@yagni-app/code-staging 1.0.6-staging.1261.1 → 1.0.7-staging.1265.1

package/README.md

@@ -183,7 +183,14 @@ 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.
+  products read. 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.
 - **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

@@ -28,10 +28,32 @@ export declare const ATTR_GEN_AI_REQUEST_MODEL = "gen_ai.request.model";
 export declare const ATTR_GEN_AI_RESPONSE_MODEL = "gen_ai.response.model";
 export declare const ATTR_GEN_AI_RESPONSE_ID = "gen_ai.response.id";
 export declare const ATTR_GEN_AI_FINISH_REASONS = "gen_ai.response.finish_reasons";
+/**
+ * Token usage. The semconv registry says `gen_ai.usage.input_tokens` "SHOULD
+ * include all types of input tokens, including cached tokens", and defines the
+ * cache counters as subsets of it. pi reports the Anthropic way (non-cached
+ * input, cache read, cache write as three disjoint numbers), so the tracker
+ * adds them up for the `gen_ai.*` layer. Consumers such as Datadog LLM
+ * Observability divide cache_read by input_tokens; feeding them the disjoint
+ * number produced cache ratios in the tens of thousands of percent.
+ *
+ * The flat Claude Code attributes (`input_tokens`, `cache_read_tokens`, …) and
+ * the `<prefix>.token.usage` metric deliberately keep Claude Code's disjoint
+ * semantics so cross-harness comparisons stay apples to apples.
+ */
 export declare const ATTR_GEN_AI_INPUT_TOKENS = "gen_ai.usage.input_tokens";
 export declare const ATTR_GEN_AI_OUTPUT_TOKENS = "gen_ai.usage.output_tokens";
-export declare const ATTR_GEN_AI_CACHE_READ_TOKENS = "gen_ai.usage.cache_read_input_tokens";
-export declare const ATTR_GEN_AI_CACHE_CREATION_TOKENS = "gen_ai.usage.cache_creation_input_tokens";
+export declare const ATTR_GEN_AI_TOTAL_TOKENS = "gen_ai.usage.total_tokens";
+export declare const ATTR_GEN_AI_CACHE_READ_TOKENS = "gen_ai.usage.cache_read.input_tokens";
+export declare const ATTR_GEN_AI_CACHE_WRITE_TOKENS = "gen_ai.usage.cache_write.input_tokens";
+/** Pre-semconv spellings emitted through 1.0.x; kept for one release so
+ * dashboards keyed on them keep working. */
+export declare const ATTR_GEN_AI_CACHE_READ_TOKENS_LEGACY = "gen_ai.usage.cache_read_input_tokens";
+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. */
+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";
 export declare const ATTR_GEN_AI_TOOL_NAME = "gen_ai.tool.name";

package/dist/extension/telemetry/attrs.js

@@ -30,10 +30,32 @@ export const ATTR_GEN_AI_REQUEST_MODEL = "gen_ai.request.model";
 export const ATTR_GEN_AI_RESPONSE_MODEL = "gen_ai.response.model";
 export const ATTR_GEN_AI_RESPONSE_ID = "gen_ai.response.id";
 export const ATTR_GEN_AI_FINISH_REASONS = "gen_ai.response.finish_reasons";
+/**
+ * Token usage. The semconv registry says `gen_ai.usage.input_tokens` "SHOULD
+ * include all types of input tokens, including cached tokens", and defines the
+ * cache counters as subsets of it. pi reports the Anthropic way (non-cached
+ * input, cache read, cache write as three disjoint numbers), so the tracker
+ * adds them up for the `gen_ai.*` layer. Consumers such as Datadog LLM
+ * Observability divide cache_read by input_tokens; feeding them the disjoint
+ * number produced cache ratios in the tens of thousands of percent.
+ *
+ * The flat Claude Code attributes (`input_tokens`, `cache_read_tokens`, …) and
+ * the `<prefix>.token.usage` metric deliberately keep Claude Code's disjoint
+ * semantics so cross-harness comparisons stay apples to apples.
+ */
 export const ATTR_GEN_AI_INPUT_TOKENS = "gen_ai.usage.input_tokens";
 export const ATTR_GEN_AI_OUTPUT_TOKENS = "gen_ai.usage.output_tokens";
-export const ATTR_GEN_AI_CACHE_READ_TOKENS = "gen_ai.usage.cache_read_input_tokens";
-export const ATTR_GEN_AI_CACHE_CREATION_TOKENS = "gen_ai.usage.cache_creation_input_tokens";
+export const ATTR_GEN_AI_TOTAL_TOKENS = "gen_ai.usage.total_tokens";
+export const ATTR_GEN_AI_CACHE_READ_TOKENS = "gen_ai.usage.cache_read.input_tokens";
+export const ATTR_GEN_AI_CACHE_WRITE_TOKENS = "gen_ai.usage.cache_write.input_tokens";
+/** Pre-semconv spellings emitted through 1.0.x; kept for one release so
+ * dashboards keyed on them keep working. */
+export const ATTR_GEN_AI_CACHE_READ_TOKENS_LEGACY = "gen_ai.usage.cache_read_input_tokens";
+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. */
+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";
 export const ATTR_GEN_AI_TOOL_NAME = "gen_ai.tool.name";

package/dist/extension/telemetry/tracker.js

@@ -20,7 +20,7 @@
 import { randomUUID } from "node:crypto";
 import { context as otelContext, SpanStatusCode, trace, } from "@opentelemetry/api";
 import { SeverityNumber } from "@opentelemetry/api-logs";
-import { ATTR_APP_ENTRYPOINT, ATTR_APP_VERSION, ATTR_ERROR_TYPE, ATTR_GEN_AI_AGENT_NAME, ATTR_GEN_AI_CACHE_CREATION_TOKENS, ATTR_GEN_AI_CACHE_READ_TOKENS, ATTR_GEN_AI_CONVERSATION_ID, ATTR_GEN_AI_FINISH_REASONS, ATTR_GEN_AI_INPUT_TOKENS, ATTR_GEN_AI_OPERATION_NAME, ATTR_GEN_AI_OUTPUT_TOKENS, ATTR_GEN_AI_PROVIDER_NAME, ATTR_GEN_AI_REQUEST_MODEL, ATTR_GEN_AI_RESPONSE_ID, ATTR_GEN_AI_RESPONSE_MODEL, ATTR_GEN_AI_SYSTEM, ATTR_GEN_AI_TOOL_CALL_ID, ATTR_GEN_AI_TOOL_NAME, ATTR_GEN_AI_TOOL_TYPE, ATTR_HTTP_STATUS_CODE, ATTR_ORGANIZATION_ID, ATTR_SESSION_ID, ATTR_TERMINAL_TYPE, ATTR_USER_EMAIL, EVENT_API_ERROR, EVENT_API_REQUEST, EVENT_ASSISTANT_RESPONSE, EVENT_PERMISSION_MODE_CHANGED, EVENT_TOOL_DECISION, EVENT_TOOL_RESULT, EVENT_USER_PROMPT, GEN_AI_PROVIDER, languageFromPath, METRIC_ACTIVE_TIME, METRIC_CODE_EDIT_DECISION, METRIC_COMMIT_COUNT, METRIC_COST_USAGE, METRIC_LINES_OF_CODE, METRIC_PULL_REQUEST_COUNT, METRIC_SESSION_COUNT, METRIC_TOKEN_USAGE, PREFIX, SPAN_INTERACTION, SPAN_LLM_REQUEST, SPAN_TOOL, SPAN_TURN, } from "./attrs.js";
+import { ATTR_APP_ENTRYPOINT, ATTR_APP_VERSION, ATTR_ERROR_TYPE, ATTR_GEN_AI_AGENT_NAME, ATTR_GEN_AI_CACHE_CREATION_TOKENS_LEGACY, ATTR_GEN_AI_CACHE_READ_TOKENS, ATTR_GEN_AI_CACHE_READ_TOKENS_LEGACY, ATTR_GEN_AI_CACHE_WRITE_TOKENS, ATTR_GEN_AI_CONVERSATION_ID, ATTR_GEN_AI_COST_ESTIMATED_TOTAL, ATTR_GEN_AI_FINISH_REASONS, ATTR_GEN_AI_INPUT_TOKENS, ATTR_GEN_AI_OPERATION_NAME, ATTR_GEN_AI_OUTPUT_TOKENS, ATTR_GEN_AI_TOTAL_TOKENS, ATTR_GEN_AI_PROVIDER_NAME, ATTR_GEN_AI_REQUEST_MODEL, ATTR_GEN_AI_RESPONSE_ID, ATTR_GEN_AI_RESPONSE_MODEL, ATTR_GEN_AI_SYSTEM, ATTR_GEN_AI_TOOL_CALL_ID, ATTR_GEN_AI_TOOL_NAME, ATTR_GEN_AI_TOOL_TYPE, ATTR_HTTP_STATUS_CODE, ATTR_ORGANIZATION_ID, ATTR_SESSION_ID, ATTR_TERMINAL_TYPE, ATTR_USER_EMAIL, EVENT_API_ERROR, EVENT_API_REQUEST, EVENT_ASSISTANT_RESPONSE, EVENT_PERMISSION_MODE_CHANGED, EVENT_TOOL_DECISION, EVENT_TOOL_RESULT, EVENT_USER_PROMPT, GEN_AI_PROVIDER, languageFromPath, METRIC_ACTIVE_TIME, METRIC_CODE_EDIT_DECISION, METRIC_COMMIT_COUNT, METRIC_COST_USAGE, METRIC_LINES_OF_CODE, METRIC_PULL_REQUEST_COUNT, METRIC_SESSION_COUNT, METRIC_TOKEN_USAGE, PREFIX, SPAN_INTERACTION, SPAN_LLM_REQUEST, SPAN_TOOL, SPAN_TURN, } from "./attrs.js";
 /** Idle cutoff for user active time: gaps longer than this are not "active". */
 export const USER_ACTIVE_IDLE_CUTOFF_MS = 5 * 60 * 1000;
 const EDIT_TOOLS = new Set(["edit", "write", "multi_edit", "notebook_edit"]);
@@ -276,19 +276,33 @@ export class SessionTelemetry {
         const output = num(usage.output);
         const cacheRead = num(usage.cacheRead);
         const cacheCreation = num(usage.cacheWrite);
+        // Claude Code always reports cost_usd (0 when it has nothing), so the flat
+        // attribute and the counter keep that shape. The vendor cost attribute is
+        // different: a 0 there renders as "$0" in cost views, so it is only set
+        // when pi actually priced the request.
+        const costKnown = typeof usage.cost === "number" || typeof usage.cost?.total === "number";
         const cost = typeof usage.cost === "number" ? usage.cost : num(usage.cost?.total);
         const stop = message.stopReason ?? "stop";
         const isError = stop === "error";
         const requestId = message.responseId ?? slot.requestId;
+        // pi's `input` excludes cached tokens (Anthropic semantics). The semconv
+        // `gen_ai.usage.input_tokens` includes them, with the cache counters as
+        // subsets; see attrs.ts. The flat Claude Code attributes below stay disjoint.
+        const inclusiveInput = input + cacheRead + cacheCreation;
         const span = slot.span;
         if (message.model)
             span.setAttribute(ATTR_GEN_AI_REQUEST_MODEL, message.model);
         span.setAttribute(ATTR_GEN_AI_RESPONSE_MODEL, model);
         span.setAttribute("model", model);
-        span.setAttribute(ATTR_GEN_AI_INPUT_TOKENS, input);
+        span.setAttribute(ATTR_GEN_AI_INPUT_TOKENS, inclusiveInput);
         span.setAttribute(ATTR_GEN_AI_OUTPUT_TOKENS, output);
+        span.setAttribute(ATTR_GEN_AI_TOTAL_TOKENS, inclusiveInput + output);
         span.setAttribute(ATTR_GEN_AI_CACHE_READ_TOKENS, cacheRead);
-        span.setAttribute(ATTR_GEN_AI_CACHE_CREATION_TOKENS, cacheCreation);
+        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);
+        if (costKnown)
+            span.setAttribute(ATTR_GEN_AI_COST_ESTIMATED_TOTAL, cost);
         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": "1.0.6-staging.1261.1",
+  "version": "1.0.7-staging.1265.1",
   "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": "4f16be93d0007e813ca15b60ac80c1ce08fd04b4"
+  "yagniSourceSha": "996a0867c1bc4859705f5aca33cb5d4e67bd154b"
 }