Split canonical ReasoningEffort::Max out of Xhigh (providers P0c-1)

OpenAI (Responses) and Anthropic (Messages) treat xhigh and max as
DISTINCT effort levels in 2026, and the Kimi K3 wire's top tier is max —
the old parse alias (max→Xhigh) conflated them. Canonical Max now exists:
parse/as_str/serde split, Messages mapping sends xhigh and max as their
own tokens (was Xhigh→"max"), and the K3 menu token max carries
canonical Max end to end.

Kimi wire is byte-identical in all four flows (menu pick, restored
legacy xhigh session, --reasoning-effort flag, /effort command) —
adversarially traced and pinned: kimi_compat's string-level xhigh→max
rename covers legacy tokens, max passes through verbatim.

From the review:
- Rollback safety: persisted reasoning_effort (session summaries, chat
  history) deserializes leniently — unknown future tokens degrade to
  None with a warning instead of hiding sessions or failing resume.
- Restore migration: a pre-split xhigh override onto a model whose menu
  offers max but not xhigh (K3) migrates once, healing display/active-row
  drift and re-persisting the live vocabulary.
- /effort max now rejects (with the offered list) on models whose menu
  lacks a max row instead of silently applying xhigh; deliberate, tested.
- The interim Responses-backend Max→xhigh downgrade (async-openai has no
  Max variant through 0.41) warns loudly; real max wiring lands with the
  OpenAI provider cycle via post-serialize body patch.
- Two rusted ignored-e2e wire pins asserted the pre-adapt reasoning_effort
  key (deleted by the body adapter since ea0ce9d); they now pin the real
  thinking.effort=max shape.

crates/codegen/kigi-shell/src/agent/handlers/model_switch.rs

@@ -110,6 +110,29 @@ pub(crate) async fn apply(
             .models_manager
             .model_supports_reasoning_effort(model_id.0.as_ref())
         {
+            // Legacy migration: pre-split sessions persisted canonical
+            // `xhigh` for models whose live menu now spells the top tier
+            // `max` (K3). The wire is identical either way (kimi_compat
+            // renames xhigh→max), but the menu has no xhigh-valued row, so
+            // display/active-row would drift from the model vocabulary and
+            // the stale token would be re-persisted forever. Migrate once.
+            let eff = if eff == kigi_sampling_types::ReasoningEffort::Xhigh
+                && !agent
+                    .models_manager
+                    .model_offers_effort(model_id.0.as_ref(), eff)
+                && agent.models_manager.model_offers_effort(
+                    model_id.0.as_ref(),
+                    kigi_sampling_types::ReasoningEffort::Max,
+                ) {
+                tracing::info!(
+                    session_id = % session_id.0,
+                    "set_session_model: migrating legacy xhigh override to max \
+                     (model menu offers max, not xhigh)"
+                );
+                kigi_sampling_types::ReasoningEffort::Max
+            } else {
+                eff
+            };
             tracing::info!(
                 session_id = % session_id.0, effort = % eff,
                 "set_session_model: applying reasoning_effort override from meta"

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

@@ -561,6 +561,22 @@ impl ModelsManager {
             .unwrap_or(false)
     }
 
+    /// Whether the model's effort menu offers this canonical value (legacy
+    /// built-in set when the menu is empty). Used to migrate pre-split
+    /// `xhigh` overrides onto `max`-vocabulary models at restore.
+    pub fn model_offers_effort(
+        &self,
+        model_id: &str,
+        effort: kigi_sampling_types::ReasoningEffort,
+    ) -> bool {
+        self.inner
+            .models
+            .read()
+            .get(model_id)
+            .map(|e| model_offers_reasoning_effort(e.info(), effort))
+            .unwrap_or(false)
+    }
+
     /// The catalog default reasoning effort for `model_id`, if the catalog
     /// pins one. Used as the final fallback when neither the session handle
     /// nor the global config sets an explicit effort, so surfaced config stays
@@ -2117,6 +2133,9 @@ pub(crate) fn resolve_model_catalog(
 /// Uses the server `reasoning_efforts` menu when present; otherwise the
 /// built-in low/medium/high/xhigh set (same as the pager legacy menu — no
 /// `none`/`minimal`).
+/// `max` is NOT in the legacy built-in set: models gain it only via an
+/// explicit server/BYOK menu entry (Kimi K3's `max` token), so an empty-menu
+/// model rejects it rather than sending a level its endpoint may 400 on.
 fn model_offers_reasoning_effort(info: &config::ModelInfo, effort: ReasoningEffort) -> bool {
     if !info.supports_reasoning_effort {
         return false;
@@ -3811,6 +3830,51 @@ mod tests {
         assert_eq!(key.0.as_ref(), "kigi");
     }
 
+    /// The restore-migration inputs: a K3-shaped menu (low/high/max tokens)
+    /// offers Max but NOT Xhigh; a legacy empty-menu model offers Xhigh but
+    /// NOT Max. `model_switch` relies on exactly this pair to migrate
+    /// pre-split `xhigh` overrides onto `max`-vocabulary models.
+    #[test]
+    fn offers_effort_distinguishes_k3_menu_from_legacy_set() {
+        let k3_wire: kigi_models::WireModel = serde_json::from_value(serde_json::json!({
+            "id": "k3",
+            "context_length": 1_048_576,
+            "supports_reasoning": true,
+            "supports_thinking_type": "only",
+            "think_efforts": {
+                "support": true,
+                "valid_efforts": ["low", "high", "max"],
+                "default_effort": "max"
+            }
+        }))
+        .unwrap();
+        let k3_cfg = crate::agent::models_fetch::platform_wire_model_to_entry(
+            kigi_models::PlatformId::KimiCode,
+            k3_wire,
+            "https://api.kimi.com/coding/v1",
+        );
+        let k3 = config::ModelInfo::from_config(&k3_cfg);
+        assert!(model_offers_reasoning_effort(&k3, ReasoningEffort::Max));
+        assert!(model_offers_reasoning_effort(&k3, ReasoningEffort::Low));
+        assert!(
+            !model_offers_reasoning_effort(&k3, ReasoningEffort::Xhigh),
+            "K3's menu has no xhigh token — the migration precondition"
+        );
+
+        let mut legacy_cfg = k3_cfg.clone();
+        legacy_cfg.reasoning_efforts = Vec::new();
+        legacy_cfg.supports_reasoning_effort = true;
+        let legacy = config::ModelInfo::from_config(&legacy_cfg);
+        assert!(model_offers_reasoning_effort(
+            &legacy,
+            ReasoningEffort::Xhigh
+        ));
+        assert!(
+            !model_offers_reasoning_effort(&legacy, ReasoningEffort::Max),
+            "legacy built-in set must not offer max (endpoint may 400)"
+        );
+    }
+
     fn test_available_keys(keys: &[&str]) -> IndexMap<acp::ModelId, acp::ModelInfo> {
         keys.iter()
             .map(|k| {

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

@@ -338,7 +338,7 @@ fn think_efforts_to_options(
         .collect()
 }
 
-fn platform_wire_model_to_entry(
+pub(crate) fn platform_wire_model_to_entry(
     platform: kigi_models::PlatformId,
     wire: kigi_models::WireModel,
     base_url: &str,
@@ -703,7 +703,9 @@ mod tests {
             "https://api.kimi.com/coding/v1",
         );
         assert!(entry.supports_reasoning_effort);
-        assert_eq!(entry.reasoning_effort, Some(ReasoningEffort::Xhigh));
+        // The wire token "max" is canonical Max since the Xhigh/Max split;
+        // kimi_compat still spells it "max" on the inference wire.
+        assert_eq!(entry.reasoning_effort, Some(ReasoningEffort::Max));
         let ids: Vec<&str> = entry
             .reasoning_efforts
             .iter()
@@ -723,7 +725,7 @@ mod tests {
             [
                 ReasoningEffort::Low,
                 ReasoningEffort::High,
-                ReasoningEffort::Xhigh
+                ReasoningEffort::Max
             ],
         );
         let max = entry

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

@@ -59,6 +59,7 @@ fn effort_label(effort: ReasoningEffort) -> String {
         ReasoningEffort::Medium => "Medium",
         ReasoningEffort::High => "High",
         ReasoningEffort::Xhigh => "X-High",
+        ReasoningEffort::Max => "Max",
     }
     .to_string()
 }