@yagni-app/code-staging 1.0.7-staging.1265.1 → 1.0.7-staging.1267.1

package/README.md

@@ -183,14 +183,20 @@ prefix instead of `claude_code`:
   `yagni_code.llm_request` and `yagni_code.tool`, carrying both the flat
   Claude Code attributes (`model`, `input_tokens`, `cost_usd`, `tool_name`)
   and the OTel GenAI conventions (`gen_ai.*`) that LLM-observability
-  products read. The two count tokens differently on purpose: the flat
+  products read. `gen_ai.operation.name` classifies each span the way those
+  products expect: `invoke_agent` on the interaction root, `chat` on the LLM
+  request, `execute_tool` on tool spans (the turn span has no semconv
+  operation and is left unset). The two count tokens differently on purpose: the flat
   `input_tokens` / `cache_read_tokens` / `cache_creation_tokens` are disjoint,
   exactly as Claude Code reports them, while `gen_ai.usage.input_tokens`
   follows the semconv and includes the cached tokens, with
   `gen_ai.usage.cache_read.input_tokens` and
   `gen_ai.usage.cache_write.input_tokens` as subsets of it.
   `gen_ai.cost.estimated_total` carries the same sell-rate cost as `cost_usd`
-  so vendor cost views work for the opaque tier id.
+  so vendor cost views work for the opaque tier id. It is an integer in
+  nanodollars (`cost_usd * 1e9`): Datadog reads that attribute unscaled into
+  its nanodollar cost metric, so a dollar value shows up a billion times too
+  small.
 - **Metrics:** `yagni_code.session.count`, `token.usage` (by `type` and
   `model`), `cost.usage`, `lines_of_code.count`, `commit.count`,
   `pull_request.count`, `code_edit_tool.decision`, `active_time.total`.

package/dist/extension/telemetry/attrs.d.ts

@@ -52,7 +52,9 @@ export declare const ATTR_GEN_AI_CACHE_READ_TOKENS_LEGACY = "gen_ai.usage.cache_
 export declare const ATTR_GEN_AI_CACHE_CREATION_TOKENS_LEGACY = "gen_ai.usage.cache_creation_input_tokens";
 /** The one cost attribute Datadog's OTLP ingestion documents. Carries the
  * same sell-rate figure as the flat `cost_usd`, so the vendor Cost view
- * populates for the opaque tier id, which no vendor price table knows. */
+ * populates for the opaque tier id, which no vendor price table knows.
+ * Unit is NANODOLLARS (integer), not dollars: Datadog copies the value
+ * unscaled into its nanodollar cost metric. */
 export declare const ATTR_GEN_AI_COST_ESTIMATED_TOTAL = "gen_ai.cost.estimated_total";
 export declare const ATTR_GEN_AI_CONVERSATION_ID = "gen_ai.conversation.id";
 export declare const ATTR_GEN_AI_AGENT_NAME = "gen_ai.agent.name";

package/dist/extension/telemetry/attrs.js

@@ -54,7 +54,9 @@ export const ATTR_GEN_AI_CACHE_READ_TOKENS_LEGACY = "gen_ai.usage.cache_read_inp
 export const ATTR_GEN_AI_CACHE_CREATION_TOKENS_LEGACY = "gen_ai.usage.cache_creation_input_tokens";
 /** The one cost attribute Datadog's OTLP ingestion documents. Carries the
  * same sell-rate figure as the flat `cost_usd`, so the vendor Cost view
- * populates for the opaque tier id, which no vendor price table knows. */
+ * populates for the opaque tier id, which no vendor price table knows.
+ * Unit is NANODOLLARS (integer), not dollars: Datadog copies the value
+ * unscaled into its nanodollar cost metric. */
 export const ATTR_GEN_AI_COST_ESTIMATED_TOTAL = "gen_ai.cost.estimated_total";
 export const ATTR_GEN_AI_CONVERSATION_ID = "gen_ai.conversation.id";
 export const ATTR_GEN_AI_AGENT_NAME = "gen_ai.agent.name";

package/dist/extension/telemetry/tracker.js

@@ -168,9 +168,13 @@ export class SessionTelemetry {
         this.interactionSequence += 1;
         const promptId = randomUUID();
         const promptLength = typeof prompt === "string" ? prompt.length : 0;
+        // `invoke_agent` is the semconv operation for the root of an agent run.
+        // Without it Datadog files the trace as a generic workflow, flags an
+        // instrumentation issue on the span, and leaves it out of per-agent views.
         const span = this.tracer.startSpan(SPAN_INTERACTION, {
             attributes: {
                 ...this.spanAttributes(),
+                [ATTR_GEN_AI_OPERATION_NAME]: "invoke_agent",
                 user_prompt_length: promptLength,
                 "interaction.sequence": this.interactionSequence,
                 "prompt.id": promptId,
@@ -301,8 +305,11 @@ export class SessionTelemetry {
         span.setAttribute(ATTR_GEN_AI_CACHE_WRITE_TOKENS, cacheCreation);
         span.setAttribute(ATTR_GEN_AI_CACHE_READ_TOKENS_LEGACY, cacheRead);
         span.setAttribute(ATTR_GEN_AI_CACHE_CREATION_TOKENS_LEGACY, cacheCreation);
+        // Datadog ingests this attribute in nanodollars (its `ml_obs.span.llm.total.cost`
+        // unit); sending dollars renders a $0.012 call as 1.2e-9 cents. Verified against
+        // a live span on 2026-09-02. `cost_usd` below stays in dollars.
         if (costKnown)
-            span.setAttribute(ATTR_GEN_AI_COST_ESTIMATED_TOTAL, cost);
+            span.setAttribute(ATTR_GEN_AI_COST_ESTIMATED_TOTAL, Math.round(cost * 1_000_000_000));
         span.setAttribute("input_tokens", input);
         span.setAttribute("output_tokens", output);
         span.setAttribute("cache_read_tokens", cacheRead);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yagni-app/code-staging",
-  "version": "1.0.7-staging.1265.1",
+  "version": "1.0.7-staging.1267.1",
   "description": "YAGNI Code: a terminal coding agent that already knows your company. One YAGNI login routes the model and grounds the agent in your team's context.",
   "license": "SEE LICENSE IN LICENSE.md",
   "author": "YAGNI, Inc. <jack@yagni.app> (https://yagni.app)",
@@ -58,5 +58,5 @@
     "turndown": "^7.2.4",
     "typebox": "^1.3.15"
   },
-  "yagniSourceSha": "996a0867c1bc4859705f5aca33cb5d4e67bd154b"
+  "yagniSourceSha": "594ebbbef9ba1c2233382184dad54d1e81a3bc0f"
 }