Wire K3 thinking-effort levels end to end (fixes 'model does not support reasoning effort')

The live /models wire (verified against api.kimi.com) marks every Kimi
Code model supports_thinking_type: "only" and gives K3 a think_efforts
block {support, valid_efforts: [low, high, max], default_effort: max} —
both of which the F4 sync discarded, hardcoding
supports_reasoning_effort: false. Every effort selection was therefore
rejected with 'current model does not support reasoning effort'.

- kigi-models: WireModel gains supports_thinking_type + WireThinkEfforts;
  "only" forces the always_thinking capability.
- models_fetch: think_efforts maps into the catalog entry — wire tokens
  stay the option ids/labels (max/Max), canonical values map via the
  ReasoningEffort parser (max → Xhigh), default_effort marks the default.
  This lights up the existing /model <model> [effort] two-phase completion
  and the /effort menu with the server's own vocabulary.
- kimi_compat: the effort level rides the wire as thinking.effort
  ({"type": "enabled", "effort": "low"} is accepted live; invalid
  levels are a 400). Only the canonical-vs-wire spelling divergence
  (xhigh → max) is renamed; levels pass through verbatim so a contract
  violation surfaces instead of being clamped away.

Live acceptance: kigi -m kimi-code/k3 --reasoning-effort max -p ... round
trips against api.kimi.com, and the refreshed models_cache.json carries
low/high/max with the max default.

crates/codegen/kigi-models/src/lib.rs

@@ -187,6 +187,28 @@ pub struct WireModel {
     pub supports_video_in: bool,
     #[serde(default)]
     pub display_name: Option<String>,
+    /// `"only"` marks always-thinking models (thinking cannot be disabled).
+    /// Verified against the live `api.kimi.com/coding/v1/models` response.
+    #[serde(default)]
+    pub supports_thinking_type: Option<String>,
+    /// Selectable thinking-effort levels, present only on models that offer
+    /// them (e.g. K3). Verified against the live `/models` response.
+    #[serde(default)]
+    pub think_efforts: Option<WireThinkEfforts>,
+}
+
+/// The `think_efforts` object of a `/models` entry. Live wire shape
+/// (api.kimi.com, 2026-07):
+/// `{"support": true, "valid_efforts": ["low", "high", "max"],
+///   "default_effort": "max"}`.
+#[derive(Debug, Clone, Default, serde::Deserialize)]
+pub struct WireThinkEfforts {
+    #[serde(default)]
+    pub support: bool,
+    #[serde(default)]
+    pub valid_efforts: Vec<String>,
+    #[serde(default)]
+    pub default_effort: Option<String>,
 }
 
 /// `GET {base}/models` response envelope.
@@ -203,14 +225,26 @@ impl WireModel {
     /// - `supports_image_in` → image_in; `supports_video_in` → video_in
     /// - id starts with `kimi-k2` → thinking + image_in + video_in
     ///
+    /// On top of that, the live wire's `supports_thinking_type: "only"`
+    /// marks a model whose thinking cannot be disabled → always_thinking.
+    ///
     /// Returned sorted + deduplicated ([`ModelCapability`]'s `Ord`).
     pub fn capabilities(&self) -> Vec<ModelCapability> {
-        derive_capabilities(
+        let mut caps = derive_capabilities(
             &self.id,
             self.supports_reasoning,
             self.supports_image_in,
             self.supports_video_in,
-        )
+        );
+        if self.supports_thinking_type.as_deref() == Some("only") {
+            for cap in [ModelCapability::Thinking, ModelCapability::AlwaysThinking] {
+                if !caps.contains(&cap) {
+                    caps.push(cap);
+                }
+            }
+            caps.sort();
+        }
+        caps
     }
 }
 
@@ -352,6 +386,63 @@ pub fn default_session_summary_model() -> &'static str {
 mod tests {
     use super::*;
 
+    /// Mirror of the live `api.kimi.com/coding/v1/models` K3 entry
+    /// (fetched 2026-07-17): `supports_thinking_type: "only"` plus a
+    /// `think_efforts` block with low/high/max and a max default.
+    #[test]
+    fn wire_model_parses_live_k3_think_efforts() {
+        let json = serde_json::json!({
+            "id": "k3",
+            "created": 1_761_264_000,
+            "object": "model",
+            "display_name": "K3",
+            "type": "model",
+            "context_length": 1_048_576,
+            "supports_reasoning": true,
+            "supports_image_in": true,
+            "supports_video_in": true,
+            "supports_thinking_type": "only",
+            "think_efforts": {
+                "support": true,
+                "valid_efforts": ["low", "high", "max"],
+                "default_effort": "max"
+            }
+        });
+        let wire: WireModel = serde_json::from_value(json).unwrap();
+        let efforts = wire.think_efforts.as_ref().unwrap();
+        assert!(efforts.support);
+        assert_eq!(efforts.valid_efforts, ["low", "high", "max"]);
+        assert_eq!(efforts.default_effort.as_deref(), Some("max"));
+        // "only" thinking type forces always_thinking on top of the
+        // supports_reasoning-derived thinking capability.
+        let caps = wire.capabilities();
+        assert!(caps.contains(&ModelCapability::Thinking));
+        assert!(caps.contains(&ModelCapability::AlwaysThinking));
+    }
+
+    /// The K2.7 entries carry `supports_thinking_type: "only"` but no
+    /// `think_efforts` — always-thinking without selectable levels.
+    #[test]
+    fn wire_model_without_think_efforts_still_always_thinking() {
+        let json = serde_json::json!({
+            "id": "kimi-for-coding",
+            "context_length": 262_144,
+            "supports_reasoning": true,
+            "supports_image_in": true,
+            "supports_video_in": true,
+            "supports_thinking_type": "only"
+        });
+        let wire: WireModel = serde_json::from_value(json).unwrap();
+        assert!(wire.think_efforts.is_none());
+        let caps = wire.capabilities();
+        assert!(caps.contains(&ModelCapability::AlwaysThinking));
+        // Sorted + deduplicated invariant holds after the "only" injection.
+        let mut sorted = caps.clone();
+        sorted.sort();
+        sorted.dedup();
+        assert_eq!(caps, sorted);
+    }
+
     #[test]
     fn platform_ids_round_trip() {
         for p in PlatformId::ALL {
@@ -463,6 +554,8 @@ mod tests {
                 supports_image_in: false,
                 supports_video_in: false,
                 display_name: None,
+                supports_thinking_type: None,
+                think_efforts: None,
             },
             WireModel {
                 id: "moonshot-v1-8k".into(),
@@ -471,6 +564,8 @@ mod tests {
                 supports_image_in: false,
                 supports_video_in: false,
                 display_name: None,
+                supports_thinking_type: None,
+                think_efforts: None,
             },
         ];
         let filtered = filter_allowed_models(PlatformId::MoonshotCn, listing.clone());

crates/codegen/kigi-sampler/src/kimi_compat.rs

@@ -32,12 +32,21 @@ pub(crate) fn adapt_chat_completions_body(body: &mut Value) {
 /// Map the OpenAI-style `reasoning_effort` knob onto Kimi's `thinking`
 /// request field and drop `reasoning_effort` from the wire.
 ///
-/// kimi-cli controls thinking exclusively through the request body's
+/// kimi-cli 1.49.0 controls thinking through the request body's
 /// `thinking: {"type": "enabled" | "disabled"}` field
 /// (packages/kosong/src/kosong/chat_provider/kimi.py:214-223 `with_thinking`:
 /// `"enabled" if effort != "off" else "disabled"`; wired by
 /// src/kimi_cli/llm.py:475-481). When no effort is configured, nothing is
 /// sent and the server default applies (llm.py:482 "leave as-is").
+///
+/// Models with selectable levels (the `/models` `think_efforts` block, e.g.
+/// K3's low/high/max) additionally take the level as `thinking.effort` —
+/// verified against the live api.kimi.com: `{"type": "enabled", "effort":
+/// "low"}` is accepted, values outside `valid_efforts` are a 400. The
+/// catalog gates efforts to that per-model list, so this layer only renames
+/// the one canonical-vs-wire divergence (`xhigh` → `max`) and passes the
+/// level through verbatim — inventing or clamping a level here would hide a
+/// real contract violation.
 fn adapt_thinking(body: &mut Value) {
     let Some(obj) = body.as_object_mut() else {
         return;
@@ -45,11 +54,22 @@ fn adapt_thinking(body: &mut Value) {
     let Some(effort) = obj.remove("reasoning_effort") else {
         return;
     };
-    let enabled = effort.as_str() != Some("none");
-    obj.insert(
-        "thinking".to_owned(),
-        serde_json::json!({ "type": if enabled { "enabled" } else { "disabled" } }),
+    let effort = effort.as_str().map(str::to_owned);
+    let enabled = effort.as_deref() != Some("none");
+    let mut thinking = serde_json::Map::new();
+    thinking.insert(
+        "type".to_owned(),
+        Value::String(if enabled { "enabled" } else { "disabled" }.to_owned()),
     );
+    if enabled && let Some(level) = effort {
+        let wire_level = if level == "xhigh" {
+            "max".to_owned()
+        } else {
+            level
+        };
+        thinking.insert("effort".to_owned(), Value::String(wire_level));
+    }
+    obj.insert("thinking".to_owned(), Value::Object(thinking));
 }
 
 /// Message-level adaptations:
@@ -262,12 +282,27 @@ mod tests {
 
     #[test]
     fn reasoning_effort_maps_to_kimi_thinking_field() {
+        // Level rides along as thinking.effort (live wire: 200 with
+        // {"type": "enabled", "effort": "low"}).
         let mut body = json!({ "model": "kimi-for-coding", "reasoning_effort": "high" });
         adapt_chat_completions_body(&mut body);
         assert_eq!(body.get("reasoning_effort"), None);
-        assert_eq!(body["thinking"], json!({ "type": "enabled" }));
+        assert_eq!(
+            body["thinking"],
+            json!({ "type": "enabled", "effort": "high" })
+        );
 
-        // kimi.py:218: "off" (our ReasoningEffort::None) → disabled.
+        // Canonical `xhigh` is spelled `max` on the Kimi wire (the K3
+        // valid_efforts vocabulary is low/high/max).
+        let mut body = json!({ "model": "k3", "reasoning_effort": "xhigh" });
+        adapt_chat_completions_body(&mut body);
+        assert_eq!(
+            body["thinking"],
+            json!({ "type": "enabled", "effort": "max" })
+        );
+
+        // kimi.py:218: "off" (our ReasoningEffort::None) → disabled, and no
+        // effort key (a disabled+effort combination would be contradictory).
         let mut body = json!({ "reasoning_effort": "none" });
         adapt_chat_completions_body(&mut body);
         assert_eq!(body["thinking"], json!({ "type": "disabled" }));

crates/codegen/kigi-sampler/tests/test_kimi_wire.rs

@@ -382,9 +382,12 @@ async fn request_carries_bearer_kigi_ua_and_kimi_dialect_body() {
     assert_eq!(body["stream"], json!(true));
     assert_eq!(body["stream_options"], json!({ "include_usage": true }));
 
-    // -- Thinking mapping (kimi.py:214-223): effort → thinking, no
-    //    reasoning_effort on the wire.
-    assert_eq!(body["thinking"], json!({ "type": "enabled" }));
+    // -- Thinking mapping (kimi.py:214-223 + live think_efforts wire):
+    //    effort → thinking {type, effort}, no reasoning_effort on the wire.
+    assert_eq!(
+        body["thinking"],
+        json!({ "type": "enabled", "effort": "high" })
+    );
     assert_eq!(body.get("reasoning_effort"), None);
 
     // -- Message adaptations.

crates/codegen/kigi-shell/src/agent/models_fetch.rs

@@ -303,12 +303,49 @@ fn fetch_one_platform_models(
 /// platforms — never key values — because raw fetched entries are persisted
 /// to the models disk cache. Config-file keys are stamped in-memory later by
 /// `resolve_model_list`'s platform-credentials layer.
+/// Map a live `think_efforts` block to catalog effort options. The wire
+/// token stays the option id/label (`"max"` → label `"Max"`) so the UI
+/// mirrors the server's vocabulary, while the canonical value maps through
+/// the [`kigi_sampling_types::ReasoningEffort`] parser (`"max"` → `Xhigh`).
+/// Unknown tokens are dropped with a warning rather than inventing a level.
+fn think_efforts_to_options(
+    think: &kigi_models::WireThinkEfforts,
+) -> Vec<kigi_sampling_types::ReasoningEffortOption> {
+    think
+        .valid_efforts
+        .iter()
+        .filter_map(|token| {
+            let value = match token.parse::<kigi_sampling_types::ReasoningEffort>() {
+                Ok(v) => v,
+                Err(error) => {
+                    tracing::warn!(%token, %error, "unknown think_efforts token; dropping");
+                    return None;
+                }
+            };
+            let mut label: String = token.clone();
+            if let Some(first) = label.get_mut(0..1) {
+                first.make_ascii_uppercase();
+            }
+            Some(kigi_sampling_types::ReasoningEffortOption {
+                id: token.clone(),
+                value,
+                label,
+                description: None,
+                default: think.default_effort.as_deref() == Some(token.as_str()),
+            })
+        })
+        .collect()
+}
+
 fn platform_wire_model_to_entry(
     platform: kigi_models::PlatformId,
     wire: kigi_models::WireModel,
     base_url: &str,
 ) -> crate::agent::config::ModelEntryConfig {
     let capabilities = wire.capabilities();
+    // Selectable thinking levels (live wire `think_efforts`, e.g. K3's
+    // low/high/max). `support: false` or absence both mean "no levels".
+    let think_efforts = wire.think_efforts.as_ref().filter(|t| t.support);
     let context_window = std::num::NonZeroU64::new(wire.context_length).unwrap_or_else(|| {
         tracing::debug!(
             model = %wire.id,
@@ -332,9 +369,13 @@ fn platform_wire_model_to_entry(
         env_key,
         api_backend: Default::default(),
         auth_scheme: None,
-        reasoning_effort: None,
-        supports_reasoning_effort: false,
-        reasoning_efforts: Vec::new(),
+        reasoning_effort: think_efforts
+            .and_then(|t| t.default_effort.as_deref())
+            .and_then(|s| s.parse().ok()),
+        supports_reasoning_effort: think_efforts.is_some(),
+        reasoning_efforts: think_efforts
+            .map(think_efforts_to_options)
+            .unwrap_or_default(),
         capabilities,
         extra_headers: IndexMap::new(),
         context_window,
@@ -634,6 +675,80 @@ mod tests {
         assert_eq!(result.model, "actual-model-id");
         assert_eq!(result.name.as_deref(), Some("Display Name"));
     }
+    /// Live-wire regression: the K3 `/models` entry (api.kimi.com, 2026-07)
+    /// must land in the catalog with selectable low/high/max efforts and a
+    /// max default — this is what feeds `/model <m> [effort]` and `/effort`.
+    #[test]
+    fn platform_entry_maps_live_k3_think_efforts() {
+        use kigi_sampling_types::ReasoningEffort;
+        let wire: kigi_models::WireModel = serde_json::from_value(serde_json::json!({
+            "id": "k3",
+            "display_name": "K3",
+            "context_length": 1_048_576,
+            "supports_reasoning": true,
+            "supports_image_in": true,
+            "supports_video_in": true,
+            "supports_thinking_type": "only",
+            "think_efforts": {
+                "support": true,
+                "valid_efforts": ["low", "high", "max"],
+                "default_effort": "max"
+            }
+        }))
+        .unwrap();
+        let entry = platform_wire_model_to_entry(
+            kigi_models::PlatformId::KimiCode,
+            wire,
+            "https://api.kimi.com/coding/v1",
+        );
+        assert!(entry.supports_reasoning_effort);
+        assert_eq!(entry.reasoning_effort, Some(ReasoningEffort::Xhigh));
+        let ids: Vec<&str> = entry
+            .reasoning_efforts
+            .iter()
+            .map(|o| o.id.as_str())
+            .collect();
+        assert_eq!(
+            ids,
+            ["low", "high", "max"],
+            "wire tokens stay the option ids"
+        );
+        assert_eq!(
+            entry
+                .reasoning_efforts
+                .iter()
+                .map(|o| o.value)
+                .collect::<Vec<_>>(),
+            [
+                ReasoningEffort::Low,
+                ReasoningEffort::High,
+                ReasoningEffort::Xhigh
+            ],
+        );
+        let max = entry
+            .reasoning_efforts
+            .iter()
+            .find(|o| o.id == "max")
+            .unwrap();
+        assert!(max.default, "max is the server default for K3");
+        assert_eq!(max.label, "Max");
+        // K2.7-style entries (no think_efforts) stay effort-less.
+        let plain: kigi_models::WireModel = serde_json::from_value(serde_json::json!({
+            "id": "kimi-for-coding",
+            "context_length": 262_144,
+            "supports_reasoning": true,
+            "supports_thinking_type": "only"
+        }))
+        .unwrap();
+        let entry = platform_wire_model_to_entry(
+            kigi_models::PlatformId::KimiCode,
+            plain,
+            "https://api.kimi.com/coding/v1",
+        );
+        assert!(!entry.supports_reasoning_effort);
+        assert!(entry.reasoning_efforts.is_empty());
+        assert!(entry.reasoning_effort.is_none());
+    }
     #[test]
     fn parse_reads_reasoning_effort_fields() {
         use kigi_sampling_types::ReasoningEffort;