fix(provider): implement proper provider-per-candidate fallback architecture

This commit fixes a critical bug in the model fallback chain where the
`provider` parameter was being completely ignored, causing all fallback
candidates to use the same provider instance configured for the primary
model. This led to requests being sent to the wrong API endpoints.

## Root Cause

The fallback callback at loop.go:522-527 was using `agent.Provider.Chat()`
for ALL candidates, ignoring the `provider` parameter passed by the fallback
chain. For example:
- Primary: cerebras/gpt-oss-120b → agent.Provider configured for api.cerebras.ai
- Fallback: ollama/llama3.2 → STILL uses cerebras provider → wrong endpoint!

Additionally, the `prompt_cache_key` parameter (an OpenAI-only feature) was
being sent to all providers, causing 422 errors from providers like Cerebras
that don't support this parameter.

## Changes

### Core Architecture Fix
1. **ProviderRegistry** (pkg/providers/registry.go): New registry that creates
   and caches provider instances per provider name (e.g., "cerebras", "ollama")
2. **AgentInstance** (pkg/agent/instance.go): Added ProviderRegistry field
3. **Fallback callback** (pkg/agent/loop.go:522-544): Now gets correct provider
   per candidate via agent.ProviderRegistry.GetProvider(providerName)

### Prompt Cache Fix
1. **Agent loop** (pkg/agent/loop.go): Removed `prompt_cache_key` from generic
   options map (lines 526, 543, 1009, 1060)
2. **OpenAI compat provider** (pkg/providers/openai_compat/provider.go:158-162):
   Changed from blacklist (exclude Gemini) to whitelist (only api.openai.com)

### Test Updates
- Updated tests to pass new ProviderRegistry parameter

## Impact

- Fallback chains with mixed providers (cerebras + ollama + openai) now work
- Providers like Cerebras no longer receive unsupported `prompt_cache_key`
- Each fallback candidate uses its correctly configured provider

Fixes: Model fallback chain sending requests to wrong API endpoints
Related: 422 errors from Cerebras due to unsupported prompt_cache_key

pkg/agent/instance.go

@@ -24,7 +24,8 @@ type AgentInstance struct {
 	MaxTokens        int
 	Temperature      float64
 	ContextWindow    int
-	Provider       providers.LLMProvider
+	Provider         providers.LLMProvider              // Default provider for backward compatibility
+	ProviderRegistry *providers.ProviderRegistry        // Registry for multi-provider fallback
 	Sessions         *session.SessionManager
 	ContextBuilder   *ContextBuilder
 	Tools            *tools.ToolRegistry
@@ -39,6 +40,7 @@ func NewAgentInstance(
 	defaults *config.AgentDefaults,
 	cfg *config.Config,
 	provider providers.LLMProvider,
+	providerRegistry *providers.ProviderRegistry,
 ) *AgentInstance {
 	workspace := resolveAgentWorkspace(agentCfg, defaults)
 	os.MkdirAll(workspace, 0o755)
@@ -109,6 +111,7 @@ func NewAgentInstance(
 		Temperature:      temperature,
 		ContextWindow:    maxTokens,
 		Provider:         provider,
+		ProviderRegistry: providerRegistry,
 		Sessions:         sessionsManager,
 		ContextBuilder:   contextBuilder,
 		Tools:            toolsRegistry,

pkg/agent/instance_resolution_test.go

@@ -133,7 +133,7 @@ func TestNewAgentInstance_ModelResolution(t *testing.T) {
 	// Use the existing mockProvider from mock_provider_test.go
 	provider := &mockProvider{}
 
-	instance := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider)
+	instance := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider, nil)
 
 	require.NotNil(t, instance)
 

pkg/agent/instance_test.go

@@ -29,7 +29,7 @@ func TestNewAgentInstance_UsesDefaultsTemperatureAndMaxTokens(t *testing.T) {
 	cfg.Agents.Defaults.Temperature = &configuredTemp
 
 	provider := &mockProvider{}
-	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider)
+	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider, nil)
 
 	if agent.MaxTokens != 1234 {
 		t.Fatalf("MaxTokens = %d, want %d", agent.MaxTokens, 1234)
@@ -61,7 +61,7 @@ func TestNewAgentInstance_DefaultsTemperatureWhenZero(t *testing.T) {
 	cfg.Agents.Defaults.Temperature = &configuredTemp
 
 	provider := &mockProvider{}
-	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider)
+	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider, nil)
 
 	if agent.Temperature != 0.0 {
 		t.Fatalf("Temperature = %f, want %f", agent.Temperature, 0.0)
@@ -87,7 +87,7 @@ func TestNewAgentInstance_DefaultsTemperatureWhenUnset(t *testing.T) {
 	}
 
 	provider := &mockProvider{}
-	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider)
+	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider, nil)
 
 	if agent.Temperature != 0.7 {
 		t.Fatalf("Temperature = %f, want %f", agent.Temperature, 0.7)

pkg/agent/loop.go

@@ -519,11 +519,26 @@ func (al *AgentLoop) runLLMIteration(
 		callLLM := func() (*providers.LLMResponse, error) {
 			if len(agent.Candidates) > 1 && al.fallback != nil {
 				fbResult, fbErr := al.fallback.Execute(ctx, agent.Candidates,
-					func(ctx context.Context, provider, model string) (*providers.LLMResponse, error) {
-						return agent.Provider.Chat(ctx, messages, providerToolDefs, model, map[string]any{
+					func(ctx context.Context, providerName, model string) (*providers.LLMResponse, error) {
+						// Get the correct provider for this candidate
+						// This is critical: different candidates may use different providers
+						// (e.g., cerebras, ollama, openai), and each requires its own
+						// provider instance with the correct API endpoint configuration.
+						var provider providers.LLMProvider
+						if agent.ProviderRegistry != nil {
+							p, err := agent.ProviderRegistry.GetProvider(providerName)
+							if err != nil {
+								return nil, fmt.Errorf("failed to get provider for %q: %w", providerName, err)
+							}
+							provider = p
+						} else {
+							// Fallback to default provider if registry not available
+							// (this shouldn't happen in normal operation)
+							provider = agent.Provider
+						}
+						return provider.Chat(ctx, messages, providerToolDefs, model, map[string]any{
 							"max_tokens":  agent.MaxTokens,
 							"temperature": agent.Temperature,
-							"prompt_cache_key": agent.ID,
 						})
 					},
 				)
@@ -540,7 +555,6 @@ func (al *AgentLoop) runLLMIteration(
 			return agent.Provider.Chat(ctx, messages, providerToolDefs, agent.Model, map[string]any{
 				"max_tokens":  agent.MaxTokens,
 				"temperature": agent.Temperature,
-				"prompt_cache_key": agent.ID,
 			})
 		}
 
@@ -1006,7 +1020,6 @@ func (al *AgentLoop) summarizeSession(agent *AgentInstance, sessionKey string) {
 			map[string]any{
 				"max_tokens":  1024,
 				"temperature": 0.3,
-				"prompt_cache_key": agent.ID,
 			},
 		)
 		if err == nil {
@@ -1057,7 +1070,6 @@ func (al *AgentLoop) summarizeBatch(
 		map[string]any{
 			"max_tokens":  1024,
 			"temperature": 0.3,
-			"prompt_cache_key": agent.ID,
 		},
 	)
 	if err != nil {

pkg/agent/registry.go

@@ -26,20 +26,23 @@ func NewAgentRegistry(
 		resolver: routing.NewRouteResolver(cfg),
 	}
 
+	// Create provider registry from config for proper fallback support
+	providerRegistry := providers.NewProviderRegistry(cfg)
+
 	agentConfigs := cfg.Agents.List
 	if len(agentConfigs) == 0 {
 		implicitAgent := &config.AgentConfig{
 			ID:      "main",
 			Default: true,
 		}
-		instance := NewAgentInstance(implicitAgent, &cfg.Agents.Defaults, cfg, provider)
+		instance := NewAgentInstance(implicitAgent, &cfg.Agents.Defaults, cfg, provider, providerRegistry)
 		registry.agents["main"] = instance
 		logger.InfoCF("agent", "Created implicit main agent (no agents.list configured)", nil)
 	} else {
 		for i := range agentConfigs {
 			ac := &agentConfigs[i]
 			id := routing.NormalizeAgentID(ac.ID)
-			instance := NewAgentInstance(ac, &cfg.Agents.Defaults, cfg, provider)
+			instance := NewAgentInstance(ac, &cfg.Agents.Defaults, cfg, provider, providerRegistry)
 			registry.agents[id] = instance
 			logger.InfoCF("agent", "Registered agent",
 				map[string]any{

pkg/providers/openai_compat/provider.go

@@ -153,10 +153,11 @@ func (p *Provider) Chat(
 	// with the same key and reuse prefix KV cache across calls.
 	// The key is typically the agent ID — stable per agent, shared across requests.
 	// See: https://platform.openai.com/docs/guides/prompt-caching
-	// Prompt caching is only supported by OpenAI-native endpoints.
-	// Gemini and other providers reject unknown fields, so skip for non-OpenAI APIs.
+	// IMPORTANT: Prompt caching is ONLY supported by OpenAI-native endpoints (api.openai.com).
+	// Gemini, Cerebras, and other providers reject unknown fields, so we must skip them.
 	if cacheKey, ok := options["prompt_cache_key"].(string); ok && cacheKey != "" {
-		if !strings.Contains(p.apiBase, "generativelanguage.googleapis.com") {
+		// Only send prompt_cache_key to actual OpenAI endpoints
+		if strings.Contains(p.apiBase, "api.openai.com") {
 			requestBody["prompt_cache_key"] = cacheKey
 		}
 	}

pkg/providers/registry.go

@@ -0,0 +1,100 @@
+// PicoClaw - Ultra-lightweight personal AI agent
+// License: MIT
+//
+// Copyright (c) 2026 PicoClaw contributors
+
+package providers
+
+import (
+	"fmt"
+	"strings"
+	"sync"
+
+	"github.com/sipeed/picoclaw/pkg/config"
+)
+
+// ProviderRegistry manages provider instances for different provider names.
+// It lazily creates providers on-demand and caches them for reuse.
+// This is essential for proper fallback behavior where different candidates
+// may use different providers (e.g., cerebras, ollama, openai).
+type ProviderRegistry struct {
+	cfg       *config.Config
+	modelList []config.ModelConfig
+	providers map[string]LLMProvider
+	mu        sync.RWMutex
+}
+
+// NewProviderRegistry creates a new provider registry with the given config.
+func NewProviderRegistry(cfg *config.Config) *ProviderRegistry {
+	var modelList []config.ModelConfig
+	if cfg != nil {
+		modelList = cfg.ModelList
+	}
+	return &ProviderRegistry{
+		cfg:       cfg,
+		modelList: modelList,
+		providers: make(map[string]LLMProvider),
+	}
+}
+
+// GetProvider returns a provider for the given provider name (e.g., "openai", "cerebras", "ollama").
+// If the provider has already been created, it returns the cached instance.
+// Otherwise, it creates a new provider from the model list config.
+func (pr *ProviderRegistry) GetProvider(providerName string) (LLMProvider, error) {
+	// Normalize provider name (lowercase)
+	providerName = strings.ToLower(providerName)
+
+	// Check cache first
+	pr.mu.RLock()
+	if provider, ok := pr.providers[providerName]; ok {
+		pr.mu.RUnlock()
+		return provider, nil
+	}
+	pr.mu.RUnlock()
+
+	// Find the model config for this provider
+	var modelCfg *config.ModelConfig
+	for i := range pr.modelList {
+		cfg := &pr.modelList[i]
+		// Extract provider from model string (e.g., "cerebras/gpt-oss-120b" -> "cerebras")
+		protocol, _ := ExtractProtocol(cfg.Model)
+		if strings.EqualFold(protocol, providerName) {
+			modelCfg = cfg
+			break
+		}
+	}
+
+	// If not found in model list, try creating with protocol-only config
+	if modelCfg == nil {
+		modelCfg = &config.ModelConfig{
+			Model: providerName + "/dummy", // Protocol is what matters
+		}
+	}
+
+	// Create the provider
+	provider, _, err := CreateProviderFromConfig(modelCfg)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create provider for %q: %w", providerName, err)
+	}
+
+	// Cache the provider
+	pr.mu.Lock()
+	pr.providers[providerName] = provider
+	pr.mu.Unlock()
+
+	return provider, nil
+}
+
+// GetDefaultProvider returns the provider for the default/primary protocol.
+// This is used for backward compatibility with code that expects a single provider.
+func (pr *ProviderRegistry) GetDefaultProvider() (LLMProvider, error) {
+	if len(pr.modelList) == 0 {
+		// No model list configured, return error
+		// Caller should handle this by using a default provider
+		return nil, fmt.Errorf("no model list configured")
+	}
+
+	// Use the first model's provider as default
+	protocol, _ := ExtractProtocol(pr.modelList[0].Model)
+	return pr.GetProvider(protocol)
+}