fix(seed): default rules try the model made for the effort first

CHANGELOG.md

@@ -7,6 +7,12 @@ All notable changes to KTMS are listed here. The format follows
 
 ## [Unreleased]
 
+### Fixed
+- The default rules of the auto-seed and `ktms init` try the model made for the effort first: on
+  high, `gemini-2.5-pro` and `claude-sonnet-5-5` before `gpt-4o`, and `gpt-4o` before
+  `gpt-4o-mini`; on medium, `gpt-4o` before the low models. They tried the models in list order,
+  so high went to `gpt-4o-mini` first. A Keystone seeded already keeps its rules.
+
 ## [1.3.3] - 2026-10-26
 
 ### Added

internal/seed/seed.go

@@ -407,6 +407,13 @@ func DefaultRules(models []*proxydb.Model, activeVendors []string) []*proxydb.Ru
 				fallback = append(fallback, m.ID)
 			}
 		}
+		// The model made for this effort comes first: on high a model that
+		// serves only high is tried before one that also serves low.
+		byFit := func(ids []string) {
+			slices.SortStableFunc(ids, func(a, b string) int { return lowestEffort(models, b) - lowestEffort(models, a) })
+		}
+		byFit(active)
+		byFit(fallback)
 		if len(active) > 0 {
 			return append(active, fallback...)
 		}
@@ -466,6 +473,23 @@ func DefaultRules(models []*proxydb.Model, activeVendors []string) []*proxydb.Ru
 	}
 }
 
+// lowestEffort ranks a model by the lowest effort it serves: 0 low, 1 medium, 2 high.
+func lowestEffort(models []*proxydb.Model, id string) int {
+	rank := map[string]int{registry.EffortLow: 0, registry.EffortMedium: 1, registry.EffortHigh: 2}
+	lowest := 2
+	for _, m := range models {
+		if m.ID != id {
+			continue
+		}
+		for _, e := range m.Efforts {
+			if r, ok := rank[e]; ok && r < lowest {
+				lowest = r
+			}
+		}
+	}
+	return lowest
+}
+
 func serves(m *proxydb.Model, effort string) bool {
 	if registry.HasCapability(m, registry.CapabilityEmbedding) {
 		return false

internal/seed/seed_test.go

@@ -1,6 +1,7 @@
 package seed
 
 import (
+	"slices"
 	"testing"
 
 	"github.com/ai-process/ktms/internal/registry"
@@ -78,3 +79,27 @@ func TestDefaultRulesChainOrder(t *testing.T) {
 		}
 	}
 }
+
+func TestDefaultRulesPutTheModelMadeForTheEffortFirst(t *testing.T) {
+	chains := map[string][]string{}
+	for _, r := range DefaultRules(DefaultModels(), []string{registry.VendorOpenAI}) {
+		chains[r.Name] = r.Use
+	}
+	want := map[string][]string{
+		"default-low":    {"gpt-4o-mini"},
+		"default-medium": {"gpt-4o", "gpt-4o-mini"},
+		"default-high":   {"gpt-4o", "gpt-4o-mini"},
+	}
+	for name, w := range want {
+		if got := chains[name]; !slices.Equal(got[:len(w)], w) {
+			t.Errorf("%s starts with %v, want %v", name, got, w)
+		}
+	}
+
+	// With every vendor, high tries the high-only models before the rest.
+	for _, r := range DefaultRules(DefaultModels(), []string{registry.VendorOpenAI, registry.VendorGoogle, registry.VendorAnthropic}) {
+		if r.Name == "default-high" && !slices.Equal(r.Use[:3], []string{"gemini-2.5-pro", "claude-sonnet-5-5", "gpt-4o"}) {
+			t.Errorf("default-high = %v", r.Use)
+		}
+	}
+}