From 42bf8f9338b4bc38e2c88734d84befb9e911ecfa Mon Sep 17 00:00:00 2001 From: Lebing Xie Date: Fri, 26 Jun 2026 15:42:22 +0800 Subject: [PATCH] Add GLM-5.2 model support --- doc/summary.md | 16 +++---- src/foundation-models.test.ts | 38 +++++++++++++++++ src/foundation-models.ts | 79 +++++++++++++++++++++++++++++++++++ src/llm-providers.test.ts | 61 +++++++++++++++++++++++++-- src/llm-providers.ts | 19 ++++++++- src/runtime-policy.test.ts | 23 ++++++++++ 6 files changed, 224 insertions(+), 12 deletions(-) diff --git a/doc/summary.md b/doc/summary.md index 13e2641..eca3af7 100644 --- a/doc/summary.md +++ b/doc/summary.md @@ -90,8 +90,8 @@ src/ ├── recovery.ts # LLM 错误分类与恢复决策：backoff/compact/continue/fail ├── terminal.ts # 终端输入输出封装：共享 readline（REPL + 权限确认共用） ├── debug-e2e.ts # 端到端调试脚本（Skill+TODO+SubAgent 协作验证） -├── foundation-models.test.ts # Profile 注册表、匹配、fallback 测试（14 个测试用例） -├── runtime-policy.test.ts # Policy 解析、env 覆盖、非法值报错测试（19 个测试用例） +├── foundation-models.test.ts # Profile 注册表、匹配、fallback 测试（17 个测试用例） +├── runtime-policy.test.ts # Policy 解析、env 覆盖、非法值报错测试（20 个测试用例） ├── runtime-policy-store.test.ts # Override 合并、reset、snapshot 测试（15 个测试用例） ├── context-budget.test.ts # 预算分配公式、总和约束、override 裁剪测试（9 个测试用例） ├── llm-adapter.test.ts # Adapter 请求构建、reasoning 回放、streaming 聚合测试（15 个测试用例） @@ -437,7 +437,7 @@ skills/ ### Foundation Model Profile 基座模型画像 (`foundation-models.ts`) - **能力驱动而非模型名驱动**：`agent.ts` 不出现 `kimi`/`deepseek` 等具体模型分支，业务层只看 `RuntimePolicy` 中的策略字段 -- **Profile Registry**：含 `generic-openai-compatible`、`kimi-k2.6`、`kimi-code`、`deepseek-v4`、`minimax-m2.7`、`minimax-m3`、`mimo-v2.5-pro`、`qwen3.7-max`、`glm-5.1` 等画像 +- **Profile Registry**：含 `generic-openai-compatible`、`kimi-k2.6`、`kimi-code`、`deepseek-v4`、`minimax-m2.7`、`minimax-m3`、`mimo-v2.5-pro`、`qwen3.7-max`、`glm-5.2`、`glm-5.1` 等画像 - **匹配优先级**：`LLM_MODEL_PROFILE` 显式指定 > exact model id > prefix > provider default > generic fallback - **硬协议字段 vs 优化提示分离**：maxTokensField、thinking requestShape、reasoning responseFields 等硬字段必须保守；context budget、compression mode 等优化提示允许合理默认 - **Profile 分级**：`verified` / `experimental` / `needs_review`，stale/high-risk profile 启动时产生 warning 但不阻断 @@ -508,7 +508,7 @@ skills/ ### LLM Provider Profile 抽象层 (`llm-providers.ts`) -- **集中 profile 表**：声明 4 个 provider（`openai_compatible`、`minimax_cn`、`kimi_platform_cn`、`kimi_code_cn`）的默认 endpoint、默认模型、key 环境变量和能力标记 +- **集中 profile 表**：声明 5 个 provider（`openai_compatible`、`minimax_cn`、`kimi_platform_cn`、`kimi_code_cn`、`zhipuai_cn`）的默认 endpoint、默认模型、key 环境变量和能力标记 - **解析优先级**：`LLM_API_KEY` / `LLM_BASE_URL` / `LLM_MODEL` 优先于 provider 默认值，兼容现有使用方式 - **启动时解析**：`resolveLLMProviderConfig()` 只读 env，不做网络请求，返回 `ResolvedLLMConfig` - **厂商差异不泄漏到 Agent 循环**：Agent、SubAgent、Async Run 只依赖 `LLMClient.chat()` @@ -725,6 +725,8 @@ skills/ | `KIMI_CODE_API_KEY` | Kimi Code CN 专用 key | `sk-kimi-...` | | `MOONSHOT_API_KEY` | Kimi Platform CN 专用 key | `sk-moonshot-...` | | `MINIMAX_CN_API_KEY` | MiniMax CN 专用 key | `sk-minimax-...` | +| `ZHIPUAI_API_KEY` | ZhipuAI CN 专用 key | `sk-zhipu-...` | +| `BIGMODEL_API_KEY` | ZhipuAI CN 备用 key | `sk-bigmodel-...` | | `LOG_LEVEL` | 日志级别 | `info` | | `COMPRESS_TOOL_OUTPUT` | 即时压缩 token 阈值 | `2000` | | `COMPRESS_DECAY_THRESHOLD` | 衰减压缩轮次阈值 | `3` | @@ -789,11 +791,11 @@ skills/ | `src/session-events.test.ts` | 5 | drain 清空、peek 不清空、顺序保持 | | `src/transcript.test.ts` | 6 | 消息分类、事件 sequence、historySequence、timing 元信息、搜索 | | `src/cache-debug.test.ts` | 7 | inspect 变化检测、system prompt 不变性、formatCacheDebugLog | -| `src/llm-providers.test.ts` | 26 | provider 解析、默认值、覆盖优先级、错误提示、能力标记 | +| `src/llm-providers.test.ts` | 31 | provider 解析、默认值、覆盖优先级、错误提示、能力标记 | | `src/config.test.ts` | 5 | loadConfig 解析 provider 字段、compression/logLevel 默认值、错误信息不泄漏 key | | `src/llm.test.ts` | 10 | non-streaming 路径、streaming content/tool_calls 聚合、llmLogger 调用 | -| `src/foundation-models.test.ts` | 14 | Profile 注册表、exact/prefix/fallback 匹配、provider 兼容校验、显式 profile、stale warning | -| `src/runtime-policy.test.ts` | 19 | Policy 默认值、env 覆盖、非法覆盖报错、协议 fallback、compression 派生 | +| `src/foundation-models.test.ts` | 17 | Profile 注册表、exact/prefix/fallback 匹配、provider 兼容校验、显式 profile、stale warning | +| `src/runtime-policy.test.ts` | 20 | Policy 默认值、env 覆盖、非法覆盖报错、协议 fallback、compression 派生 | | `src/context-budget.test.ts` | 9 | 三种模式预算分配、总和约束、override 处理、裁剪优先级、极小预算边界 | | `src/runtime-policy-store.test.ts` | 15 | Override 合并、reset、snapshot、非法更新报错 | | `src/llm-adapter.test.ts` | 15 | 请求构建、reasoning 占位、streaming 聚合、max token 字段、usage 解析 | diff --git a/src/foundation-models.test.ts b/src/foundation-models.test.ts index ec53837..8486082 100644 --- a/src/foundation-models.test.ts +++ b/src/foundation-models.test.ts @@ -74,6 +74,34 @@ describe("resolveFoundationModelProfile", () => { expect(result.id).toBe("deepseek-v4"); }); + it("matches exact model id for glm-5.2", () => { + const result = resolveFoundationModelProfile({ + provider: "zhipuai_cn", + model: "glm-5.2", + }); + expect(result.id).toBe("glm-5.2"); + expect(result.provider).toBe("zhipuai_cn"); + expect(result.limits.contextWindowTokens).toBe(1000000); + expect(result.limits.effectiveContextBudgetTokens).toBe(750000); + expect(result.thinking.defaultMode).toBe("adaptive"); + expect(result.reasoning.responseFields).toEqual(["reasoning_content"]); + expect(result.cache.supported).toBe(false); + }); + + it("matches GLM-5.2 aliases and prefixes conservatively", () => { + const exactAlias = resolveFoundationModelProfile({ + provider: "zhipuai_cn", + model: "GLM5.2", + }); + const prefixAlias = resolveFoundationModelProfile({ + provider: "zhipuai_cn", + model: "glm-5.2-long-context", + }); + + expect(exactAlias.id).toBe("glm-5.2"); + expect(prefixAlias.id).toBe("glm-5.2"); + }); + it("uses explicit profile id when provided", () => { const result = resolveFoundationModelProfile({ provider: "kimi_platform_cn", @@ -83,6 +111,15 @@ describe("resolveFoundationModelProfile", () => { expect(result.id).toBe("kimi-k2.6"); }); + it("returns zhipuai default profile before generic for unknown zhipu model", () => { + const result = resolveFoundationModelProfile({ + provider: "zhipuai_cn", + model: "glm-custom-alias", + }); + expect(result.id).toBe("glm-5.2"); + expect(result.provider).toBe("zhipuai_cn"); + }); + it("throws for unknown explicit profile id", () => { expect(() => resolveFoundationModelProfile({ @@ -147,6 +184,7 @@ describe("getRegisteredModelProfileIds", () => { expect(ids).toContain("minimax-m3"); expect(ids).toContain("mimo-v2.5-pro"); expect(ids).toContain("qwen3.7-max"); + expect(ids).toContain("glm-5.2"); expect(ids).toContain("glm-5.1"); }); }); diff --git a/src/foundation-models.ts b/src/foundation-models.ts index ff76a6c..6ea9af2 100644 --- a/src/foundation-models.ts +++ b/src/foundation-models.ts @@ -781,6 +781,85 @@ const modelProfiles: FoundationModelProfile[] = [ }, }, + // ------------------------------------------------------------------------- + // GLM-5.2 + // ------------------------------------------------------------------------- + { + id: "glm-5.2", + displayName: "GLM-5.2", + provider: "zhipuai_cn", + match: { + exactModelIds: ["glm-5.2", "GLM-5.2", "glm5.2", "GLM5.2"], + modelIdPrefixes: ["glm-5.2", "GLM-5.2"], + }, + protocol: { + preferred: "openai-chat-completions", + fallbacks: [], + implemented: ["openai-chat-completions"], + }, + limits: { + contextWindowTokens: 1000000, + effectiveContextBudgetTokens: 750000, + longContextThresholdTokens: 512000, + maxOutputTokens: 65536, + maxTokensField: "max_tokens", + }, + thinking: { + supported: true, + defaultMode: "adaptive", + efforts: ["default"], + enableForAgenticTasks: true, + disableForSimpleChat: true, + requestShape: "extra_body_thinking", + }, + reasoning: { + returned: true, + mustReplayWithToolCalls: false, + preserveRawAssistantMessage: true, + responseFields: ["reasoning_content"], + streamingDeltaFields: ["reasoning_content"], + }, + tools: { + supported: true, + supportsToolChoiceRequired: false, + allowedToolChoiceModes: ["auto", "none"], + streamingArguments: true, + multimodalToolResults: false, + }, + cache: { + supported: false, + automatic: false, + exposesUsage: false, + usageFields: {}, + }, + modalities: { + text: true, + image: false, + video: false, + audio: false, + }, + optimizationHints: { + bestFor: ["coding", "long_horizon_agent", "large_context"], + defaultCompressionMode: "long_context", + prefersStreaming: true, + goodForSubagents: false, + }, + knownQuirks: [ + "官方资料确认 1M context 与 reasoning_content，但本仓尚未做 live smoke test", + "max output 和 cache usage 字段未在当前资料中核实，先使用保守上限并关闭 cache telemetry", + "thinking effort 的具体枚举未核实，暂不暴露 low/medium/high 等非 default 覆盖", + ], + documentation: { + sourceUrls: [ + "https://github.com/zai-org/glm-5/blob/main/README.md", + "https://github.com/metaglm/zhipuai-sdk-python-v4", + ], + verifiedAt: "2026-06-26", + updateRisk: "high", + status: "experimental", + }, + }, + // ------------------------------------------------------------------------- // GLM-5.1 // ------------------------------------------------------------------------- diff --git a/src/llm-providers.test.ts b/src/llm-providers.test.ts index 96c3474..99ec443 100644 --- a/src/llm-providers.test.ts +++ b/src/llm-providers.test.ts @@ -180,6 +180,49 @@ describe("minimax_cn provider", () => { }); }); +// ============================================================ +// zhipuai_cn 解析 +// ============================================================ + +describe("zhipuai_cn provider", () => { + it("只设置 ZHIPUAI_API_KEY 时可解析成功", () => { + const config = resolveLLMProviderConfig({ + LLM_PROVIDER: "zhipuai_cn", + ZHIPUAI_API_KEY: "sk-zhipu-test", + }); + expect(config.provider).toBe("zhipuai_cn"); + expect(config.apiKey).toBe("sk-zhipu-test"); + expect(config.baseURL).toBe("https://open.bigmodel.cn/api/paas/v4/"); + expect(config.model).toBe("glm-5.2"); + }); + + it("BIGMODEL_API_KEY 也可作为 fallback", () => { + const config = resolveLLMProviderConfig({ + LLM_PROVIDER: "zhipuai_cn", + BIGMODEL_API_KEY: "sk-bigmodel-fallback", + }); + expect(config.apiKey).toBe("sk-bigmodel-fallback"); + }); + + it("LLM_MODEL 优先于 GLM-5.2 默认模型", () => { + const config = resolveLLMProviderConfig({ + LLM_PROVIDER: "zhipuai_cn", + ZHIPUAI_API_KEY: "sk-zhipu-test", + LLM_MODEL: "glm-5.2-proxy", + }); + expect(config.model).toBe("glm-5.2-proxy"); + }); + + it("声明 supportsThinking 和 prefersStreaming 能力", () => { + const config = resolveLLMProviderConfig({ + LLM_PROVIDER: "zhipuai_cn", + ZHIPUAI_API_KEY: "sk-zhipu-test", + }); + expect(config.capabilities.supportsThinking).toBe(true); + expect(config.capabilities.prefersStreaming).toBe(true); + }); +}); + // ============================================================ // 启发式推断 // ============================================================ @@ -213,6 +256,16 @@ describe("baseURL 启发式推断", () => { expect(config.provider).toBe("kimi_platform_cn"); }); + it("未设置 LLM_PROVIDER 时，从 LLM_BASE_URL 推断为 zhipuai_cn", () => { + const config = resolveLLMProviderConfig({ + LLM_API_KEY: "sk-test", + LLM_BASE_URL: "https://open.bigmodel.cn/api/paas/v4/", + LLM_MODEL: "glm-5.2", + }); + expect(config.provider).toBe("zhipuai_cn"); + expect(config.capabilities.supportsThinking).toBe(true); + }); + it("不匹配的 baseURL 回退到 openai_compatible", () => { const config = resolveLLMProviderConfig({ LLM_API_KEY: "sk-test", @@ -261,15 +314,15 @@ describe("错误提示", () => { it("apiKey 缺失错误提示包含 provider id 和候选 key env", () => { expect(() => resolveLLMProviderConfig({ - LLM_PROVIDER: "kimi_code_cn", + LLM_PROVIDER: "zhipuai_cn", }), - ).toThrow('provider "kimi_code_cn"'); + ).toThrow('provider "zhipuai_cn"'); expect(() => resolveLLMProviderConfig({ - LLM_PROVIDER: "kimi_code_cn", + LLM_PROVIDER: "zhipuai_cn", }), - ).toThrow("LLM_API_KEY, KIMI_CODE_API_KEY"); + ).toThrow("LLM_API_KEY, ZHIPUAI_API_KEY, BIGMODEL_API_KEY"); }); it("错误信息不泄漏已有 key 值", () => { diff --git a/src/llm-providers.ts b/src/llm-providers.ts index 712890a..f9c2c89 100644 --- a/src/llm-providers.ts +++ b/src/llm-providers.ts @@ -23,7 +23,8 @@ export type LLMProviderId = | "openai_compatible" | "minimax_cn" | "kimi_platform_cn" - | "kimi_code_cn"; + | "kimi_code_cn" + | "zhipuai_cn"; /** * Provider 能力标记 @@ -136,6 +137,22 @@ const providerProfiles: Record = { supportsThinking: false, }, }, + zhipuai_cn: { + id: "zhipuai_cn", + displayName: "ZhipuAI CN", + protocol: "openai-chat-completions", + // 智谱官方 SDK 文档使用 open.bigmodel.cn 的 v4 基础路径。 + // 这里作为可运行默认值；企业代理或自建网关仍通过 LLM_BASE_URL 覆盖。 + defaultBaseURL: "https://open.bigmodel.cn/api/paas/v4/", + defaultModel: "glm-5.2", + apiKeyEnvNames: ["ZHIPUAI_API_KEY", "BIGMODEL_API_KEY"], + capabilities: { + supportsTools: true, + supportsToolChoiceRequired: false, + prefersStreaming: true, + supportsThinking: true, + }, + }, }; /** diff --git a/src/runtime-policy.test.ts b/src/runtime-policy.test.ts index baa6a40..8482074 100644 --- a/src/runtime-policy.test.ts +++ b/src/runtime-policy.test.ts @@ -73,6 +73,29 @@ describe("resolveRuntimePolicy defaults", () => { expect(policy.context.compressionMode).toBe("balanced"); }); + it("derives correct defaults for glm-5.2", () => { + const profile = resolveFoundationModelProfile({ + provider: "zhipuai_cn", + model: "glm-5.2", + }); + const policy = resolveRuntimePolicy(profile, "glm-5.2"); + + expect(policy.modelProfileId).toBe("glm-5.2"); + expect(policy.context.contextWindowTokens).toBe(1000000); + expect(policy.context.effectiveBudgetTokens).toBe(750000); + expect(policy.context.longContextThresholdTokens).toBe(512000); + expect(policy.context.compressionMode).toBe("long_context"); + expect(policy.request.prefersStreaming).toBe(true); + expect(policy.request.thinkingMode).toBe("adaptive"); + expect(policy.request.extraBody).toEqual({ + thinking: { type: "auto" }, + }); + expect(policy.reasoning.responseFields).toEqual(["reasoning_content"]); + expect(policy.tools.streamingArguments).toBe(true); + expect(policy.cache.supported).toBe(false); + expect(policy.telemetry.recordCacheTokens).toBe(false); + }); + it("derives long_context compression with relaxed thresholds", () => { const profile = getProfile("deepseek-v4"); const policy = resolveRuntimePolicy(profile, "deepseek-v4");