fix(agent): keep multikey cooldown isolation in candidate keying

pkg/agent/instance.go

@@ -6,6 +6,7 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
+	"strconv"
 	"strings"
 
 	"github.com/sipeed/picoclaw/pkg/config"
@@ -239,7 +240,7 @@ func applyCooldownKeys(cfg *config.Config, candidates []providers.FallbackCandid
 	copy(resolved, candidates)
 
 	for i := range resolved {
-		resolved[i].CooldownKey = resolveCooldownKey(strategyLookup, resolved[i])
+		resolved[i].CooldownKey = resolveCooldownKey(strategyLookup, resolved[i], resolved)
 	}
 
 	return resolved
@@ -267,7 +268,11 @@ func buildCooldownStrategyLookup(cfg *config.Config) map[string]string {
 	return strategyLookup
 }
 
-func resolveCooldownKey(strategyLookup map[string]string, candidate providers.FallbackCandidate) string {
+func resolveCooldownKey(
+	strategyLookup map[string]string,
+	candidate providers.FallbackCandidate,
+	candidates []providers.FallbackCandidate,
+) string {
 	if strings.TrimSpace(candidate.Provider) == "" {
 		if strings.TrimSpace(candidate.Model) == "" {
 			return ""
@@ -276,6 +281,9 @@ func resolveCooldownKey(strategyLookup map[string]string, candidate providers.Fa
 	}
 
 	candidateKey := providers.ModelKey(candidate.Provider, candidate.Model)
+	if belongsToMultiKeySet(candidate, candidates) {
+		return candidateKey
+	}
 	if strategyLookup[candidateKey] == "model" {
 		return candidateKey
 	}
@@ -283,6 +291,46 @@ func resolveCooldownKey(strategyLookup map[string]string, candidate providers.Fa
 	return candidate.Provider
 }
 
+func belongsToMultiKeySet(
+	candidate providers.FallbackCandidate,
+	candidates []providers.FallbackCandidate,
+) bool {
+	provider := providers.NormalizeProvider(candidate.Provider)
+	model := strings.ToLower(strings.TrimSpace(candidate.Model))
+	if provider == "" || model == "" {
+		return false
+	}
+
+	baseModel, isReplica := multiKeyBaseModel(model)
+	if isReplica {
+		return true
+	}
+
+	for _, other := range candidates {
+		if providers.NormalizeProvider(other.Provider) != provider {
+			continue
+		}
+		otherBase, otherIsReplica := multiKeyBaseModel(other.Model)
+		if otherIsReplica && otherBase == baseModel {
+			return true
+		}
+	}
+
+	return false
+}
+
+func multiKeyBaseModel(model string) (string, bool) {
+	normalized := strings.ToLower(strings.TrimSpace(model))
+	idx := strings.LastIndex(normalized, "__key_")
+	if idx <= 0 {
+		return normalized, false
+	}
+	if _, err := strconv.Atoi(normalized[idx+len("__key_"):]); err != nil {
+		return normalized, false
+	}
+	return normalized[:idx], true
+}
+
 // resolveAgentWorkspace determines the workspace directory for an agent.
 func resolveAgentWorkspace(agentCfg *config.AgentConfig, defaults *config.AgentDefaults) string {
 	if agentCfg != nil && strings.TrimSpace(agentCfg.Workspace) != "" {

pkg/agent/instance_test.go

@@ -172,11 +172,11 @@ func TestNewAgentInstance_ResolvePerModelCooldownKeys(t *testing.T) {
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
 				Workspace:      tmpDir,
-				Model:          "router-a",
+				ModelName:      "router-a",
 				ModelFallbacks: []string{"router-b", "shared-provider"},
 			},
 		},
-		ModelList: []config.ModelConfig{
+		ModelList: []*config.ModelConfig{
 			{
 				ModelName:        "router-a",
 				Model:            "litellm/openai/gpt-4o-mini",
@@ -218,7 +218,7 @@ func TestNewAgentInstance_ResolveLightModelPerModelCooldownKeys(t *testing.T) {
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
 				Workspace: tmpDir,
-				Model:     "primary-model",
+				ModelName: "primary-model",
 				Routing: &config.RoutingConfig{
 					Enabled:    true,
 					LightModel: "light-model",
@@ -226,7 +226,7 @@ func TestNewAgentInstance_ResolveLightModelPerModelCooldownKeys(t *testing.T) {
 				},
 			},
 		},
-		ModelList: []config.ModelConfig{
+		ModelList: []*config.ModelConfig{
 			{
 				ModelName: "primary-model",
 				Model:     "litellm/openai/gpt-4o-mini",
@@ -255,6 +255,41 @@ func TestNewAgentInstance_ResolveLightModelPerModelCooldownKeys(t *testing.T) {
 	}
 }
 
+func TestNewAgentInstance_ResolveMultiKeyCooldownKeys(t *testing.T) {
+	tmpDir := t.TempDir()
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:      tmpDir,
+				ModelName:      "glm-main",
+				ModelFallbacks: []string{"glm-replica"},
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName: "glm-main",
+				Model:     "zhipu/glm-4.7",
+			},
+			{
+				ModelName: "glm-replica",
+				Model:     "zhipu/glm-4.7__key_1",
+			},
+		},
+	}
+
+	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, &mockProvider{})
+	if len(agent.Candidates) != 2 {
+		t.Fatalf("len(Candidates) = %d, want 2", len(agent.Candidates))
+	}
+	if got := agent.Candidates[0].CooldownKey; got != "zhipu/glm-4.7" {
+		t.Fatalf("candidate[0] cooldown key = %q, want %q", got, "zhipu/glm-4.7")
+	}
+	if got := agent.Candidates[1].CooldownKey; got != "zhipu/glm-4.7__key_1" {
+		t.Fatalf("candidate[1] cooldown key = %q, want %q", got, "zhipu/glm-4.7__key_1")
+	}
+}
+
 func TestNewAgentInstance_AllowsMediaTempDirForReadListAndExec(t *testing.T) {
 	workspace := t.TempDir()
 	mediaDir := media.TempDir()

pkg/config/config.go

@@ -667,10 +667,10 @@ type ModelConfig struct {
 	Workspace   string `json:"workspace,omitempty"`    // Workspace path for CLI-based providers
 
 	// Optional optimizations
-	RPM            int            `json:"rpm,omitempty"`              // Requests per minute limit
-	MaxTokensField string         `json:"max_tokens_field,omitempty"` // Field name for max tokens (e.g., "max_completion_tokens")
-	RequestTimeout int            `json:"request_timeout,omitempty"`
-	ThinkingLevel  string         `json:"thinking_level,omitempty"` // Extended thinking: off|low|medium|high|xhigh|adaptive
+	RPM            int    `json:"rpm,omitempty"`              // Requests per minute limit
+	MaxTokensField string `json:"max_tokens_field,omitempty"` // Field name for max tokens (e.g., "max_completion_tokens")
+	RequestTimeout int    `json:"request_timeout,omitempty"`
+	ThinkingLevel  string `json:"thinking_level,omitempty"` // Extended thinking: off|low|medium|high|xhigh|adaptive
 	// CooldownStrategy controls the scope of cooldown tracking for this model.
 	// - "provider" (default): cooldown is shared across all models in the provider
 	// - "model" or "per-model": cooldown is isolated to this specific model
@@ -696,7 +696,6 @@ func (c *ModelConfig) APIKey() string {
 func (c *ModelConfig) IsVirtual() bool {
 	return c.isVirtual
 }
-}
 
 // Validate checks if the ModelConfig has all required fields.
 func (c *ModelConfig) Validate() error {
@@ -737,6 +736,7 @@ func (c *ModelConfig) SetAPIKey(value string) {
 		c.APIKeys = append(c.APIKeys, NewSecureString(value))
 	}
 }
+
 type GatewayConfig struct {
 	Host      string `json:"host"                env:"PICOCLAW_GATEWAY_HOST"`
 	Port      int    `json:"port"                env:"PICOCLAW_GATEWAY_PORT"`