fix(codex): point Zhipu GLM presets at the official Responses endpoint (#6957)

Zhipu documents three separate base URLs per site (Anthropic /api/anthropic,
Chat /api/coding/paas/v4, Responses /api/v1) and warns that the wrong one
cannot use Coding Plan quota. Codex direct-connect speaks Responses, so the
presets must target /api/v1; the 400 in #6944 (`unknown variant custom`)
came from the Chat endpoint's strict legacy gateway.

- Switch the Zhipu GLM / Zhipu GLM en presets to apiFormat openai_responses
  (NativeResponses catalog: shell_command edits, no freeform apply_patch)
- Point base_url/endpointCandidates at /api/v1 and mirror the official
  models.json: glm-5.3 (1M context, low/high/max, default max) plus
  glm-5-turbo on the CN site; drop glm-5.2 and the `none` level
- Host fallback: rows saved with a stale openai_chat apiFormat but a native
  Responses base_url get the NativeResponses catalog without a re-save;
  Chat-endpoint paths (/paas/v4, /chat/completions) are excluded
- Match vendor hosts on DNS label boundaries via a shared helper, so `z.ai`
  no longer captures xyz.ai; add bigmodel.cn / z.ai and the glm prefix to
  the Codex web_search reject list

Existing Zhipu providers still point at the Chat endpoint; re-import the
preset to pick up /api/v1.

Fixes #6944

Co-authored-by: Jason <farion1231@gmail.com>

src-tauri/src/codex_config.rs

@@ -68,6 +68,13 @@ const CODEX_WEB_SEARCH_REJECT_HOSTS: &[&str] = &[
     "longcat.chat",   // Meituan LongCat (api.longcat.chat)
     "minimax.io",     // MiniMax global (api.minimax.io)
     "minimaxi.com",   // MiniMax CN (api.minimaxi.com)
+    // Zhipu GLM CN / global (open.bigmodel.cn, api.z.ai): the native Responses
+    // gateway's tool-type enum is `function | web_search_preview |
+    // code_interpreter | mcp` (verbatim from the #6944 400 body) — Codex's
+    // `web_search` hosted tool is not in it. Matched on host labels (see
+    // `codex_url_host_matches_any`), so `xyz.ai` never collides with `z.ai`.
+    "bigmodel.cn",
+    "z.ai",
 ];
 
 /// Brand prefixes of models whose native gateways reject `web_search`, matched
@@ -75,7 +82,41 @@ const CODEX_WEB_SEARCH_REJECT_HOSTS: &[&str] = &[
 /// `MiniMaxAI/MiniMax-M3` are caught. Exact brand names (not a fuzzy heuristic),
 /// so a supporting gateway is never wrongly matched.
 const CODEX_WEB_SEARCH_REJECT_MODEL_PREFIXES: &[&str] =
-    &["mimo", "longcat", "minimax", "qwen3-coder"];
+    &["mimo", "longcat", "minimax", "qwen3-coder", "glm"];
+
+/// Host component of a base URL (or a bare host), lowercased, without scheme,
+/// userinfo, port, path or query. Tolerates the loose forms users paste into
+/// the provider form (`example.com`, `https://user@Example.com:8443/v1`).
+pub(crate) fn codex_url_host(url_or_host: &str) -> String {
+    let trimmed = url_or_host.trim();
+    let rest = trimmed
+        .split_once("://")
+        .map_or(trimmed, |(_scheme, rest)| rest);
+    let authority = rest.split(['/', '?', '#']).next().unwrap_or("");
+    let host_port = authority.rsplit('@').next().unwrap_or(authority);
+    let host = if let Some(ipv6) = host_port.strip_prefix('[') {
+        ipv6.split(']').next().unwrap_or(ipv6)
+    } else {
+        host_port.split(':').next().unwrap_or(host_port)
+    };
+    host.trim_end_matches('.').to_ascii_lowercase()
+}
+
+/// Whether the URL's host IS one of `hosts` or a subdomain of it, matched on
+/// DNS label boundaries. Vendor host lists must go through this rather than a
+/// substring `contains`: a 4-char entry like `z.ai` would otherwise also match
+/// `api.xyz.ai` / `viz.ai` and silently push an unrelated provider onto a
+/// vendor-specific code path.
+pub(crate) fn codex_url_host_matches_any(url_or_host: &str, hosts: &[&str]) -> bool {
+    let host = codex_url_host(url_or_host);
+    if host.is_empty() {
+        return false;
+    }
+    hosts.iter().any(|candidate| {
+        let candidate = candidate.trim_start_matches('.').to_ascii_lowercase();
+        host == candidate || host.ends_with(&format!(".{candidate}"))
+    })
+}
 
 /// Top-level `model` id from a Codex `config.toml`.
 fn codex_top_level_model(config_text: &str) -> Option<String> {
@@ -92,11 +133,7 @@ fn codex_top_level_model(config_text: &str) -> Option<String> {
 /// the live `config.toml`, so it applies to existing providers without a re-save.
 fn codex_native_gateway_rejects_web_search(config_text: &str) -> bool {
     if let Some(base_url) = extract_codex_base_url(config_text) {
-        let base_url = base_url.to_ascii_lowercase();
-        if CODEX_WEB_SEARCH_REJECT_HOSTS
-            .iter()
-            .any(|host| base_url.contains(host))
-        {
+        if codex_url_host_matches_any(&base_url, CODEX_WEB_SEARCH_REJECT_HOSTS) {
             return true;
         }
     }
@@ -7250,6 +7287,8 @@ web_search = "disabled"
             ("LongCat-2.0", "https://api.longcat.chat/openai/v1"),
             ("MiniMax-M3", "https://api.minimax.io/v1"),
             ("MiniMax-M3", "https://api.minimaxi.com/v1"),
+            ("glm-5.3", "https://open.bigmodel.cn/api/v1"),
+            ("glm-5.3", "https://api.z.ai/api/v1"),
         ] {
             assert!(
                 codex_native_gateway_rejects_web_search(&cfg(model, host)),
@@ -7263,6 +7302,7 @@ web_search = "disabled"
             ("MiniMax-M3", "https://api.siliconflow.cn/v1"),
             ("MiniMaxAI/MiniMax-M3", "https://api.siliconflow.cn/v1"),
             ("mimo-v2.5-pro", "https://some-aggregator.example/v1"),
+            ("zai-org/glm-5.3", "https://some-aggregator.example/v1"),
             (
                 "qwen/qwen3-coder-plus",
                 "https://some-aggregator.example/v1",
@@ -7296,6 +7336,12 @@ web_search = "disabled"
                 "https://ark.cn-beijing.volces.com/api/v3",
             ),
             ("Pro/moonshotai/Kimi-K2.6", "https://api.siliconflow.cn/v1"),
+            // Host-label matching: `z.ai` / `bigmodel.cn` must not swallow
+            // unrelated domains that merely contain them as a substring.
+            ("gpt-5.5", "https://api.xyz.ai/v1"),
+            ("gpt-5.5", "https://viz.ai/v1"),
+            ("gpt-5.5", "https://notbigmodel.cn/v1"),
+            ("gpt-5.5", "https://z.ai.example.com/v1"),
         ] {
             assert!(
                 !codex_native_gateway_rejects_web_search(&cfg(model, host)),
@@ -7304,6 +7350,34 @@ web_search = "disabled"
         }
     }
 
+    #[test]
+    fn url_host_matcher_uses_label_boundaries() {
+        let hosts = &["z.ai", "bigmodel.cn"];
+        for url in [
+            "https://api.z.ai/api/v1",
+            "https://open.bigmodel.cn/api/v1",
+            "https://Open.BigModel.cn/api/coding/paas/v4",
+            "https://user:pw@api.z.ai:8443/api/v1?x=1#f",
+            "z.ai",
+            "api.z.ai.",
+        ] {
+            assert!(codex_url_host_matches_any(url, hosts), "{url}");
+        }
+        for url in [
+            "https://api.xyz.ai/v1",
+            "https://viz.ai/v1",
+            "https://z.ai.example.com/v1",
+            "https://notbigmodel.cn/v1",
+            "https://example.com/z.ai/v1",
+            "https://example.com/?next=https://api.z.ai",
+            "",
+        ] {
+            assert!(!codex_url_host_matches_any(url, hosts), "{url}");
+        }
+        assert_eq!(codex_url_host("https://[::1]:8080/v1"), "::1");
+        assert_eq!(codex_url_host("HTTP://Example.COM:80"), "example.com");
+    }
+
     #[test]
     fn resolve_catalog_path_returns_none_when_config_missing_field() {
         let base = PathBuf::from("/tmp/.codex");

src-tauri/src/proxy/providers/codex.rs

@@ -333,6 +333,47 @@ pub fn is_codex_official_provider(provider: &Provider) -> bool {
     is_fixed_official_id || provider.category.as_deref() == Some("official")
 }
 
+/// Vendors whose OFFICIAL Codex integration is a native `/responses` gateway that
+/// rejects Codex's freeform custom tools (`apply_patch` with `type: "custom"`,
+/// #6944). Same vendor set as `CODEX_WEB_SEARCH_REJECT_HOSTS` in `codex_config`
+/// (kept separate: that list also gates aggregators by model brand). Matched on
+/// host labels via `codex_url_host_matches_any`, never by substring.
+const CODEX_NATIVE_RESPONSES_HOSTS: &[&str] = &[
+    "bigmodel.cn",
+    "z.ai",
+    "xiaomimimo.com",
+    "minimaxi.com",
+    "minimax.io",
+    "longcat.chat",
+];
+
+/// Path markers of a listed vendor's OpenAI *Chat Completions* endpoint, which is
+/// NOT its Responses endpoint. Zhipu documents three separate base URLs per site
+/// (Anthropic `/api/anthropic`, Chat `/api/coding/paas/v4` + pay-as-you-go
+/// `/api/paas/v4`, Responses `/api/v1`) and warns that the wrong one cannot use
+/// Coding Plan quota. A stored provider still pointing at a Chat path is a
+/// pre-2026-09 Chat-route record: it keeps its `ProxyChat` catalog (the proxy
+/// route converts it correctly; direct connect fails loudly with the #6944 400
+/// until the preset is re-imported) instead of being silently steered onto the
+/// wrong endpoint with a native catalog.
+const CODEX_NATIVE_RESPONSES_CHAT_PATH_MARKERS: &[&str] = &["/paas/v4"];
+
+/// Whether `base_url` points at a listed vendor's native Responses gateway, so a
+/// provider whose stored `apiFormat` predates the preset's switch to
+/// `openai_responses` still gets the `NativeResponses` catalog without a re-save.
+pub fn is_codex_native_responses_url(base_url: &str) -> bool {
+    if !crate::codex_config::codex_url_host_matches_any(base_url, CODEX_NATIVE_RESPONSES_HOSTS) {
+        return false;
+    }
+    if is_chat_completions_url(base_url) {
+        return false;
+    }
+    let lower = base_url.to_ascii_lowercase();
+    !CODEX_NATIVE_RESPONSES_CHAT_PATH_MARKERS
+        .iter()
+        .any(|marker| lower.contains(marker))
+}
+
 /// Resolve the model-catalog tool profile for a Codex provider using the SAME
 /// Anthropic detection as the proxy router ([`codex_provider_uses_anthropic`]), so the
 /// generated catalog never disagrees with the routed transform. A provider whose
@@ -355,9 +396,49 @@ pub fn resolve_codex_catalog_tool_profile(
     if codex_provider_uses_anthropic(provider) {
         return CodexCatalogToolProfile::Anthropic;
     }
-    CodexCatalogToolProfile::from_api_format(
-        provider.meta.as_ref().and_then(|m| m.api_format.as_deref()),
-    )
+
+    // Defensive fallback for providers saved in SQLite before their preset
+    // switched to `openai_responses` (the #6944 reporter reinstalled to no
+    // effect precisely because the stale `apiFormat` lives in the DB row): a
+    // base_url on a listed vendor's native Responses gateway forces the
+    // NativeResponses catalog. Chat-endpoint paths are deliberately excluded —
+    // see `CODEX_NATIVE_RESPONSES_CHAT_PATH_MARKERS`.
+    if let Some(base_url) = provider
+        .settings_config
+        .get("config")
+        .and_then(|v| v.as_str())
+        .and_then(extract_codex_base_url_from_toml)
+        .or_else(|| {
+            provider
+                .settings_config
+                .get("base_url")
+                .or_else(|| provider.settings_config.get("baseURL"))
+                .and_then(|v| v.as_str())
+                .map(ToString::to_string)
+        })
+    {
+        if is_codex_native_responses_url(&base_url) {
+            return CodexCatalogToolProfile::NativeResponses;
+        }
+    }
+
+    let api_format = provider
+        .meta
+        .as_ref()
+        .and_then(|m| m.api_format.as_deref())
+        .or_else(|| {
+            provider
+                .settings_config
+                .get("api_format")
+                .and_then(|v| v.as_str())
+        })
+        .or_else(|| {
+            provider
+                .settings_config
+                .get("apiFormat")
+                .and_then(|v| v.as_str())
+        });
+    CodexCatalogToolProfile::from_api_format(api_format)
 }
 
 /// Extract the real upstream model configured for a Codex provider.
@@ -1457,6 +1538,106 @@ wire_api = "anthropic"
             resolve_codex_catalog_tool_profile(&chat),
             CodexCatalogToolProfile::ProxyChat
         );
+
+        // Host fallback (#6944): a DB row saved while the Zhipu preset was still
+        // `openai_chat` but whose base_url is the vendor's native Responses
+        // gateway (`/api/v1`) resolves to NativeResponses without a re-save —
+        // whether the URL lives in the TOML or in settings `baseURL`.
+        let legacy_chat_toml = |base_url: &str| {
+            create_provider(json!({
+                "apiFormat": "openai_chat",
+                "config": format!(
+                    "model_provider = \"custom\"\n\n[model_providers.custom]\nname = \"zhipu_glm\"\nbase_url = \"{base_url}\"\nwire_api = \"responses\"\n"
+                )
+            }))
+        };
+        for base_url in [
+            "https://open.bigmodel.cn/api/v1",
+            "https://api.z.ai/api/v1",
+            "https://Open.BigModel.cn/api/v1/",
+        ] {
+            assert_eq!(
+                resolve_codex_catalog_tool_profile(&legacy_chat_toml(base_url)),
+                CodexCatalogToolProfile::NativeResponses,
+                "{base_url}"
+            );
+        }
+        let legacy_chat_settings = create_provider(json!({
+            "apiFormat": "openai_chat",
+            "baseURL": "https://api.z.ai/api/v1"
+        }));
+        assert_eq!(
+            resolve_codex_catalog_tool_profile(&legacy_chat_settings),
+            CodexCatalogToolProfile::NativeResponses
+        );
+
+        // The vendor's Chat Completions endpoints are NOT its Responses gateway
+        // (Zhipu: `/api/coding/paas/v4` Coding Plan, `/api/paas/v4` pay-as-you-go)
+        // — a stale row there keeps ProxyChat so it is never steered onto the
+        // endpoint Zhipu documents as unable to use Coding Plan quota.
+        for base_url in [
+            "https://open.bigmodel.cn/api/coding/paas/v4",
+            "https://api.z.ai/api/coding/paas/v4",
+            "https://open.bigmodel.cn/api/paas/v4",
+            "https://open.bigmodel.cn/api/paas/v4/chat/completions",
+        ] {
+            assert_eq!(
+                resolve_codex_catalog_tool_profile(&legacy_chat_toml(base_url)),
+                CodexCatalogToolProfile::ProxyChat,
+                "{base_url}"
+            );
+        }
+
+        // Host matching is on DNS labels: `xyz.ai` must not be captured by the
+        // 4-char `z.ai` entry (it would lose apply_patch and web_search).
+        for base_url in [
+            "https://api.xyz.ai/v1",
+            "https://viz.ai/v1",
+            "https://z.ai.example.com/v1",
+        ] {
+            assert_eq!(
+                resolve_codex_catalog_tool_profile(&legacy_chat_toml(base_url)),
+                CodexCatalogToolProfile::ProxyChat,
+                "{base_url}"
+            );
+        }
+
+        // An explicit `openai_responses` on an unlisted host is untouched by the
+        // fallback (it only ever widens toward NativeResponses for listed hosts).
+        let explicit_native = create_provider(json!({
+            "apiFormat": "openai_responses",
+            "baseURL": "https://api.xyz.ai/v1"
+        }));
+        assert_eq!(
+            resolve_codex_catalog_tool_profile(&explicit_native),
+            CodexCatalogToolProfile::NativeResponses
+        );
+    }
+
+    #[test]
+    fn native_responses_url_guard_matches_gateway_not_chat_paths() {
+        for url in [
+            "https://open.bigmodel.cn/api/v1",
+            "https://api.z.ai/api/v1",
+            "https://api.xiaomimimo.com/v1",
+            "https://token-plan-cn.xiaomimimo.com/v1",
+            "https://api.minimaxi.com/v1",
+            "https://api.minimax.io/v1",
+            "https://api.longcat.chat/openai/v1",
+        ] {
+            assert!(is_codex_native_responses_url(url), "{url}");
+        }
+        for url in [
+            "https://open.bigmodel.cn/api/coding/paas/v4",
+            "https://api.z.ai/api/coding/paas/v4",
+            "https://open.bigmodel.cn/api/paas/v4",
+            "https://api.minimaxi.com/v1/chat/completions",
+            "https://api.xyz.ai/v1",
+            "https://api.deepseek.com",
+            "",
+        ] {
+            assert!(!is_codex_native_responses_url(url), "{url}");
+        }
     }
 
     #[test]

src/config/codexProviderPresets.ts

@@ -1172,32 +1172,45 @@ requires_openai_auth = true`,
     websiteUrl: "https://open.bigmodel.cn",
     apiKeyUrl: "https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII",
     auth: generateThirdPartyAuth(""),
+    // 智谱三端点分立（docs.bigmodel.cn/cn/coding-plan/tool/others）：Anthropic
+    // /api/anthropic、OpenAI Chat /api/coding/paas/v4、OpenAI Responses /api/v1，
+    // 并明示「错误配置端点将导致无法使用 GLM Coding Plan 套餐额度」。Codex 直连
+    // 发的是 Responses wire，base_url 必须是 /api/v1；Chat 端点上的 /responses
+    // 是严格旧网关（拒 type=custom 工具 → #6944 的 400）
     config: generateThirdPartyConfig(
       "zhipu_glm",
-      "https://open.bigmodel.cn/api/coding/paas/v4",
-      "glm-5.2",
+      "https://open.bigmodel.cn/api/v1",
+      "glm-5.3",
     ),
-    endpointCandidates: ["https://open.bigmodel.cn/api/coding/paas/v4"],
-    apiFormat: "openai_chat",
+    endpointCandidates: ["https://open.bigmodel.cn/api/v1"],
+    // 官方 Codex 接入页（docs.bigmodel.cn/cn/coding-plan/tool/codex，2026-09-04
+    // 核对）：wire_api=responses 对自家 /api/v1，与 MiMo/MiniMax 同为原生直连
+    // → NativeResponses profile（shell_command 编辑、不发 freeform apply_patch；
+    // 官方目录虽声明 freeform，无真机验证前按保守口径，不引入 400 风险）
+    apiFormat: "openai_responses",
+    // 档位/上下文/模态照抄官方 models.json：glm-5.3 low/high/max 默认 max；
+    // glm-5-turbo 官方档位为空、默认 max——cc-switch 表达不了空档位（回落会得到
+    // 模板 none/high，none 在原生直连下没有转换层兜底、会原样发给严格网关），
+    // 按官方默认收成单档 max。两模型 input_modalities=["text"]、并行工具调用 true
     modelCatalog: modelCatalog([
-      // Chat 路由 supportsEffort:false：档位值不进 wire，none=注入
-      // thinking:{type:"disabled"} 关思考，其余档一律等价于开思考。只暴露真实
-      // 两态；不填的话 gpt5_5 模板默认 low/medium/high/xhigh 全是假差异档，
-      // 且没有 none，用户在 Codex 里反而关不掉思考
       {
-        model: "glm-5.2",
-        displayName: "GLM-5.2",
-        contextWindow: 200000,
-        reasoningLevels: ["none", "high"],
+        model: "glm-5.3",
+        displayName: "GLM-5.3",
+        contextWindow: 1048576,
+        inputModalities: ["text"],
+        supportsParallelToolCalls: true,
+        reasoningLevels: ["low", "high", "max"],
+        defaultReasoningLevel: "max",
+      },
+      {
+        model: "glm-5-turbo",
+        displayName: "GLM-5-Turbo",
+        contextWindow: 204800,
+        inputModalities: ["text"],
+        supportsParallelToolCalls: true,
+        reasoningLevels: ["max"],
       },
     ]),
-    codexChatReasoning: {
-      supportsThinking: true,
-      supportsEffort: false,
-      thinkingParam: "thinking",
-      effortParam: "none",
-      outputFormat: "reasoning_content",
-    },
     category: "cn_official",
     icon: "zhipu",
     iconColor: "#0F62FE",
@@ -1207,32 +1220,26 @@ requires_openai_auth = true`,
     websiteUrl: "https://z.ai",
     apiKeyUrl: "https://z.ai/subscribe?ic=8JVLJQFSKB",
     auth: generateThirdPartyAuth(""),
+    // 国际站同上（docs.z.ai/devpack/tool/others + devpack/tool/codex，2026-09-04
+    // 核对）：Responses 端点 /api/v1，官方 models.json 仅列 glm-5.3
     config: generateThirdPartyConfig(
       "zhipu_glm_en",
-      "https://api.z.ai/api/coding/paas/v4",
-      "glm-5.2",
+      "https://api.z.ai/api/v1",
+      "glm-5.3",
     ),
-    endpointCandidates: ["https://api.z.ai/api/coding/paas/v4"],
-    apiFormat: "openai_chat",
+    endpointCandidates: ["https://api.z.ai/api/v1"],
+    apiFormat: "openai_responses",
     modelCatalog: modelCatalog([
-      // Chat 路由 supportsEffort:false：档位值不进 wire，none=注入
-      // thinking:{type:"disabled"} 关思考，其余档一律等价于开思考。只暴露真实
-      // 两态；不填的话 gpt5_5 模板默认 low/medium/high/xhigh 全是假差异档，
-      // 且没有 none，用户在 Codex 里反而关不掉思考
       {
-        model: "glm-5.2",
-        displayName: "GLM-5.2",
-        contextWindow: 200000,
-        reasoningLevels: ["none", "high"],
+        model: "glm-5.3",
+        displayName: "GLM-5.3",
+        contextWindow: 1048576,
+        inputModalities: ["text"],
+        supportsParallelToolCalls: true,
+        reasoningLevels: ["low", "high", "max"],
+        defaultReasoningLevel: "max",
       },
     ]),
-    codexChatReasoning: {
-      supportsThinking: true,
-      supportsEffort: false,
-      thinkingParam: "thinking",
-      effortParam: "none",
-      outputFormat: "reasoning_content",
-    },
     category: "cn_official",
     icon: "zhipu",
     iconColor: "#0F62FE",

tests/config/codexChatProviderPresets.test.ts

@@ -10,22 +10,8 @@ const expectedChatPresets = new Map<
   string,
   { baseUrl: string; contextWindows: Record<string, number> }
 >([
-  // 火山 Agent Plan / Coding Plan 与 BytePlus 国际站（coding/v3）均已切
+  // 火山 Agent Plan / Coding Plan 与 BytePlus 国际站（coding/v3）、智谱 GLM 均已切
   // 原生 Responses，见下方 native 清单
-  [
-    "Zhipu GLM",
-    {
-      baseUrl: "https://open.bigmodel.cn/api/coding/paas/v4",
-      contextWindows: { "glm-5.2": 200000 },
-    },
-  ],
-  [
-    "Zhipu GLM en",
-    {
-      baseUrl: "https://api.z.ai/api/coding/paas/v4",
-      contextWindows: { "glm-5.2": 200000 },
-    },
-  ],
   [
     "Baidu Qianfan Coding Plan",
     {
@@ -168,7 +154,7 @@ describe("Codex Chat provider presets", () => {
   it("uses native Responses API for migrated CN providers without local route mapping", () => {
     const nativeResponsesPresets = new Map<
       string,
-      { contextWindows: Record<string, number> }
+      { baseUrl?: string; contextWindows: Record<string, number> }
     >([
       // 官方 Codex 文档确认 Agent Plan /api/plan/v3 与 Coding Plan
       // /api/coding/v3 均支持 Responses API；BytePlus 国际站 coding/v3
@@ -218,6 +204,25 @@ describe("Codex Chat provider presets", () => {
           },
         },
       ],
+      // 智谱三端点分立（Anthropic /api/anthropic、Chat /api/coding/paas/v4、
+      // Responses /api/v1），官方明示错误端点无法使用 Coding Plan 套餐额度——
+      // 原生 Responses 预设必须锁在 /api/v1（#6944；docs.bigmodel.cn/cn/coding-plan/
+      // tool/codex 与 docs.z.ai/devpack/tool/codex 自带 models.json，2026-09-04 核对：
+      // 国内站 glm-5.3 + glm-5-turbo，国际站仅 glm-5.3）
+      [
+        "Zhipu GLM",
+        {
+          baseUrl: "https://open.bigmodel.cn/api/v1",
+          contextWindows: { "glm-5.3": 1048576, "glm-5-turbo": 204800 },
+        },
+      ],
+      [
+        "Zhipu GLM en",
+        {
+          baseUrl: "https://api.z.ai/api/v1",
+          contextWindows: { "glm-5.3": 1048576 },
+        },
+      ],
     ]);
 
     for (const [name, expected] of nativeResponsesPresets) {
@@ -225,6 +230,15 @@ describe("Codex Chat provider presets", () => {
 
       expect(preset, `${name} preset`).toBeDefined();
       expect(preset?.apiFormat).toBe("openai_responses");
+      if (expected.baseUrl) {
+        // 直连预设的 base_url 必须是厂商的 Responses 端点本身（不是同站的
+        // Chat 端点）；endpointCandidates 与主地址同路径档
+        expect(extractCodexBaseUrl(preset?.config)).toBe(expected.baseUrl);
+        expect(preset?.endpointCandidates).toContain(expected.baseUrl);
+        expect(extractCodexModelName(preset?.config)).toBe(
+          preset?.modelCatalog?.[0]?.model,
+        );
+      }
       // 原生 Responses 预设现在带 modelCatalog：cc-switch 直连时据此生成
       // ~/.codex 的 model-catalogs.json（shell_command 编辑、不发 freeform
       // apply_patch）。带 catalog 不再强制开“本地路由映射”——前端已按

tests/config/codexReasoningLevelPresets.test.ts

@@ -93,10 +93,15 @@ describe("Codex preset pre-filled reasoning levels", () => {
     ["Xiaomi MiMo", "mimo-v2.5", ["none", "high"]],
     ["Xiaomi MiMo Token Plan (China)", "mimo-v2.5-pro", ["none", "high"]],
     ["Xiaomi MiMo Token Plan (China)", "mimo-v2.5", ["none", "high"]],
-    // GLM 走 Chat 路由（supportsEffort:false）：none=真实关思考开关，其余档
-    // 等价开思考；只暴露两态，顺带补上模板四档里缺失的 none（关思考入口）
-    ["Zhipu GLM", "glm-5.2", ["none", "high"]],
-    ["Zhipu GLM en", "glm-5.2", ["none", "high"]],
+    // 智谱官方 Codex 接入页自带 models.json（docs.bigmodel.cn/cn/coding-plan/tool/
+    // codex、docs.z.ai/devpack/tool/codex，2026-09-04 核对）：glm-5.3 档位
+    // low/high/max、默认 max（≠ 后端回落的模板默认 high，故显式声明）；
+    // glm-5-turbo 官方档位为空、默认 max——cc-switch 表达不了空档位（会回落
+    // 到模板 none/high，而 none 在原生直连下没有转换层兜底、会原样发给严格
+    // 网关），按官方默认收成单档 max
+    ["Zhipu GLM", "glm-5.3", ["low", "high", "max"], "max"],
+    ["Zhipu GLM", "glm-5-turbo", ["max"]],
+    ["Zhipu GLM en", "glm-5.3", ["low", "high", "max"], "max"],
     // SiliconFlow .com 的 M3：平台级 enable_thinking 布尔开关（后端按平台
     // 推断兜底），M3 官方可关思考 → 两态
     ["SiliconFlow en", "MiniMaxAI/MiniMax-M3", ["none", "high"]],