@poncho-ai/harness 0.61.1 → 0.61.2

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.61.1 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.61.2 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
 > node scripts/embed-docs.js && tsup src/index.ts --format esm --dts
 
 [embed-docs] Generated poncho-docs.ts with 4 topics
@@ -8,9 +8,9 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js            580.50 KB
+ESM dist/index.js            581.91 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 227ms
+ESM ⚡️ Build success in 239ms
 DTS Build start
-DTS ⚡️ Build success in 7778ms
+DTS ⚡️ Build success in 7703ms
 DTS dist/index.d.ts 110.56 KB

package/CHANGELOG.md

@@ -1,5 +1,23 @@
 # @poncho-ai/harness
 
+## 0.61.2
+
+### Patch Changes
+
+- [`c881cd2`](https://github.com/cesr/poncho-ai/commit/c881cd2cafd4cb70f9614601dd233aafac334b8c) Thanks [@cesr](https://github.com/cesr)! - Telemetry: make context compaction observable and tag subagent runs.
+  - When auto-compaction fires mid-run, stamp the active `invoke_agent` span with
+    `poncho.compaction.occurred` / `tokens_before` / `tokens_after` /
+    `context_window` / `trigger`. Compaction rewrites the message prefix — which
+    wipes the per-model prompt cache and forces the next turn to re-create a large
+    cached span cold — so "which turns compacted, and what did that cost" was
+    previously invisible at the span level. No-op when telemetry is suppressed
+    (no active span).
+  - Subagent runs spawned by the orchestrator now carry
+    `telemetryAttributes` (`poncho.run.kind: "subagent"`, `latitude.tags`,
+    and a `latitude.metadata` link to the parent conversation) so subagent spend
+    segments out of the parent chat/job on the trace instead of being
+    indistinguishable from it.
+
 ## 0.61.1
 
 ### Patch Changes

package/dist/index.js

@@ -11461,6 +11461,13 @@ ${textContent}` };
                 const tokensAfterCompaction = estimateTotalTokens(systemPrompt, messages, toolDefsJsonForEstimate);
                 latestContextTokens = tokensAfterCompaction;
                 toolOutputEstimateSinceModel = 0;
+                trace.getActiveSpan()?.setAttributes({
+                  "poncho.compaction.occurred": true,
+                  "poncho.compaction.tokens_before": effectiveTokens,
+                  "poncho.compaction.tokens_after": tokensAfterCompaction,
+                  "poncho.compaction.context_window": contextWindow,
+                  "poncho.compaction.trigger": compactionConfig.trigger
+                });
                 yield pushEvent({
                   type: "compaction:completed",
                   tokensBefore: effectiveTokens,
@@ -13707,7 +13714,17 @@ var AgentOrchestrator = class {
         messages: harnessMessages,
         abortSignal: childAbortController.signal,
         // Inherit the parent run's telemetry choice (e.g. incognito).
-        suppressTelemetry: conversation.subagentMeta?.suppressTelemetry
+        suppressTelemetry: conversation.subagentMeta?.suppressTelemetry,
+        // Distinguish subagent turns from the parent chat/job on the trace:
+        // `poncho.run.kind` is the queryable attribute PonchOS already sets on
+        // foreground/job runs, and `latitude.tags` (JSON-array string) is what
+        // Latitude reads into its filterable tags field, so subagent spend can
+        // be segmented out. `latitude.metadata` links back to the parent.
+        telemetryAttributes: {
+          "poncho.run.kind": "subagent",
+          "latitude.tags": '["subagent"]',
+          ...conversation.parentConversationId ? { "latitude.metadata": JSON.stringify({ parentConversationId: conversation.parentConversationId }) } : {}
+        }
       })) {
         if (event.type === "run:started") {
           latestRunId = event.runId;
@@ -14200,7 +14217,14 @@ ${resultBody}`,
         messages: continuationMessages,
         abortSignal: childAbortController.signal,
         // Inherit the parent run's telemetry choice (e.g. incognito).
-        suppressTelemetry: conversation.subagentMeta?.suppressTelemetry
+        suppressTelemetry: conversation.subagentMeta?.suppressTelemetry,
+        // Tag the subagent continuation the same way as the initial run (see
+        // the fan-out call above) so its spend segments out of the parent's.
+        telemetryAttributes: {
+          "poncho.run.kind": "subagent",
+          "latitude.tags": '["subagent"]',
+          "latitude.metadata": JSON.stringify({ parentConversationId })
+        }
       })) {
         if (event.type === "run:started") {
           const active = this.activeConversationRuns.get(conversationId);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.61.1",
+  "version": "0.61.2",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/harness.ts

@@ -3115,6 +3115,21 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
               const tokensAfterCompaction = estimateTotalTokens(systemPrompt, messages, toolDefsJsonForEstimate);
               latestContextTokens = tokensAfterCompaction;
               toolOutputEstimateSinceModel = 0;
+              // Stamp the compaction onto the active OTel span (the invoke_agent
+              // root, active here because run() executes inside its context) so
+              // observability backends can find compacted turns and measure the
+              // cost they cause — a compaction rewrites the message prefix, which
+              // wipes the per-model prompt cache and forces the next turn to
+              // re-create a large cached span cold. Without this, "which turns
+              // compacted, and what did it cost" is invisible at the span level.
+              // No-op when telemetry is off (no active span).
+              trace.getActiveSpan()?.setAttributes({
+                "poncho.compaction.occurred": true,
+                "poncho.compaction.tokens_before": effectiveTokens,
+                "poncho.compaction.tokens_after": tokensAfterCompaction,
+                "poncho.compaction.context_window": contextWindow,
+                "poncho.compaction.trigger": compactionConfig.trigger,
+              });
               yield pushEvent({
                 type: "compaction:completed",
                 tokensBefore: effectiveTokens,

package/src/orchestrator/orchestrator.ts

@@ -862,6 +862,18 @@ export class AgentOrchestrator {
         abortSignal: childAbortController.signal,
         // Inherit the parent run's telemetry choice (e.g. incognito).
         suppressTelemetry: conversation.subagentMeta?.suppressTelemetry,
+        // Distinguish subagent turns from the parent chat/job on the trace:
+        // `poncho.run.kind` is the queryable attribute PonchOS already sets on
+        // foreground/job runs, and `latitude.tags` (JSON-array string) is what
+        // Latitude reads into its filterable tags field, so subagent spend can
+        // be segmented out. `latitude.metadata` links back to the parent.
+        telemetryAttributes: {
+          "poncho.run.kind": "subagent",
+          "latitude.tags": "[\"subagent\"]",
+          ...(conversation.parentConversationId
+            ? { "latitude.metadata": JSON.stringify({ parentConversationId: conversation.parentConversationId }) }
+            : {}),
+        },
       })) {
         if (event.type === "run:started") {
           latestRunId = event.runId;
@@ -1452,6 +1464,13 @@ export class AgentOrchestrator {
         abortSignal: childAbortController.signal,
         // Inherit the parent run's telemetry choice (e.g. incognito).
         suppressTelemetry: conversation.subagentMeta?.suppressTelemetry,
+        // Tag the subagent continuation the same way as the initial run (see
+        // the fan-out call above) so its spend segments out of the parent's.
+        telemetryAttributes: {
+          "poncho.run.kind": "subagent",
+          "latitude.tags": "[\"subagent\"]",
+          "latitude.metadata": JSON.stringify({ parentConversationId }),
+        },
       })) {
         if (event.type === "run:started") {
           const active = this.activeConversationRuns.get(conversationId);