pi-cache-optimizer 2.4.3 → 2.4.4

package/README.md

@@ -35,6 +35,17 @@ This release keeps the original DeepSeek behavior and adds read-only stats adapt
 |---|---|---|---|
 | DeepSeek | Model id/name contains `deepseek` | `DS cache` | Pi `usage.cacheRead`/`usage.input`, or raw `prompt_cache_hit_tokens`, `prompt_cache_miss_tokens`, `prompt_tokens` when visible |
 | OpenAI-family | Model id/name contains conservative OpenAI-family tokens such as `gpt-`, `chatgpt`, `o1`, `o3`, `o4`, or `o5` | `OpenAI cache` | Pi-normalized usage, or raw `prompt_tokens_details.cached_tokens` / `input_tokens_details.cached_tokens` with prompt/input totals |
+| Kimi / Moonshot | Model id/name contains `kimi` | `Kimi cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| Qwen / Alibaba | Model id/name contains `qwen` | `Qwen cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| GLM / Zhipu | Model id/name contains `glm` | `GLM cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| MiniMax | Model id/name contains `minimax` | `MiniMax cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| Hunyuan / Tencent | Model id/name contains `hunyuan` | `Hunyuan cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| Mistral | Model id/name contains `mistral`, `mixtral`, or `codestral` | `Mistral cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| xAI / Grok | Model id/name contains `grok`, or pattern `xai` with safe boundaries | `Grok cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| Meta / Llama | Model id/name contains `llama` | `Llama cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| NVIDIA Nemotron | Model id/name contains `nemotron` | `Nemotron cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| Cohere / Command | Model id/name contains `cohere` or `command-r` | `Cohere cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
+| Yi / 零一万物 | Model id/name contains `yi-`, `01-ai`, `zero-one`, or pattern `yi` with safe boundaries | `Yi cache` | Pi-normalized usage, or raw OpenAI-shaped fields when visible |
 | Anthropic / Claude | Model id/name contains `anthropic` or `claude` | `Claude cache` | Pi-normalized usage, or raw `cache_read_input_tokens`, `cache_creation_input_tokens`, `input_tokens` |
 | Gemini / Vertex | Model id/name contains `gemini` or `vertex` | `Gemini cache` | Pi-normalized usage, or raw Gemini/Vertex cached-content token metadata when visible |
 

package/README.zh-CN.md

@@ -38,6 +38,17 @@
 |---|---|---|---|
 | DeepSeek | model id/name 包含 `deepseek` | `DS cache` | Pi `usage.cacheRead`/`usage.input`，或可见 raw 字段 `prompt_cache_hit_tokens`、`prompt_cache_miss_tokens`、`prompt_tokens` |
 | OpenAI-family | model id/name 包含保守 OpenAI-family token，例如 `gpt-`、`chatgpt`、`o1`、`o3`、`o4` 或 `o5` | `OpenAI cache` | Pi 归一化 usage，或可见 raw 字段 `prompt_tokens_details.cached_tokens` / `input_tokens_details.cached_tokens` 及 prompt/input total |
+| Kimi / Moonshot | model id/name 包含 `kimi` | `Kimi cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| Qwen / Alibaba | model id/name 包含 `qwen` | `Qwen cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| GLM / Zhipu | model id/name 包含 `glm` | `GLM cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| MiniMax | model id/name 包含 `minimax` | `MiniMax cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| Hunyuan / Tencent | model id/name 包含 `hunyuan` | `Hunyuan cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| Mistral | model id/name 包含 `mistral`、`mixtral` 或 `codestral` | `Mistral cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| xAI / Grok | model id/name 包含 `grok`，或安全边界内 `xai` 模式 | `Grok cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| Meta / Llama | model id/name 包含 `llama` | `Llama cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| NVIDIA Nemotron | model id/name 包含 `nemotron` | `Nemotron cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| Cohere / Command | model id/name 包含 `cohere` 或 `command-r` | `Cohere cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
+| Yi / 零一万物 | model id/name 包含 `yi-`、`01-ai`、`zero-one`，或安全边界内 `yi` 模式 | `Yi cache` | Pi 归一化 usage，或可见 OpenAI 形状字段 |
 | Anthropic / Claude | model id/name 包含 `anthropic` 或 `claude` | `Claude cache` | Pi 归一化 usage，或可见 raw 字段 `cache_read_input_tokens`、`cache_creation_input_tokens`、`input_tokens` |
 | Gemini / Vertex | model id/name 包含 `gemini` 或 `vertex` | `Gemini cache` | Pi 归一化 usage，或可见 Gemini/Vertex cached-content token metadata |
 

package/index.ts

@@ -80,6 +80,7 @@ const MIN_STABLE_CANDIDATE_LENGTH = 8;
 
 const ASSISTANT_MESSAGE_MODEL_TOKEN_KEYS = ["model", "name"];
 const OPENAI_REASONING_MODEL_PATTERN = /(^|[/\s:_-])o[1345]($|[-_.:/\s])/;
+const XAI_MODEL_PATTERN = /(^|[/\s:_-])xai($|[-_.:/\s])/;
 
 type CacheCompat = {
   sendSessionAffinityHeaders?: boolean;
@@ -672,6 +673,62 @@ function isHunyuanLikeAssistantMessage(message: unknown, model: PiModel | undefi
   return modelOrAssistantMessageHas(message, model, ["hunyuan"]);
 }
 
+// ── Additional OpenAI-compatible model detection ──────────────────
+
+function isMistralLikeModel(model: PiModel | undefined): boolean {
+  return hasAnyTokenContaining(getModelIdNameTokenValues(model), ["mistral", "mixtral", "codestral"]);
+}
+function isMistralLikeAssistantMessage(message: unknown, model: PiModel | undefined): boolean {
+  return modelOrAssistantMessageHas(message, model, ["mistral", "mixtral", "codestral"]);
+}
+
+function isGrokLikeModel(model: PiModel | undefined): boolean {
+  const tokens = getModelIdNameTokenValues(model);
+  return hasAnyTokenContaining(tokens, ["grok"]) || tokens.some((t) => XAI_MODEL_PATTERN.test(t));
+}
+function isGrokLikeAssistantMessage(message: unknown, model: PiModel | undefined): boolean {
+  const allTokens = [
+    ...getModelIdNameTokenValues(model),
+    ...getAssistantMessageModelTokenValues(message),
+  ];
+  return hasAnyTokenContaining(allTokens, ["grok"]) || allTokens.some((t) => XAI_MODEL_PATTERN.test(t));
+}
+
+function isLlamaLikeModel(model: PiModel | undefined): boolean {
+  return hasAnyTokenContaining(getModelIdNameTokenValues(model), ["llama"]);
+}
+function isLlamaLikeAssistantMessage(message: unknown, model: PiModel | undefined): boolean {
+  return modelOrAssistantMessageHas(message, model, ["llama"]);
+}
+
+function isNemotronLikeModel(model: PiModel | undefined): boolean {
+  return hasAnyTokenContaining(getModelIdNameTokenValues(model), ["nemotron"]);
+}
+function isNemotronLikeAssistantMessage(message: unknown, model: PiModel | undefined): boolean {
+  return modelOrAssistantMessageHas(message, model, ["nemotron"]);
+}
+
+function isCohereLikeModel(model: PiModel | undefined): boolean {
+  return hasAnyTokenContaining(getModelIdNameTokenValues(model), ["cohere", "command-r"]);
+}
+function isCohereLikeAssistantMessage(message: unknown, model: PiModel | undefined): boolean {
+  return modelOrAssistantMessageHas(message, model, ["cohere", "command-r"]);
+}
+
+const YI_MODEL_PATTERN = /(^|[\/\s:_-])yi($|[\-_.:\/\s])/;
+
+function isYiLikeModel(model: PiModel | undefined): boolean {
+  const tokens = getModelIdNameTokenValues(model);
+  return hasAnyTokenContaining(tokens, ["yi-", "01-ai", "zero-one"]) || tokens.some((t) => YI_MODEL_PATTERN.test(t));
+}
+function isYiLikeAssistantMessage(message: unknown, model: PiModel | undefined): boolean {
+  const allTokens = [
+    ...getModelIdNameTokenValues(model),
+    ...getAssistantMessageModelTokenValues(message),
+  ];
+  return hasAnyTokenContaining(allTokens, ["yi-", "01-ai", "zero-one"]) || allTokens.some((t) => YI_MODEL_PATTERN.test(t));
+}
+
 // ── Model key ──────────────────────────────────────────────────────
 
 function modelKey(model: PiModel): string {
@@ -1136,6 +1193,109 @@ const CACHE_PROVIDER_ADAPTERS: CacheProviderAdapter[] = [
       return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
     },
   },
+  // ── More OpenAI-compatible adapters ──────────────────────────
+  {
+    id: "openai" as CacheProviderId,
+    label: "Mistral cache",
+    matchesModel: isMistralLikeModel,
+    matchesAssistantMessage(message, model) {
+      if (!isAssistantMessage(message)) return false;
+      return isMistralLikeAssistantMessage(message, model);
+    },
+    normalizeUsage(message) {
+      return normalizeWithFallback(message, getOpenAIRawUsage);
+    },
+    warningText(model) {
+      const missing = describeMissingOpenAICompatibleProxyCompat(model);
+      if (missing.length === 0) return undefined;
+      return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
+    },
+  },
+  {
+    id: "openai" as CacheProviderId,
+    label: "Grok cache",
+    matchesModel: isGrokLikeModel,
+    matchesAssistantMessage(message, model) {
+      if (!isAssistantMessage(message)) return false;
+      return isGrokLikeAssistantMessage(message, model);
+    },
+    normalizeUsage(message) {
+      return normalizeWithFallback(message, getOpenAIRawUsage);
+    },
+    warningText(model) {
+      const missing = describeMissingOpenAICompatibleProxyCompat(model);
+      if (missing.length === 0) return undefined;
+      return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
+    },
+  },
+  {
+    id: "openai" as CacheProviderId,
+    label: "Llama cache",
+    matchesModel: isLlamaLikeModel,
+    matchesAssistantMessage(message, model) {
+      if (!isAssistantMessage(message)) return false;
+      return isLlamaLikeAssistantMessage(message, model);
+    },
+    normalizeUsage(message) {
+      return normalizeWithFallback(message, getOpenAIRawUsage);
+    },
+    warningText(model) {
+      const missing = describeMissingOpenAICompatibleProxyCompat(model);
+      if (missing.length === 0) return undefined;
+      return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
+    },
+  },
+  {
+    id: "openai" as CacheProviderId,
+    label: "Nemotron cache",
+    matchesModel: isNemotronLikeModel,
+    matchesAssistantMessage(message, model) {
+      if (!isAssistantMessage(message)) return false;
+      return isNemotronLikeAssistantMessage(message, model);
+    },
+    normalizeUsage(message) {
+      return normalizeWithFallback(message, getOpenAIRawUsage);
+    },
+    warningText(model) {
+      const missing = describeMissingOpenAICompatibleProxyCompat(model);
+      if (missing.length === 0) return undefined;
+      return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
+    },
+  },
+  {
+    id: "openai" as CacheProviderId,
+    label: "Cohere cache",
+    matchesModel: isCohereLikeModel,
+    matchesAssistantMessage(message, model) {
+      if (!isAssistantMessage(message)) return false;
+      return isCohereLikeAssistantMessage(message, model);
+    },
+    normalizeUsage(message) {
+      return normalizeWithFallback(message, getOpenAIRawUsage);
+    },
+    warningText(model) {
+      const missing = describeMissingOpenAICompatibleProxyCompat(model);
+      if (missing.length === 0) return undefined;
+      return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
+    },
+  },
+  {
+    id: "openai" as CacheProviderId,
+    label: "Yi cache",
+    matchesModel: isYiLikeModel,
+    matchesAssistantMessage(message, model) {
+      if (!isAssistantMessage(message)) return false;
+      return isYiLikeAssistantMessage(message, model);
+    },
+    normalizeUsage(message) {
+      return normalizeWithFallback(message, getOpenAIRawUsage);
+    },
+    warningText(model) {
+      const missing = describeMissingOpenAICompatibleProxyCompat(model);
+      if (missing.length === 0) return undefined;
+      return buildOpenAIProxyCompatWarningText(modelKey(model), missing);
+    },
+  },
 ];
 
 function selectAdapterForModel(model: PiModel | undefined): CacheProviderAdapter | undefined {
@@ -1398,6 +1558,19 @@ export const __internals_for_tests = {
   isMiniMaxLikeAssistantMessage,
   isHunyuanLikeModel,
   isHunyuanLikeAssistantMessage,
+  // Additional OpenAI-compatible model detection
+  isMistralLikeModel,
+  isMistralLikeAssistantMessage,
+  isGrokLikeModel,
+  isGrokLikeAssistantMessage,
+  isLlamaLikeModel,
+  isLlamaLikeAssistantMessage,
+  isNemotronLikeModel,
+  isNemotronLikeAssistantMessage,
+  isCohereLikeModel,
+  isCohereLikeAssistantMessage,
+  isYiLikeModel,
+  isYiLikeAssistantMessage,
   buildOpenAIProxyCompatWarningText,
   getModelIdNameTokenValues,
   getAssistantMessageModelTokenValues,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-cache-optimizer",
-  "version": "2.4.3",
+  "version": "2.4.4",
   "description": "Pi extension that improves provider-side KV/prompt cache hit rates (DeepSeek, OpenAI, Claude, Gemini) by reordering the system prompt, requesting long retention, and showing footer cache stats. Renamed from pi-deepseek-cache-optimizer.",
   "keywords": [
     "pi-package",