@fonoster/community 0.30.0 → 0.31.0

package/dist/adapters.js

@@ -63,6 +63,9 @@ exports.ADAPTERS = {
     openai: { llm: openAiCompatible("openai") },
     groq: { llm: openAiCompatible("groq") },
     ollama: { keyless: true, llm: openAiCompatible("ollama") },
+    // Noise cancelling (`nc`): served by the appbridge's aiCoustics filter with the provider's key,
+    // so the entry only makes the adapter known to the registry and the gateway config.
+    aicoustics: {},
     fake: {
         keyless: true,
         tts: (o) => {

package/dist/llm.d.ts

@@ -33,12 +33,26 @@ export type ChatMessage = {
     tool_call_id: string;
     content: string;
 };
+/** Token counts as the vendor reports them for one request (add-usage-metering). */
+export type TokenUsage = {
+    inputTokens: number;
+    outputTokens: number;
+    /** Input tokens served from the vendor's prompt cache (part of inputTokens). */
+    cachedTokens: number;
+};
+/**
+ * A stream ends with at most one `usage` chunk, after the text and tool calls, when the vendor
+ * reports usage. Its absence means the vendor reported none (or the stream was cut short).
+ */
 export type ThinkChunk = {
     type: "text";
     text: string;
 } | {
     type: "tool_calls";
     calls: ToolCall[];
+} | {
+    type: "usage";
+    usage: TokenUsage;
 };
 export type ToolSpec = {
     name: string;

package/dist/openai/index.js

@@ -53,7 +53,9 @@ class OpenAiCompatibleThink {
     stream(messages, tools, signal) {
         return __asyncGenerator(this, arguments, function* stream_1() {
             var _a, _b, _c, _d, _e, _f, _g, _h;
-            const body = Object.assign(Object.assign({ model: this.model, stream: true, temperature: this.options.temperature, max_tokens: this.options.maxTokens, messages: messages.map(toOpenAi) }, (tools.length
+            const body = Object.assign(Object.assign({ model: this.model, stream: true, 
+                // The final chunk carries the request's token usage (add-usage-metering D6).
+                stream_options: { include_usage: true }, temperature: this.options.temperature, max_tokens: this.options.maxTokens, messages: messages.map(toOpenAi) }, (tools.length
                 ? { tools: tools.map((t) => ({ type: "function", function: t })) }
                 : {})), (this.options.jsonMode
                 ? { response_format: { type: "json_object" } }
@@ -79,6 +81,7 @@ class OpenAiCompatibleThink {
             if (!res.ok || !res.body)
                 throw new llm_1.ThinkError("provider_unavailable", `language model answered HTTP ${res.status}`);
             const calls = new Map();
+            let usage;
             const decoder = new util_1.TextDecoder();
             let buffer = "";
             const reader = res.body.getReader();
@@ -105,7 +108,10 @@ class OpenAiCompatibleThink {
                     const data = line.slice(5).trim();
                     if (data === "[DONE]")
                         continue;
-                    const delta = (_c = (_b = (_a = JSON.parse(data).choices) === null || _a === void 0 ? void 0 : _a[0]) === null || _b === void 0 ? void 0 : _b.delta) !== null && _c !== void 0 ? _c : {};
+                    const parsed = JSON.parse(data);
+                    if (parsed.usage)
+                        usage = toTokenUsage(parsed.usage);
+                    const delta = (_c = (_b = (_a = parsed.choices) === null || _a === void 0 ? void 0 : _a[0]) === null || _b === void 0 ? void 0 : _b.delta) !== null && _c !== void 0 ? _c : {};
                     if (delta.content)
                         yield yield __await({ type: "text", text: delta.content });
                     for (const tc of (_d = delta.tool_calls) !== null && _d !== void 0 ? _d : []) {
@@ -124,10 +130,21 @@ class OpenAiCompatibleThink {
             }
             if (calls.size)
                 yield yield __await({ type: "tool_calls", calls: [...calls.values()] });
+            if (usage)
+                yield yield __await({ type: "usage", usage });
         });
     }
 }
 exports.OpenAiCompatibleThink = OpenAiCompatibleThink;
+/** OpenAI's usage object as TokenUsage; cached tokens are optional in OpenAI-compatible APIs. */
+function toTokenUsage(u) {
+    var _a, _b, _c, _d;
+    return {
+        inputTokens: (_a = u.prompt_tokens) !== null && _a !== void 0 ? _a : 0,
+        outputTokens: (_b = u.completion_tokens) !== null && _b !== void 0 ? _b : 0,
+        cachedTokens: (_d = (_c = u.prompt_tokens_details) === null || _c === void 0 ? void 0 : _c.cached_tokens) !== null && _d !== void 0 ? _d : 0
+    };
+}
 function toOpenAi(m) {
     var _a;
     if (m.role === "assistant" && ((_a = m.tool_calls) === null || _a === void 0 ? void 0 : _a.length)) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@fonoster/community",
-  "version": "0.30.0",
+  "version": "0.31.0",
   "description": "Vendor adapters (STT, TTS, LLM) behind shared interfaces, one subpath per vendor",
   "author": "Pedro Sanders <psanders@fonoster.com>",
   "homepage": "https://github.com/fonoster/fonoster#readme",
@@ -59,8 +59,7 @@
   },
   "scripts": {
     "prebuild": "rimraf ./dist tsconfig.tsbuildinfo",
-    "build": "tsc -b tsconfig.json",
-    "clean": "rimraf ./dist node_modules tsconfig.tsbuildinfo"
+    "build": "tsc -b tsconfig.json"
   },
   "dependencies": {
     "ws": "^8.18.0"
@@ -78,5 +77,5 @@
   "bugs": {
     "url": "https://github.com/fonoster/fonoster/issues"
   },
-  "gitHead": "08c9cfd7932675b6b16111daa4d03f14b4e287d1"
+  "gitHead": "f0a5102dddeea0f70f659234108805bc10690f2e"
 }