@zhushanwen/pi-subagent-workflow 8.14.7 → 9.0.1

package/src/injectors/__tests__/engine-section-stability.test.ts

@@ -1,353 +0,0 @@
-// engine-section-stability.test.ts —— [engine-awareness U4 stability-guard]
-//
-// 字节稳定守护（设计 docs/design/subagent-engine-awareness-injection.md（已删除，git 可追溯）§3.3 D7 /
-// §4 A8 的代码级前置）：engine 段延续注入纪律——确定性渲染、链尾位置，引擎切换只断
-// system prompt 尾部 cache 前缀（与 provider models 段变更同判，不新增 cache 破坏面，
-// cache-probe 前缀指纹归因兼容）。
-//
-// 三层覆盖：
-// 1. 段渲染确定性：pi / zcode / ghost 三形态多次独立调用逐字节相等（toBe 级）；
-// 2. 引擎切换只变尾部：provider models 段前缀不变，变化只发生在 engine 段位置（尾部）；
-// 3. 段序守护：engine 段恒位于 provider models 段之后（恒链尾）。双保险实现：
-//    a) handler 链模拟——真实导入 setupModelListInjector + 复刻 engine handler
-//       （D7-④ 起为 engine-awareness.ts 的 setupEngineAwarenessInjector）的渲染
-//       拼装（检测编排部分归 engine-awareness.test.ts），模拟 pi 的
-//       before_agent_start 链式叠加语义；
-//    b) 源码级结构断言——index.ts 锚定注册序（四个 setup 调用先后），engine-awareness.ts
-//       锚定拼装形态。
-//
-// 源码级断言的脆弱性权衡（刻意为之）：engine handler 的生产实现装配依赖
-// ModelConfigService 单例与 sessionState 存取器（getModelConfigService() 在测试环境
-// 恒 null，导入后 handler 直接早退），链模拟仍以复刻形态锚定拼装保真；而
-// before_agent_start 多 handler 的段序由注册序唯一决定，「注册序」这一事实只存在于
-// index.ts 源码中——源码结构断言是唯一能锚定它的方式。重构挪动注册位置导致本断言
-// 失败属预期行为（提醒同步更新链模拟与守护锚点），不是误报。
-
-import { readFileSync } from "node:fs";
-import { fileURLToPath } from "node:url";
-
-import { afterEach, beforeEach, describe, expect, it } from "vitest";
-
-import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
-
-// C5①：formatModelList/ModelEntry 下沉 core barrel（setupModelListInjector 留 injector）
-import { formatModelList, type ModelEntry } from "@zhushanwen/subagent-core";
-import { MODEL_LIST_GUIDE, setupModelListInjector } from "../model-list-injector.ts";
-import type { EnginePort } from "@zhushanwen/subagent-core";
-import { clearEngines, registerEngine } from "@zhushanwen/subagent-core/execution/engine/registry.ts";
-import {
-	buildEngineModelsPromptAppend,
-	buildSubagentEngineSection,
-} from "@zhushanwen/subagent-core";
-
-// ── 测试数据 ────────────────────────────────────────────
-
-/** pi registry 风格条目（provider models 段数据源；同时充当 ctx.modelRegistry 快照）。 */
-const PROVIDER_ENTRIES: ModelEntry[] = [
-	{
-		provider: "zai-coding-cn",
-		id: "glm-5.3",
-		name: "GLM 5.3",
-		reasoning: true,
-		input: ["text"],
-		contextWindow: 200_000,
-	},
-	{
-		provider: "minimax-cn",
-		id: "MiniMax-M3",
-		name: "MiniMax M3",
-		reasoning: true,
-		input: ["text", "image"],
-		contextWindow: 1_000_000,
-	},
-];
-
-/** zcode 引擎清单（v2 registry 风格 id）。 */
-const ZCODE_MODELS: Array<{ id: string; name?: string }> = [
-	{ id: "builtin:bigmodel-coding-plan/GLM-5.3", name: "GLM-5.3" },
-	{ id: "builtin:bigmodel-coding-plan/GLM-5.3-Flash" },
-];
-
-/** 模拟 pi 核心 system prompt（engine 追加段之外的全部内容）。 */
-const BASE = "You are a coding agent.\n\n# System\nCore prompt body.";
-
-/**
- * engine handler 的追加段拼装（与 engine-awareness.ts setupEngineAwarenessInjector
- * 内 handler 同款：状态段 + 清单段、空段剔除、\n\n 连接）。复刻保真由下方「段序守护」
- * describe 的源码锚点断言保证。
- */
-function composeEngineAppend(defaultEngine: string | undefined): string {
-	return [buildSubagentEngineSection(defaultEngine), buildEngineModelsPromptAppend(defaultEngine)]
-		.filter((part) => part !== "")
-		.join("\n\n");
-}
-
-/** index.ts 尾部追加模板的分隔符（`${event.systemPrompt}\n\n${append}`）。 */
-const APPEND_SEPARATOR = "\n\n";
-
-/** 合成一个 turn 的 system prompt 尾部：BASE + provider models 段 + engine 追加段。 */
-function composeTurn(defaultEngine: string | undefined): string {
-	const providerSection = formatModelList(PROVIDER_ENTRIES, { guide: MODEL_LIST_GUIDE });
-	const afterProvider = providerSection === "" ? BASE : BASE + providerSection;
-	const append = composeEngineAppend(defaultEngine);
-	return append === "" ? afterProvider : `${afterProvider}${APPEND_SEPARATOR}${append}`;
-}
-
-/** 最大公共前缀长度（逐码元比较；用于「分叉点不早于稳定头部」断言）。 */
-function commonPrefixLength(a: string, b: string): number {
-	const n = Math.min(a.length, b.length);
-	let i = 0;
-	while (i < n && a[i] === b[i]) i++;
-	return i;
-}
-
-/** 最小 fake 引擎（照 model-prompt.test.ts 同款形态）：仅 listModels 参与注入链。 */
-function fakeEngine(id: string, models: Array<{ id: string; name?: string }>): EnginePort {
-	return {
-		id,
-		capabilities: () => ({ conversation: "unsupported", steer: "unsupported", sandbox: "none" }),
-		probe: async () => ({ ok: true, engineVersion: "test" }),
-		run: async () => {
-			throw new Error("not used in this test");
-		},
-		read: async () => ({ engineId: id, turns: [], source: "outcome-only" }),
-		// 每次调用返回新数组实例（模拟 registry 每 turn 现值）——渲染必须与实例无关
-		listModels: () => models.map((m) => ({ ...m })),
-	};
-}
-
-beforeEach(() => {
-	clearEngines();
-	registerEngine("zcode", () => fakeEngine("zcode", ZCODE_MODELS));
-});
-
-afterEach(() => {
-	clearEngines();
-});
-
-// ── 1. 段渲染确定性（D7）────────────────────────────────
-
-describe("段渲染确定性（D7：同输入恒同输出）", () => {
-	it("pi 形态：状态段多次独立调用逐字节相等；清单段恒空串", () => {
-		// pi 形态不依赖注册表（直接走缺省分支），registry 中有 zcode 不影响
-		const first = buildSubagentEngineSection("pi");
-		const second = buildSubagentEngineSection("pi");
-		const third = buildSubagentEngineSection("pi");
-		expect(first).toBe(second);
-		expect(second).toBe(third);
-		// 非空自检：确认 toBe 断言真的在比对段内容而非空串
-		expect(first).toContain("<current_subagent_engine>");
-		expect(buildEngineModelsPromptAppend("pi")).toBe("");
-	});
-
-	it("zcode 形态：状态段 + 清单段多次独立调用逐字节相等（listModels 每次返回新数组实例）", () => {
-		const s1 = buildSubagentEngineSection("zcode");
-		const s2 = buildSubagentEngineSection("zcode");
-		const m1 = buildEngineModelsPromptAppend("zcode");
-		const m2 = buildEngineModelsPromptAppend("zcode");
-		expect(s1).toBe(s2);
-		expect(m1).toBe(m2);
-		expect(m1).toContain("<available_zcode_models>");
-		expect(m1).toContain("<id>builtin:bigmodel-coding-plan/GLM-5.3</id>");
-	});
-
-	it("ghost 形态（未注册引擎）：警告段多次独立调用逐字节相等", () => {
-		const first = buildSubagentEngineSection("ghost");
-		const second = buildSubagentEngineSection("ghost");
-		expect(first).toBe(second);
-		expect(first).toContain("engine 'ghost' is not registered");
-		expect(buildEngineModelsPromptAppend("ghost")).toBe("");
-	});
-
-	it("形态间干扰不产生状态残留：zcode → pi → zcode，两次 zcode 输出逐字节相等", () => {
-		const before = buildSubagentEngineSection("zcode");
-		// 穿插渲染其他形态（含清单段消费 listModels），再回到 zcode
-		buildSubagentEngineSection("pi");
-		buildEngineModelsPromptAppend("zcode");
-		buildSubagentEngineSection("ghost");
-		expect(buildSubagentEngineSection("zcode")).toBe(before);
-	});
-
-	it("注册表重建（clearEngines + 同 id 同清单重注册）后同形态输出不变", () => {
-		const firstSection = buildSubagentEngineSection("zcode");
-		const firstAppend = buildEngineModelsPromptAppend("zcode");
-		clearEngines();
-		registerEngine("zcode", () => fakeEngine("zcode", ZCODE_MODELS));
-		expect(buildSubagentEngineSection("zcode")).toBe(firstSection);
-		expect(buildEngineModelsPromptAppend("zcode")).toBe(firstAppend);
-	});
-
-	it("组合拼装（index.ts 同款：状态段 + 清单段 join）同引擎多次调用逐字节相等", () => {
-		for (const engine of ["zcode", "pi", "ghost"]) {
-			expect(composeEngineAppend(engine)).toBe(composeEngineAppend(engine));
-		}
-	});
-});
-
-// ── 2. 引擎切换只变尾部（A8 前置）───────────────────────
-
-/**
- * 断言「从 from 切到 to 时变化只发生在尾部 engine 段」：
- *   - provider models 段及其之前的头部逐字节不变（公共前缀 toBe 级断言）；
- *   - 剥离尾部 engine 追加段后两串余部逐字节相等且恰为稳定头部；
- *   - 分叉点不早于稳定头部末尾（差异绝不侵入 provider models 段）；
- *   - 切换后的引擎追加段恰居 prompt 尾部。
- */
-function expectOnlyTailDiffers(from: string | undefined, to: string | undefined): void {
-	const stableHead = BASE + formatModelList(PROVIDER_ENTRIES, { guide: MODEL_LIST_GUIDE });
-	const before = composeTurn(from);
-	const after = composeTurn(to);
-	const beforeAppend = composeEngineAppend(from);
-	const afterAppend = composeEngineAppend(to);
-
-	expect(before).not.toBe(after);
-	expect(before.startsWith(stableHead)).toBe(true);
-	expect(after.startsWith(stableHead)).toBe(true);
-	expect(before.slice(0, before.length - beforeAppend.length - APPEND_SEPARATOR.length)).toBe(
-		stableHead,
-	);
-	expect(after.slice(0, after.length - afterAppend.length - APPEND_SEPARATOR.length)).toBe(
-		stableHead,
-	);
-	expect(commonPrefixLength(before, after)).toBeGreaterThanOrEqual(stableHead.length);
-	expect(after.endsWith(afterAppend)).toBe(true);
-}
-
-describe("引擎切换只变尾部（A8 前置：变化只断尾部 cache 前缀）", () => {
-	it("zcode → pi：provider models 段前缀不变，变化只发生在 engine 段（尾部）", () => {
-		expectOnlyTailDiffers("zcode", "pi");
-	});
-
-	it("pi → zcode（反向切换）：同款只变尾部", () => {
-		expectOnlyTailDiffers("pi", "zcode");
-	});
-
-	it("zcode → ghost（配置手误形态）：G4 降级段同样只替换尾部，不破坏前缀", () => {
-		expectOnlyTailDiffers("zcode", "ghost");
-	});
-});
-
-// ── 3. 段序守护：engine 段恒链尾 ────────────────────────
-
-/** 链模拟用 handler 形态（参数收窄到运行时实际消费的字段）。 */
-type ChainHandler = (event: { systemPrompt: string }, ctx: unknown) => unknown;
-
-/** fake ExtensionAPI：捕获 pi.on 注册的 handler（保持注册序）。 */
-function capturePi(): { handlers: ChainHandler[]; api: ExtensionAPI } {
-	const handlers: ChainHandler[] = [];
-	const api = {
-		on: (_event: string, handler: (...args: unknown[]) => unknown) => {
-			// on 的 mock 签名参数为 unknown[]，收窄到两参形态（pi 调用约定恒两参）
-			handlers.push(((event: unknown, ctx: unknown) => handler(event, ctx)) as ChainHandler);
-		},
-	} as unknown as ExtensionAPI;
-	return { handlers, api };
-}
-
-/** 模拟 pi 的 before_agent_start 串联语义：按注册序执行，前序返回的 systemPrompt 作为后序输入。 */
-async function runHandlerChain(handlers: ChainHandler[], initialPrompt: string): Promise<string> {
-	let prompt = initialPrompt;
-	const ctx = { modelRegistry: { getAvailable: () => PROVIDER_ENTRIES } };
-	for (const handler of handlers) {
-		const result: unknown = await handler({ systemPrompt: prompt }, ctx);
-		// 运行时收窄：handler 返回 { systemPrompt: string } 时替换链上 prompt，否则保持
-		if (typeof result === "object" && result !== null && "systemPrompt" in result) {
-			const sp: unknown = (result as { systemPrompt: unknown }).systemPrompt;
-			if (typeof sp === "string") prompt = sp;
-		}
-	}
-	return prompt;
-}
-
-describe("段序守护：engine 段恒链尾（D7）", () => {
-	it("handler 链模拟：按 index.ts 注册序执行后，engine 段位于 provider models 段之后且居尾", async () => {
-		const { handlers, api } = capturePi();
-		// 真实导入 model list injector（provider models 段的生产渲染路径）
-		setupModelListInjector(api);
-		// 复刻 engine handler（engine-awareness.ts setupEngineAwarenessInjector）的渲染
-		// 拼装部分（检测编排部分归 engine-awareness.test.ts）
-		const defaultEngine = "zcode";
-		handlers.push((event) => {
-			const append = composeEngineAppend(defaultEngine);
-			return { systemPrompt: `${event.systemPrompt}${APPEND_SEPARATOR}${append}` };
-		});
-
-		const prompt = await runHandlerChain(handlers, BASE);
-
-		const providerIdx = prompt.indexOf("<available_provider_models>");
-		const statusIdx = prompt.indexOf("<current_subagent_engine>");
-		const zcodeListIdx = prompt.indexOf("<available_zcode_models>");
-		expect(providerIdx).toBeGreaterThanOrEqual(0);
-		// 段序硬约束：engine 状态段与引擎清单段都在 provider models 段之后
-		expect(statusIdx).toBeGreaterThan(providerIdx);
-		expect(zcodeListIdx).toBeGreaterThan(providerIdx);
-		// 恒链尾：engine 追加段是 prompt 的最后内容
-		expect(prompt.endsWith(`${APPEND_SEPARATOR}${composeEngineAppend(defaultEngine)}`)).toBe(true);
-		// model list handler 注入形态锚定（BASE + injection 无分隔，生产 handler 同款）
-		expect(prompt.startsWith(BASE + formatModelList(PROVIDER_ENTRIES, { guide: MODEL_LIST_GUIDE }))).toBe(true);
-	});
-
-	it("对照：注册序颠倒（engine 先、model list 后）时段序断言必失败——守护有判别力", async () => {
-		const { handlers, api } = capturePi();
-		const defaultEngine = "zcode";
-		handlers.push((event) => {
-			const append = composeEngineAppend(defaultEngine);
-			return { systemPrompt: `${event.systemPrompt}${APPEND_SEPARATOR}${append}` };
-		});
-		setupModelListInjector(api);
-
-		const prompt = await runHandlerChain(handlers, BASE);
-
-		// 颠倒后 engine 段跑到 provider models 段之前——证明上一条 it 的断言非平凡
-		const providerIdx = prompt.indexOf("<available_provider_models>");
-		const statusIdx = prompt.indexOf("<current_subagent_engine>");
-		expect(statusIdx).toBeGreaterThanOrEqual(0);
-		expect(statusIdx).toBeLessThan(providerIdx);
-	});
-
-	it("源码级守护：engine handler 注册序在 model list injector 之后（段序由注册序唯一决定）", () => {
-		const src = readFileSync(fileURLToPath(new URL("../../index.ts", import.meta.url)), "utf8");
-		const eaSrc = readFileSync(fileURLToPath(new URL("../engine-awareness.ts", import.meta.url)), "utf8");
-		const lineOf = (needle: string): number => {
-			const idx = src.indexOf(needle);
-			if (idx < 0) throw new Error(`index.ts 中未找到守护锚点：${needle}`);
-			return src.slice(0, idx).split("\n").length;
-		};
-
-		// 注册序：subagents → workflows → provider models → engine handler（链尾，D7）。
-		// D7-④：index.ts 不再有直接 pi.on("before_agent_start") 注册——四段注入全部经
-		// setup* 函数（engine 段经 setupEngineAwarenessInjector）。「零直接注册」断言让
-		// 「链尾」语义精确成立；未来若新增直接注册需人工复核段序。
-		expect(src.split('pi.on("before_agent_start"')).toHaveLength(1);
-		const lineSubagents = lineOf("setupSubagentListInjector(pi);");
-		const lineWorkflows = lineOf("setupWorkflowListInjector(pi);");
-		const lineModels = lineOf("setupModelListInjector(pi);");
-		const lineEngineSetup = lineOf("setupEngineAwarenessInjector(pi,");
-		expect(lineSubagents).toBeLessThan(lineWorkflows);
-		expect(lineWorkflows).toBeLessThan(lineModels);
-		expect(lineModels).toBeLessThan(lineEngineSetup);
-
-		// engine handler 的唯一注册点在 engine-awareness.ts（setupEngineAwarenessInjector
-		// 内恰一处 pi.on("before_agent_start")）。
-		expect(eaSrc.split('pi.on("before_agent_start"')).toHaveLength(2);
-
-		// engine handler 内渲染调用序：状态段在清单段之前（append 内段序：状态段 → 清单段）。
-		// 两个调用在 engine-awareness.ts 中位于同一行（数组字面量），行号无法区分先后——
-		// 用字符偏移断言
-		const offsetStatusCall = eaSrc.indexOf("buildSubagentEngineSection(defaultEngine)");
-		const offsetModelsCall = eaSrc.indexOf("buildEngineModelsPromptAppend(defaultEngine)");
-		if (offsetStatusCall < 0 || offsetModelsCall < 0) {
-			throw new Error("engine-awareness.ts 中未找到 engine 段渲染调用锚点");
-		}
-		expect(offsetStatusCall).toBeGreaterThan(eaSrc.indexOf('pi.on("before_agent_start"'));
-		expect(offsetStatusCall).toBeLessThan(offsetModelsCall);
-
-		// 链模拟复刻保真锚点：append 拼装形态（空段剔除 + \n\n 连接 + 尾部追加模板）。
-		// 这些锚点保证 composeEngineAppend 与生产拼装（engine-awareness.ts）的一致性——
-		// 生产形态变化会先在这里红，提醒同步更新本文件全部复刻点。
-		expect(eaSrc).toContain('[buildSubagentEngineSection(defaultEngine), buildEngineModelsPromptAppend(defaultEngine)]');
-		expect(eaSrc).toContain('.filter((part) => part !== "")');
-		expect(eaSrc).toContain('.join("\\n\\n")');
-		expect(eaSrc).toContain("${event.systemPrompt}\\n\\n${append}");
-	});
-});

package/src/injectors/__tests__/model-list-injector.test.ts

@@ -1,136 +0,0 @@
-// model-list-injector 单测
-//
-// 覆盖（与 workflow-list-injector.test.ts 对称，但本 injector 无文件发现/模块级
-// 缓存，纯函数比重更大）：
-// 1. formatModelList：排序稳定性（registry 返回顺序不保证，输出必须字节稳定）、
-//    caps 推导（reasoning/vision）、空 caps 省略、XML 转义、空列表不注入
-// 2. setupModelListInjector handler：注入 append、空列表返回 undefined、
-//    registry 异常不阻断（fail-safe）
-
-import { describe, expect, it, vi } from "vitest";
-
-import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
-
-// C5①：formatModelList + ModelEntry 口径下沉 core barrel——测试改从 barrel 取实现，
-// guide 用 pi 宿主注入文案
-import { formatModelList, type ModelEntry } from "@zhushanwen/subagent-core";
-import { MODEL_LIST_GUIDE } from "../model-list-injector";
-
-// ── 测试数据工厂 ────────────────────────────────────────
-
-function entry(overrides: Partial<ModelEntry> = {}): ModelEntry {
-	return {
-		provider: "zai-coding-cn",
-		id: "glm-5.2",
-		name: "GLM 5.2",
-		reasoning: true,
-		input: ["text"],
-		contextWindow: 200_000,
-		...overrides,
-	};
-}
-
-// ── 纯函数：formatModelList ─────────────────────────────
-
-describe("formatModelList", () => {
-	it("空列表返回空串（不注入）", () => {
-		expect(formatModelList([], { guide: MODEL_LIST_GUIDE })).toBe("");
-	});
-
-	it("渲染 provider/modelId + name + caps + contextWindow", () => {
-		const out = formatModelList([
-			entry({ provider: "p1", id: "m1", name: "Model One" }),
-		], { guide: MODEL_LIST_GUIDE });
-		expect(out).toContain("<available_provider_models>");
-		expect(out).toContain("<model><id>p1/m1</id><name>Model One</name><caps>reasoning</caps><contextWindow>200000</contextWindow></model>");
-		expect(out).toContain("</available_provider_models>");
-	});
-
-	it("input 含 image 时 caps 加 vision；无能力时省略 caps 段", () => {
-		const out = formatModelList([
-			entry({ id: "vision-m", input: ["text", "image"] }),
-			entry({ id: "plain-m", reasoning: false }),
-		], { guide: MODEL_LIST_GUIDE });
-		expect(out).toContain("<id>zai-coding-cn/vision-m</id>");
-		expect(out).toContain("<caps>reasoning,vision</caps>");
-		// plain-m：reasoning=false 且无 image → 无 caps 段
-		expect(out).toMatch(/<id>zai-coding-cn\/plain-m<\/id><name>[^<]*<\/name><contextWindow>/);
-	});
-
-	it("按 (provider, id) 排序——输入乱序输出仍字节稳定（KV cache 前提）", () => {
-		const models = [
-			entry({ provider: "b-prov", id: "z-model" }),
-			entry({ provider: "a-prov", id: "y-model" }),
-			entry({ provider: "a-prov", id: "x-model" }),
-		];
-		const out1 = formatModelList([...models], { guide: MODEL_LIST_GUIDE });
-		const out2 = formatModelList([...models].reverse(), { guide: MODEL_LIST_GUIDE });
-		expect(out1).toBe(out2);
-		// 排序断言：a-prov/x-model 在最前，b-prov 在最后
-		const ids = [...out1.matchAll(/<id>([^<]+)<\/id>/g)].map((m) => m[1]);
-		expect(ids).toEqual(["a-prov/x-model", "a-prov/y-model", "b-prov/z-model"]);
-	});
-
-	it("排序为码点序而非 locale 序（跨环境字节一致契约，禁 localeCompare）", () => {
-		// 判别样本：码点序 "B"(0x42) < "a"(0x61)，而多数 locale 的 localeCompare
-		// 会把 "a-model" 排在 "B-model" 前——本断言在 localeCompare 实现下必挂
-		const out = formatModelList([
-			entry({ provider: "p", id: "a-model" }),
-			entry({ provider: "p", id: "B-model" }),
-		], { guide: MODEL_LIST_GUIDE });
-		const ids = [...out.matchAll(/<id>([^<]+)<\/id>/g)].map((m) => m[1]);
-		expect(ids).toEqual(["p/B-model", "p/a-model"]);
-	});
-
-	it("name 含 XML 特殊字符时转义（防注入段结构破坏）", () => {
-		const out = formatModelList([
-			entry({ provider: "p", id: "m", name: `A<&>"'B` }),
-		], { guide: MODEL_LIST_GUIDE });
-		expect(out).toContain("<name>A&lt;&amp;&gt;&quot;&apos;B</name>");
-	});
-});
-
-// ── handler 行为：setupModelListInjector ────────────────
-
-describe("setupModelListInjector", () => {
-	type Handler = (event: unknown, ctx: unknown) => Promise<unknown>;
-
-	async function setupWithRegistry(models: ModelEntry[], fail = false): Promise<Handler> {
-		const mod = await import("../model-list-injector"); // handler 注册面（含 MODEL_LIST_GUIDE 注入）
-		const handlers: Record<string, Handler> = {};
-		const pi = {
-			on: (name: string, fn: Handler) => {
-				handlers[name] = fn;
-			},
-		} as unknown as ExtensionAPI;
-		mod.setupModelListInjector(pi);
-		const registry = fail
-			? { getAvailable: vi.fn(() => { throw new Error("registry boom"); }) }
-			: { getAvailable: vi.fn(() => models) };
-		const handler = handlers["before_agent_start"];
-		if (!handler) throw new Error("before_agent_start handler not registered");
-		return (event: unknown) => handler(event, { modelRegistry: registry });
-	}
-
-	it("注册 before_agent_start handler，注入 append 到 systemPrompt 尾部", async () => {
-		const handler = await setupWithRegistry([entry()]);
-		const result = (await handler({ systemPrompt: "BASE" }, {})) as {
-			systemPrompt: string;
-		};
-		expect(result.systemPrompt.startsWith("BASE")).toBe(true);
-		expect(result.systemPrompt).toContain("<available_provider_models>");
-		expect(result.systemPrompt).toContain("zai-coding-cn/glm-5.2");
-	});
-
-	it("空模型列表返回 undefined（不返回 systemPrompt，不干预链）", async () => {
-		const handler = await setupWithRegistry([]);
-		const result = await handler({ systemPrompt: "BASE" }, {});
-		expect(result).toBeUndefined();
-	});
-
-	it("registry 异常被吞掉（fail-safe，不阻断 agent turn）", async () => {
-		const handler = await setupWithRegistry([], true);
-		const result = await handler({ systemPrompt: "BASE" }, {});
-		expect(result).toBeUndefined();
-	});
-});