fix(routing): 对齐缓存亲和力硬规则与审计

2026-07-24 23:51:27 +08:00
parent d7a0ec56f5
commit 152885bbb6
10 changed files with 540 additions and 94 deletions
@@ -1746,6 +1746,7 @@ WHERE m.platform_id = $1::uuid
 	}

 	restartModel := "worker-restart-" + suffixText
+	createSimulationTextBaseModel(t, server.URL, loginResponse.AccessToken, restartModel)
 	createSimulationTextPlatformModel(
 		t,
 		server.URL,
@@ -1864,6 +1865,7 @@ func TestCacheAffinityRoutingFlow(t *testing.T) {

 	model := "cache-affinity-smoke-" + suffixText
 	cacheKey := "cache-affinity-key-" + suffixText
+	createSimulationTextBaseModel(t, server.URL, loginResponse.AccessToken, model)
 	lowPlatform, lowPlatformModelID := createSimulationTextPlatformModel(t, server.URL, loginResponse.AccessToken, "cache-low-"+suffixText, "Cache Affinity Low Priority", model, 20, map[string]any{
 		"rules": []map[string]any{
 			{"metric": "concurrent", "limit": 1, "leaseTtlSeconds": 120},
@@ -1880,8 +1882,31 @@ func TestCacheAffinityRoutingFlow(t *testing.T) {
 		}
 	}

-	highPlatform, _ := createSimulationTextPlatformModel(t, server.URL, loginResponse.AccessToken, "cache-high-"+suffixText, "Cache Affinity High Priority", model, 1, nil)
-	_ = highPlatform
+	tools := []any{map[string]any{
+		"type": "function",
+		"function": map[string]any{
+			"name":        "lookup_weather",
+			"description": "Look up current weather",
+			"parameters":  map[string]any{"type": "object", "properties": map[string]any{"city": map[string]any{"type": "string"}}},
+		},
+	}}
+	initialToolMessages := []any{
+		map[string]any{"role": "system", "content": "Use tools when helpful."},
+		map[string]any{"role": "user", "content": "Weather in Hangzhou?"},
+	}
+	for index := 0; index < 3; index++ {
+		detail := runCacheAffinityChatTaskWithBody(t, ctx, testPool, server.URL, apiKeyResponse.Secret, model, "tool-prime-"+strconv.Itoa(index)+"-"+suffixText, map[string]any{
+			"tools":           tools,
+			"messages":        initialToolMessages,
+			"simulationUsage": map[string]any{"inputTokens": 1000, "cachedInputTokens": 750, "outputTokens": 20},
+		})
+		if detail.Status != "succeeded" || len(detail.Attempts) != 1 || detail.Attempts[0].PlatformName != "Cache Affinity Low Priority" {
+			t.Fatalf("tool-history priming request should use the only platform, got %+v", detail)
+		}
+	}
+
+	peerPlatform, _ := createSimulationTextPlatformModel(t, server.URL, loginResponse.AccessToken, "cache-peer-"+suffixText, "Cache Affinity Peer", model, 20, nil)
+	_ = peerPlatform
 	sameKeyDetail := runCacheAffinityChatTask(t, ctx, testPool, server.URL, apiKeyResponse.Secret, model, cacheKey, "same-key-"+suffixText, map[string]any{
 		"inputTokens":       1000,
 		"cachedInputTokens": 700,
@@ -1899,12 +1924,36 @@ func TestCacheAffinityRoutingFlow(t *testing.T) {
 		t.Fatalf("same cache key task detail should expose simulated usage, got %+v", sameKeyDetail.Usage)
 	}

+	toolHistoryDetail := runCacheAffinityChatTaskWithBody(t, ctx, testPool, server.URL, apiKeyResponse.Secret, model, "tool-history-"+suffixText, map[string]any{
+		"tools": tools,
+		"messages": append(append([]any{}, initialToolMessages...),
+			map[string]any{"role": "assistant", "tool_calls": []any{map[string]any{
+				"id":   "call_weather",
+				"type": "function",
+				"function": map[string]any{
+					"name":      "lookup_weather",
+					"arguments": `{"city":"Hangzhou"}`,
+				},
+			}}},
+			map[string]any{"role": "tool", "tool_call_id": "call_weather", "content": `{"temperature":31}`},
+		),
+		"simulationUsage": map[string]any{"inputTokens": 1400, "cachedInputTokens": 900, "outputTokens": 30},
+	})
+	if toolHistoryDetail.Status != "succeeded" || len(toolHistoryDetail.Attempts) != 1 || toolHistoryDetail.Attempts[0].PlatformName != "Cache Affinity Low Priority" {
+		t.Fatalf("tool continuation should retain the cached platform, got %+v", toolHistoryDetail)
+	}
+	if !boolFromTestMap(toolHistoryDetail.Attempts[0].Metrics, "cacheAffinityMatched") ||
+		intFromTestAny(toolHistoryDetail.Attempts[0].Metrics["cacheAffinityMatchedPrefixDepth"]) < 2 ||
+		intFromTestAny(toolHistoryDetail.Attempts[0].Metrics["cacheAffinityCandidateCount"]) < 1 {
+		t.Fatalf("tool continuation should expose prefix and candidate metrics, got %+v", toolHistoryDetail.Attempts[0].Metrics)
+	}
+
 	newKeyDetail := runCacheAffinityChatTask(t, ctx, testPool, server.URL, apiKeyResponse.Secret, model, "cache-affinity-new-"+suffixText, "new-key-"+suffixText, map[string]any{
 		"inputTokens":       1000,
 		"cachedInputTokens": 0,
 		"outputTokens":      20,
 	})
-	if newKeyDetail.Status != "succeeded" || len(newKeyDetail.Attempts) != 1 || newKeyDetail.Attempts[0].PlatformName != "Cache Affinity High Priority" {
+	if newKeyDetail.Status != "succeeded" || len(newKeyDetail.Attempts) != 1 || newKeyDetail.Attempts[0].PlatformName != "Cache Affinity Peer" {
 		t.Fatalf("new cache key should fall back to base priority, got %+v", newKeyDetail)
 	}

@@ -1914,9 +1963,12 @@ func TestCacheAffinityRoutingFlow(t *testing.T) {
 		"cachedInputTokens": 0,
 		"outputTokens":      20,
 	})
-	if fullAvoidedDetail.Status != "succeeded" || len(fullAvoidedDetail.Attempts) != 1 || fullAvoidedDetail.Attempts[0].PlatformName != "Cache Affinity High Priority" {
+	if fullAvoidedDetail.Status != "succeeded" || len(fullAvoidedDetail.Attempts) != 1 || fullAvoidedDetail.Attempts[0].PlatformName != "Cache Affinity Peer" {
 		t.Fatalf("full cached platform should be avoided before cache affinity boost, got %+v", fullAvoidedDetail)
 	}
+	if fullAvoidedDetail.Attempts[0].Metrics["cacheAffinityOverrideReason"] != "capacity_tier_unavailable" {
+		t.Fatalf("full cached platform override reason should be visible, got %+v", fullAvoidedDetail.Attempts[0].Metrics)
+	}

 	var requestCount int
 	var inputTokens int
@@ -1926,7 +1978,7 @@ func TestCacheAffinityRoutingFlow(t *testing.T) {
 	SELECT request_count::int, input_tokens::int, cached_input_tokens::int, ema_hit_ratio::float8
 	FROM gateway_cache_affinity_stats
 	WHERE client_id = $1
-	ORDER BY updated_at DESC
+	ORDER BY request_count DESC, cached_input_tokens DESC, updated_at DESC
 	LIMIT 1`, lowPlatform.PlatformKey+":text_generate:"+model).Scan(&requestCount, &inputTokens, &cachedTokens, &emaHitRatio); err != nil {
 		t.Fatalf("read cache affinity stats: %v", err)
 	}
@@ -2214,6 +2266,20 @@ type cacheAffinityTaskDetail struct {
 	} `json:"attempts"`
 }

+func createSimulationTextBaseModel(t *testing.T, baseURL string, adminToken string, model string) {
+	t.Helper()
+	doJSON(t, baseURL, http.MethodPost, "/api/admin/catalog/base-models", adminToken, map[string]any{
+		"providerKey":       "openai",
+		"canonicalModelKey": "openai:" + model,
+		"invocationName":    model,
+		"providerModelName": model,
+		"modelType":         []string{"text_generate"},
+		"displayName":       model,
+		"capabilities":      map[string]any{"originalTypes": []string{"text_generate"}},
+		"metadata":          map[string]any{"source": "cache-affinity-integration-test"},
+	}, http.StatusCreated, nil)
+}
+
 func createSimulationTextPlatformModel(t *testing.T, baseURL string, adminToken string, platformKey string, platformName string, model string, priority int, rateLimitPolicy map[string]any) (cacheAffinityPlatformFixture, string) {
 	t.Helper()
 	var platform cacheAffinityPlatformFixture
@@ -2230,7 +2296,7 @@ func createSimulationTextPlatformModel(t *testing.T, baseURL string, adminToken
 		t.Fatalf("cache affinity platform was not created: %+v", platform)
 	}
 	payload := map[string]any{
-		"canonicalModelKey": "openai:gpt-4o-mini",
+		"canonicalModelKey": "openai:" + model,
 		"modelName":         model,
 		"providerModelName": model,
 		"modelAlias":        model,
@@ -2251,17 +2317,27 @@ func createSimulationTextPlatformModel(t *testing.T, baseURL string, adminToken
 }

 func runCacheAffinityChatTask(t *testing.T, ctx context.Context, pool *pgxpool.Pool, baseURL string, token string, model string, cacheAffinityKey string, marker string, simulationUsage map[string]any) cacheAffinityTaskDetail {
+	t.Helper()
+	return runCacheAffinityChatTaskWithBody(t, ctx, pool, baseURL, token, model, marker, map[string]any{
+		"cacheAffinityKey": cacheAffinityKey,
+		"simulationUsage":  simulationUsage,
+		"messages":         []any{map[string]any{"role": "user", "content": "cache affinity route"}},
+	})
+}
+
+func runCacheAffinityChatTaskWithBody(t *testing.T, ctx context.Context, pool *pgxpool.Pool, baseURL string, token string, model string, marker string, body map[string]any) cacheAffinityTaskDetail {
 	t.Helper()
 	var detail cacheAffinityTaskDetail
-	doAPIV1ChatCompletionAndLoadTask(t, ctx, pool, baseURL, token, map[string]any{
+	requestBody := map[string]any{
 		"model":                model,
 		"runMode":              "simulation",
 		"simulation":           true,
 		"simulationDurationMs": 5,
-		"cacheAffinityKey":     cacheAffinityKey,
-		"simulationUsage":      simulationUsage,
-		"messages":             []map[string]any{{"role": "user", "content": "cache affinity route"}},
-	}, marker, http.StatusOK, nil, &detail)
+	}
+	for key, value := range body {
+		requestBody[key] = value
+	}
+	doAPIV1ChatCompletionAndLoadTask(t, ctx, pool, baseURL, token, requestBody, marker, http.StatusOK, nil, &detail)
 	return detail
 }

@@ -13,6 +13,8 @@ type cacheAffinityKeys struct {
 	Record  []string
 }

+const cacheAffinityMinimumStableMessages = 2
+
 func buildCacheAffinityKey(kind string, modelType string, body map[string]any) string {
 	return buildCacheAffinityKeys(kind, modelType, body).Primary
 }
@@ -37,48 +39,28 @@ func buildCacheAffinityKeys(kind string, modelType string, body map[string]any)
 		return cacheAffinityKeys{}
 	}
 	basePayload := map[string]any{
-		"kind":      kind,
-		"modelType": modelType,
+		"version": "v2",
 	}
 	if tools, ok := body["tools"]; ok {
 		basePayload["tools"] = tools
 	}
-	if instructions, ok := body["instructions"]; ok {
-		basePayload["instructions"] = instructions
-	}
 	messageHashes := cacheAffinityMessageHashes(body)
-	if len(messageHashes) > 0 {
+	if len(messageHashes) >= cacheAffinityMinimumStableMessages {
 		baseHash := stableJSONHash(basePayload)
-		keys := make([]string, 0, len(messageHashes))
-		for index := range messageHashes {
-			keyPayload := map[string]any{
-				"base":          baseHash,
-				"messageHashes": messageHashes[:index+1],
-				"prefixLength":  index + 1,
-			}
-			keys = append(keys, "prompt_lcp:"+stableJSONHash(keyPayload))
-		}
+		keys := cacheAffinityPrefixKeys("prompt_lcp_v2:", baseHash, messageHashes)
 		if len(keys) == 0 {
 			return cacheAffinityKeys{}
 		}
-		lookup := make([]string, 0, len(keys))
-		for index := len(keys) - 1; index >= 0; index-- {
-			lookup = append(lookup, keys[index])
-		}
+		lookup := reverseCacheAffinityKeys(keys)
+		legacyKeys := legacyCacheAffinityPrefixKeys(kind, modelType, body)
+		lookup = normalizedCacheAffinityKeyList(append(lookup, reverseCacheAffinityKeys(legacyKeys)...))
 		return cacheAffinityKeys{
 			Primary: keys[len(keys)-1],
 			Lookup:  lookup,
 			Record:  keys,
 		}
 	}
-	if input, ok := body["input"]; ok {
-		basePayload["input"] = input
-	}
-	if len(basePayload) <= 2 {
-		return cacheAffinityKeys{}
-	}
-	key := "prompt:" + stableJSONHash(basePayload)
-	return cacheAffinityKeys{Primary: key, Lookup: []string{key}, Record: []string{key}}
+	return cacheAffinityKeys{}
 }

 func cacheAffinityPromptHashSupported(kind string, modelType string) bool {
@@ -95,6 +77,146 @@ func cacheAffinityPromptHashSupported(kind string, modelType string) bool {
 }

 func cacheAffinityMessageHashes(body map[string]any) []string {
+	if hashes := stringListFromAny(body["messageHashes"]); len(hashes) > 0 {
+		return hashes
+	}
+	if hashes := stringListFromAny(body["message_hashes"]); len(hashes) > 0 {
+		return hashes
+	}
+	messages := cacheAffinityStableMessages(body)
+	if len(messages) == 0 {
+		return nil
+	}
+	hashes := make([]string, 0, len(messages))
+	for _, message := range messages {
+		hash := stableJSONHash(message)
+		if hash == "" {
+			continue
+		}
+		hashes = append(hashes, hash)
+	}
+	return hashes
+}
+
+func cacheAffinityStableMessages(body map[string]any) []any {
+	messages := make([]any, 0)
+	if instructions := strings.TrimSpace(stringFromAny(body["instructions"])); instructions != "" {
+		messages = append(messages, map[string]any{"role": "system", "content": instructions})
+	}
+	if rawMessages, ok := body["messages"].([]any); ok {
+		for _, message := range rawMessages {
+			if normalized := normalizeCacheAffinityMessage(message); normalized != nil {
+				messages = append(messages, normalized)
+			}
+		}
+		return messages
+	}
+	switch input := body["input"].(type) {
+	case string:
+		if strings.TrimSpace(input) != "" {
+			messages = append(messages, map[string]any{"role": "user", "content": input})
+		}
+	case []any:
+		for _, message := range input {
+			if normalized := normalizeCacheAffinityMessage(message); normalized != nil {
+				messages = append(messages, normalized)
+			}
+		}
+	}
+	return messages
+}
+
+func normalizeCacheAffinityMessage(value any) any {
+	message, ok := value.(map[string]any)
+	if !ok {
+		if text := strings.TrimSpace(stringFromAny(value)); text != "" {
+			return map[string]any{"role": "user", "content": text}
+		}
+		return nil
+	}
+	messageType := strings.TrimSpace(stringFromAny(message["type"]))
+	switch messageType {
+	case "function_call":
+		return map[string]any{
+			"role": "assistant",
+			"tool_calls": []any{map[string]any{
+				"id":   firstNonEmptyCacheAffinityString(message["call_id"], message["id"]),
+				"type": "function",
+				"function": map[string]any{
+					"name":      message["name"],
+					"arguments": message["arguments"],
+				},
+			}},
+		}
+	case "function_call_output":
+		return map[string]any{
+			"role":         "tool",
+			"tool_call_id": firstNonEmptyCacheAffinityString(message["call_id"], message["id"]),
+			"content":      message["output"],
+		}
+	}
+	normalized := make(map[string]any, len(message))
+	for key, item := range message {
+		switch key {
+		case "id", "status":
+			continue
+		case "type":
+			if messageType == "message" {
+				continue
+			}
+		}
+		normalized[key] = item
+	}
+	if role := strings.TrimSpace(stringFromAny(normalized["role"])); role != "" {
+		normalized["role"] = role
+	}
+	return normalized
+}
+
+func firstNonEmptyCacheAffinityString(values ...any) string {
+	for _, value := range values {
+		if text := strings.TrimSpace(stringFromAny(value)); text != "" {
+			return text
+		}
+	}
+	return ""
+}
+
+func cacheAffinityPrefixKeys(prefix string, baseHash string, messageHashes []string) []string {
+	if len(messageHashes) < cacheAffinityMinimumStableMessages {
+		return nil
+	}
+	keys := make([]string, 0, len(messageHashes)-cacheAffinityMinimumStableMessages+1)
+	for prefixLength := cacheAffinityMinimumStableMessages; prefixLength <= len(messageHashes); prefixLength++ {
+		keyPayload := map[string]any{
+			"base":          baseHash,
+			"messageHashes": messageHashes[:prefixLength],
+			"prefixLength":  prefixLength,
+		}
+		keys = append(keys, prefix+stableJSONHash(keyPayload))
+	}
+	return keys
+}
+
+func legacyCacheAffinityPrefixKeys(kind string, modelType string, body map[string]any) []string {
+	messageHashes := legacyCacheAffinityMessageHashes(body)
+	if len(messageHashes) < cacheAffinityMinimumStableMessages {
+		return nil
+	}
+	basePayload := map[string]any{
+		"kind":      kind,
+		"modelType": modelType,
+	}
+	if tools, ok := body["tools"]; ok {
+		basePayload["tools"] = tools
+	}
+	if instructions, ok := body["instructions"]; ok {
+		basePayload["instructions"] = instructions
+	}
+	return cacheAffinityPrefixKeys("prompt_lcp:", stableJSONHash(basePayload), messageHashes)
+}
+
+func legacyCacheAffinityMessageHashes(body map[string]any) []string {
 	if hashes := stringListFromAny(body["messageHashes"]); len(hashes) > 0 {
 		return hashes
 	}
@@ -102,20 +224,41 @@ func cacheAffinityMessageHashes(body map[string]any) []string {
 		return hashes
 	}
 	messages, ok := body["messages"].([]any)
-	if !ok || len(messages) == 0 {
+	if !ok {
 		return nil
 	}
 	hashes := make([]string, 0, len(messages))
 	for _, message := range messages {
 		raw, err := json.Marshal(message)
-		if err != nil || len(raw) == 0 {
-			continue
+		if err == nil && len(raw) > 0 {
+			hashes = append(hashes, sha256Text(string(raw)))
 		}
-		hashes = append(hashes, sha256Text(string(raw)))
 	}
 	return hashes
 }

+func reverseCacheAffinityKeys(keys []string) []string {
+	reversed := make([]string, 0, len(keys))
+	for index := len(keys) - 1; index >= 0; index-- {
+		reversed = append(reversed, keys[index])
+	}
+	return reversed
+}
+
+func normalizedCacheAffinityKeyList(keys []string) []string {
+	seen := map[string]bool{}
+	out := make([]string, 0, len(keys))
+	for _, key := range keys {
+		key = strings.TrimSpace(key)
+		if key == "" || seen[key] {
+			continue
+		}
+		seen[key] = true
+		out = append(out, key)
+	}
+	return out
+}
+
 func stableJSONHash(value any) string {
 	raw, err := json.Marshal(value)
 	if err != nil || len(raw) == 0 {
@@ -39,7 +39,7 @@ func TestBuildCacheAffinityKeyFallsBackToPromptHash(t *testing.T) {

 	first := buildCacheAffinityKey("chat.completions", "text_generate", body)
 	second := buildCacheAffinityKey("chat.completions", "text_generate", body)
-	if first == "" || first != second || !strings.HasPrefix(first, "prompt_lcp:") {
+	if first == "" || first != second || !strings.HasPrefix(first, "prompt_lcp_v2:") {
 		t.Fatalf("expected stable prompt LCP cache affinity key, first=%q second=%q", first, second)
 	}
 }
@@ -54,10 +54,10 @@ func TestBuildCacheAffinityKeysUsesMessagePrefixLCPKeys(t *testing.T) {
 		"tools":         []any{map[string]any{"type": "function", "function": map[string]any{"name": "lookup"}}},
 	})

-	if !strings.HasPrefix(first.Primary, "prompt_lcp:") {
+	if !strings.HasPrefix(first.Primary, "prompt_lcp_v2:") {
 		t.Fatalf("expected LCP prompt key, got %+v", first)
 	}
-	if len(first.Record) != 2 || len(next.Lookup) != 4 {
+	if len(first.Record) != 1 || len(next.Record) != 3 || len(next.Lookup) != 6 {
 		t.Fatalf("expected per-message prefix keys, first=%+v next=%+v", first, next)
 	}
 	found := false
@@ -71,3 +71,81 @@ func TestBuildCacheAffinityKeysUsesMessagePrefixLCPKeys(t *testing.T) {
 		t.Fatalf("next lookup keys should include first request prefix key, first=%+v next=%+v", first, next)
 	}
 }
+
+func TestBuildCacheAffinityKeysRequiresTwoStableMessages(t *testing.T) {
+	systemOnly := buildCacheAffinityKeys("chat.completions", "text_generate", map[string]any{
+		"messages": []any{
+			map[string]any{"role": "system", "content": "shared system prompt"},
+		},
+	})
+	firstConversation := buildCacheAffinityKeys("chat.completions", "text_generate", map[string]any{
+		"messages": []any{
+			map[string]any{"role": "system", "content": "shared system prompt"},
+			map[string]any{"role": "user", "content": "conversation A"},
+		},
+	})
+	secondConversation := buildCacheAffinityKeys("chat.completions", "text_generate", map[string]any{
+		"messages": []any{
+			map[string]any{"role": "system", "content": "shared system prompt"},
+			map[string]any{"role": "user", "content": "conversation B"},
+		},
+	})
+
+	if systemOnly.Primary != "" || len(systemOnly.Lookup) != 0 || len(systemOnly.Record) != 0 {
+		t.Fatalf("shared system prompt alone must not create affinity, got %+v", systemOnly)
+	}
+	if firstConversation.Primary == "" || secondConversation.Primary == "" {
+		t.Fatalf("two stable messages should create affinity, first=%+v second=%+v", firstConversation, secondConversation)
+	}
+	for _, firstKey := range firstConversation.Lookup {
+		for _, secondKey := range secondConversation.Lookup {
+			if firstKey == secondKey {
+				t.Fatalf("different conversations sharing only a system prompt must not intersect: key=%s", firstKey)
+			}
+		}
+	}
+}
+
+func TestBuildCacheAffinityKeysNormalizesResponsesToolHistory(t *testing.T) {
+	tools := []any{map[string]any{
+		"type": "function",
+		"function": map[string]any{
+			"name": "lookup_weather",
+		},
+	}}
+	chat := buildCacheAffinityKeys("chat.completions", "text_generate", map[string]any{
+		"tools": tools,
+		"messages": []any{
+			map[string]any{"role": "system", "content": "Use tools when helpful."},
+			map[string]any{"role": "user", "content": "Weather in Hangzhou?"},
+		},
+	})
+	responses := buildCacheAffinityKeys("responses", "text_generate", map[string]any{
+		"tools":        tools,
+		"instructions": "Use tools when helpful.",
+		"input":        "Weather in Hangzhou?",
+	})
+	continued := buildCacheAffinityKeys("responses", "text_generate", map[string]any{
+		"tools":        tools,
+		"instructions": "Use tools when helpful.",
+		"input": []any{
+			map[string]any{"role": "user", "content": "Weather in Hangzhou?"},
+			map[string]any{"type": "function_call", "call_id": "call_weather", "name": "lookup_weather", "arguments": `{"city":"Hangzhou"}`},
+			map[string]any{"type": "function_call_output", "call_id": "call_weather", "output": `{"temperature":31}`},
+		},
+	})
+
+	if chat.Primary == "" || chat.Primary != responses.Primary {
+		t.Fatalf("chat and responses should share the V2 protocol-independent prefix, chat=%+v responses=%+v", chat, responses)
+	}
+	found := false
+	for _, key := range continued.Lookup {
+		if key == responses.Primary {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Fatalf("responses tool continuation should retain its stable prefix, initial=%+v continued=%+v", responses, continued)
+	}
+}
@@ -170,6 +170,14 @@ func attemptMetrics(candidate store.RuntimeModelCandidate, attemptNo int, simula
 		metrics["cacheAffinityCachedInputTokens"] = candidate.CacheAffinity.CachedInputTokens
 		metrics["cacheAffinityBoost"] = candidate.CacheAffinity.Boost
 		metrics["cacheAffinityApplied"] = candidate.CacheAffinity.Applied
+		metrics["cacheAffinityMatched"] = candidate.CacheAffinity.Applied
+		metrics["cacheAffinityCandidateCount"] = candidate.CacheAffinity.CandidateCount
+		if candidate.CacheAffinity.MatchedPrefixDepth > 0 {
+			metrics["cacheAffinityMatchedPrefixDepth"] = candidate.CacheAffinity.MatchedPrefixDepth
+		}
+		if candidate.CacheAffinity.OverrideReason != "" {
+			metrics["cacheAffinityOverrideReason"] = candidate.CacheAffinity.OverrideReason
+		}
 	}
 	if attemptNo > 0 {
 		metrics["attempt"] = attemptNo
@@ -13,16 +13,19 @@ func TestAttemptMetricsIncludesCacheAffinityCoefficient(t *testing.T) {
 		PlatformID:       "deepseek-platform",
 		PlatformPriority: 100,
 		CacheAffinity: store.RuntimeCandidateCacheAffinity{
-			Key:               "conversation:test",
-			RequestCount:      2,
-			CachedInputTokens: 7296,
-			EMAHitRatio:       0.91,
-			LastHitRatio:      0.92,
-			Confidence:        1,
-			Score:             1.74,
-			Boost:             20,
-			AdjustedPriority:  80,
-			Applied:           true,
+			Key:                "prompt_lcp_v2:test",
+			RequestCount:       2,
+			CachedInputTokens:  7296,
+			EMAHitRatio:        0.91,
+			LastHitRatio:       0.92,
+			Confidence:         1,
+			Score:              1.74,
+			Boost:              20,
+			AdjustedPriority:   80,
+			MatchedPrefixDepth: 4,
+			CandidateCount:     2,
+			OverrideReason:     "different_effective_priority_tier",
+			Applied:            true,
 		},
 	}, 1, false)

@@ -35,4 +38,10 @@ func TestAttemptMetricsIncludesCacheAffinityCoefficient(t *testing.T) {
 	if metrics["cacheAffinityHitRatio"] != 0.91 || metrics["cacheAffinityLastHitRatio"] != 0.92 {
 		t.Fatalf("expected hit ratios in attempt metrics, got %+v", metrics)
 	}
+	if metrics["cacheAffinityMatched"] != true ||
+		metrics["cacheAffinityMatchedPrefixDepth"] != 4 ||
+		metrics["cacheAffinityCandidateCount"] != 2 ||
+		metrics["cacheAffinityOverrideReason"] != "different_effective_priority_tier" {
+		t.Fatalf("expected cache affinity routing diagnostics in attempt metrics, got %+v", metrics)
+	}
 }
@@ -500,14 +500,15 @@ func applyRuntimeCandidateCacheAffinity(candidate *RuntimeModelCandidate, option
 		key = options.CacheAffinityKey
 	}
 	affinity := RuntimeCandidateCacheAffinity{
-		Key:               key,
-		RequestCount:      input.RequestCount,
-		InputTokens:       input.InputTokens,
-		CachedInputTokens: input.CachedInputTokens,
-		EMAHitRatio:       boundedRatio(input.EMAHitRatio),
-		LastHitRatio:      boundedRatio(input.LastHitRatio),
-		LastObservedUnix:  input.LastObservedUnix,
-		AdjustedPriority:  float64(candidate.PlatformPriority),
+		Key:                key,
+		RequestCount:       input.RequestCount,
+		InputTokens:        input.InputTokens,
+		CachedInputTokens:  input.CachedInputTokens,
+		EMAHitRatio:        boundedRatio(input.EMAHitRatio),
+		LastHitRatio:       boundedRatio(input.LastHitRatio),
+		LastObservedUnix:   input.LastObservedUnix,
+		AdjustedPriority:   float64(candidate.PlatformPriority),
+		MatchedPrefixDepth: cacheAffinityMatchedPrefixDepth(key, options.CacheAffinityKeys),
 	}
 	minSamples := cacheAffinityMinSamples(options.CacheAffinityPolicy)
 	hasObservedCachedHit := affinity.CachedInputTokens > 0 || affinity.LastHitRatio > 0
@@ -538,6 +539,32 @@ func applyRuntimeCandidateCacheAffinity(candidate *RuntimeModelCandidate, option
 	candidate.CacheAffinity = affinity
 }

+func cacheAffinityMatchedPrefixDepth(key string, lookup []string) int {
+	prefix := ""
+	switch {
+	case strings.HasPrefix(key, "prompt_lcp_v2:"):
+		prefix = "prompt_lcp_v2:"
+	case strings.HasPrefix(key, "prompt_lcp:"):
+		prefix = "prompt_lcp:"
+	default:
+		return 0
+	}
+	matchingKeys := make([]string, 0, len(lookup))
+	for _, candidateKey := range lookup {
+		if strings.HasPrefix(candidateKey, prefix) {
+			matchingKeys = append(matchingKeys, candidateKey)
+		}
+	}
+	for index, candidateKey := range matchingKeys {
+		if candidateKey == key {
+			return len(matchingKeys) - index + cacheAffinityMinimumPrefixDepth - 1
+		}
+	}
+	return 0
+}
+
+const cacheAffinityMinimumPrefixDepth = 2
+
 func cacheAffinityPolicyEnabled(policy map[string]any, modelType string) bool {
 	if enabled, ok := policy["enabled"].(bool); ok && !enabled {
 		return false
@@ -712,6 +739,9 @@ func sortRuntimeModelCandidates(items []RuntimeModelCandidate) {
 		if aFull != bFull {
 			return !aFull
 		}
+		if items[i].PlatformPriority != items[j].PlatformPriority {
+			return items[i].PlatformPriority < items[j].PlatformPriority
+		}
 		if items[i].CacheAffinity.Applied != items[j].CacheAffinity.Applied {
 			return items[i].CacheAffinity.Applied
 		}
@@ -746,6 +776,32 @@ func sortRuntimeModelCandidates(items []RuntimeModelCandidate) {
 		}
 		return false
 	})
+	appliedCount := 0
+	for index := range items {
+		if items[index].CacheAffinity.Applied {
+			appliedCount++
+		}
+	}
+	for index := range items {
+		items[index].CacheAffinity.CandidateCount = appliedCount
+	}
+	if len(items) == 0 || items[0].CacheAffinity.Applied {
+		return
+	}
+	for index := 1; index < len(items); index++ {
+		if !items[index].CacheAffinity.Applied {
+			continue
+		}
+		switch {
+		case runtimeCandidateFull(items[index]) && !runtimeCandidateFull(items[0]):
+			items[0].CacheAffinity.OverrideReason = "capacity_tier_unavailable"
+		case items[index].PlatformPriority != items[0].PlatformPriority:
+			items[0].CacheAffinity.OverrideReason = "different_effective_priority_tier"
+		}
+		if items[0].CacheAffinity.OverrideReason != "" {
+			break
+		}
+	}
 }

 func runtimeCandidateFull(candidate RuntimeModelCandidate) bool {
@@ -120,7 +120,7 @@ func TestRuntimeCandidateSortingPrefersCacheAffinityWithinAvailableCandidates(t
 		},
 		{
 			PlatformID:       "cache-affinity",
-			PlatformPriority: 20,
+			PlatformPriority: 10,
 		},
 	}
 	applyRuntimeCandidateCacheAffinity(&candidates[0], ListModelCandidatesOptions{
@@ -148,7 +148,7 @@ func TestRuntimeCandidateSortingPrefersCacheAffinityWithinAvailableCandidates(t
 	}
 }

-func TestRuntimeCandidateSortingPromotesObservedCacheHitAbovePriority(t *testing.T) {
+func TestRuntimeCandidateSortingKeepsEffectivePriorityAheadOfCacheAffinity(t *testing.T) {
 	candidates := []RuntimeModelCandidate{
 		{
 			PlatformID:       "volces-priority",
@@ -177,14 +177,17 @@ func TestRuntimeCandidateSortingPromotesObservedCacheHitAbovePriority(t *testing

 	sortRuntimeModelCandidates(candidates)

-	if candidates[0].PlatformID != "deepseek-cache-hit" || !candidates[0].CacheAffinity.Applied {
-		t.Fatalf("expected observed cache hit to outrank platform priority, got %+v", candidates)
+	if candidates[0].PlatformID != "volces-priority" || candidates[0].CacheAffinity.Applied {
+		t.Fatalf("expected effective priority tier to remain first, got %+v", candidates)
 	}
-	if candidates[0].CacheAffinity.Score <= 0 || candidates[0].CacheAffinity.Boost <= 0 {
-		t.Fatalf("expected observed cache hit to carry score and boost, got %+v", candidates[0].CacheAffinity)
+	if candidates[0].CacheAffinity.OverrideReason != "different_effective_priority_tier" {
+		t.Fatalf("expected priority override reason on selected candidate, got %+v", candidates[0].CacheAffinity)
 	}
-	if candidates[1].PlatformID != "volces-priority" || candidates[1].CacheAffinity.Applied {
-		t.Fatalf("expected priority-only candidate second, got %+v", candidates)
+	if candidates[1].PlatformID != "deepseek-cache-hit" || !candidates[1].CacheAffinity.Applied {
+		t.Fatalf("expected cache affinity candidate to remain visible as fallback, got %+v", candidates)
+	}
+	if candidates[1].CacheAffinity.Score <= 0 || candidates[1].CacheAffinity.Boost <= 0 {
+		t.Fatalf("expected observed cache hit to retain score and boost diagnostics, got %+v", candidates[1].CacheAffinity)
 	}
 }

@@ -196,7 +199,7 @@ func TestRuntimeCandidateSortingOrdersMultipleCacheAffinityCandidatesByScore(t *
 		},
 		{
 			PlatformID:       "higher-cache-score",
-			PlatformPriority: 50,
+			PlatformPriority: 1,
 		},
 	}
 	policy := map[string]any{"enabled": true, "minSamples": 1, "maxPriorityBoost": 20, "modelTypes": []any{"text_generate"}}
@@ -260,6 +263,29 @@ func TestRuntimeCandidateSortingKeepsFullCacheAffinityCandidateAvoided(t *testin
 	if candidates[1].PlatformID != "cache-affinity-full" || !candidates[1].LoadAvoided {
 		t.Fatalf("expected full cache-affinity candidate to remain avoided fallback, got %+v", candidates)
 	}
+	if candidates[0].CacheAffinity.OverrideReason != "capacity_tier_unavailable" {
+		t.Fatalf("expected capacity override reason on selected candidate, got %+v", candidates[0].CacheAffinity)
+	}
+}
+
+func TestCacheAffinityMatchedPrefixDepthIgnoresLegacyCompatibilityKeys(t *testing.T) {
+	lookup := []string{
+		"prompt_lcp_v2:depth-four",
+		"prompt_lcp_v2:depth-three",
+		"prompt_lcp_v2:depth-two",
+		"prompt_lcp:depth-four",
+		"prompt_lcp:depth-three",
+		"prompt_lcp:depth-two",
+	}
+	if depth := cacheAffinityMatchedPrefixDepth("prompt_lcp_v2:depth-three", lookup); depth != 3 {
+		t.Fatalf("V2 matched prefix depth=%d want=3", depth)
+	}
+	if depth := cacheAffinityMatchedPrefixDepth("prompt_lcp:depth-two", lookup); depth != 2 {
+		t.Fatalf("legacy matched prefix depth=%d want=2", depth)
+	}
+	if depth := cacheAffinityMatchedPrefixDepth("explicit:key", lookup); depth != 0 {
+		t.Fatalf("explicit key prefix depth=%d want=0", depth)
+	}
 }

 func TestDefaultRunnerPriorityDemotePolicyUsesAutoMode(t *testing.T) {
@@ -178,18 +178,21 @@ type RuntimeModelCandidate struct {
 }

 type RuntimeCandidateCacheAffinity struct {
-	Key               string
-	RequestCount      int
-	InputTokens       int
-	CachedInputTokens int
-	EMAHitRatio       float64
-	LastHitRatio      float64
-	LastObservedUnix  float64
-	Confidence        float64
-	Score             float64
-	Boost             float64
-	AdjustedPriority  float64
-	Applied           bool
+	Key                string
+	RequestCount       int
+	InputTokens        int
+	CachedInputTokens  int
+	EMAHitRatio        float64
+	LastHitRatio       float64
+	LastObservedUnix   float64
+	Confidence         float64
+	Score              float64
+	Boost              float64
+	AdjustedPriority   float64
+	MatchedPrefixDepth int
+	CandidateCount     int
+	OverrideReason     string
+	Applied            bool
 }

 type RuntimeCandidateLoadMetrics struct {
@@ -128,7 +128,7 @@ WHERE task_id=$1::uuid
 	}
 }

-func TestTaskAttemptDropsDuplicateJSONSnapshots(t *testing.T) {
+func TestTaskAttemptKeepsOnlyCompactDiagnostics(t *testing.T) {
 	db := billingV2IntegrationStore(t)
 	ctx := context.Background()
 	user := &auth.User{ID: "attempt-minimal-" + uuid.NewString()}
@@ -157,8 +157,8 @@ func TestTaskAttemptDropsDuplicateJSONSnapshots(t *testing.T) {
 		Status:           "failed",
 		Retryable:        true,
 		StatusCode:       503,
-		Usage:            map[string]any{"output_tokens": 100},
-		Metrics:          map[string]any{"trace": []any{map[string]any{"large": "duplicate"}}},
+		Usage:            map[string]any{"output_tokens": 100, "raw": strings.Repeat("duplicate", 100)},
+		Metrics:          map[string]any{"trace": []any{map[string]any{"large": "duplicate"}}, "cacheAffinityMatched": true, "cacheAffinityMatchedPrefixDepth": 3},
 		ResponseSnapshot: map[string]any{"result": "duplicate"},
 		ErrorCode:        "upstream_error",
 		ErrorMessage:     strings.Repeat("错误", 4096),
@@ -173,9 +173,16 @@ func TestTaskAttemptDropsDuplicateJSONSnapshots(t *testing.T) {
 		t.Fatalf("attempts=%d", len(attempts))
 	}
 	attempt := attempts[0]
-	if len(attempt.RequestSnapshot) != 0 || len(attempt.ResponseSnapshot) != 0 ||
-		len(attempt.Usage) != 0 || len(attempt.Metrics) != 0 || len(attempt.PricingSnapshot) != 0 {
-		t.Fatalf("attempt contains duplicate JSON: %+v", attempt)
+	if len(attempt.RequestSnapshot) != 0 || len(attempt.ResponseSnapshot) != 0 || len(attempt.PricingSnapshot) != 0 {
+		t.Fatalf("attempt contains duplicate snapshots: %+v", attempt)
+	}
+	if len(attempt.Usage) != 1 || taskAttemptMetricInt(attempt.Usage, "output_tokens") != 100 {
+		t.Fatalf("attempt compact usage mismatch: %+v", attempt.Usage)
+	}
+	if len(attempt.Metrics) != 2 ||
+		attempt.Metrics["cacheAffinityMatched"] != true ||
+		taskAttemptMetricInt(attempt.Metrics, "cacheAffinityMatchedPrefixDepth") != 3 {
+		t.Fatalf("attempt compact metrics mismatch: %+v", attempt.Metrics)
 	}
 	if attempt.StatusCode != 503 || len(attempt.ErrorMessage) > 2048 {
 		t.Fatalf("attempt status/error not preserved safely: %+v", attempt)
@@ -996,14 +996,16 @@ func (s *Store) FinishTaskAttempt(ctx context.Context, input FinishTaskAttemptIn
 	if statusCode == 0 {
 		statusCode = taskAttemptMetricInt(input.Metrics, "statusCode")
 	}
+	usageJSON, _ := json.Marshal(sanitizeJSONForStorage(minimalTaskAttemptUsage(input.Usage)))
+	metricsJSON, _ := json.Marshal(sanitizeJSONForStorage(minimalTaskAttemptMetrics(input.Metrics)))
 	_, err := s.pool.Exec(ctx, `
 UPDATE gateway_task_attempts
 SET status = $2::text,
    retryable = $3,
    request_id = NULLIF($4::text, ''),
    status_code = NULLIF($5::int, 0),
-    usage = '{}'::jsonb,
-    metrics = '{}'::jsonb,
+    usage = $11::jsonb,
+    metrics = $12::jsonb,
    response_snapshot = '{}'::jsonb,
    pricing_snapshot = '{}'::jsonb,
    request_fingerprint = NULL,
@@ -1024,10 +1026,45 @@ WHERE id = $1::uuid`,
 		input.ResponseDurationMS,
 		input.ErrorCode,
 		truncateUTF8Bytes(input.ErrorMessage, 2048),
+		usageJSON,
+		metricsJSON,
 	)
 	return err
 }

+func minimalTaskAttemptUsage(usage map[string]any) map[string]any {
+	return whitelistedTaskAttemptMap(usage, []string{
+		"inputTokens", "promptTokens", "input_tokens", "prompt_tokens",
+		"cachedInputTokens", "cachedPromptTokens", "cached_input_tokens", "cached_tokens",
+		"cachedInputTokensKnown",
+		"outputTokens", "completionTokens", "output_tokens", "completion_tokens",
+		"totalTokens", "total_tokens",
+	})
+}
+
+func minimalTaskAttemptMetrics(metrics map[string]any) map[string]any {
+	return whitelistedTaskAttemptMap(metrics, []string{
+		"platformPriority", "currentPriority", "loadRatio", "loadAvoided",
+		"cacheAffinityKey", "cacheAdjustedPriority", "cacheAffinitySamples",
+		"cacheAffinityScore", "cacheAffinityConfidence", "cacheAffinityHitRatio",
+		"cacheAffinityEMAHitRatio", "cacheAffinityLastHitRatio",
+		"cacheAffinityCachedInputTokens", "cacheAffinityBoost",
+		"cacheAffinityApplied", "cacheAffinityMatched",
+		"cacheAffinityCandidateCount", "cacheAffinityMatchedPrefixDepth",
+		"cacheAffinityOverrideReason",
+	})
+}
+
+func whitelistedTaskAttemptMap(input map[string]any, keys []string) map[string]any {
+	out := map[string]any{}
+	for _, key := range keys {
+		if value, ok := input[key]; ok {
+			out[key] = value
+		}
+	}
+	return out
+}
+
 func (s *Store) FinishTaskSuccess(ctx context.Context, input FinishTaskSuccessInput) (GatewayTask, error) {
 	resultReport := sanitizeJSONForStorageWithReport(minimalTaskResult(input.Result))
 	if resultReport.BinaryCount > 0 {
@@ -1041,6 +1078,8 @@ func (s *Store) FinishTaskSuccess(ctx context.Context, input FinishTaskSuccessIn
 	billingsJSON, _ := json.Marshal(sanitizeJSONForStorage(input.Billings))
 	usageJSON, _ := json.Marshal(sanitizeJSONForStorage(emptyObjectIfNil(input.Usage)))
 	metricsJSON, _ := json.Marshal(sanitizeJSONForStorage(emptyObjectIfNil(input.Metrics)))
+	attemptUsageJSON, _ := json.Marshal(sanitizeJSONForStorage(minimalTaskAttemptUsage(input.Usage)))
+	attemptMetricsJSON, _ := json.Marshal(sanitizeJSONForStorage(minimalTaskAttemptMetrics(input.Metrics)))
 	billingSummaryJSON, _ := json.Marshal(sanitizeJSONForStorage(emptyObjectIfNil(input.BillingSummary)))
 	pricingSnapshotJSON, _ := json.Marshal(sanitizeJSONForStorage(emptyObjectIfNil(input.PricingSnapshot)))
 	finalChargeAmount := strings.TrimSpace(input.FinalChargeAmountText)
@@ -1056,8 +1095,8 @@ SET status = 'succeeded',
    retryable = false,
    request_id = NULLIF($2, ''),
    status_code = NULL,
-    usage = '{}'::jsonb,
-    metrics = '{}'::jsonb,
+    usage = $6::jsonb,
+    metrics = $7::jsonb,
    response_snapshot = '{}'::jsonb,
    pricing_snapshot = '{}'::jsonb,
    request_fingerprint = NULL,
@@ -1072,6 +1111,7 @@ SET status = 'succeeded',
 WHERE id = $1::uuid`,
 				input.AttemptID, input.RequestID,
 				nullableTime(input.ResponseStartedAt), nullableTime(input.ResponseFinishedAt), input.ResponseDurationMS,
+				attemptUsageJSON, attemptMetricsJSON,
 			); err != nil {
 				return err
 			}