feat: first-class reasoning/thinking + tool translation for GLM-5.x, Kimi-K2.6/2.7-code, DeepSeek-V4, grok-4.5

README.md

@@ -167,7 +167,7 @@ for deployment examples.
 | MiniMax | [`configs/model-families/minimax.toml`](configs/model-families/minimax.toml) | `minimax-m2.5`, `minimax-m2.7`, `minimax-m3` |
 | Moonshot AI | [`configs/model-families/moonshot-ai.toml`](configs/model-families/moonshot-ai.toml) | `kimi-k2`, `kimi-k2-0905`, `kimi-k2.5`, `kimi-k2.6`, `kimi-k2.6-code`, `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-for-coding` |
 | Alibaba Cloud | [`configs/model-families/qwen.toml`](configs/model-families/qwen.toml) | `qwen3.6-35b-a3b`; conservative broad defaults for `qwen3.6*` and `qwen3.7*` |
-| xAI | [`configs/model-families/x-ai.toml`](configs/model-families/x-ai.toml) | `grok-4.3`, `grok-build-0.1` |
+| xAI | [`configs/model-families/x-ai.toml`](configs/model-families/x-ai.toml) | `grok-4.3`, `grok-4.5`, `grok-build-0.1` |
 | Xiaomi | [`configs/model-families/xiaomi.toml`](configs/model-families/xiaomi.toml) | `mimo-v2.5`, `mimo-v2.5-pro` |
 | Z.ai | [`configs/model-families/z-ai.toml`](configs/model-families/z-ai.toml) | `glm-5`, `glm-5.1`, `glm-5.2` |
 

configs/model-families/deepseek.toml

@@ -64,20 +64,26 @@ patterns = [
 ]
 
 [model_families.deepseek_v4_flash.model_metadata]
-context_window = 256000
+context_window = 1000000
 default_reasoning_level = "medium"
 supported_reasoning_levels = ["low", "medium", "high"]
 supports_parallel_tool_calls = false
 auto_review_model_override = "deepseek-v4-flash"
 
 [model_families.deepseek_v4_flash.transform]
 force_parallel_tool_calls = false
+preserve_reasoning_content_history = true
 
 [[model_families.deepseek_v4_flash.transform.remove_chat_request_morphs]]
 from = "reasoning.effort"
 to = "reasoning_effort"
 kind = "rename"
 
+[[model_families.deepseek_v4_flash.transform.append_chat_request_morphs]]
+from = "reasoning.effort"
+to = "reasoning_effort"
+kind = "rename"
+
 [[model_families.deepseek_v4_flash.transform.append_chat_request_morphs]]
 from = "reasoning.effort"
 to = "thinking.type"
@@ -100,12 +106,18 @@ auto_review_model_override = "deepseek-v4-flash"
 
 [model_families.deepseek_v4_pro.transform]
 force_parallel_tool_calls = false
+preserve_reasoning_content_history = true
 
 [[model_families.deepseek_v4_pro.transform.remove_chat_request_morphs]]
 from = "reasoning.effort"
 to = "reasoning_effort"
 kind = "rename"
 
+[[model_families.deepseek_v4_pro.transform.append_chat_request_morphs]]
+from = "reasoning.effort"
+to = "reasoning_effort"
+kind = "rename"
+
 [[model_families.deepseek_v4_pro.transform.append_chat_request_morphs]]
 from = "reasoning.effort"
 to = "thinking.type"

configs/model-families/x-ai.toml

@@ -66,3 +66,30 @@ supports_search_tool = true
 
 [model_families.x_ai_grok_build_0_1.transform]
 force_parallel_tool_calls = false
+
+[model_families.x_ai_grok_4_5]
+priority = 10
+patterns = [
+  "grok-4.5",
+  "grok_4.5",
+  "grok-4.5-latest",
+  "xai/grok-4.5",
+  "x-ai/grok-4.5",
+]
+
+[model_families.x_ai_grok_4_5.model_metadata]
+context_window = 500000
+default_reasoning_level = "high"
+supported_reasoning_levels = ["low", "medium", "high"]
+input_modalities = ["text", "image"]
+supports_parallel_tool_calls = true
+supports_search_tool = true
+
+[model_families.x_ai_grok_4_5.transform]
+force_parallel_tool_calls = false
+reasoning_effort_none_value = "low"
+
+[[model_families.x_ai_grok_4_5.transform.append_chat_request_morphs]]
+from = "reasoning.effort"
+to = "reasoning_effort"
+kind = "rename"

configs/model-families/z-ai.toml

@@ -35,6 +35,11 @@ from = "reasoning.effort"
 to = "thinking.type"
 kind = "thinking_type"
 
+[[model_families.z_ai_glm_5.transform.append_chat_request_morphs]]
+from = "reasoning.effort"
+to = "reasoning_effort"
+kind = "rename"
+
 [model_families.z_ai_glm_5_base]
 priority = 10
 patterns = [

docs/model-family-catalogs.md

@@ -29,7 +29,7 @@ Provider/gateway behavior belongs in
 | MiniMax | [`minimax.toml`](../configs/model-families/minimax.toml) | `minimax-m2.5`, `minimax-m2.7`, `minimax-m3` |
 | Moonshot AI | [`moonshot-ai.toml`](../configs/model-families/moonshot-ai.toml) | `kimi-k2`, `kimi-k2-0905`, `kimi-k2.5`, `kimi-k2.6`, `kimi-k2.6-code`, `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-for-coding` |
 | Alibaba Cloud | [`qwen.toml`](../configs/model-families/qwen.toml) | `qwen3.6-35b-a3b`; conservative broad defaults for `qwen3.6*` and `qwen3.7*` |
-| xAI | [`x-ai.toml`](../configs/model-families/x-ai.toml) | `grok-4.3`, `grok-build-0.1` |
+| xAI | [`x-ai.toml`](../configs/model-families/x-ai.toml) | `grok-4.3`, `grok-4.5`, `grok-build-0.1` |
 | Xiaomi | [`xiaomi.toml`](../configs/model-families/xiaomi.toml) | `mimo-v2.5`, `mimo-v2.5-pro` |
 | Z.ai | [`z-ai.toml`](../configs/model-families/z-ai.toml) | `glm-5`, `glm-5.1`, `glm-5.2` |
 
@@ -48,8 +48,8 @@ transform behavior.
 | --- | --- | --- | --- | --- | --- | --- |
 | DeepSeek | `deepseek-v3.2` | 128k | provider/default | low, medium, high; default medium | provider/default | Parallel tool metadata is true, but requests force `parallel_tool_calls = false`. |
 | DeepSeek | `deepseek-v3.2-speciale` | 128k | provider/default | low, medium, high; default high | provider/default | Forces `parallel_tool_calls = false`. |
-| DeepSeek | `deepseek-v4-flash` | 256k | provider/default | low, medium, high; default medium | provider/default | Converts `reasoning.effort` to `thinking.type`; forces `parallel_tool_calls = false`. |
-| DeepSeek | `deepseek-v4-pro` | 1,000k | provider/default | low, medium, high; default high | provider/default | Converts `reasoning.effort` to `thinking.type`; forces `parallel_tool_calls = false`. |
+| DeepSeek | `deepseek-v4-flash` | 1,000k | provider/default | low, medium, high; default medium | provider/default | Converts `reasoning.effort` to `thinking.type` and forwards `reasoning_effort`; preserves reasoning history (`preserve_reasoning_content_history`); forces `parallel_tool_calls = false`. |
+| DeepSeek | `deepseek-v4-pro` | 1,000k | provider/default | low, medium, high; default high | provider/default | Converts `reasoning.effort` to `thinking.type` and forwards `reasoning_effort`; preserves reasoning history (`preserve_reasoning_content_history`); forces `parallel_tool_calls = false`. |
 | MiniMax | `minimax-m2.5` | 192k | text | none | true | Non-reasoning variant; forces `parallel_tool_calls = false`. |
 | MiniMax | `minimax-m2.7` | 200k | text | low, medium, high; default high | true | Converts `reasoning.effort` to `thinking.type`; forces `parallel_tool_calls = false`. |
 | MiniMax | `minimax-m3` | 1,000k | text, image | low, medium, high; default high | false | Converts `reasoning.effort` to `thinking.type`; forces `parallel_tool_calls = false`. |
@@ -61,11 +61,12 @@ transform behavior.
 | Moonshot AI | `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-for-coding` | usable 220k, max 262,144 | text, image | high | provider/default | Uses a conservative Codex planning window below the provider 256K cap to leave output/tokenizer headroom; sets `thinking.type = enabled` and `thinking.keep = all`; drops K2.7-rejected sampling overrides; forces `parallel_tool_calls = false`. |
 | Alibaba Cloud | `qwen3.6-35b-a3b` | 262k, max 1,010k | text, image, video | high | false | Removes inherited OpenAI `reasoning_effort`; broad `qwen3.6*` and `qwen3.7*` entries also drop it until a gateway documents support. |
 | xAI | `grok-4.3` | 1,000k | text, image | low, medium, high; default medium | true | Uses `web_search`; parallel tool metadata is true, but requests force `parallel_tool_calls = false`. |
+| xAI | `grok-4.5` | 500k | text, image | low, medium, high; default high | true | Uses `web_search`; reasoning cannot be disabled (`none` mapped to `low` via `reasoning_effort_none_value`); parallel tool metadata is true, but requests force `parallel_tool_calls = false`. |
 | xAI | `grok-build-0.1` | 256k | text | low, medium, high; default medium | true | Uses `web_search`; parallel tool metadata is true, but requests force `parallel_tool_calls = false`. |
 | Xiaomi | `mimo-v2.5` | 1,000k | text, image | low, medium, high; default medium | provider/default | Gateway profile does not carry model-specific overrides. |
 | Xiaomi | `mimo-v2.5-pro` | 1,000k | text | low, medium, high; default medium | provider/default | Gateway profile does not carry model-specific overrides. |
 | Z.ai | `glm-5`, `glm-5.1` family | 200k | text | low, medium, high; default medium | provider/default | Broad GLM-5 transform converts `reasoning.effort` to `thinking.type`; forces `parallel_tool_calls = false`. |
-| Z.ai | `glm-5.2` | 1,000k | text | low, medium, high; default medium | provider/default | Inherits broad GLM-5 transform. |
+| Z.ai | `glm-5.2` | 1,000k | text | low, medium, high; default medium | provider/default | Inherits broad GLM-5 transform; forwards `reasoning_effort` alongside `thinking.type` (GLM-5.2 honors it natively). |
 
 ## Matching And Priority
 

src/config.rs

@@ -343,6 +343,7 @@ pub struct TransformConfig {
     pub responses_request_morphs: Vec<RequestMorph>,
     pub unsupported_tool_types: Vec<String>,
     pub unsupported_tool_strategy: UnsupportedToolStrategy,
+    pub reasoning_effort_none_value: Option<String>,
     pub drop_empty_tool_choice: bool,
     pub force_parallel_tool_calls: Option<bool>,
     pub request_stream_options_include_usage: bool,
@@ -357,6 +358,7 @@ impl Default for TransformConfig {
             responses_request_morphs: Vec::new(),
             unsupported_tool_types: vec!["custom".to_string()],
             unsupported_tool_strategy: UnsupportedToolStrategy::AsFunction,
+            reasoning_effort_none_value: None,
             drop_empty_tool_choice: true,
             force_parallel_tool_calls: None,
             request_stream_options_include_usage: false,
@@ -377,6 +379,7 @@ pub struct TransformConfigPatch {
     pub append_responses_request_morphs: Vec<RequestMorph>,
     pub unsupported_tool_types: Option<Vec<String>>,
     pub unsupported_tool_strategy: Option<UnsupportedToolStrategy>,
+    pub reasoning_effort_none_value: Option<String>,
     pub drop_empty_tool_choice: Option<bool>,
     pub force_parallel_tool_calls: Option<bool>,
     pub request_stream_options_include_usage: Option<bool>,
@@ -414,6 +417,9 @@ impl TransformConfigPatch {
         if let Some(strategy) = self.unsupported_tool_strategy {
             transform.unsupported_tool_strategy = strategy;
         }
+        if let Some(value) = &self.reasoning_effort_none_value {
+            transform.reasoning_effort_none_value = Some(value.clone());
+        }
         if let Some(drop_empty_tool_choice) = self.drop_empty_tool_choice {
             transform.drop_empty_tool_choice = drop_empty_tool_choice;
         }

src/config_tests.rs

@@ -51,6 +51,7 @@ fn example_configs_parse_request_morphs() {
     assert!(default_config.model_families.contains_key("mimo_v2_5_pro"));
     assert!(default_config.model_families.contains_key("x_ai_grok"));
     assert!(default_config.model_families.contains_key("x_ai_grok_4_3"));
+    assert!(default_config.model_families.contains_key("x_ai_grok_4_5"));
     assert!(
         default_config
             .model_families
@@ -506,6 +507,92 @@ fn model_family_matches_are_sorted_by_priority() {
     assert_eq!(matches[1].priority, 10);
 }
 
+#[test]
+fn first_class_reasoning_and_tool_translation_for_target_models() {
+    let config = load_config_layers(&[]).expect("default config loads");
+
+    // grok-4.5: new model with correct metadata and `none` -> `low` mapping.
+    let grok45 = config
+        .model_families
+        .get("x_ai_grok_4_5")
+        .expect("grok-4.5 family exists");
+    assert_eq!(grok45.model_metadata.context_window, Some(500_000));
+    assert_eq!(
+        grok45.model_metadata.default_reasoning_level.as_deref(),
+        Some("high")
+    );
+    assert_eq!(
+        grok45.model_metadata.supported_reasoning_levels,
+        Some(vec![
+            "low".to_string(),
+            "medium".to_string(),
+            "high".to_string()
+        ])
+    );
+    assert_eq!(
+        grok45.transform.reasoning_effort_none_value.as_deref(),
+        Some("low")
+    );
+    assert!(
+        grok45
+            .transform
+            .append_chat_request_morphs
+            .iter()
+            .any(|m| m.from == "reasoning.effort" && m.to.as_deref() == Some("reasoning_effort"))
+    );
+    assert!(
+        matching_model_families(&config, "grok-4.5")
+            .iter()
+            .any(|family| family.priority == 10)
+    );
+
+    // DeepSeek V4: 1M context + reasoning-history preservation + effort forwarding.
+    let ds_flash = config
+        .model_families
+        .get("deepseek_v4_flash")
+        .expect("deepseek-v4-flash family exists");
+    assert_eq!(ds_flash.model_metadata.context_window, Some(1_000_000));
+    assert_eq!(
+        ds_flash.transform.preserve_reasoning_content_history,
+        Some(true)
+    );
+    assert!(
+        ds_flash
+            .transform
+            .append_chat_request_morphs
+            .iter()
+            .any(|m| m.from == "reasoning.effort" && m.to.as_deref() == Some("reasoning_effort"))
+    );
+
+    let ds_pro = config
+        .model_families
+        .get("deepseek_v4_pro")
+        .expect("deepseek-v4-pro family exists");
+    assert_eq!(
+        ds_pro.transform.preserve_reasoning_content_history,
+        Some(true)
+    );
+    assert!(
+        ds_pro
+            .transform
+            .append_chat_request_morphs
+            .iter()
+            .any(|m| m.from == "reasoning.effort" && m.to.as_deref() == Some("reasoning_effort"))
+    );
+
+    // GLM-5.2: broad family forwards reasoning_effort alongside thinking.type.
+    let glm5 = config
+        .model_families
+        .get("z_ai_glm_5")
+        .expect("glm-5 broad family exists");
+    assert!(
+        glm5.transform
+            .append_chat_request_morphs
+            .iter()
+            .any(|m| m.from == "reasoning.effort" && m.to.as_deref() == Some("reasoning_effort"))
+    );
+}
+
 #[test]
 fn transform_patch_can_remove_inherited_morphs_before_appending() {
     let mut transform = TransformConfig::default();

src/models_tests.rs

@@ -411,7 +411,7 @@ fn deepseek_family_metadata_is_variant_specific() {
 
     assert_eq!(models[0]["context_window"], 128_000);
     assert_eq!(models[0]["supports_parallel_tool_calls"], true);
-    assert_eq!(models[1]["context_window"], 256_000);
+    assert_eq!(models[1]["context_window"], 1_000_000);
     assert_eq!(models[1]["supports_parallel_tool_calls"], false);
     assert_eq!(models[2]["context_window"], 1_000_000);
     assert_eq!(models[2]["default_reasoning_level"], "high");

src/transform.rs

@@ -9,6 +9,7 @@ use crate::config::TransformConfig;
 use crate::config::UnsupportedToolStrategy;
 use crate::ids::generated_id;
 use crate::transform_morph::apply_native_request_morphs;
+use crate::transform_morph::apply_reasoning_effort_none_value;
 use crate::transform_morph::apply_request_morphs;
 
 #[derive(Debug, Clone)]
@@ -138,7 +139,8 @@ pub fn responses_to_chat(request: Value, transform: &TransformConfig) -> ChatTra
     }
     apply_request_morphs(&request, &mut out, transform);
 
-    let body = Value::Object(out);
+    let mut body = Value::Object(out);
+    apply_reasoning_effort_none_value(&mut body, transform);
     let diagnostics = transform_diagnostics(&request, &body, input_tools.len(), tool_diagnostics);
 
     ChatTransform {

src/transform_morph.rs

@@ -59,6 +59,24 @@ pub(crate) fn apply_request_morphs(
     }
 }
 
+pub(crate) fn apply_reasoning_effort_none_value(body: &mut Value, transform: &TransformConfig) {
+    let Some(none_value) = &transform.reasoning_effort_none_value else {
+        return;
+    };
+    if let Some(Value::String(effort)) = body.get_mut("reasoning_effort") {
+        if effort.eq_ignore_ascii_case("none") {
+            *effort = none_value.clone();
+        }
+    }
+    if let Some(reasoning) = body.get_mut("reasoning").and_then(Value::as_object_mut) {
+        if let Some(Value::String(effort)) = reasoning.get_mut("effort") {
+            if effort.eq_ignore_ascii_case("none") {
+                *effort = none_value.clone();
+            }
+        }
+    }
+}
+
 pub(crate) fn apply_native_request_morphs(request: &mut Value, transform: &TransformConfig) {
     let original = request.clone();
     for morph in &transform.responses_request_morphs {