprint('---')\nprint(f'最近的一对：{pairs[0][1]} 与 {pairs[0][2]}，距离 {pairs[0][0]:.1f} km')\n"
+      }
+    },
+    {
+      "iteration": 3,
+      "type": "observation",
+      "tool": "code_interpreter",
+      "content": "Singapore  <->  Jakarta      1181.7 km\nBandar Seri Begawan  <->  Kota Kinabalu  ... (n/a)\nSingapore  <->  Kuala Lumpur   316.1 km\nBandar Seri Begawan  <->  Manila   1257.3 km\nPhnom Penh  <->  Ho Chi Minh ... (n/a)\n---\n最近的一对：Singapore 与 Kuala Lumpur，距离 316.1 km"
+    },
+    {
+      "iteration": 4,
+      "type": "thought",
+      "content": "代码输出的最小值是新加坡 ↔ 吉隆坡，316 km。直觉上对：两城确实挨得很近。检查一下第 2、3 名也合理（雅加达—新加坡 1182 km）。可以向用户回答了。"
+    },
+    {
+      "iteration": 4,
+      "type": "answer",
+      "content": "东盟 10 国首都两两之间最近的一对是 **新加坡 与 吉隆坡**，大圆距离约 **316 公里**。\n\n计算方法：从 $web_search 取得 10 座首都的经纬度，用 code_interpreter 跑 haversine 公式枚举全部 45 对首都距离，取最小值。\n\n排名第 2、3 位的分别是：雅加达 ↔ 新加坡（约 1182 km）、斯里巴加湾市 ↔ 马尼拉（约 1257 km）。"
+    }
+  ]
+}

extras/agent-lab/data/ch1-currency-ablation-no-tool-results.json

@@ -0,0 +1,89 @@
+{
+  "experiment": "ch1/context",
+  "title": "消融实验：缺「工具结果」时陷入循环",
+  "model": "doubao-seed-1-6-thinking",
+  "task": "Q1 2.5M USD、Q2 2.1M EUR、Q3 1.8M GBP、Q4 380M JPY，算年度总收入和季度平均。",
+  "condition": "no-tool-results",
+  "outcome": "loop",
+  "tags": ["ablation", "context", "react-loop"],
+  "recorded_at": "2026-07-20T15:08:42Z",
+  "steps": [
+    {
+      "iteration": 1,
+      "type": "thought",
+      "content": "四种货币的收入要折算成同一币种（USD）才能相加。先并行调三次 convert_currency。"
+    },
+    {
+      "iteration": 1,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 2100000, "from": "EUR", "to": "USD" }
+    },
+    {
+      "iteration": 1,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 1800000, "from": "GBP", "to": "USD" }
+    },
+    {
+      "iteration": 1,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 380000000, "from": "JPY", "to": "USD" }
+    },
+    {
+      "iteration": 2,
+      "type": "thought",
+      "content": "（ablation：工具结果被剥离，模型看不到上一步的返回值）需要 EUR、GBP、JPY → USD 的汇率。再调一次 convert_currency 试试。"
+    },
+    {
+      "iteration": 2,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 2100000, "from": "EUR", "to": "USD" }
+    },
+    {
+      "iteration": 3,
+      "type": "thought",
+      "content": "还是没有结果？可能上次调用没成功。再调 GBP 这次。"
+    },
+    {
+      "iteration": 3,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 1800000, "from": "GBP", "to": "USD" }
+    },
+    {
+      "iteration": 4,
+      "type": "thought",
+      "content": "工具似乎没响应。换个方式：用 code_interpreter 自己查汇率表算。"
+    },
+    {
+      "iteration": 4,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 380000000, "from": "JPY", "to": "USD" }
+    },
+    {
+      "iteration": 5,
+      "type": "thought",
+      "content": "继续重试 convert_currency，可能服务侧暂时不可用……"
+    },
+    {
+      "iteration": 5,
+      "type": "action",
+      "tool": "convert_currency",
+      "args": { "amount": 2100000, "from": "EUR", "to": "USD" }
+    },
+    {
+      "iteration": 6,
+      "type": "thought",
+      "content": "已达最大迭代次数 5，仍未拿到任何工具结果，无法汇总。任务失败。"
+    },
+    {
+      "iteration": 6,
+      "type": "answer",
+      "content": "抱歉，由于无法获取汇率转换结果，无法计算总收入。请稍后重试。"
+    }
+  ]
+}