<!DOCTYPE html>
<html lang="zh-CN">
<head>
<meta charset="UTF-8">
<title>ctx 模块 — atomcode 上下文管理架构</title>
<style>
:root {
--bg: #fafafa;
--panel: #ffffff;
--ink: #1a1a1a;
--muted: #666;
--line: #e5e5e5;
--line-strong: #bbb;
--accent: #0066cc;
--warn: #c25;
--green: #2a7;
--red: #c33;
--code-bg: #f4f4f4;
--added: #e6f4e6;
--removed: #fde;
--changed: #fff4d6;
}
* { box-sizing: border-box; }
body {
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI",
"Helvetica Neue", "PingFang SC", "Hiragino Sans GB",
"Microsoft YaHei", sans-serif;
line-height: 1.65;
color: var(--ink);
background: var(--bg);
margin: 0;
padding: 0;
font-size: 15px;
}
.container {
max-width: 1080px;
margin: 0 auto;
padding: 36px 48px 80px;
}
header {
border-bottom: 3px solid var(--ink);
padding-bottom: 20px;
margin-bottom: 36px;
}
header h1 {
margin: 0 0 8px;
font-size: 30px;
letter-spacing: -0.02em;
}
header .sub {
color: var(--muted);
font-size: 14px;
}
header .sub code {
background: var(--code-bg);
padding: 1px 6px;
border-radius: 3px;
font-size: 13px;
}
nav {
background: var(--panel);
border: 1px solid var(--line);
padding: 14px 20px;
margin-bottom: 40px;
border-radius: 4px;
font-size: 14px;
}
nav ol {
margin: 0;
padding-left: 20px;
columns: 2;
column-gap: 24px;
}
nav a { color: var(--accent); text-decoration: none; }
nav a:hover { text-decoration: underline; }
h2 {
font-size: 24px;
margin-top: 44px;
margin-bottom: 14px;
padding-bottom: 8px;
border-bottom: 2px solid var(--line-strong);
}
h3 {
font-size: 18px;
margin-top: 28px;
margin-bottom: 10px;
}
h4 {
font-size: 14px;
margin-top: 18px;
margin-bottom: 8px;
color: var(--muted);
text-transform: uppercase;
letter-spacing: 0.05em;
font-weight: 600;
}
p { margin: 12px 0; }
code {
font-family: "SF Mono", Menlo, Consolas, monospace;
font-size: 13px;
background: var(--code-bg);
padding: 1px 5px;
border-radius: 3px;
}
pre {
background: #1e1e1e;
color: #d4d4d4;
padding: 14px 18px;
border-radius: 6px;
overflow-x: auto;
font-size: 13px;
line-height: 1.55;
margin: 14px 0;
}
pre code {
background: none;
padding: 0;
color: inherit;
}
.kw { color: #569cd6; }
.ty { color: #4ec9b0; }
.fn { color: #dcdcaa; }
.cm { color: #6a9955; }
.st { color: #ce9178; }
.nm { color: #d4d4d4; }
table {
width: 100%;
border-collapse: collapse;
margin: 14px 0;
font-size: 14px;
}
table th {
background: #f0f0f0;
text-align: left;
padding: 9px 12px;
border-bottom: 2px solid var(--line-strong);
font-weight: 600;
}
table td {
padding: 8px 12px;
border-bottom: 1px solid var(--line);
vertical-align: top;
}
table tr:nth-child(even) { background: #fafafa; }
td.num { text-align: right; font-variant-numeric: tabular-nums; }
.tag {
display: inline-block;
padding: 2px 8px;
border-radius: 3px;
font-size: 11px;
font-weight: 600;
letter-spacing: 0.03em;
text-transform: uppercase;
margin-right: 4px;
}
.tag.new { background: #d4f4d4; color: #060; }
.tag.move { background: #fff4d6; color: #840; }
.tag.modify { background: #eef5ff; color: #036; }
.tag.delete { background: #fde; color: #800; }
.panel {
background: var(--panel);
border: 1px solid var(--line);
border-radius: 6px;
padding: 16px 20px;
margin: 14px 0;
}
.panel.info { background: #eef5ff; border-left: 4px solid var(--accent); }
.panel.ok { background: #e9f7e9; border-left: 4px solid var(--green); }
.panel.warn { background: #fff8e6; border-left: 4px solid var(--warn); }
.callout {
background: #fff8e6;
border-left: 4px solid var(--warn);
padding: 12px 16px;
margin: 18px 0;
border-radius: 0 4px 4px 0;
font-size: 14px;
}
.callout .label {
font-weight: 600;
color: var(--warn);
text-transform: uppercase;
font-size: 11px;
letter-spacing: 0.08em;
margin-right: 6px;
}
.seq {
background: var(--panel);
border: 1px solid var(--line);
border-radius: 6px;
padding: 18px 16px;
margin: 18px 0;
overflow-x: auto;
}
.seq svg { display: block; margin: 0 auto; }
.seq .caption {
font-size: 13px;
color: var(--muted);
text-align: center;
margin-top: 10px;
}
.twocol {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 18px;
margin: 16px 0;
}
footer {
margin-top: 70px;
padding-top: 20px;
border-top: 1px solid var(--line);
font-size: 13px;
color: var(--muted);
}
footer a { color: var(--accent); text-decoration: none; }
@media print {
body { font-size: 11px; }
nav { break-after: page; }
h2 { break-before: page; }
}
</style>
</head>
<body>
<div class="container">
<header>
<h1>ctx 模块</h1>
<div class="sub">
atomcode 上下文管理架构 · 分支 <code>feat/ctx-module</code>
· <code>crates/atomcode-core/src/ctx/</code>
</div>
</header>
<nav>
<ol>
<li><a href="#arch">架构设计</a></li>
<li><a href="#modules">模块结构</a></li>
<li><a href="#sequences">关键时序</a></li>
<li><a href="#examples">实例：加自定义 ctx</a></li>
</ol>
</nav>
<h2 id="arch">1. 架构设计</h2>
<h3>1.1 一句话</h3>
<p><strong>所有模型共享一个 trait <code>CtxBuilder</code>，每个模型可选择实现自己的 context 构造逻辑；没实现 → 走 <code>DefaultCtx</code>。</strong></p>
<h3>1.2 设计原则</h3>
<table>
<tr>
<th style="width: 140px;">原则</th>
<th>含义</th>
</tr>
<tr>
<td><strong>单一 trait</strong></td>
<td>不拆 role、不做 facade 聚合。<code>CtxBuilder</code> 5 个方法，一个 impl 一套完整行为</td>
</tr>
<tr>
<td><strong>impl 自主</strong></td>
<td>trait 只定接口、不强制流程。<code>build_messages</code> 里爱怎么构造消息都行——改 system prompt、砍 tool schema、自定义 turn_reminder、自创压缩规则</td>
</tr>
<tr>
<td><strong>默认兜底</strong></td>
<td><code>DefaultCtx</code> 完全等价原有 <code>Conversation::to_provider_messages_budgeted</code> 行为。规则表未命中的模型永远走它</td>
</tr>
<tr>
<td><strong>YAGNI</strong></td>
<td>trait 只有 5 个方法、数据契约只返回 <code>(Vec<Message>, ContextStats)</code>。不预留字段、不预留枚举变体，等真需要再加</td>
</tr>
<tr>
<td><strong>按 provider 构造</strong></td>
<td>注册入口 <code>for_provider(&ProviderConfig)</code>，impl 从 config 里捕获 ctx_window / max_tokens / base_url 等所需字段</td>
</tr>
<tr>
<td><strong>跟随模型切换</strong></td>
<td><code>ReloadConfig</code> 重建 ctx；<code>build_system_prompt()</code> 已经每轮重建、按 provider 读字段——两层自然协同，无需特殊握手</td>
</tr>
</table>
<h3>1.3 Trait 长这样</h3>
<pre><code><span class="kw">pub trait</span> <span class="ty">CtxBuilder</span>: <span class="ty">Send</span> + <span class="ty">Sync</span> {
<span class="cm">/// 构造这一轮发送给 LLM 的消息数组 + 统计。</span>
<span class="cm">/// impl 完全自主决定怎么组装——system prompt 要不要变、</span>
<span class="cm">/// tool_result 要不要压、turn_reminder 要不要注入。</span>
<span class="kw">fn</span> <span class="fn">build_messages</span>(
&<span class="kw">self</span>,
conv: &<span class="ty">Conversation</span>,
system_prompt: &<span class="kw">str</span>,
) -> (<span class="ty">Vec</span><<span class="ty">Message</span>>, <span class="ty">ContextStats</span>);
<span class="cm">/// 是否触发压缩(每轮末检查)。默认：不压。</span>
<span class="kw">fn</span> <span class="fn">needs_compression</span>(&<span class="kw">self</span>, _: &<span class="ty">Conversation</span>, _: <span class="kw">usize</span>) -> <span class="kw">bool</span> { <span class="kw">false</span> }
<span class="cm">/// 压缩计划(待摘要文本, 删几条)；None = 不压。</span>
<span class="kw">fn</span> <span class="fn">compression_plan</span>(&<span class="kw">self</span>, _: &<span class="ty">Conversation</span>) -> <span class="ty">Option</span><(<span class="ty">String</span>, <span class="kw">usize</span>)> { <span class="kw">None</span> }
<span class="cm">/// 单条 tool 输出截断。</span>
<span class="kw">fn</span> <span class="fn">truncate_tool_output</span>(&<span class="kw">self</span>, result: &<span class="kw">mut</span> <span class="ty">ToolResult</span>, tool_name: &<span class="kw">str</span>);
<span class="cm">/// 策略名(日志/调试)。</span>
<span class="kw">fn</span> <span class="fn">name</span>(&<span class="kw">self</span>) -> &<span class="st">'static</span> <span class="kw">str</span>;
}</code></pre>
<div class="panel info">
<strong>为什么 <code>build_messages</code> 一个方法就够</strong>——
它直接拿到 <code>system_prompt: &str</code>（来自 agent::prompt 层，已按
provider 重建）。impl 想改 system prompt 自己变换即可；想自定义
tool_result 压缩、turn_reminder 注入时机、replace_stale_reads 策略，
全部在 <code>build_messages</code> 函数体内决定。<br><br>
<strong>不需要</strong>额外的 <code>transform_system_prompt</code> /
<code>build_cold_zone</code> / <code>build_hot_history</code> 钩子——
那些都是没证据支撑的过度设计。真需要时再加。
</div>
<h3>1.4 注册入口</h3>
<pre><code><span class="cm">/// 给定 provider config 返回对应 CtxBuilder。</span>
<span class="cm">/// 添加自定义模型：在 match 里加一条 arm。</span>
<span class="cm">/// 没命中 → DefaultCtx(保留当前行为)。</span>
<span class="kw">pub fn</span> <span class="fn">for_provider</span>(provider: &<span class="ty">ProviderConfig</span>) -> <span class="ty">Box</span><<span class="kw">dyn</span> <span class="ty">CtxBuilder</span>> {
<span class="cm">// 自定义规则示例(目前 feat/ctx-module 上是空的)：</span>
<span class="cm">// if provider.provider_type == "ollama" {</span>
<span class="cm">// return Box::new(OllamaCtx::new(provider));</span>
<span class="cm">// }</span>
<span class="cm">// if provider.model.starts_with("claude-") {</span>
<span class="cm">// return Box::new(ClaudeCtx::new(provider));</span>
<span class="cm">// }</span>
<span class="ty">Box</span>::<span class="fn">new</span>(<span class="ty">DefaultCtx</span>::<span class="fn">new</span>(provider))
}</code></pre>
<h2 id="modules">2. 模块结构</h2>
<h3>2.1 文件布局</h3>
<pre><code>crates/atomcode-core/src/
├── ctx/ <span class="cm">← 新增模块</span>
│ ├── mod.rs <span class="cm">── CtxBuilder trait + re-exports</span>
│ ├── resolver.rs <span class="cm">── for_provider() 规则表</span>
│ ├── render.rs <span class="cm">── 默认 render / compression 策略</span>
│ ├── default.rs <span class="cm">── DefaultCtx = ctx::render 薄封装</span>
│ ├── ollama.rs <span class="cm">── OllamaCtx 小窗口本地模型</span>
│ └── truncate.rs <span class="cm">── 从 turn/truncation.rs 迁入</span>
│
├── conversation/mod.rs <span class="cm">── 纯数据层(add_* / apply_compression + 字段)</span>
├── agent/mod.rs <span class="cm">── AgentLoop 加 ctx 字段 + 4 call site 改走 self.ctx.*</span>
├── config/mod.rs <span class="cm">── Config::default_context_window() helper</span>
└── turn/
├── mod.rs <span class="cm">── 去掉 pub mod truncation</span>
└── runner.rs <span class="cm">── 调 ctx::render::build_messages</span></code></pre>
<div class="panel info">
<strong>依赖方向(严格 DAG)</strong>: <code>agent / turn</code> → <code>ctx::{default, ollama, resolver, render, truncate}</code> → <code>conversation</code>(叶子)。<code>conversation</code> 不反向引用 <code>ctx</code>,生产和测试代码都是(grep 验证)。
</div>
<h3>2.2 新增 / 修改详情</h3>
<table>
<tr>
<th>文件</th>
<th>类型</th>
<th>职责</th>
<th class="num">代码量</th>
</tr>
<tr>
<td><code>ctx/mod.rs</code></td>
<td><span class="tag new">new</span></td>
<td>CtxBuilder trait 定义 + re-exports(routing 逻辑在 resolver.rs)</td>
<td class="num">~75 行</td>
</tr>
<tr>
<td><code>ctx/resolver.rs</code></td>
<td><span class="tag new">new</span></td>
<td><code>for_provider</code> 规则表,添加新模型在此登记</td>
<td class="num">~90 行</td>
</tr>
<tr>
<td><code>ctx/render.rs</code></td>
<td><span class="tag new">new</span></td>
<td>默认 render 管道:build_messages / needs_compression / build_compression_content + 4 个私有 helper (microcompact / replace_stale_reads / clean_message_pipeline / sanitize_messages),从 Conversation impl 搬出</td>
<td class="num">~720 行(含测试)</td>
</tr>
<tr>
<td><code>ctx/default.rs</code></td>
<td><span class="tag new">new</span></td>
<td><code>DefaultCtx</code> 薄封装(全部 delegate 到 ctx::render) + 基础测试</td>
<td class="num">~115 行</td>
</tr>
<tr>
<td><code>ctx/ollama.rs</code></td>
<td><span class="tag new">new</span></td>
<td><code>OllamaCtx</code> 小窗口优化:35% 压缩阈值 + 工具输出 2K-6K 字节 belt + UTF-8 安全截断</td>
<td class="num">~260 行</td>
</tr>
<tr>
<td><code>ctx/truncate.rs</code></td>
<td><span class="tag move">moved</span></td>
<td>per-tool 截断（整体 git rename，内容不变）</td>
<td class="num">521 行</td>
</tr>
<tr>
<td><code>conversation/mod.rs</code></td>
<td><span class="tag modify">modified</span></td>
<td>删 dead summary 4 方法 + render 3 方法 + 4 个 helper → 迁到 ctx::render; 保留 add_* / apply_compression / 字段 / KEEP_MESSAGES pub(crate) const</td>
<td class="num">≈ 2000 → 715 行</td>
</tr>
<tr>
<td><code>agent/mod.rs</code></td>
<td><span class="tag modify">modified</span></td>
<td>加 <code>ctx</code> 字段 · <code>AgentLoop::new</code> 初始化 · <code>ReloadConfig</code> 同步重建 · 4 处 call site 改走 <code>self.ctx.*</code> · 删 <code>maybe_summarize_old_turns</code></td>
<td class="num">+43 / −70</td>
</tr>
<tr>
<td><code>config/mod.rs</code></td>
<td><span class="tag modify">modified</span></td>
<td>加 <code>Config::default_context_window()</code> helper</td>
<td class="num">+11</td>
</tr>
<tr>
<td><code>turn/truncation.rs</code></td>
<td><span class="tag delete">deleted</span></td>
<td>整体迁到 <code>ctx/truncate.rs</code></td>
<td class="num">−524</td>
</tr>
<tr>
<td><code>agent/tool_dispatch.rs</code></td>
<td><span class="tag modify">modified</span></td>
<td>import 改走 <code>ctx::truncate</code> + DRY context_window</td>
<td class="num">±3</td>
</tr>
<tr>
<td><code>turn/runner.rs</code></td>
<td><span class="tag modify">modified</span></td>
<td>DRY context_window lookup</td>
<td class="num">±5</td>
</tr>
<tr>
<td><code>atomcode-tui/src/app.rs</code></td>
<td><span class="tag modify">modified</span></td>
<td>DRY context_window lookup</td>
<td class="num">±1</td>
</tr>
<tr style="border-top: 2px solid var(--line-strong); font-weight: 600;">
<td colspan="3">净变化</td>
<td class="num">+354 / −268</td>
</tr>
</table>
<h3>2.3 公共 API</h3>
<h4>导出(<code>crates/atomcode-core/src/ctx/mod.rs</code>)</h4>
<pre><code><span class="cm">// 类型</span>
<span class="kw">pub trait</span> <span class="ty">CtxBuilder</span> { ... } <span class="cm">// 单一 trait</span>
<span class="kw">pub use</span> default::<span class="ty">DefaultCtx</span>; <span class="cm">// 默认实现</span>
<span class="cm">// 入口函数</span>
<span class="kw">pub fn</span> <span class="fn">for_provider</span>(provider: &<span class="ty">ProviderConfig</span>) -> <span class="ty">Box</span><<span class="kw">dyn</span> <span class="ty">CtxBuilder</span>>;
<span class="cm">// 子模块</span>
<span class="kw">pub mod</span> truncate; <span class="cm">// 搬家来的截断工具函数</span></code></pre>
<h4>AgentLoop 字段</h4>
<pre><code><span class="kw">pub struct</span> <span class="ty">AgentLoop</span> {
<span class="cm">// ... 原有字段 ...</span>
<span class="kw">pub</span> config: <span class="ty">Config</span>,
<span class="kw">pub</span> ctx: <span class="ty">Box</span><<span class="kw">dyn</span> crate::ctx::<span class="ty">CtxBuilder</span>>, <span class="cm">// ← 新字段</span>
<span class="cm">// ...</span>
}</code></pre>
<h2 id="sequences">3. 关键时序</h2>
<h3>3.1 启动：config → CtxBuilder</h3>
<div class="seq">
<svg width="780" height="250" xmlns="http://www.w3.org/2000/svg" font-family="SF Mono, Menlo, monospace" font-size="11">
<g stroke="#bbb" stroke-width="1">
<line x1="90" y1="40" x2="90" y2="230"/>
<line x1="260" y1="40" x2="260" y2="230"/>
<line x1="440" y1="40" x2="440" y2="230"/>
<line x1="640" y1="40" x2="640" y2="230"/>
</g>
<g fill="#eef5ff" stroke="#036">
<rect x="35" y="16" width="110" height="28" rx="3"/>
<rect x="205" y="16" width="110" height="28" rx="3"/>
<rect x="385" y="16" width="110" height="28" rx="3"/>
<rect x="585" y="16" width="110" height="28" rx="3"/>
</g>
<g fill="#036" text-anchor="middle" font-weight="600">
<text x="90" y="34">CLI / TUI</text>
<text x="260" y="34">AgentLoop::new</text>
<text x="440" y="34">ctx::for_provider</text>
<text x="640" y="34">DefaultCtx::new</text>
</g>
<g fill="#1a1a1a">
<line x1="90" y1="70" x2="258" y2="70" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr)"/>
<text x="174" y="65">AgentLoop::new(config, provider, tools, ...)</text>
<line x1="260" y1="105" x2="438" y2="105" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr)"/>
<text x="349" y="100">for_provider(&provider_config)</text>
<text x="450" y="128" fill="#666">match 规则表(当前空) → fallthrough</text>
<line x1="440" y1="158" x2="638" y2="158" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr)"/>
<text x="540" y="153">DefaultCtx::new(provider)</text>
<line x1="638" y1="188" x2="442" y2="188" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG)"/>
<text x="540" y="183" fill="#2a7">DefaultCtx { ctx_window: 128_000 }</text>
<line x1="438" y1="215" x2="262" y2="215" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG)"/>
<text x="350" y="210" fill="#2a7">Box<dyn CtxBuilder></text>
</g>
<defs>
<marker id="arr" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#1a1a1a"/>
</marker>
<marker id="arrG" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#2a7"/>
</marker>
</defs>
</svg>
<div class="caption">图 3.1 — <code>AgentLoop::new</code> 只调一次 <code>for_provider</code>，拿到 <code>Box<dyn CtxBuilder></code> 存在 <code>self.ctx</code>。整个 session 期间除非 <code>ReloadConfig</code>，否则不再换。</div>
</div>
<h3>3.2 每轮：render 与 LLM 调用</h3>
<div class="seq">
<svg width="820" height="310" xmlns="http://www.w3.org/2000/svg" font-family="SF Mono, Menlo, monospace" font-size="11">
<g stroke="#bbb" stroke-width="1">
<line x1="80" y1="40" x2="80" y2="290"/>
<line x1="230" y1="40" x2="230" y2="290"/>
<line x1="400" y1="40" x2="400" y2="290"/>
<line x1="560" y1="40" x2="560" y2="290"/>
<line x1="720" y1="40" x2="720" y2="290"/>
</g>
<g fill="#eef5ff" stroke="#036">
<rect x="30" y="16" width="100" height="28" rx="3"/>
<rect x="175" y="16" width="110" height="28" rx="3"/>
<rect x="340" y="16" width="120" height="28" rx="3"/>
<rect x="505" y="16" width="110" height="28" rx="3"/>
<rect x="665" y="16" width="110" height="28" rx="3"/>
</g>
<g fill="#036" text-anchor="middle" font-weight="600">
<text x="80" y="34">user input</text>
<text x="230" y="34">AgentLoop</text>
<text x="400" y="34">self.ctx</text>
<text x="560" y="34">TurnRunner</text>
<text x="720" y="34">LlmProvider</text>
</g>
<g fill="#1a1a1a">
<line x1="80" y1="68" x2="228" y2="68" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr2)"/>
<text x="154" y="63">run_turn()</text>
<text x="240" y="90" fill="#666">system_prompt = build_system_prompt() <span class="cm">// 已按 provider 重建</span></text>
<line x1="230" y1="115" x2="398" y2="115" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr2)"/>
<text x="314" y="110">build_messages(&conv, &system_prompt)</text>
<g fill="#666" font-size="10">
<text x="408" y="135">── DefaultCtx: 透传 conv.to_provider_messages_budgeted</text>
<text x="408" y="148">── 自定义 impl: 自由变换 system_prompt + 组装 messages</text>
</g>
<line x1="398" y1="168" x2="232" y2="168" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG2)"/>
<text x="314" y="163" fill="#2a7">(Vec<Message>, ContextStats)</text>
<line x1="230" y1="195" x2="558" y2="195" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr2)"/>
<text x="394" y="190">turn_runner.run(messages, ...)</text>
<line x1="560" y1="222" x2="718" y2="222" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr2)"/>
<text x="639" y="217">chat_stream(messages)</text>
<line x1="718" y1="249" x2="562" y2="249" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG2)"/>
<text x="639" y="244" fill="#2a7">StreamEvent::Delta / ToolCall ...</text>
<line x1="560" y1="278" x2="402" y2="278" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr2)"/>
<text x="481" y="273">truncate_tool_output(&mut r, name)</text>
</g>
<defs>
<marker id="arr2" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#1a1a1a"/>
</marker>
<marker id="arrG2" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#2a7"/>
</marker>
</defs>
</svg>
<div class="caption">图 3.2 — AgentLoop 每轮：先让 agent::prompt 构 base system_prompt（已按当前 provider），再交给 <code>self.ctx.build_messages</code> 组装最终消息数组。TurnRunner 拿结果跑 LLM。工具结果回流走 <code>self.ctx.truncate_tool_output</code>。</div>
</div>
<h3>3.3 压缩：策略决策 + 数据层应用</h3>
<div class="seq">
<svg width="820" height="300" xmlns="http://www.w3.org/2000/svg" font-family="SF Mono, Menlo, monospace" font-size="11">
<g stroke="#bbb" stroke-width="1">
<line x1="90" y1="40" x2="90" y2="280"/>
<line x1="260" y1="40" x2="260" y2="280"/>
<line x1="440" y1="40" x2="440" y2="280"/>
<line x1="620" y1="40" x2="620" y2="280"/>
</g>
<g fill="#eef5ff" stroke="#036">
<rect x="40" y="16" width="100" height="28" rx="3"/>
<rect x="210" y="16" width="100" height="28" rx="3"/>
<rect x="385" y="16" width="110" height="28" rx="3"/>
<rect x="565" y="16" width="110" height="28" rx="3"/>
</g>
<g fill="#036" text-anchor="middle" font-weight="600">
<text x="90" y="34">AgentLoop</text>
<text x="260" y="34">self.ctx</text>
<text x="440" y="34">Conversation</text>
<text x="620" y="34">LlmProvider</text>
</g>
<g fill="#1a1a1a">
<line x1="90" y1="68" x2="258" y2="68" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr3)"/>
<text x="174" y="63">needs_compression(&conv, sys_tok)</text>
<line x1="258" y1="92" x2="92" y2="92" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG3)"/>
<text x="174" y="87" fill="#2a7">true</text>
<line x1="90" y1="118" x2="258" y2="118" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr3)"/>
<text x="174" y="113">compression_plan(&conv)</text>
<line x1="258" y1="142" x2="92" y2="142" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG3)"/>
<text x="174" y="137" fill="#2a7">Some((待摘要文本, 删 N 条))</text>
<line x1="90" y1="170" x2="618" y2="170" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr3)"/>
<text x="354" y="165">chat_stream(summarize prompt)</text>
<line x1="618" y1="194" x2="92" y2="194" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG3)"/>
<text x="354" y="189" fill="#2a7">summary(失败时 fallback 到机械)</text>
<line x1="90" y1="224" x2="438" y2="224" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr3)"/>
<text x="264" y="219">apply_compression(N, summary)</text>
<g fill="#666" font-size="10">
<text x="450" y="242">├─ drain 最早 N 条</text>
<text x="450" y="258">├─ cold_summaries FIFO push</text>
<text x="450" y="274">└─ turn_tracker 重索引</text>
</g>
</g>
<defs>
<marker id="arr3" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#1a1a1a"/>
</marker>
<marker id="arrG3" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#2a7"/>
</marker>
</defs>
</svg>
<div class="caption">图 3.3 — 压缩分两步：<strong>ctx 决策</strong>（是否压、压什么 content、删几条）+ <strong>agent 执行</strong>（真调 LLM 摘要 + Conversation 应用变更）。关键分工：ctx 不调 LLM 不动数据，Conversation 不懂策略。</div>
</div>
<h3>3.4 切模型：ReloadConfig 同步重建 ctx</h3>
<div class="seq">
<svg width="800" height="270" xmlns="http://www.w3.org/2000/svg" font-family="SF Mono, Menlo, monospace" font-size="11">
<g stroke="#bbb" stroke-width="1">
<line x1="100" y1="40" x2="100" y2="250"/>
<line x1="280" y1="40" x2="280" y2="250"/>
<line x1="480" y1="40" x2="480" y2="250"/>
<line x1="660" y1="40" x2="660" y2="250"/>
</g>
<g fill="#fff4d6" stroke="#840">
<rect x="45" y="16" width="110" height="28" rx="3"/>
<rect x="225" y="16" width="110" height="28" rx="3"/>
<rect x="425" y="16" width="110" height="28" rx="3"/>
<rect x="605" y="16" width="110" height="28" rx="3"/>
</g>
<g fill="#840" text-anchor="middle" font-weight="600">
<text x="100" y="34">TUI /model</text>
<text x="280" y="34">AgentLoop</text>
<text x="480" y="34">ctx::for_provider</text>
<text x="660" y="34">build_system_prompt</text>
</g>
<g fill="#1a1a1a">
<line x1="100" y1="68" x2="278" y2="68" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr4)"/>
<text x="189" y="63">AgentCommand::ReloadConfig(new_config)</text>
<g fill="#666" font-size="10">
<text x="290" y="90">旧 provider ≠ 新 → clear conversation</text>
<text x="290" y="104">替换 turn_runner.provider / config</text>
</g>
<line x1="280" y1="126" x2="478" y2="126" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr4)"/>
<text x="379" y="121">for_provider(&new_provider)</text>
<line x1="478" y1="148" x2="282" y2="148" stroke="#2a7" stroke-width="1.2" stroke-dasharray="4 3" marker-end="url(#arrG4)"/>
<text x="380" y="143" fill="#2a7">Box<dyn CtxBuilder>(新模型的)</text>
<text x="290" y="170" fill="#666">self.ctx = <新的 Box></text>
<text x="100" y="204" fill="#666">...下一轮 turn...</text>
<line x1="280" y1="228" x2="658" y2="228" stroke="#1a1a1a" stroke-width="1.2" marker-end="url(#arr4)"/>
<text x="469" y="223">build_system_prompt() <span class="cm">// 读新 provider</span></text>
</g>
<defs>
<marker id="arr4" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#1a1a1a"/>
</marker>
<marker id="arrG4" viewBox="0 0 10 10" refX="10" refY="5" markerWidth="6" markerHeight="6" orient="auto">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#2a7"/>
</marker>
</defs>
</svg>
<div class="caption">图 3.4 — 用户 <code>/model</code> 切模型时：<code>AgentLoop</code> 一行 <code>self.ctx = for_provider(...)</code> 完成策略层切换。下一轮 turn 开始时 <code>build_system_prompt()</code> 本来就是每轮重建的，自然读到新 provider。两层自动协同。</div>
</div>
<h2 id="examples">4. 实例：如何加自定义 ctx</h2>
<p>本节 3 个真实例子，展示不同粒度的定制。全部可复制粘贴。</p>
<h3>4.1 例：OllamaCtx — 小窗口激进截断</h3>
<p><strong>需求</strong>：本地 llama3/qwen 在 8K 窗口下跑。标准 tool schema 太长（占 2K+），bash 输出占 32K 上限太宽松。需要整体裁。</p>
<h4>步骤 1：新建 <code>crates/atomcode-core/src/ctx/ollama.rs</code></h4>
<pre><code><span class="kw">use super</span>::<span class="ty">CtxBuilder</span>;
<span class="kw">use crate</span>::config::provider::<span class="ty">ProviderConfig</span>;
<span class="kw">use crate</span>::conversation::{<span class="ty">ContextStats</span>, <span class="ty">Conversation</span>};
<span class="kw">use crate</span>::conversation::message::<span class="ty">Message</span>;
<span class="kw">use crate</span>::tool::<span class="ty">ToolResult</span>;
<span class="kw">pub struct</span> <span class="ty">OllamaCtx</span> {
ctx_window: <span class="kw">usize</span>,
}
<span class="kw">impl</span> <span class="ty">OllamaCtx</span> {
<span class="kw">pub fn</span> <span class="fn">new</span>(p: &<span class="ty">ProviderConfig</span>) -> <span class="ty">Self</span> {
<span class="ty">Self</span> { ctx_window: p.context_window.<span class="fn">max</span>(<span class="nm">4000</span>) }
}
}
<span class="kw">impl</span> <span class="ty">CtxBuilder</span> <span class="kw">for</span> <span class="ty">OllamaCtx</span> {
<span class="kw">fn</span> <span class="fn">build_messages</span>(
&<span class="kw">self</span>,
conv: &<span class="ty">Conversation</span>,
system_prompt: &<span class="kw">str</span>,
) -> (<span class="ty">Vec</span><<span class="ty">Message</span>>, <span class="ty">ContextStats</span>) {
<span class="cm">// 1. 砍 system prompt 里的 tool schema 长描述(每个工具只留名称+一行 usage)</span>
<span class="kw">let</span> trimmed = <span class="fn">strip_verbose_tool_schemas</span>(system_prompt);
<span class="cm">// 2. 调 legacy 底层走窗口裁切，拿到基础消息数组</span>
conv.<span class="fn">to_provider_messages_budgeted</span>(&trimmed, <span class="nm">self</span>.ctx_window)
}
<span class="kw">fn</span> <span class="fn">needs_compression</span>(&<span class="kw">self</span>, conv: &<span class="ty">Conversation</span>, sys_tokens: <span class="kw">usize</span>) -> <span class="kw">bool</span> {
<span class="cm">// 小窗口更激进：40% 就触发(默认 50%)</span>
<span class="kw">let</span> total: <span class="kw">usize</span> = sys_tokens
+ conv.messages.<span class="fn">iter</span>().<span class="fn">map</span>(|m| m.<span class="fn">estimate_tokens</span>()).<span class="fn">sum</span>::<<span class="kw">usize</span>>();
total > <span class="nm">self</span>.ctx_window * <span class="nm">40</span> / <span class="nm">100</span>
}
<span class="kw">fn</span> <span class="fn">compression_plan</span>(&<span class="kw">self</span>, conv: &<span class="ty">Conversation</span>) -> <span class="ty">Option</span><(<span class="ty">String</span>, <span class="kw">usize</span>)> {
<span class="kw">let</span> (content, n) = conv.<span class="fn">build_compression_content</span>();
<span class="kw">if</span> content.<span class="fn">is_empty</span>() || n == <span class="nm">0</span> { <span class="kw">return</span> <span class="kw">None</span>; }
<span class="ty">Some</span>((content, n))
}
<span class="kw">fn</span> <span class="fn">truncate_tool_output</span>(&<span class="kw">self</span>, result: &<span class="kw">mut</span> <span class="ty">ToolResult</span>, tool_name: &<span class="kw">str</span>) {
<span class="cm">// Ollama 下 tool_result 更紧：4K 字节上限(DefaultCtx 是 32K)</span>
<span class="kw">super</span>::truncate::<span class="fn">truncate_output</span>(result, tool_name, <span class="nm">self</span>.ctx_window);
<span class="kw">const</span> SMALL_CAP: <span class="kw">usize</span> = <span class="nm">4000</span>;
<span class="kw">if</span> result.output.<span class="fn">len</span>() > SMALL_CAP {
<span class="kw">let</span> <span class="kw">mut</span> cap = SMALL_CAP;
<span class="kw">while</span> cap > <span class="nm">0</span> && !result.output.<span class="fn">is_char_boundary</span>(cap) { cap -= <span class="nm">1</span>; }
result.output.<span class="fn">truncate</span>(cap);
result.output.<span class="fn">push_str</span>(<span class="st">"\n[... truncated by OllamaCtx ...]"</span>);
}
}
<span class="kw">fn</span> <span class="fn">name</span>(&<span class="kw">self</span>) -> &<span class="st">'static</span> <span class="kw">str</span> { <span class="st">"ollama"</span> }
}
<span class="kw">fn</span> <span class="fn">strip_verbose_tool_schemas</span>(sys: &<span class="kw">str</span>) -> <span class="ty">String</span> {
<span class="cm">// 业务实现：按正则砍掉 === TOOL SCHEMAS === 到下一个 === 之间的大段文字，</span>
<span class="cm">// 只保留每个工具名 + 一行 usage。具体取决于 system prompt 模板结构。</span>
sys.<span class="fn">to_string</span>() <span class="cm">// 示意</span>
}</code></pre>
<h4>步骤 2：在 <code>ctx/mod.rs</code> 声明 + <code>ctx/resolver.rs</code> 注册</h4>
<pre><code><span class="cm">// ctx/mod.rs: 加一行</span>
<span class="kw">pub mod</span> ollama;
<span class="cm">// ctx/resolver.rs: 在 for_provider 里加分支</span>
<span class="kw">pub fn</span> <span class="fn">for_provider</span>(provider: &<span class="ty">ProviderConfig</span>) -> <span class="ty">Box</span><<span class="kw">dyn</span> <span class="ty">CtxBuilder</span>> {
<span class="kw">if</span> provider.provider_type == <span class="st">"ollama"</span> {
<span class="kw">return</span> <span class="ty">Box</span>::<span class="fn">new</span>(super::ollama::<span class="ty">OllamaCtx</span>::<span class="fn">new</span>(provider));
}
<span class="cm">// ...其它规则...</span>
<span class="ty">Box</span>::<span class="fn">new</span>(<span class="ty">DefaultCtx</span>::<span class="fn">new</span>(provider))
}</code></pre>
<h4>就这三步。</h4>
<p>用户配 <code>provider_type = "ollama"</code> → 自动走 <code>OllamaCtx</code>。其它 provider 不受影响。</p>
<h3>4.2 例：ClaudeCtx — 稳定 cache 前缀</h3>
<p><strong>需求</strong>：Claude 启用 prompt cache 时，希望 system + cold_zone 前缀跨轮稳定（便于 cache 命中）。需要确保 <code>build_messages</code> 永远先出 system，再出 cold_summaries 作为连续 System 消息，再出 hot messages。</p>
<pre><code><span class="kw">pub struct</span> <span class="ty">ClaudeCtx</span> {
ctx_window: <span class="kw">usize</span>,
max_output_tokens: <span class="ty">Option</span><<span class="kw">usize</span>>,
}
<span class="kw">impl</span> <span class="ty">ClaudeCtx</span> {
<span class="kw">pub fn</span> <span class="fn">new</span>(p: &<span class="ty">ProviderConfig</span>) -> <span class="ty">Self</span> {
<span class="ty">Self</span> {
ctx_window: p.context_window.<span class="fn">max</span>(<span class="nm">128_000</span>),
max_output_tokens: p.max_tokens,
}
}
}
<span class="kw">impl</span> <span class="ty">CtxBuilder</span> <span class="kw">for</span> <span class="ty">ClaudeCtx</span> {
<span class="kw">fn</span> <span class="fn">build_messages</span>(
&<span class="kw">self</span>,
conv: &<span class="ty">Conversation</span>,
system_prompt: &<span class="kw">str</span>,
) -> (<span class="ty">Vec</span><<span class="ty">Message</span>>, <span class="ty">ContextStats</span>) {
<span class="cm">// Claude 下 ctx_window 够大，直接用 legacy 窗口裁切</span>
<span class="kw">let</span> (<span class="kw">mut</span> messages, stats) =
conv.<span class="fn">to_provider_messages_budgeted</span>(system_prompt, <span class="nm">self</span>.ctx_window);
<span class="cm">// 将来 provider 层想读 cache 位置时可以这样:</span>
<span class="cm">// 1. 找 messages 里最后一条 System(system + cold_zone 块末尾)</span>
<span class="cm">// 2. 这个位置就是 cache_control: ephemeral 该插的地方</span>
<span class="cm">// 3. Claude provider payload assembler 自己去定位即可,不需要 ctx 返</span>
(messages, stats)
}
<span class="kw">fn</span> <span class="fn">needs_compression</span>(&<span class="kw">self</span>, conv: &<span class="ty">Conversation</span>, sys_tokens: <span class="kw">usize</span>) -> <span class="kw">bool</span> {
conv.<span class="fn">needs_compression</span>(sys_tokens, <span class="nm">self</span>.ctx_window)
}
<span class="kw">fn</span> <span class="fn">compression_plan</span>(&<span class="kw">self</span>, conv: &<span class="ty">Conversation</span>) -> <span class="ty">Option</span><(<span class="ty">String</span>, <span class="kw">usize</span>)> {
<span class="kw">let</span> (c, n) = conv.<span class="fn">build_compression_content</span>();
<span class="kw">if</span> c.<span class="fn">is_empty</span>() || n == <span class="nm">0</span> { <span class="kw">None</span> } <span class="kw">else</span> { <span class="ty">Some</span>((c, n)) }
}
<span class="kw">fn</span> <span class="fn">truncate_tool_output</span>(&<span class="kw">self</span>, r: &<span class="kw">mut</span> <span class="ty">ToolResult</span>, name: &<span class="kw">str</span>) {
<span class="kw">super</span>::truncate::<span class="fn">truncate_output</span>(r, name, <span class="nm">self</span>.ctx_window);
}
<span class="kw">fn</span> <span class="fn">name</span>(&<span class="kw">self</span>) -> &<span class="st">'static</span> <span class="kw">str</span> { <span class="st">"claude"</span> }
}</code></pre>
<p>注册：</p>
<pre><code><span class="kw">if</span> provider.model.<span class="fn">starts_with</span>(<span class="st">"claude-"</span>) {
<span class="kw">return</span> <span class="ty">Box</span>::<span class="fn">new</span>(claude::<span class="ty">ClaudeCtx</span>::<span class="fn">new</span>(provider));
}</code></pre>
<h3>4.3 例：CustomFineTunedCtx — 完全替换 system prompt</h3>
<p><strong>需求</strong>：用户有个微调模型，训练时 system prompt 用的是自定义模板（比如 <code><|im_start|>system\n...<|im_end|></code>）。atomcode 默认的 system prompt 不能直接用，得替换。</p>
<pre><code><span class="kw">pub struct</span> <span class="ty">CustomFineTunedCtx</span> {
ctx_window: <span class="kw">usize</span>,
}
<span class="kw">impl</span> <span class="ty">CtxBuilder</span> <span class="kw">for</span> <span class="ty">CustomFineTunedCtx</span> {
<span class="kw">fn</span> <span class="fn">build_messages</span>(
&<span class="kw">self</span>,
conv: &<span class="ty">Conversation</span>,
_base_system: &<span class="kw">str</span>, <span class="cm">// ← 忽略传入的 base</span>
) -> (<span class="ty">Vec</span><<span class="ty">Message</span>>, <span class="ty">ContextStats</span>) {
<span class="cm">// 完全替换成训练时匹配的模板</span>
<span class="kw">const</span> MY_TEMPLATE: &<span class="kw">str</span> = <span class="st">"<|im_start|>system\n\
You are a helpful coding assistant. Tools: read_file, edit_file, bash.\n\
Respond with JSON tool calls.\
<|im_end|>"</span>;
conv.<span class="fn">to_provider_messages_budgeted</span>(MY_TEMPLATE, <span class="nm">self</span>.ctx_window)
}
<span class="cm">// ... 其他方法用默认 / 透传到 Conversation ...</span>
<span class="kw">fn</span> <span class="fn">truncate_tool_output</span>(&<span class="kw">self</span>, r: &<span class="kw">mut</span> <span class="ty">ToolResult</span>, name: &<span class="kw">str</span>) {
<span class="kw">super</span>::truncate::<span class="fn">truncate_output</span>(r, name, <span class="nm">self</span>.ctx_window);
}
<span class="kw">fn</span> <span class="fn">name</span>(&<span class="kw">self</span>) -> &<span class="st">'static</span> <span class="kw">str</span> { <span class="st">"custom-finetuned"</span> }
}</code></pre>
<p>注册：</p>
<pre><code><span class="kw">if</span> provider.model == <span class="st">"my-finetune-v1"</span> {
<span class="kw">return</span> <span class="ty">Box</span>::<span class="fn">new</span>(custom_ft::<span class="ty">CustomFineTunedCtx</span>::<span class="fn">new</span>(provider));
}</code></pre>
<h3>4.4 约定总结</h3>
<table>
<tr>
<th style="width: 30%;">要定制的维度</th>
<th style="width: 35%;">在 impl 里改</th>
<th>例</th>
</tr>
<tr>
<td>system prompt 文本变换</td>
<td><code>build_messages</code> 开头变换 <code>system_prompt</code> 参数</td>
<td>OllamaCtx 砍工具 schema</td>
</tr>
<tr>
<td>system prompt 完全替换</td>
<td><code>build_messages</code> 忽略参数、用自己模板</td>
<td>CustomFineTunedCtx 训练格式</td>
</tr>
<tr>
<td>触发压缩的阈值</td>
<td>重写 <code>needs_compression</code></td>
<td>OllamaCtx 40% 阈值</td>
</tr>
<tr>
<td>压缩后保留多少 / 怎么摘要</td>
<td>重写 <code>compression_plan</code></td>
<td>自定义保留策略</td>
</tr>
<tr>
<td>工具输出裁切强度</td>
<td>重写 <code>truncate_tool_output</code></td>
<td>OllamaCtx 4K 上限</td>
</tr>
<tr>
<td>消息层直接定制（加入额外 marker、调顺序）</td>
<td><code>build_messages</code> 拿到 legacy 输出后再处理</td>
<td>ClaudeCtx 未来加 cache breakpoint 元数据</td>
</tr>
<tr>
<td>工具集按模型筛选</td>
<td>目前不在 ctx 范围（在 <code>ToolRegistry</code> 层）</td>
<td>—</td>
</tr>
</table>
<div class="panel ok">
<strong>一致性保证</strong>：未命中任何规则的 provider 走 <code>DefaultCtx</code>，后者完全委托给 <code>ctx::render</code>——与重构前行为一致(render 模块的代码就是 2026-04 从 <code>Conversation</code> impl 平移出来的,byte-equivalent)。<code>ctx/render.rs</code> 里搬过来的 12 个 budgeted/sanitize 回归测试直接守护这层行为。</div>
<h2 id="tests">附录 · 测试覆盖</h2>
<table>
<tr>
<th>文件</th>
<th>测试</th>
<th style="width: 60px;">数量</th>
</tr>
<tr>
<td><code>ctx/default.rs</code></td>
<td>name / ctx_window clamp / empty conv render / compression None</td>
<td class="num">4</td>
</tr>
<tr>
<td><code>ctx/ollama.rs</code></td>
<td>name / ctx clamp / tool_output_cap 公式 / 截断生效 / CJK boundary / 35% 阈值早触发 / compression None / 非空渲染</td>
<td class="num">7</td>
</tr>
<tr>
<td><code>ctx/render.rs</code></td>
<td>budgeted 路径 10 项(empty / recent / under_80pct / drops_oldest / keeps_latest / never_system_only / emergency / drops_no_summary / preserves_order / cold_zone_compression) + sanitize 2 项</td>
<td class="num">12</td>
</tr>
<tr>
<td><code>ctx/resolver.rs</code></td>
<td>ollama 命中 / unknown 回 default / 任意 model 配 ollama provider 都走 OllamaCtx</td>
<td class="num">3</td>
</tr>
<tr>
<td><code>ctx/truncate.rs</code></td>
<td>原 <code>turn/truncation.rs</code> 单测整体迁入</td>
<td class="num">10</td>
</tr>
<tr style="font-weight: 600;">
<td colspan="2">合计</td>
<td class="num">36</td>
</tr>
</table>
<p>整 lib 套件 <code>cargo test -p atomcode-core --lib</code>: <strong>310 passed</strong>, 1 pre-existing fail (<code>self_update::is_newer_semver</code>, 与 ctx 无关)。</p>
<footer>
<p>
<strong>关联文档</strong><br>
· <code>docs/architecture.md</code> — atomcode 整体架构<br>
· 源码入口 <code>crates/atomcode-core/src/ctx/mod.rs</code><br>
· 源码 <code>crates/atomcode-core/src/ctx/</code> on branch <code>feat/ctx-module</code>
</p>
<p>
文档生成于 2026-04-20 · atomcode v4.19.0
</p>
</footer>
</div>
</body>
</html>
