thatgfsj-code 2.2.9 → 3.0.1

package/CHANGELOG.md

@@ -4,77 +4,25 @@ All notable changes to **Thatgfsj Code** are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/) and the project adheres to
 [Semantic Versioning](https://semver.org/).
 
-## [2.2.8 / 产品 0.4.5] - 2026-07-06  - 边界测试驱动的 5 处修复
+## [3.0.0 / 产品 0.5.0] - 2026-08-08  - Reasonix 风格 Prompt Caching 架构
 
-> 双版本号方案:
+> **重大版本变更**：流式协议 `AsyncGenerator<string>` → `AsyncGenerator<StreamChunk>`。
+> 自定义依赖 `LLMService.chatStream` / `App.streamResponse` 的下游代码需要适配。
 >
-> | 位置 | 版本 | 含义 |
-> |---|---|---|
-> | `package.json` `"version"` | `2.2.8` | npm 包版本,正常递增 (+0.0.1) |
-> | `src/cmd/index.tsx` `.version(...)` | `0.4.5` | **产品版本** (`gfcode --version` 看到的) |
-
-### Fixed (边界测试 `tests/edge-*.mjs` 捕获的真实 bug)
-
-**Bug 1: ToolCall 在 `result === null` 时误渲染** — 旧代码
-`tool.result !== undefined ? ... : null` 把 `null` 当有效结果,空字符串
-被显示。改成 `tool.result !== undefined && tool.result !== null`,
-`null` 现在进入 pending 状态,显示 ⏳ running。
-
-**Bug 2: `SessionManager.truncate(N)` 不生效** — compactor 的
-`preserveRecent` 默认 10,比 `maxMessages` 大,所以 `truncate(2)`
-触发 compact 时,`others.length <= preserveRecent` 分支返回原消息。
-改成 `truncate(N)` 时同步 `preserveRecent = N - 1`,让 max 边界真
-的生效。
-
-**Bug 3: thinking 正则漏掉 `<thinking>` delimiter** — 用户实际会话
-里模型用的是 `<thinking>...</thinking>` 不是 `<<<think>...</think>`。
-加上 `<thinking>` / `<THINK>` / `[think]` 三种额外格式,总共 6 个
-delimiter。
-
-**Bug 4: pollution 过滤 STRONG 模式过宽** — `/\[已中断\]/` 在消息
-**任何位置** 都触发 drop,导致 300 字的中段提到这个标记的合法
-解释型消息被误删。改成所有 STRONG 也受 length/question/temporal
-三个门控,长消息中段出现的引用不会被误杀。
-
-**Bug 5: 1.0.4 源码里没有 `npm test` script** — `prepublishOnly`
-引用了它,但 1.0.4 tarball 缺这个 script,导致 release 后用户跑
-`npm test` 会失败。加上 npm test 聚合脚本 + `prepublishOnly`
-显式跑 build + test。
+> 参考：DeepSeek-Reasonix (esengine/DeepSeek-Reasonix) 的"不可变前缀 + 追加日志 + 易变草稿"三支柱。
 
 ### Added
-
-- **5 套边界测试** (`tests/edge-*.mjs`),共 **118 个用例**,覆盖
-  - thinking compression: 嵌套 delimiter、无 closing tag、大小写、
-    whitespace-only、超长输入、混合 `bcdf` (Case O)
-  - pollution filter: 大小写、unicode 变种、`?` / 中文 `？` 时间
-    副词、`null byte`、markdown 装饰
-  - ToolCall 渲染: `null` vs `undefined`, `\\r\\n`, 10000 字符
-    单行,边界 8/9 行
-  - SessionManager: 空 user msg、1MB msg、`getMessages()` 不可变性、
-    `truncate(N)` 边界
-  - CLI 参数: -V, --version, -h, --show-thinking, -m, -i, 1000 char
-    prompt, unicode prompt, 含 shell 元字符的 prompt
-- **`npm test`** 脚本 (聚合 8 套测试, 181 / 181 pass)
-- **`prepublishOnly`** (build + test,确保发布前必过)
-- **`tests/` 加入 `.gitignore`** — 之前的 tarball 误把 dev-time
-  smoke 测试发到 npm 了
-
-### Test Summary
-
-```
-smoke-thinking:    28 / 28  pass
-smoke-tool:        18 / 18  pass
-smoke-pollution:   17 / 17  pass
-edge-thinking:     28 / 28  pass
-edge-tool:         25 / 25  pass
-edge-pollution:    23 / 23  pass
-edge-session:      27 / 27  pass
-edge-cli:          15 / 15  pass
-                  ────────────────
-total:           181 / 181  pass
-```
-
----
+- **M1 — 流式协议重构**：干掉 `@@TOOL@@` 哨兵字符串；`StreamChunk` 升级为结构化 union（text / tool_calls / thinking / usage）。
+- **M2 — Anthropic prompt cache**：`system` 改为 block 数组，最后一块挂 `cache_control: { type: 'ephemeral' }`；工具列表最后一个也挂 marker；新增 `anthropic-beta: prompt-caching-2024-07-31` header；修「第二条 system 消息被丢弃」bug。
+- **M3 — 缓存统计与 UI**：`CacheStatsStore` 持久化到 `~/.thatgfsj/cache-stats.json`；TUI Header 显示 `⚡ 命中率 87%`；新增 `/cache` 命令显示 24h 命中率曲线。
+- **M4 — 高级特性**：`VolatileScratch` 隔离 thinking；Tool-call Repair（append `[TOOL_REPAIR]` 而非修改既有消息）；`smartModel.shouldDowngrade` 智能路由；`Config.cache` 字段 + InitWizard 两步缓存配置。
+- **OpenAI provider 缓存友好**：`stableStringify` 保证 JSON 字节级稳定；流末尾 yield `{ type: 'usage' }` chunk。
+- **28 个 vitest 单元测试**（`tests/cache/`）。
+
+### Changed
+- `SessionManager.autoCompact` 不再 mutate messages，改为通过 `onSuggestNewSession` 回调通知用户调 `/new`。Reasonix 原则：超长开新会话，保护上游缓存前缀。
+- `system prompt` 片段顺序调整：NWT history 和 Date 移到末尾（易变区），前 6 段是 Immutable Prefix。
+- `LLMService.fromConfig` 接收 `Config.cache` 透传给 provider。
 
 ## [2.2.7 / 产品 0.4.4] - 2026-07-06  - 烟囱测试驱动的 anti-pollution 过滤收紧
 

package/dist/app/index.d.ts

@@ -1,6 +1,10 @@
 /**
  * App - Core application singleton
  * Simplified: directly uses LLMService (which has built-in agent loop)
+ *
+ * v3.0.0+: streamResponse yields structured StreamChunk
+ *   - runPrompt streams { type: 'text' } chunks to stdout and captures the
+ *     final usage for cache stats recording.
  */
 import { ConfigManager } from '../config/index.js';
 import { LLMService } from '../llm/index.js';
@@ -9,7 +13,8 @@ import { ToolRegistry } from '../tools/index.js';
 import { HookManager } from '../hooks/index.js';
 import { SystemPromptBuilder } from '../prompts/index.js';
 import { SkillRegistry } from '../skills/index.js';
-import type { ChatMessage } from '../types.js';
+import { CacheStatsStore } from '../cache/stats.js';
+import type { ChatMessage, ChatResponse, StreamChunk, Usage } from '../types.js';
 export declare class App {
     config: ConfigManager;
     llm: LLMService;
@@ -19,7 +24,13 @@ export declare class App {
     prompts: SystemPromptBuilder;
     skills: SkillRegistry;
     /**
-     * v2.2.5 (product 0.4.2): toggle <think> block compression. Default
+     * v3.0.0: persistent cache stats store. The single source of truth for
+     * cache hit-rate and estimated savings, surfaced through the TUI Header
+     * and the /cache command.
+     */
+    cacheStats: CacheStatsStore;
+    /**
+     * v2.2.5 (product 0.4.2): toggle  block compression. Default
      * true. Toggled by `--show-thinking` on the CLI or `/thinking on|off`
      * in the REPL.
      */
@@ -29,8 +40,17 @@ export declare class App {
     /**
      * Stream a response for the current session messages.
      * The LLMService handles the full agent loop internally.
+     *
+     * Yields structured StreamChunks. Returns the final ChatResponse (with usage
+     * if the provider reported it) so the caller can record cache stats.
+     *
+     * Implementation note: we drain the inner stream manually so the final
+     * ChatResponse returned by LLMService.chatStream is propagated as this
+     * generator's return value. Using yield* doesn't carry the return value
+     * through TS's AsyncGenerator<T, R> type inference in this version of
+     * TypeScript, so we wrap with an inner for-await and explicit return.
      */
-    streamResponse(messages?: ChatMessage[]): AsyncGenerator<string, void>;
+    streamResponse(messages?: ChatMessage[]): AsyncGenerator<StreamChunk, ChatResponse>;
     /**
      * Run a single prompt (non-interactive mode)
      *
@@ -38,10 +58,13 @@ export declare class App {
      * uses addMessageSafe, which drops the message if it contains
      * pollution markers like "[已中断]".
      *
-     * v2.2.5 (product 0.4.2): persistence also strips <think> blocks
+     * v2.2.5 (product 0.4.2): persistence also strips  blocks
      * (and similar reasoning delimiters) when showThinking is false,
      * so the conversation log stays compact.
+     *
+     * v3.0.0: yields structured StreamChunks; final usage is captured
+     * into onUsage callback for cache stats persistence.
      */
-    runPrompt(prompt: string): Promise<string>;
+    runPrompt(prompt: string, onUsage?: (usage: Usage) => void): Promise<string>;
 }
 //# sourceMappingURL=index.d.ts.map

package/dist/app/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/app/index.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACrD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAC;AAC1D,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AAEnD,OAAO,KAAK,EAAE,WAAW,EAAgB,MAAM,aAAa,CAAC;AAE7D,qBAAa,GAAG;IACd,MAAM,EAAE,aAAa,CAAC;IACtB,GAAG,EAAE,UAAU,CAAC;IAChB,OAAO,EAAE,cAAc,CAAC;IACxB,KAAK,EAAE,YAAY,CAAC;IACpB,KAAK,EAAE,WAAW,CAAC;IACnB,OAAO,EAAE,mBAAmB,CAAC;IAC7B,MAAM,EAAE,aAAa,CAAC;IACtB;;;;OAIG;IACH,YAAY,EAAE,OAAO,CAAS;IAE9B,OAAO;WAkBM,MAAM,IAAI,OAAO,CAAC,GAAG,CAAC;IA+BnC;;;OAGG;IACI,cAAc,CAAC,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,cAAc,CAAC,MAAM,EAAE,IAAI,CAAC;IAO7E;;;;;;;;;;OAUG;IACG,SAAS,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;CAsBjD"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/app/index.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACrD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAC;AAC1D,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,eAAe,EAAE,MAAM,mBAAmB,CAAC;AAEpD,OAAO,KAAK,EAAE,WAAW,EAAE,YAAY,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAEjF,qBAAa,GAAG;IACd,MAAM,EAAE,aAAa,CAAC;IACtB,GAAG,EAAE,UAAU,CAAC;IAChB,OAAO,EAAE,cAAc,CAAC;IACxB,KAAK,EAAE,YAAY,CAAC;IACpB,KAAK,EAAE,WAAW,CAAC;IACnB,OAAO,EAAE,mBAAmB,CAAC;IAC7B,MAAM,EAAE,aAAa,CAAC;IACtB;;;;OAIG;IACH,UAAU,EAAE,eAAe,CAAC;IAC5B;;;;OAIG;IACH,YAAY,EAAE,OAAO,CAAS;IAE9B,OAAO;WAoBM,MAAM,IAAI,OAAO,CAAC,GAAG,CAAC;IAyCnC;;;;;;;;;;;;OAYG;IACI,cAAc,CAAC,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,cAAc,CAAC,WAAW,EAAE,YAAY,CAAC;IA2B1F;;;;;;;;;;;;;OAaG;IACG,SAAS,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,KAAK,KAAK,IAAI,GAAG,OAAO,CAAC,MAAM,CAAC;CAgCnF"}

package/dist/app/index.js

@@ -1,6 +1,10 @@
 /**
  * App - Core application singleton
  * Simplified: directly uses LLMService (which has built-in agent loop)
+ *
+ * v3.0.0+: streamResponse yields structured StreamChunk
+ *   - runPrompt streams { type: 'text' } chunks to stdout and captures the
+ *     final usage for cache stats recording.
  */
 import { ConfigManager } from '../config/index.js';
 import { LLMService } from '../llm/index.js';
@@ -9,6 +13,7 @@ import { ToolRegistry } from '../tools/index.js';
 import { HookManager } from '../hooks/index.js';
 import { SystemPromptBuilder } from '../prompts/index.js';
 import { SkillRegistry } from '../skills/index.js';
+import { CacheStatsStore } from '../cache/stats.js';
 import { compressThinking } from '../utils/thinking.js';
 export class App {
     config;
@@ -19,12 +24,18 @@ export class App {
     prompts;
     skills;
     /**
-     * v2.2.5 (product 0.4.2): toggle <think> block compression. Default
+     * v3.0.0: persistent cache stats store. The single source of truth for
+     * cache hit-rate and estimated savings, surfaced through the TUI Header
+     * and the /cache command.
+     */
+    cacheStats;
+    /**
+     * v2.2.5 (product 0.4.2): toggle  block compression. Default
      * true. Toggled by `--show-thinking` on the CLI or `/thinking on|off`
      * in the REPL.
      */
     showThinking = false;
-    constructor(config, llm, session, tools, hooks, prompts, skills) {
+    constructor(config, llm, session, tools, hooks, prompts, skills, cacheStats) {
         this.config = config;
         this.llm = llm;
         this.session = session;
@@ -32,12 +43,21 @@ export class App {
         this.hooks = hooks;
         this.prompts = prompts;
         this.skills = skills;
+        this.cacheStats = cacheStats;
     }
     static async create() {
         const config = await ConfigManager.load();
         const aiConfig = config.getAIConfig();
         const llm = LLMService.fromConfig(aiConfig);
+        const cacheStats = new CacheStatsStore();
         const session = new SessionManager(config.get().contextLength || 50);
+        // v3.0.0: do not mutate messages when context grows — instead, surface
+        // a "consider /new" toast via onSuggestNewSession. The TUI wires this
+        // up in app.tsx; in CLI single-prompt mode it's a no-op (one-shot).
+        session.onSuggestNewSession = (info) => {
+            console.warn(`\n  ⚠️  上下文较长（${info.currentLength}/${info.max}）。` +
+                `建议调 /new 开新会话（NWT 已自动归档历史）\n`);
+        };
         const tools = new ToolRegistry();
         const hooks = new HookManager();
         const skills = new SkillRegistry();
@@ -56,17 +76,47 @@ export class App {
             skillsPrompt: skills.getActivePrompts(),
         });
         session.addMessage('system', prompts.build());
-        return new App(config, llm, session, tools, hooks, prompts, skills);
+        return new App(config, llm, session, tools, hooks, prompts, skills, cacheStats);
     }
     /**
      * Stream a response for the current session messages.
      * The LLMService handles the full agent loop internally.
+     *
+     * Yields structured StreamChunks. Returns the final ChatResponse (with usage
+     * if the provider reported it) so the caller can record cache stats.
+     *
+     * Implementation note: we drain the inner stream manually so the final
+     * ChatResponse returned by LLMService.chatStream is propagated as this
+     * generator's return value. Using yield* doesn't carry the return value
+     * through TS's AsyncGenerator<T, R> type inference in this version of
+     * TypeScript, so we wrap with an inner for-await and explicit return.
      */
     async *streamResponse(messages) {
         const msgs = messages || this.session.getMessages();
-        for await (const chunk of this.llm.chatStream(msgs)) {
-            yield chunk;
+        const inner = this.llm.chatStream(msgs);
+        const debugUsage = !!process.env.GFCODE_DEBUG_USAGE;
+        let next = await inner.next();
+        while (!next.done) {
+            // Forward chunks unchanged, but capture usage into the persistent
+            // cache stats store so the TUI Header / /cache command can read it.
+            if (next.value && next.value.type === 'usage') {
+                try {
+                    this.cacheStats.record(next.value.usage);
+                }
+                catch { /* best-effort */ }
+                if (debugUsage) {
+                    // v3.0.0 DEBUG: dump raw usage fields to stderr so the user can
+                    // confirm whether the upstream provider/relay forwards cache stats.
+                    // Enable with: GFCODE_DEBUG_USAGE=1 gfcode ...
+                    process.stderr.write('[debug_usage] ' + JSON.stringify(next.value.usage) + '\n');
+                }
+            }
+            yield next.value;
+            next = await inner.next();
         }
+        // The generator's return value (ChatResponse with usage) is propagated
+        // to callers via `for await ... await streamResponse.next()` semantics.
+        return next.value;
     }
     /**
      * Run a single prompt (non-interactive mode)
@@ -75,17 +125,26 @@ export class App {
      * uses addMessageSafe, which drops the message if it contains
      * pollution markers like "[已中断]".
      *
-     * v2.2.5 (product 0.4.2): persistence also strips <think> blocks
+     * v2.2.5 (product 0.4.2): persistence also strips  blocks
      * (and similar reasoning delimiters) when showThinking is false,
      * so the conversation log stays compact.
+     *
+     * v3.0.0: yields structured StreamChunks; final usage is captured
+     * into onUsage callback for cache stats persistence.
      */
-    async runPrompt(prompt) {
+    async runPrompt(prompt, onUsage) {
         this.session.addMessage('user', prompt);
         let fullResponse = '';
+        let finalUsage;
         try {
             for await (const chunk of this.streamResponse()) {
-                process.stdout.write(chunk);
-                fullResponse += chunk;
+                if (chunk.type === 'text' && chunk.content) {
+                    process.stdout.write(chunk.content);
+                    fullResponse += chunk.content;
+                }
+                else if (chunk.type === 'usage') {
+                    finalUsage = chunk.usage;
+                }
             }
         }
         catch (err) {
@@ -95,9 +154,16 @@ export class App {
             throw err;
         }
         console.log();
-        // v2.2.5: compress <think> blocks before persisting.
+        // v2.2.5: compress  blocks before persisting.
         const toPersist = compressThinking(fullResponse, this.showThinking);
         this.session.addMessageSafe('assistant', toPersist);
+        // v3.0.0: forward usage to caller (CLI single-shot mode records stats too)
+        if (finalUsage && onUsage) {
+            try {
+                onUsage(finalUsage);
+            }
+            catch { /* best-effort */ }
+        }
         return fullResponse;
     }
 }

package/dist/app/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/app/index.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACrD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAC;AAC1D,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAGxD,MAAM,OAAO,GAAG;IACd,MAAM,CAAgB;IACtB,GAAG,CAAa;IAChB,OAAO,CAAiB;IACxB,KAAK,CAAe;IACpB,KAAK,CAAc;IACnB,OAAO,CAAsB;IAC7B,MAAM,CAAgB;IACtB;;;;OAIG;IACH,YAAY,GAAY,KAAK,CAAC;IAE9B,YACE,MAAqB,EACrB,GAAe,EACf,OAAuB,EACvB,KAAmB,EACnB,KAAkB,EAClB,OAA4B,EAC5B,MAAqB;QAErB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;QACf,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACvB,CAAC;IAED,MAAM,CAAC,KAAK,CAAC,MAAM;QACjB,MAAM,MAAM,GAAG,MAAM,aAAa,CAAC,IAAI,EAAE,CAAC;QAC1C,MAAM,QAAQ,GAAG,MAAM,CAAC,WAAW,EAAE,CAAC;QAEtC,MAAM,GAAG,GAAG,UAAU,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,OAAO,GAAG,IAAI,cAAc,CAAC,MAAM,CAAC,GAAG,EAAE,CAAC,aAAa,IAAI,EAAE,CAAC,CAAC;QACrE,MAAM,KAAK,GAAG,IAAI,YAAY,EAAE,CAAC;QACjC,MAAM,KAAK,GAAG,IAAI,WAAW,EAAE,CAAC;QAChC,MAAM,MAAM,GAAG,IAAI,aAAa,EAAE,CAAC;QAEnC,kCAAkC;QAClC,GAAG,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;QAEhC,yBAAyB;QACzB,MAAM,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;QACjC,IAAI,OAAO,EAAE,CAAC;YACZ,MAAM,OAAO,CAAC,OAAO,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,CAAC;QAC5C,CAAC;QAED,yCAAyC;QACzC,MAAM,OAAO,GAAG,IAAI,mBAAmB,CAAC;YACtC,GAAG,EAAE,OAAO,CAAC,GAAG,EAAE;YAClB,KAAK,EAAE,KAAK,CAAC,IAAI,EAAE;YACnB,cAAc,EAAE,KAAK;YACrB,YAAY,EAAE,MAAM,CAAC,gBAAgB,EAAE;SACxC,CAAC,CAAC;QACH,OAAO,CAAC,UAAU,CAAC,QAAQ,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC;QAE9C,OAAO,IAAI,GAAG,CAAC,MAAM,EAAE,GAAG,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;IACtE,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,CAAC,cAAc,CAAC,QAAwB;QAC5C,MAAM,IAAI,GAAG,QAAQ,IAAI,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC;QACpD,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,IAAI,CAAC,EAAE,CAAC;YACpD,MAAM,KAAK,CAAC;QACd,CAAC;IACH,CAAC;IAED;;;;;;;;;;OAUG;IACH,KAAK,CAAC,SAAS,CAAC,MAAc;QAC5B,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;QAExC,IAAI,YAAY,GAAG,EAAE,CAAC;QACtB,IAAI,CAAC;YACH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,IAAI,CAAC,cAAc,EAAE,EAAE,CAAC;gBAChD,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;gBAC5B,YAAY,IAAI,KAAK,CAAC;YACxB,CAAC;QACH,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,2DAA2D;YAC3D,4DAA4D;YAC5D,UAAU;YACV,MAAM,GAAG,CAAC;QACZ,CAAC;QAED,OAAO,CAAC,GAAG,EAAE,CAAC;QACd,qDAAqD;QACrD,MAAM,SAAS,GAAG,gBAAgB,CAAC,YAAY,EAAE,IAAI,CAAC,YAAY,CAAC,CAAC;QACpE,IAAI,CAAC,OAAO,CAAC,cAAc,CAAC,WAAW,EAAE,SAAS,CAAC,CAAC;QACpD,OAAO,YAAY,CAAC;IACtB,CAAC;CACF"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/app/index.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACrD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAC;AAC1D,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,eAAe,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,EAAE,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAGxD,MAAM,OAAO,GAAG;IACd,MAAM,CAAgB;IACtB,GAAG,CAAa;IAChB,OAAO,CAAiB;IACxB,KAAK,CAAe;IACpB,KAAK,CAAc;IACnB,OAAO,CAAsB;IAC7B,MAAM,CAAgB;IACtB;;;;OAIG;IACH,UAAU,CAAkB;IAC5B;;;;OAIG;IACH,YAAY,GAAY,KAAK,CAAC;IAE9B,YACE,MAAqB,EACrB,GAAe,EACf,OAAuB,EACvB,KAAmB,EACnB,KAAkB,EAClB,OAA4B,EAC5B,MAAqB,EACrB,UAA2B;QAE3B,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;QACf,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;IAC/B,CAAC;IAED,MAAM,CAAC,KAAK,CAAC,MAAM;QACjB,MAAM,MAAM,GAAG,MAAM,aAAa,CAAC,IAAI,EAAE,CAAC;QAC1C,MAAM,QAAQ,GAAG,MAAM,CAAC,WAAW,EAAE,CAAC;QAEtC,MAAM,GAAG,GAAG,UAAU,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,UAAU,GAAG,IAAI,eAAe,EAAE,CAAC;QACzC,MAAM,OAAO,GAAG,IAAI,cAAc,CAAC,MAAM,CAAC,GAAG,EAAE,CAAC,aAAa,IAAI,EAAE,CAAC,CAAC;QACrE,uEAAuE;QACvE,sEAAsE;QACtE,oEAAoE;QACpE,OAAO,CAAC,mBAAmB,GAAG,CAAC,IAAI,EAAE,EAAE;YACrC,OAAO,CAAC,IAAI,CACV,iBAAiB,IAAI,CAAC,aAAa,IAAI,IAAI,CAAC,GAAG,IAAI;gBACnD,8BAA8B,CAC/B,CAAC;QACJ,CAAC,CAAC;QACF,MAAM,KAAK,GAAG,IAAI,YAAY,EAAE,CAAC;QACjC,MAAM,KAAK,GAAG,IAAI,WAAW,EAAE,CAAC;QAChC,MAAM,MAAM,GAAG,IAAI,aAAa,EAAE,CAAC;QAEnC,kCAAkC;QAClC,GAAG,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;QAEhC,yBAAyB;QACzB,MAAM,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;QACjC,IAAI,OAAO,EAAE,CAAC;YACZ,MAAM,OAAO,CAAC,OAAO,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,CAAC;QAC5C,CAAC;QAED,yCAAyC;QACzC,MAAM,OAAO,GAAG,IAAI,mBAAmB,CAAC;YACtC,GAAG,EAAE,OAAO,CAAC,GAAG,EAAE;YAClB,KAAK,EAAE,KAAK,CAAC,IAAI,EAAE;YACnB,cAAc,EAAE,KAAK;YACrB,YAAY,EAAE,MAAM,CAAC,gBAAgB,EAAE;SACxC,CAAC,CAAC;QACH,OAAO,CAAC,UAAU,CAAC,QAAQ,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC;QAE9C,OAAO,IAAI,GAAG,CAAC,MAAM,EAAE,GAAG,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,CAAC;IAClF,CAAC;IAED;;;;;;;;;;;;OAYG;IACH,KAAK,CAAC,CAAC,cAAc,CAAC,QAAwB;QAC5C,MAAM,IAAI,GAAG,QAAQ,IAAI,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC;QACpD,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC;QACxC,MAAM,UAAU,GAAG,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,kBAAkB,CAAC;QACpD,IAAI,IAAI,GAAG,MAAM,KAAK,CAAC,IAAI,EAAE,CAAC;QAC9B,OAAO,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;YAClB,kEAAkE;YAClE,oEAAoE;YACpE,IAAI,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;gBAC9C,IAAI,CAAC;oBAAC,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;gBAAC,CAAC;gBAAC,MAAM,CAAC,CAAC,iBAAiB,CAAC,CAAC;gBAC7E,IAAI,UAAU,EAAE,CAAC;oBACf,gEAAgE;oBAChE,oEAAoE;oBACpE,+CAA+C;oBAC/C,OAAO,CAAC,MAAM,CAAC,KAAK,CAClB,gBAAgB,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,GAAG,IAAI,CAC3D,CAAC;gBACJ,CAAC;YACH,CAAC;YACD,MAAM,IAAI,CAAC,KAAK,CAAC;YACjB,IAAI,GAAG,MAAM,KAAK,CAAC,IAAI,EAAE,CAAC;QAC5B,CAAC;QACD,uEAAuE;QACvE,wEAAwE;QACxE,OAAO,IAAI,CAAC,KAAK,CAAC;IACpB,CAAC;IAED;;;;;;;;;;;;;OAaG;IACH,KAAK,CAAC,SAAS,CAAC,MAAc,EAAE,OAAgC;QAC9D,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;QAExC,IAAI,YAAY,GAAG,EAAE,CAAC;QACtB,IAAI,UAA6C,CAAC;QAClD,IAAI,CAAC;YACH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,IAAI,CAAC,cAAc,EAAE,EAAE,CAAC;gBAChD,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC;oBAC3C,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;oBACpC,YAAY,IAAI,KAAK,CAAC,OAAO,CAAC;gBAChC,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;oBAClC,UAAU,GAAG,KAAK,CAAC,KAAK,CAAC;gBAC3B,CAAC;YACH,CAAC;QACH,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,2DAA2D;YAC3D,4DAA4D;YAC5D,UAAU;YACV,MAAM,GAAG,CAAC;QACZ,CAAC;QAED,OAAO,CAAC,GAAG,EAAE,CAAC;QACd,8CAA8C;QAC9C,MAAM,SAAS,GAAG,gBAAgB,CAAC,YAAY,EAAE,IAAI,CAAC,YAAY,CAAC,CAAC;QACpE,IAAI,CAAC,OAAO,CAAC,cAAc,CAAC,WAAW,EAAE,SAAS,CAAC,CAAC;QAEpD,2EAA2E;QAC3E,IAAI,UAAU,IAAI,OAAO,EAAE,CAAC;YAC1B,IAAI,CAAC;gBAAC,OAAO,CAAC,UAAU,CAAC,CAAC;YAAC,CAAC;YAAC,MAAM,CAAC,CAAC,iBAAiB,CAAC,CAAC;QAC1D,CAAC;QACD,OAAO,YAAY,CAAC;IACtB,CAAC;CACF"}

package/dist/cache/fingerprint.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Cache fingerprints — small, stable hashes used to detect when the
+ * "cacheable prefix" of an LLM request changes between rounds.
+ *
+ * Why this exists:
+ *   Anthropic prompt cache and DeepSeek automatic prefix cache both break
+ *   when the request body changes upstream of the cache breakpoint. To
+ *   debug "why is my hit-rate 0%", we want a quick fingerprint of the
+ *   things that *should* be cacheable (tools + system prefix) so we can
+ *   log it on every request and compare across rounds.
+ *
+ *   The fingerprint itself is NOT used to decide anything (the upstream
+ *   provider does the actual cache lookup). It's purely a debug /
+ *   observability tool — same role as `reasoning_content` in the message
+ *   schema: captured for transparency, not for control flow.
+ *
+ * Implementation:
+ *   - Uses stableStringify (sibling module) so logically equal objects
+ *     always produce the same JSON.
+ *   - Hashes with sha256 (Node built-in `crypto`), truncated to 16 chars.
+ *     16 hex chars = 64 bits, which is plenty for collision-resistance
+ *     across a single user's session.
+ */
+import type { Tool } from '../tools/types.js';
+import type { SystemSegment } from '../prompts/index.js';
+/**
+ * Hash a stable JSON representation of `value` and return a 16-char hex prefix.
+ * Recursively normalizes object keys (sorted) so insertion order does not
+ * affect the output.
+ */
+export declare function fingerprint(value: unknown): string;
+/**
+ * Compute a fingerprint of the registered tool schemas.
+ *
+ * We deliberately pick only `name`, `description`, and `inputSchema` — the
+ * three fields that matter for the upstream prompt cache. Other Tool
+ * metadata (version strings, descriptions of internal handlers, etc) is
+ * ignored because it never reaches the wire.
+ */
+export declare function fingerprintTools(tools: Tool[]): string;
+/**
+ * Compute a fingerprint of the system prompt's *immutable* portion.
+ *
+ * Volatile segments (NWT history, current time) are deliberately excluded
+ * because they change between rounds and would defeat the fingerprint's
+ * purpose as a "did the cacheable prefix change?" signal.
+ */
+export declare function fingerprintSystemPrefix(segments: SystemSegment[]): string;
+//# sourceMappingURL=fingerprint.d.ts.map

package/dist/cache/fingerprint.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"fingerprint.d.ts","sourceRoot":"","sources":["../../src/cache/fingerprint.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAGH,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,mBAAmB,CAAC;AAC9C,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAEzD;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAUlD;AAED;;;;;;;GAOG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,IAAI,EAAE,GAAG,MAAM,CAOtD;AAED;;;;;;GAMG;AACH,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,aAAa,EAAE,GAAG,MAAM,CAGzE"}

package/dist/cache/fingerprint.js

@@ -0,0 +1,100 @@
+/**
+ * Cache fingerprints — small, stable hashes used to detect when the
+ * "cacheable prefix" of an LLM request changes between rounds.
+ *
+ * Why this exists:
+ *   Anthropic prompt cache and DeepSeek automatic prefix cache both break
+ *   when the request body changes upstream of the cache breakpoint. To
+ *   debug "why is my hit-rate 0%", we want a quick fingerprint of the
+ *   things that *should* be cacheable (tools + system prefix) so we can
+ *   log it on every request and compare across rounds.
+ *
+ *   The fingerprint itself is NOT used to decide anything (the upstream
+ *   provider does the actual cache lookup). It's purely a debug /
+ *   observability tool — same role as `reasoning_content` in the message
+ *   schema: captured for transparency, not for control flow.
+ *
+ * Implementation:
+ *   - Uses stableStringify (sibling module) so logically equal objects
+ *     always produce the same JSON.
+ *   - Hashes with sha256 (Node built-in `crypto`), truncated to 16 chars.
+ *     16 hex chars = 64 bits, which is plenty for collision-resistance
+ *     across a single user's session.
+ */
+import { createHash } from 'crypto';
+/**
+ * Hash a stable JSON representation of `value` and return a 16-char hex prefix.
+ * Recursively normalizes object keys (sorted) so insertion order does not
+ * affect the output.
+ */
+export function fingerprint(value) {
+    // stableStringify is imported lazily to avoid a circular dep — fingerprint
+    // is consumed by both providers (which sit above utils/) and the cache
+    // stats store (which sits next to fingerprint). Top-level import works
+    // because utils has no other consumers besides fingerprint, so we just
+    // import statically.
+    return createHash('sha256')
+        .update(stableStringifyLocal(value))
+        .digest('hex')
+        .slice(0, 16);
+}
+/**
+ * Compute a fingerprint of the registered tool schemas.
+ *
+ * We deliberately pick only `name`, `description`, and `inputSchema` — the
+ * three fields that matter for the upstream prompt cache. Other Tool
+ * metadata (version strings, descriptions of internal handlers, etc) is
+ * ignored because it never reaches the wire.
+ */
+export function fingerprintTools(tools) {
+    const minimal = tools.map(t => ({
+        name: t.name,
+        description: t.description,
+        inputSchema: t.inputSchema,
+    }));
+    return fingerprint({ tools: minimal });
+}
+/**
+ * Compute a fingerprint of the system prompt's *immutable* portion.
+ *
+ * Volatile segments (NWT history, current time) are deliberately excluded
+ * because they change between rounds and would defeat the fingerprint's
+ * purpose as a "did the cacheable prefix change?" signal.
+ */
+export function fingerprintSystemPrefix(segments) {
+    const immutable = segments.filter(s => !s.volatile);
+    return fingerprint({ system: immutable.map(s => ({ name: s.name, content: s.content })) });
+}
+// -- Local copy of stableStringify to avoid a circular dep. This must stay
+//    byte-for-byte identical to src/utils/stableStringify.ts. If you change
+//    one, change both.
+function stableStringifyLocal(value) {
+    if (value === null)
+        return 'null';
+    if (value === undefined)
+        return 'null';
+    const t = typeof value;
+    if (t === 'string')
+        return JSON.stringify(value);
+    if (t === 'number')
+        return Number.isFinite(value) ? String(value) : 'null';
+    if (t === 'boolean')
+        return value ? 'true' : 'false';
+    if (Array.isArray(value)) {
+        return '[' + value.map((v) => stableStringifyLocal(v)).join(',') + ']';
+    }
+    if (t === 'object') {
+        const obj = value;
+        const keys = Object.keys(obj).sort();
+        const parts = [];
+        for (const k of keys) {
+            const v = obj[k];
+            if (v === undefined)
+                continue;
+            parts.push(JSON.stringify(k) + ':' + stableStringifyLocal(v));
+        }
+        return '{' + parts.join(',') + '}';
+    }
+    return 'null';
+}
+//# sourceMappingURL=fingerprint.js.map

package/dist/cache/fingerprint.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fingerprint.js","sourceRoot":"","sources":["../../src/cache/fingerprint.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,QAAQ,CAAC;AAIpC;;;;GAIG;AACH,MAAM,UAAU,WAAW,CAAC,KAAc;IACxC,2EAA2E;IAC3E,uEAAuE;IACvE,uEAAuE;IACvE,uEAAuE;IACvE,qBAAqB;IACrB,OAAO,UAAU,CAAC,QAAQ,CAAC;SACxB,MAAM,CAAC,oBAAoB,CAAC,KAAK,CAAC,CAAC;SACnC,MAAM,CAAC,KAAK,CAAC;SACb,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;AAClB,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,gBAAgB,CAAC,KAAa;IAC5C,MAAM,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;QAC9B,IAAI,EAAE,CAAC,CAAC,IAAI;QACZ,WAAW,EAAE,CAAC,CAAC,WAAW;QAC1B,WAAW,EAAE,CAAC,CAAC,WAAW;KAC3B,CAAC,CAAC,CAAC;IACJ,OAAO,WAAW,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,CAAC,CAAC;AACzC,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,uBAAuB,CAAC,QAAyB;IAC/D,MAAM,SAAS,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;IACpD,OAAO,WAAW,CAAC,EAAE,MAAM,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC;AAC7F,CAAC;AAED,2EAA2E;AAC3E,4EAA4E;AAC5E,uBAAuB;AAEvB,SAAS,oBAAoB,CAAC,KAAc;IAC1C,IAAI,KAAK,KAAK,IAAI;QAAE,OAAO,MAAM,CAAC;IAClC,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,MAAM,CAAC;IACvC,MAAM,CAAC,GAAG,OAAO,KAAK,CAAC;IACvB,IAAI,CAAC,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;IACjD,IAAI,CAAC,KAAK,QAAQ;QAAE,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAe,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;IACrF,IAAI,CAAC,KAAK,SAAS;QAAE,OAAO,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC;IACrD,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QACzB,OAAO,GAAG,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC;IACzE,CAAC;IACD,IAAI,CAAC,KAAK,QAAQ,EAAE,CAAC;QACnB,MAAM,GAAG,GAAG,KAAgC,CAAC;QAC7C,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;QACrC,MAAM,KAAK,GAAa,EAAE,CAAC;QAC3B,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;YACrB,MAAM,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC;YACjB,IAAI,CAAC,KAAK,SAAS;gBAAE,SAAS;YAC9B,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,GAAG,GAAG,GAAG,oBAAoB,CAAC,CAAC,CAAC,CAAC,CAAC;QAChE,CAAC;QACD,OAAO,GAAG,GAAG,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC;IACrC,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC"}

package/dist/cache/index.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Cache subsystem barrel. Re-exports the public surface used by the TUI,
+ * the App singleton, and the LLM service.
+ *
+ * Modules:
+ *   - fingerprint.ts: stable hashes for tools + immutable system prefix
+ *   - stats.ts:       CacheStatsStore — per-round hit/miss + cost savings
+ *   - volatile.ts:    VolatileScratch — per-round scratch that never reaches the API
+ *   - smartModel.ts:  smart-model routing hook (simple → mini, complex → main)
+ */
+export { fingerprint, fingerprintTools, fingerprintSystemPrefix } from './fingerprint.js';
+export { CacheStatsStore, estimateSavingsCNY, type CacheStats, type CacheSnapshot } from './stats.js';
+export { VolatileScratch } from './volatile.js';
+export { shouldDowngrade, type SmartModelDecision } from './smartModel.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/cache/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/cache/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,WAAW,EAAE,gBAAgB,EAAE,uBAAuB,EAAE,MAAM,kBAAkB,CAAC;AAC1F,OAAO,EAAE,eAAe,EAAE,kBAAkB,EAAE,KAAK,UAAU,EAAE,KAAK,aAAa,EAAE,MAAM,YAAY,CAAC;AACtG,OAAO,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAChD,OAAO,EAAE,eAAe,EAAE,KAAK,kBAAkB,EAAE,MAAM,iBAAiB,CAAC"}

package/dist/cache/index.js

@@ -0,0 +1,15 @@
+/**
+ * Cache subsystem barrel. Re-exports the public surface used by the TUI,
+ * the App singleton, and the LLM service.
+ *
+ * Modules:
+ *   - fingerprint.ts: stable hashes for tools + immutable system prefix
+ *   - stats.ts:       CacheStatsStore — per-round hit/miss + cost savings
+ *   - volatile.ts:    VolatileScratch — per-round scratch that never reaches the API
+ *   - smartModel.ts:  smart-model routing hook (simple → mini, complex → main)
+ */
+export { fingerprint, fingerprintTools, fingerprintSystemPrefix } from './fingerprint.js';
+export { CacheStatsStore, estimateSavingsCNY } from './stats.js';
+export { VolatileScratch } from './volatile.js';
+export { shouldDowngrade } from './smartModel.js';
+//# sourceMappingURL=index.js.map

package/dist/cache/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/cache/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,WAAW,EAAE,gBAAgB,EAAE,uBAAuB,EAAE,MAAM,kBAAkB,CAAC;AAC1F,OAAO,EAAE,eAAe,EAAE,kBAAkB,EAAuC,MAAM,YAAY,CAAC;AACtG,OAAO,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAChD,OAAO,EAAE,eAAe,EAA2B,MAAM,iBAAiB,CAAC"}

package/dist/cache/smartModel.d.ts

@@ -0,0 +1,40 @@
+/**
+ * Smart-model routing (P3 of the Reasonix plan).
+ *
+ * Idea: classify the user's input as "simple" or "complex" and route
+ * simple queries to a cheaper / faster model while keeping the main model
+ * for complex work. The classification is intentionally trivial — prompt
+ * length + presence of recent tool calls — because the cost of a
+ * misclassification (a slightly worse answer on a simple query) is much
+ * lower than the cost of running a frontier model on every greeting.
+ *
+ * Trade-off: each routing decision adds one model swap (different
+ * `model` field on the wire). If the cache prefix is keyed on the model
+ * string, the cache miss rate can spike on every simple query. To keep
+ * the prefix stable, callers should pass a *family* (e.g. "anthropic" or
+ * "deepseek") rather than a specific model name when forwarding the
+ * prompt to the downstream provider — that's outside the scope of this
+ * hook; this module only answers the question "should I downgrade?".
+ *
+ * v3.0.0: shipped as a stub. Real decision logic is conservative
+ * (downgrade only when prompt is short AND no recent tool activity).
+ * Future iterations can add heuristic or classifier-based routing.
+ */
+import type { ChatMessage } from '../types.js';
+export interface SmartModelDecision {
+    /** Should we route this turn to a cheaper model? */
+    downgrade: boolean;
+    /** Why we made this call — useful for `/cache` debug output. */
+    reason: string;
+}
+/**
+ * Heuristic: downgrade when
+ *   1. The latest user message is short (< 200 chars), AND
+ *   2. The last 3 messages contain no assistant tool_calls.
+ *
+ * This covers greetings, one-line questions, and simple follow-ups
+ * ("thanks", "ok", "explain this one line") while routing anything
+ * that triggered a tool to the main model.
+ */
+export declare function shouldDowngrade(messages: ChatMessage[], lastUserInput: string): SmartModelDecision;
+//# sourceMappingURL=smartModel.d.ts.map

package/dist/cache/smartModel.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"smartModel.d.ts","sourceRoot":"","sources":["../../src/cache/smartModel.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAE/C,MAAM,WAAW,kBAAkB;IACjC,oDAAoD;IACpD,SAAS,EAAE,OAAO,CAAC;IACnB,gEAAgE;IAChE,MAAM,EAAE,MAAM,CAAC;CAChB;AAED;;;;;;;;GAQG;AACH,wBAAgB,eAAe,CAAC,QAAQ,EAAE,WAAW,EAAE,EAAE,aAAa,EAAE,MAAM,GAAG,kBAAkB,CAUlG"}

package/dist/cache/smartModel.js

@@ -0,0 +1,43 @@
+/**
+ * Smart-model routing (P3 of the Reasonix plan).
+ *
+ * Idea: classify the user's input as "simple" or "complex" and route
+ * simple queries to a cheaper / faster model while keeping the main model
+ * for complex work. The classification is intentionally trivial — prompt
+ * length + presence of recent tool calls — because the cost of a
+ * misclassification (a slightly worse answer on a simple query) is much
+ * lower than the cost of running a frontier model on every greeting.
+ *
+ * Trade-off: each routing decision adds one model swap (different
+ * `model` field on the wire). If the cache prefix is keyed on the model
+ * string, the cache miss rate can spike on every simple query. To keep
+ * the prefix stable, callers should pass a *family* (e.g. "anthropic" or
+ * "deepseek") rather than a specific model name when forwarding the
+ * prompt to the downstream provider — that's outside the scope of this
+ * hook; this module only answers the question "should I downgrade?".
+ *
+ * v3.0.0: shipped as a stub. Real decision logic is conservative
+ * (downgrade only when prompt is short AND no recent tool activity).
+ * Future iterations can add heuristic or classifier-based routing.
+ */
+/**
+ * Heuristic: downgrade when
+ *   1. The latest user message is short (< 200 chars), AND
+ *   2. The last 3 messages contain no assistant tool_calls.
+ *
+ * This covers greetings, one-line questions, and simple follow-ups
+ * ("thanks", "ok", "explain this one line") while routing anything
+ * that triggered a tool to the main model.
+ */
+export function shouldDowngrade(messages, lastUserInput) {
+    if (lastUserInput.length > 200) {
+        return { downgrade: false, reason: 'prompt-too-long' };
+    }
+    const recent = messages.slice(-3);
+    const hasRecentToolCalls = recent.some(m => m.tool_calls && m.tool_calls.length > 0);
+    if (hasRecentToolCalls) {
+        return { downgrade: false, reason: 'recent-tool-call' };
+    }
+    return { downgrade: true, reason: 'short-no-tools' };
+}
+//# sourceMappingURL=smartModel.js.map

package/dist/cache/smartModel.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"smartModel.js","sourceRoot":"","sources":["../../src/cache/smartModel.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAWH;;;;;;;;GAQG;AACH,MAAM,UAAU,eAAe,CAAC,QAAuB,EAAE,aAAqB;IAC5E,IAAI,aAAa,CAAC,MAAM,GAAG,GAAG,EAAE,CAAC;QAC/B,OAAO,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,EAAE,iBAAiB,EAAE,CAAC;IACzD,CAAC;IACD,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IAClC,MAAM,kBAAkB,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,UAAU,IAAI,CAAC,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IACrF,IAAI,kBAAkB,EAAE,CAAC;QACvB,OAAO,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,EAAE,kBAAkB,EAAE,CAAC;IAC1D,CAAC;IACD,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,gBAAgB,EAAE,CAAC;AACvD,CAAC"}