@@ -897,6 +897,8 @@ The existing region detector still probes DuckDuckGo once at startup through `Pr
 `saved_models` is a named registry; `default_model_name` points into it. Each entry needs `name` (must NOT start with `(` — reserved for synthesized `(session:<id>)`), `provider`, `model`. OpenAI entries also need `base_url` + `api_key`. On load, array order is oldest-to-newest: for duplicate names, every occurrence except the last is atomically persisted as `<name>-<6-lowercase-hex-hash>`, while the last keeps the original name. Duplicate-repair write failures remain fatal. A non-empty dangling `default_model_name` is instead replaced with the first saved-model name before validation and atomically written back; a write failure is logged but startup continues with the in-memory fallback and retries on the next load. If the main logger is not initialized yet, this repair silently opens `~/.acecode/logs/config-<date>.log`; an active logger is never redirected. This repair does not rewrite an empty default and emits no UI message. Successful full loads and saves refresh a restricted `config-backups/last-good/config.json` snapshot. Later JSON parse/type/range/semantic failures archive the exact invalid bytes under `config-backups/invalid`, atomically restore and validate that snapshot, and persist a one-time Web/Desktop warning; recovery logs contain paths and categories only. Without a valid prior snapshot, startup keeps the invalid active file and follows the existing fatal boundary.
 `acecode configure` upserts one named `saved_models` entry from the selected provider/model and sets it as `default_model_name`; normal startup no longer derives a selectable model from top-level provider fields.
 
+Each saved model may carry an optional `sampling` object (`temperature` [0,2], `top_p` (0,1], `top_k` ≥ 1, `min_p` [0,1]). Unset keys are never sent, so a profile without sampling produces byte-identical requests. Support is per provider/protocol and validated in `validate_model_sampling_support` (`saved_models.cpp`): Chat Completions sends all four (`top_k`/`min_p` are vLLM/SGLang extensions the official OpenAI API rejects), Responses only `temperature`/`top_p`, Anthropic rejects `min_p` and caps `temperature` at 1, managed Copilot/Grok reject sampling entirely — a value that would be silently dropped is a validation error, not a no-op. Edits replace the whole object (`null`/`{}` clears; a provider switch clears it). Sampling is part of the provider construction fingerprint, so running sessions rebuild on the next turn after an edit. Settings > Models > 高级设置 shows only the fields the selected provider/protocol supports (`modelSamplingFieldSupport` in `web/src/lib/modelSettings.js`). Added for the intranet wizard-ai gateway, whose GLM model ships recommended values (top_p 0.95 / top_k 20 / temperature 1 / min_p 0) that ACECode previously never sent.
+
 ### Model profile resolution
 
 At startup and on every `--resume` / `/resume`, `src/adapters/provider/model_resolver.cpp` resolves the effective `ModelEntry` in three layers. CLI `-r` is not a `--resume` short alias; it starts the TUI normally, runs the startup hook path, then opens the same `/resume` picker immediately.
 
@@ -3019,7 +3019,7 @@ cannot be disabled and return `409 {"error":"HOOK_MANAGED"}`.
 原值和 `has_api_key` 布尔值，供编辑表单默认遮罩回填；调用方必须把整个响应视为
 敏感数据，日志与错误消息仍不得包含密钥。高级字段有值时会原样返回，包括
 `endpoint_mode`、`api_protocol`、`max_output_tokens`、`capabilities_source`、`reasoning`、
-`request_headers`、`context_window` 与 `stream_timeout_ms`。
+`sampling`、`request_headers`、`context_window` 与 `stream_timeout_ms`。
 
 ### `POST /api/models`
 
@@ -3048,10 +3048,22 @@ cannot be disabled and return `409 {"error":"HOOK_MANAGED"}`.
     "default_effort": "medium",
     "effort": "high",
     "supports_max_tokens": false
-  }
+  },
+  "sampling": {"temperature": 1, "top_p": 0.95, "top_k": 20, "min_p": 0}
 }
 ```
 
+`sampling` 是可选的采样参数对象，四个键都可省略或为 `null`（视为未设置）：
+`temperature` ∈ [0, 2]、`top_p` ∈ (0, 1]、`top_k` 为正整数、`min_p` ∈ [0, 1]。
+未设置的键不会出现在发往模型服务的请求体里，沿用服务端默认值；整个对象省略或
+没有任何已设置的键时，请求体与不支持该功能的版本完全一致。支持范围按 Provider /
+API 协议区分：OpenAI Chat Completions 四个都发送（`top_k` / `min_p` 是 vLLM、
+SGLang 等推理服务的扩展字段，官方 OpenAI 接口会拒绝）；Responses 只支持
+`temperature` / `top_p`；Anthropic 支持 `temperature`（上限 1）、`top_p`、`top_k`，
+不支持 `min_p`；Copilot / Grok 托管 Provider 不接受采样参数。越界值或当前 Provider /
+协议不支持的键返回 `INVALID_SAMPLING`，托管 Provider 返回 `UNSUPPORTED_MODEL_OPTION`。
+响应只返回已设置的键。
+
 `api_protocol` 仅适用于 `provider:"openai"`，接受 `chat_completions` 或
 `responses`；省略时，目录身份 `models_dev_provider_id:"acemodel"` 使用 Responses，
 其他 OpenAI 配置继续使用 Chat Completions。模型设置的 **API 协议** 下拉框可
@@ -3075,9 +3087,10 @@ cannot be disabled and return `409 {"error":"HOOK_MANAGED"}`.
 路径仍区分大小写。
 
 省略 `base_url`、`context_window`、`stream_timeout_ms`、`capabilities`、
-`endpoint_mode`、`api_protocol`、`max_output_tokens`、`capabilities_source`、`reasoning` 或
-`request_headers` 会保留原值。显式空 `request_headers` 会清空请求头；对应高级
-字段传 `null` 时按各字段合同清除。外部登录器留下的 legacy `readonly:true` 只是
+`endpoint_mode`、`api_protocol`、`max_output_tokens`、`capabilities_source`、`reasoning`、
+`sampling` 或 `request_headers` 会保留原值。显式空 `request_headers` 会清空请求头；
+`sampling` 传入时整体替换（旧值中未出现的键会被清掉），传 `null` 或 `{}` 清除全部
+采样参数，切换到其他 Provider 时同样清除；其余高级字段传 `null` 时按各字段合同清除。外部登录器留下的 legacy `readonly:true` 只是
 兼容元数据，不阻止编辑。响应包含 `api_key` 原值与 `has_api_key`，并仅允许经
 Web 认证的调用方读取。
 
 
@@ -516,6 +516,11 @@ nlohmann::json AnthropicProvider::build_request_body(
     body["max_tokens"] = request_options_.max_output_tokens.value_or(
         kDefaultMaxTokens);
     if (stream) body["stream"] = true;
+    // 采样参数只在配置了时发送。Anthropic 不支持 min_p,配置校验已拒绝。
+    const auto& sampling = request_options_.sampling;
+    if (sampling.temperature.has_value()) body["temperature"] = *sampling.temperature;
+    if (sampling.top_p.has_value()) body["top_p"] = *sampling.top_p;
+    if (sampling.top_k.has_value()) body["top_k"] = *sampling.top_k;
 
     if (request_options_.reasoning_protocol == ReasoningWireProtocol::Anthropic &&
         request_options_.reasoning.has_value() &&
 
@@ -898,6 +898,14 @@ nlohmann::json OpenAiCompatProvider::build_request_body(
     if (request_options_.max_output_tokens.has_value()) {
         body["max_tokens"] = *request_options_.max_output_tokens;
     }
+    // 采样参数只在用户为这个模型配置了时发送;未配置时请求体与旧版本一致。
+    // top_k / min_p 不是 OpenAI 官方字段(vLLM / SGLang 支持),配置校验已保证
+    // Responses 协议不会带上它们(Responses 转换也只保留 temperature / top_p)。
+    const auto& sampling = request_options_.sampling;
+    if (sampling.temperature.has_value()) body["temperature"] = *sampling.temperature;
+    if (sampling.top_p.has_value()) body["top_p"] = *sampling.top_p;
+    if (sampling.top_k.has_value()) body["top_k"] = *sampling.top_k;
+    if (sampling.min_p.has_value()) body["min_p"] = *sampling.min_p;
     if (request_options_.reasoning_protocol == ReasoningWireProtocol::OpenRouter &&
         request_options_.reasoning.has_value() &&
         request_options_.reasoning->supported) {
 
@@ -14,6 +14,7 @@
 #include <algorithm>
 #include <cctype>
 #include <cstdint>
+#include <cstring>
 #include <map>
 #include <memory>
 #include <random>
@@ -52,6 +53,7 @@ ProviderRequestOptions request_options_from_entry(const ModelProfile& entry) {
     options.api_protocol = model_profile_api_protocol(entry);
     options.max_output_tokens = entry.max_output_tokens;
     options.reasoning = entry.reasoning;
+    options.sampling = entry.sampling;
     if (entry.provider == "anthropic" && entry.reasoning.has_value()) {
         options.reasoning_protocol = ReasoningWireProtocol::Anthropic;
     } else if (entry.provider == "openai" && entry.reasoning.has_value() &&
@@ -88,6 +90,24 @@ void append_optional_int(std::string& out, const std::optional<T>& value) {
     }
 }
 
+void append_optional_double(std::string& out,
+                            const std::optional<double>& value) {
+    append_bool(out, value.has_value());
+    if (!value.has_value()) return;
+    // 按位编码:0.95 与 0.9500000001 必须得到不同指纹,不能先格式化成文本。
+    std::uint64_t bits = 0;
+    static_assert(sizeof(bits) == sizeof(double), "double must be 64-bit");
+    std::memcpy(&bits, &*value, sizeof(bits));
+    append_u64(out, bits);
+}
+
+void append_sampling(std::string& out, const ModelSamplingOptions& sampling) {
+    append_optional_double(out, sampling.temperature);
+    append_optional_double(out, sampling.top_p);
+    append_optional_int(out, sampling.top_k);
+    append_optional_double(out, sampling.min_p);
+}
+
 void append_optional_bool(std::string& out,
                           const std::optional<bool>& value) {
     append_bool(out, value.has_value());
@@ -142,6 +162,7 @@ std::string canonical_plan_bytes(const EffectiveProviderBuildPlan& plan) {
         append_u64(out, static_cast<std::uint64_t>(
             plan.request_options.reasoning_protocol));
         append_reasoning(out, plan.request_options.reasoning);
+        append_sampling(out, plan.request_options.sampling);
     }
     append_bool(out, plan.applies_vision_routing);
     if (plan.applies_vision_routing) {
@@ -209,18 +230,20 @@ std::optional<EffectiveProviderBuildPlan> effective_plan_from_entry(
     } else if (entry.provider == "copilot") {
         if (!entry.base_url.empty() || !entry.api_key.empty() ||
             !entry.request_headers.empty() || entry.endpoint_mode.has_value() ||
-            entry.max_output_tokens.has_value() || entry.reasoning.has_value()) {
+            entry.max_output_tokens.has_value() || entry.reasoning.has_value() ||
+            !entry.sampling.empty()) {
             LOG_WARN("[provider_factory] refusing unsupported Copilot endpoint, "
-                     "credential, output, or reasoning customization");
+                     "credential, output, reasoning, or sampling customization");
             return std::nullopt;
         }
     } else if (entry.provider == "grok") {
         if (!entry.base_url.empty() || !entry.api_key.empty() ||
             !entry.request_headers.empty() || entry.endpoint_mode.has_value() ||
             entry.max_output_tokens.has_value() || entry.reasoning.has_value() ||
-            entry.stream_timeout_ms.has_value()) {
+            entry.stream_timeout_ms.has_value() || !entry.sampling.empty()) {
             LOG_WARN("[provider_factory] refusing unsupported Grok endpoint, "
-                     "credential, timeout, output, or reasoning customization");
+                     "credential, timeout, output, reasoning, or sampling "
+                     "customization");
             return std::nullopt;
         }
     } else if (entry.provider == "codex") {
 
@@ -22,6 +22,8 @@ struct ProviderRequestOptions {
     std::optional<int> max_output_tokens;
     std::optional<ModelReasoningOptions> reasoning;
     ReasoningWireProtocol reasoning_protocol = ReasoningWireProtocol::None;
+    // 空 = 一个采样字段都不发,请求体与旧版本逐字节一致。
+    ModelSamplingOptions sampling;
 };
 
 } // namespace acecode
@@ -90,6 +90,9 @@ nlohmann::json entry_to_json(const ModelProfile& entry) {
     if (entry.reasoning.has_value()) {
         o["reasoning"] = model_reasoning_options_to_json(*entry.reasoning);
     }
+    if (!entry.sampling.empty()) {
+        o["sampling"] = model_sampling_options_to_json(entry.sampling);
+    }
     if (!entry.request_headers.empty()) {
         o["request_headers"] = entry.request_headers;
     }
@@ -388,6 +391,22 @@ std::optional<SavedModelDraft> parse_model_draft(const nlohmann::json& body,
             }
         }
     }
+    if (body.contains("sampling")) {
+        // null / {} 都表示清除;对象整体替换已保存的采样参数。
+        d.sampling_supplied = true;
+        if (body["sampling"].is_null()) {
+            d.sampling.reset();
+        } else {
+            std::string sampling_error;
+            auto sampling = parse_model_sampling_options(body["sampling"],
+                                                         sampling_error);
+            if (!sampling.has_value()) {
+                err = "field 'sampling' is invalid: " + sampling_error;
+            } else {
+                d.sampling = *sampling;
+            }
+        }
+    }
     if (body.contains("request_headers")) {
         d.request_headers_supplied = true;
         auto parsed = parse_request_headers_json(body["request_headers"],
 
@@ -2693,6 +2693,9 @@ nlohmann::json build_config_json(const AppConfig& cfg) {
         if (e.reasoning.has_value()) {
             ej["reasoning"] = model_reasoning_options_to_json(*e.reasoning);
         }
+        if (!e.sampling.empty()) {
+            ej["sampling"] = model_sampling_options_to_json(e.sampling);
+        }
         if (!e.request_headers.empty()) {
             nlohmann::json headers = nlohmann::json::object();
             for (const auto& [k, v] : e.request_headers) headers[k] = v;