@glassflow-ai/rius 0.4.0 → 0.5.0

package/dist/index.cjs

@@ -356,7 +356,7 @@ function kindAttributes(kind) {
   if (operation !== void 0) attributes[GEN_AI_OPERATION_NAME] = operation;
   return attributes;
 }
-var TRACER_NAME, SERVICE_INSTANCE_ID, OPENINFERENCE_SPAN_KIND, INPUT_VALUE, OUTPUT_VALUE, SESSION_ID, WORKSPACE_ROUTE, GEN_AI_OPERATION_NAME, GEN_AI_PROVIDER_NAME, GEN_AI_REQUEST_MODEL, GEN_AI_RESPONSE_MODEL, GEN_AI_USAGE_INPUT_TOKENS, GEN_AI_USAGE_OUTPUT_TOKENS, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS, GEN_AI_INPUT_MESSAGES, GEN_AI_OUTPUT_MESSAGES, GEN_AI_RESPONSE_FINISH_REASONS, GEN_AI_TOOL_NAME, GEN_AI_REQUEST_PREFIX, MCP_RESULT_TYPE, GEN_AI_FIRST_TOKEN_EVENT, SpanKind, OPERATION_BY_KIND, CONTENT_ATTRIBUTES, CONTENT_ATTRIBUTE_PREFIXES, CONTENT_ATTRIBUTE_SUFFIXES, GLASSFLOW_SPAN_PENDING, PENDING_IDENTITY_ATTRIBUTES, PENDING_IDENTITY_PREFIXES;
+var TRACER_NAME, SERVICE_INSTANCE_ID, OPENINFERENCE_SPAN_KIND, INPUT_VALUE, OUTPUT_VALUE, SESSION_ID, WORKSPACE_ROUTE, GEN_AI_OPERATION_NAME, GEN_AI_PROVIDER_NAME, GEN_AI_REQUEST_MODEL, GEN_AI_REQUEST_REASONING_LEVEL, GEN_AI_RESPONSE_MODEL, GEN_AI_USAGE_INPUT_TOKENS, GEN_AI_USAGE_OUTPUT_TOKENS, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS, GEN_AI_USAGE_REASONING_OUTPUT_TOKENS, GEN_AI_INPUT_MESSAGES, GEN_AI_OUTPUT_MESSAGES, GEN_AI_RESPONSE_FINISH_REASONS, GEN_AI_TOOL_NAME, GEN_AI_REQUEST_PREFIX, MCP_RESULT_TYPE, GEN_AI_FIRST_TOKEN_EVENT, SpanKind, OPERATION_BY_KIND, CONTENT_ATTRIBUTES, CONTENT_ATTRIBUTE_PREFIXES, CONTENT_ATTRIBUTE_SUFFIXES, GLASSFLOW_SPAN_PENDING, PENDING_IDENTITY_ATTRIBUTES, PENDING_IDENTITY_PREFIXES;
 var init_semconv = __esm({
   "src/semconv.ts"() {
     "use strict";
@@ -371,11 +371,13 @@ var init_semconv = __esm({
     GEN_AI_OPERATION_NAME = "gen_ai.operation.name";
     GEN_AI_PROVIDER_NAME = "gen_ai.provider.name";
     GEN_AI_REQUEST_MODEL = "gen_ai.request.model";
+    GEN_AI_REQUEST_REASONING_LEVEL = "gen_ai.request.reasoning.level";
     GEN_AI_RESPONSE_MODEL = "gen_ai.response.model";
     GEN_AI_USAGE_INPUT_TOKENS = "gen_ai.usage.input_tokens";
     GEN_AI_USAGE_OUTPUT_TOKENS = "gen_ai.usage.output_tokens";
     GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS = "gen_ai.usage.cache_read.input_tokens";
     GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS = "gen_ai.usage.cache_write.input_tokens";
+    GEN_AI_USAGE_REASONING_OUTPUT_TOKENS = "gen_ai.usage.reasoning.output_tokens";
     GEN_AI_INPUT_MESSAGES = "gen_ai.input.messages";
     GEN_AI_OUTPUT_MESSAGES = "gen_ai.output.messages";
     GEN_AI_RESPONSE_FINISH_REASONS = "gen_ai.response.finish_reasons";
@@ -1332,6 +1334,15 @@ init_semconv();
 init_serde();
 init_spans();
 var Generation = class extends Observation {
+  /**
+   * The provider passed at creation; drives the Anthropic input-token summing
+   * in {@link setUsage}. A bare `new Generation(span)` has none and never sums.
+   */
+  constructor(span, provider) {
+    super(span);
+    this.provider = provider;
+  }
+  provider;
   firstTokenRecorded = false;
   setInput(value) {
     this.span.setAttribute(GEN_AI_INPUT_MESSAGES, toAttributeValue(value));
@@ -1346,15 +1357,20 @@ var Generation = class extends Observation {
     return this;
   }
   /**
-   * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is.
-   * Per the GenAI conventions, `inputTokens` is the total including cached
-   * tokens (the cache counts are subsets of it); providers that report an
-   * exclusive `inputTokens` (e.g. Anthropic) are detected and normalized by
-   * the backend, so no client-side arithmetic is needed.
+   * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is;
+   * never pre-add anything. Per the GenAI conventions, `inputTokens` is the
+   * total including cached tokens (the cache counts are subsets of it).
+   * Anthropic's API reports `input_tokens` excluding the cache counts, and
+   * the conventions require the instrumentation to do the summing, so when
+   * the generation's provider is `"anthropic"` the emitted total is
+   * `inputTokens` plus both cache counts. Every other provider is recorded
+   * verbatim.
    */
   setUsage(usage) {
     if (usage.inputTokens !== void 0) {
-      this.span.setAttribute(GEN_AI_USAGE_INPUT_TOKENS, usage.inputTokens);
+      const sums = this.provider?.toLowerCase() === "anthropic";
+      const total = sums ? usage.inputTokens + (usage.cacheReadInputTokens ?? 0) + (usage.cacheWriteInputTokens ?? 0) : usage.inputTokens;
+      this.span.setAttribute(GEN_AI_USAGE_INPUT_TOKENS, total);
     }
     if (usage.outputTokens !== void 0) {
       this.span.setAttribute(GEN_AI_USAGE_OUTPUT_TOKENS, usage.outputTokens);
@@ -1365,6 +1381,9 @@ var Generation = class extends Observation {
     if (usage.cacheWriteInputTokens !== void 0) {
       this.span.setAttribute(GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS, usage.cacheWriteInputTokens);
     }
+    if (usage.reasoningOutputTokens !== void 0) {
+      this.span.setAttribute(GEN_AI_USAGE_REASONING_OUTPUT_TOKENS, usage.reasoningOutputTokens);
+    }
     return this;
   }
   /**
@@ -1404,17 +1423,20 @@ function configure2(generation, options) {
   for (const [key, value] of Object.entries(options.modelParameters ?? {})) {
     generation.setAttribute(`${GEN_AI_REQUEST_PREFIX}${key}`, value);
   }
+  if (options.reasoningLevel !== void 0) {
+    generation.setAttribute(GEN_AI_REQUEST_REASONING_LEVEL, options.reasoningLevel);
+  }
   if (options.input !== void 0) generation.setInput(options.input);
   return generation;
 }
 function startGeneration(name, options = {}) {
   const span = getTracer().startSpan(name, { attributes: attributesFor(options) });
-  return configure2(new Generation(span), options);
+  return configure2(new Generation(span, options.provider), options);
 }
 function startAsCurrentGeneration(name, optionsOrFn, maybeFn) {
   const [options, fn] = typeof optionsOrFn === "function" ? [{}, optionsOrFn] : [optionsOrFn, maybeFn];
   return getTracer().startActiveSpan(name, { attributes: attributesFor(options) }, async (span) => {
-    const generation = configure2(new Generation(span), options);
+    const generation = configure2(new Generation(span, options.provider), options);
     try {
       return await fn(generation);
     } catch (error) {
@@ -1451,7 +1473,7 @@ init_semconv();
 init_session();
 init_spans();
 init_workspace();
-var VERSION = "0.4.0";
+var VERSION = "0.5.0";
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   Generation,

package/dist/index.d.cts

@@ -223,19 +223,36 @@ interface GenerationOptions {
      * so use the provider's own parameter names.
      */
     modelParameters?: Record<string, unknown>;
+    /**
+     * Requested reasoning/thinking effort level
+     * (`gen_ai.request.reasoning.level`), e.g. OpenAI's `reasoning.effort`
+     * values. Provider-defined string, recorded verbatim. A first-class option
+     * because the `modelParameters` pass-through would spell the key
+     * `gen_ai.request.reasoning_level`, which is not the convention's name.
+     */
+    reasoningLevel?: string;
 }
 /** An LLM call. Content uses gen_ai message keys, never input.value. */
 declare class Generation extends Observation {
+    private readonly provider?;
     private firstTokenRecorded;
+    /**
+     * The provider passed at creation; drives the Anthropic input-token summing
+     * in {@link setUsage}. A bare `new Generation(span)` has none and never sums.
+     */
+    constructor(span: ConstructorParameters<typeof Observation>[0], provider?: string | undefined);
     setInput(value: unknown): this;
     setOutput(value: unknown): this;
     setModel(model: string): this;
     /**
-     * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is.
-     * Per the GenAI conventions, `inputTokens` is the total including cached
-     * tokens (the cache counts are subsets of it); providers that report an
-     * exclusive `inputTokens` (e.g. Anthropic) are detected and normalized by
-     * the backend, so no client-side arithmetic is needed.
+     * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is;
+     * never pre-add anything. Per the GenAI conventions, `inputTokens` is the
+     * total including cached tokens (the cache counts are subsets of it).
+     * Anthropic's API reports `input_tokens` excluding the cache counts, and
+     * the conventions require the instrumentation to do the summing, so when
+     * the generation's provider is `"anthropic"` the emitted total is
+     * `inputTokens` plus both cache counts. Every other provider is recorded
+     * verbatim.
      */
     setUsage(usage: {
         inputTokens?: number;
@@ -247,6 +264,13 @@ declare class Generation extends Observation {
          * (called "cache creation" by Anthropic).
          */
         cacheWriteInputTokens?: number;
+        /**
+         * Output tokens spent on reasoning / extended thinking. A subset of
+         * `outputTokens`, never in addition to it: providers already include
+         * reasoning tokens in the output total, so pass both as reported and
+         * do no arithmetic.
+         */
+        reasoningOutputTokens?: number;
     }): this;
     /**
      * Why generation stopped (`gen_ai.response.finish_reasons`), e.g. `"stop"`,
@@ -315,6 +339,6 @@ declare function observe<F extends (...args: never[]) => unknown>(fn: F, options
 declare function withSession<T>(fn: (sessionId: string) => T): T;
 declare function withSession<T>(sessionId: string, fn: (sessionId: string) => T): T;
 
-declare const VERSION = "0.4.0";
+declare const VERSION = "0.5.0";
 
 export { Generation, type GenerationBody, type GenerationOptions, type InitOptions, type Mask, Observation, type ObserveOptions, RiusClient, type RiusOptions, type SpanBody, SpanKind, type SpanOptions, VERSION, type WorkspaceExporterFactory, getTracer, init, observe, registerWorkspace, startAsCurrentGeneration, startAsCurrentSpan, startGeneration, startSpan, withSession, withWorkspace };

package/dist/index.d.ts

@@ -223,19 +223,36 @@ interface GenerationOptions {
      * so use the provider's own parameter names.
      */
     modelParameters?: Record<string, unknown>;
+    /**
+     * Requested reasoning/thinking effort level
+     * (`gen_ai.request.reasoning.level`), e.g. OpenAI's `reasoning.effort`
+     * values. Provider-defined string, recorded verbatim. A first-class option
+     * because the `modelParameters` pass-through would spell the key
+     * `gen_ai.request.reasoning_level`, which is not the convention's name.
+     */
+    reasoningLevel?: string;
 }
 /** An LLM call. Content uses gen_ai message keys, never input.value. */
 declare class Generation extends Observation {
+    private readonly provider?;
     private firstTokenRecorded;
+    /**
+     * The provider passed at creation; drives the Anthropic input-token summing
+     * in {@link setUsage}. A bare `new Generation(span)` has none and never sums.
+     */
+    constructor(span: ConstructorParameters<typeof Observation>[0], provider?: string | undefined);
     setInput(value: unknown): this;
     setOutput(value: unknown): this;
     setModel(model: string): this;
     /**
-     * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is.
-     * Per the GenAI conventions, `inputTokens` is the total including cached
-     * tokens (the cache counts are subsets of it); providers that report an
-     * exclusive `inputTokens` (e.g. Anthropic) are detected and normalized by
-     * the backend, so no client-side arithmetic is needed.
+     * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is;
+     * never pre-add anything. Per the GenAI conventions, `inputTokens` is the
+     * total including cached tokens (the cache counts are subsets of it).
+     * Anthropic's API reports `input_tokens` excluding the cache counts, and
+     * the conventions require the instrumentation to do the summing, so when
+     * the generation's provider is `"anthropic"` the emitted total is
+     * `inputTokens` plus both cache counts. Every other provider is recorded
+     * verbatim.
      */
     setUsage(usage: {
         inputTokens?: number;
@@ -247,6 +264,13 @@ declare class Generation extends Observation {
          * (called "cache creation" by Anthropic).
          */
         cacheWriteInputTokens?: number;
+        /**
+         * Output tokens spent on reasoning / extended thinking. A subset of
+         * `outputTokens`, never in addition to it: providers already include
+         * reasoning tokens in the output total, so pass both as reported and
+         * do no arithmetic.
+         */
+        reasoningOutputTokens?: number;
     }): this;
     /**
      * Why generation stopped (`gen_ai.response.finish_reasons`), e.g. `"stop"`,
@@ -315,6 +339,6 @@ declare function observe<F extends (...args: never[]) => unknown>(fn: F, options
 declare function withSession<T>(fn: (sessionId: string) => T): T;
 declare function withSession<T>(sessionId: string, fn: (sessionId: string) => T): T;
 
-declare const VERSION = "0.4.0";
+declare const VERSION = "0.5.0";
 
 export { Generation, type GenerationBody, type GenerationOptions, type InitOptions, type Mask, Observation, type ObserveOptions, RiusClient, type RiusOptions, type SpanBody, SpanKind, type SpanOptions, VERSION, type WorkspaceExporterFactory, getTracer, init, observe, registerWorkspace, startAsCurrentGeneration, startAsCurrentSpan, startGeneration, startSpan, withSession, withWorkspace };

package/dist/index.js

@@ -343,7 +343,7 @@ function kindAttributes(kind) {
   if (operation !== void 0) attributes[GEN_AI_OPERATION_NAME] = operation;
   return attributes;
 }
-var TRACER_NAME, SERVICE_INSTANCE_ID, OPENINFERENCE_SPAN_KIND, INPUT_VALUE, OUTPUT_VALUE, SESSION_ID, WORKSPACE_ROUTE, GEN_AI_OPERATION_NAME, GEN_AI_PROVIDER_NAME, GEN_AI_REQUEST_MODEL, GEN_AI_RESPONSE_MODEL, GEN_AI_USAGE_INPUT_TOKENS, GEN_AI_USAGE_OUTPUT_TOKENS, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS, GEN_AI_INPUT_MESSAGES, GEN_AI_OUTPUT_MESSAGES, GEN_AI_RESPONSE_FINISH_REASONS, GEN_AI_TOOL_NAME, GEN_AI_REQUEST_PREFIX, MCP_RESULT_TYPE, GEN_AI_FIRST_TOKEN_EVENT, SpanKind, OPERATION_BY_KIND, CONTENT_ATTRIBUTES, CONTENT_ATTRIBUTE_PREFIXES, CONTENT_ATTRIBUTE_SUFFIXES, GLASSFLOW_SPAN_PENDING, PENDING_IDENTITY_ATTRIBUTES, PENDING_IDENTITY_PREFIXES;
+var TRACER_NAME, SERVICE_INSTANCE_ID, OPENINFERENCE_SPAN_KIND, INPUT_VALUE, OUTPUT_VALUE, SESSION_ID, WORKSPACE_ROUTE, GEN_AI_OPERATION_NAME, GEN_AI_PROVIDER_NAME, GEN_AI_REQUEST_MODEL, GEN_AI_REQUEST_REASONING_LEVEL, GEN_AI_RESPONSE_MODEL, GEN_AI_USAGE_INPUT_TOKENS, GEN_AI_USAGE_OUTPUT_TOKENS, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS, GEN_AI_USAGE_REASONING_OUTPUT_TOKENS, GEN_AI_INPUT_MESSAGES, GEN_AI_OUTPUT_MESSAGES, GEN_AI_RESPONSE_FINISH_REASONS, GEN_AI_TOOL_NAME, GEN_AI_REQUEST_PREFIX, MCP_RESULT_TYPE, GEN_AI_FIRST_TOKEN_EVENT, SpanKind, OPERATION_BY_KIND, CONTENT_ATTRIBUTES, CONTENT_ATTRIBUTE_PREFIXES, CONTENT_ATTRIBUTE_SUFFIXES, GLASSFLOW_SPAN_PENDING, PENDING_IDENTITY_ATTRIBUTES, PENDING_IDENTITY_PREFIXES;
 var init_semconv = __esm({
   "src/semconv.ts"() {
     "use strict";
@@ -358,11 +358,13 @@ var init_semconv = __esm({
     GEN_AI_OPERATION_NAME = "gen_ai.operation.name";
     GEN_AI_PROVIDER_NAME = "gen_ai.provider.name";
     GEN_AI_REQUEST_MODEL = "gen_ai.request.model";
+    GEN_AI_REQUEST_REASONING_LEVEL = "gen_ai.request.reasoning.level";
     GEN_AI_RESPONSE_MODEL = "gen_ai.response.model";
     GEN_AI_USAGE_INPUT_TOKENS = "gen_ai.usage.input_tokens";
     GEN_AI_USAGE_OUTPUT_TOKENS = "gen_ai.usage.output_tokens";
     GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS = "gen_ai.usage.cache_read.input_tokens";
     GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS = "gen_ai.usage.cache_write.input_tokens";
+    GEN_AI_USAGE_REASONING_OUTPUT_TOKENS = "gen_ai.usage.reasoning.output_tokens";
     GEN_AI_INPUT_MESSAGES = "gen_ai.input.messages";
     GEN_AI_OUTPUT_MESSAGES = "gen_ai.output.messages";
     GEN_AI_RESPONSE_FINISH_REASONS = "gen_ai.response.finish_reasons";
@@ -1304,6 +1306,15 @@ init_semconv();
 init_serde();
 init_spans();
 var Generation = class extends Observation {
+  /**
+   * The provider passed at creation; drives the Anthropic input-token summing
+   * in {@link setUsage}. A bare `new Generation(span)` has none and never sums.
+   */
+  constructor(span, provider) {
+    super(span);
+    this.provider = provider;
+  }
+  provider;
   firstTokenRecorded = false;
   setInput(value) {
     this.span.setAttribute(GEN_AI_INPUT_MESSAGES, toAttributeValue(value));
@@ -1318,15 +1329,20 @@ var Generation = class extends Observation {
     return this;
   }
   /**
-   * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is.
-   * Per the GenAI conventions, `inputTokens` is the total including cached
-   * tokens (the cache counts are subsets of it); providers that report an
-   * exclusive `inputTokens` (e.g. Anthropic) are detected and normalized by
-   * the backend, so no client-side arithmetic is needed.
+   * Token usage (`gen_ai.usage.*`). Pass provider-reported values as-is;
+   * never pre-add anything. Per the GenAI conventions, `inputTokens` is the
+   * total including cached tokens (the cache counts are subsets of it).
+   * Anthropic's API reports `input_tokens` excluding the cache counts, and
+   * the conventions require the instrumentation to do the summing, so when
+   * the generation's provider is `"anthropic"` the emitted total is
+   * `inputTokens` plus both cache counts. Every other provider is recorded
+   * verbatim.
    */
   setUsage(usage) {
     if (usage.inputTokens !== void 0) {
-      this.span.setAttribute(GEN_AI_USAGE_INPUT_TOKENS, usage.inputTokens);
+      const sums = this.provider?.toLowerCase() === "anthropic";
+      const total = sums ? usage.inputTokens + (usage.cacheReadInputTokens ?? 0) + (usage.cacheWriteInputTokens ?? 0) : usage.inputTokens;
+      this.span.setAttribute(GEN_AI_USAGE_INPUT_TOKENS, total);
     }
     if (usage.outputTokens !== void 0) {
       this.span.setAttribute(GEN_AI_USAGE_OUTPUT_TOKENS, usage.outputTokens);
@@ -1337,6 +1353,9 @@ var Generation = class extends Observation {
     if (usage.cacheWriteInputTokens !== void 0) {
       this.span.setAttribute(GEN_AI_USAGE_CACHE_WRITE_INPUT_TOKENS, usage.cacheWriteInputTokens);
     }
+    if (usage.reasoningOutputTokens !== void 0) {
+      this.span.setAttribute(GEN_AI_USAGE_REASONING_OUTPUT_TOKENS, usage.reasoningOutputTokens);
+    }
     return this;
   }
   /**
@@ -1376,17 +1395,20 @@ function configure2(generation, options) {
   for (const [key, value] of Object.entries(options.modelParameters ?? {})) {
     generation.setAttribute(`${GEN_AI_REQUEST_PREFIX}${key}`, value);
   }
+  if (options.reasoningLevel !== void 0) {
+    generation.setAttribute(GEN_AI_REQUEST_REASONING_LEVEL, options.reasoningLevel);
+  }
   if (options.input !== void 0) generation.setInput(options.input);
   return generation;
 }
 function startGeneration(name, options = {}) {
   const span = getTracer().startSpan(name, { attributes: attributesFor(options) });
-  return configure2(new Generation(span), options);
+  return configure2(new Generation(span, options.provider), options);
 }
 function startAsCurrentGeneration(name, optionsOrFn, maybeFn) {
   const [options, fn] = typeof optionsOrFn === "function" ? [{}, optionsOrFn] : [optionsOrFn, maybeFn];
   return getTracer().startActiveSpan(name, { attributes: attributesFor(options) }, async (span) => {
-    const generation = configure2(new Generation(span), options);
+    const generation = configure2(new Generation(span, options.provider), options);
     try {
       return await fn(generation);
     } catch (error) {
@@ -1423,7 +1445,7 @@ init_semconv();
 init_session();
 init_spans();
 init_workspace();
-var VERSION = "0.4.0";
+var VERSION = "0.5.0";
 export {
   Generation,
   Observation,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@glassflow-ai/rius",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "OpenTelemetry-native tracing for AI agents and LLM applications",
   "keywords": [
     "opentelemetry",