@poncho-ai/harness 0.59.9 → 0.59.11

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.59.9 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.59.11 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            558.75 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 233ms
+ESM dist/index.js            560.62 KB
+ESM ⚡️ Build success in 239ms
 DTS Build start
-DTS ⚡️ Build success in 7974ms
+DTS ⚡️ Build success in 7569ms
 DTS dist/index.d.ts 102.06 KB

package/CHANGELOG.md

@@ -1,5 +1,28 @@
 # @poncho-ai/harness
 
+## 0.59.11
+
+### Patch Changes
+
+- [`7464ad0`](https://github.com/cesr/poncho-ai/commit/7464ad04448095b34c4b1cbd52da559cf6bd6024) Thanks [@cesr](https://github.com/cesr)! - The user's memory file gets its own 1-hour Anthropic cache breakpoint.
+  It previously rode the uncached dynamic system tail (with todos + time),
+  which re-wrote the memory block — typically the bulk of a new
+  conversation's one-time cache cost — on every cold prefix, despite memory
+  only changing on explicit writes. System prompt is now three tiers:
+  static (1h), memory (1h), volatile todos+time (uncached).
+
+## 0.59.10
+
+### Patch Changes
+
+- [`fad3918`](https://github.com/cesr/poncho-ai/commit/fad3918302114f76a29080cf28e9c003c61ef0d9) Thanks [@cesr](https://github.com/cesr)! - Stamp `session.id` / `user.id` on EVERY span, not just the invoke_agent
+  root. Observability backends resolve a span's identity from its own
+  attributes — Latitude's console session/conversation views key on the LLM
+  generation spans, so root-only attributes grouped the API-level trace but
+  left the console showing one session per turn and no user. The identity now
+  rides the OTel Context and an IdentityAttributeSpanProcessor injects it
+  into every descendant span (LLM steps, tool executions) at start.
+
 ## 0.59.9
 
 ### Patch Changes

package/dist/index.js

@@ -8745,7 +8745,7 @@ var createSubagentTools = (manager) => [
 ];
 
 // src/harness.ts
-import { trace, context as otelContext, SpanStatusCode, SpanKind, diag, DiagConsoleLogger, DiagLogLevel } from "@opentelemetry/api";
+import { trace, context as otelContext, createContextKey, SpanStatusCode, SpanKind, diag, DiagConsoleLogger, DiagLogLevel } from "@opentelemetry/api";
 import { NodeTracerProvider, BatchSpanProcessor } from "@opentelemetry/sdk-trace-node";
 import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-http";
 
@@ -8928,6 +8928,28 @@ var telemetryLog2 = createLogger7("telemetry");
 var costLog = createLogger7("cost");
 var mcpLog2 = createLogger7("mcp");
 var modelLog = createLogger7("model");
+var TELEMETRY_SESSION_ID_KEY = createContextKey("poncho.telemetry.session_id");
+var TELEMETRY_USER_ID_KEY = createContextKey("poncho.telemetry.user_id");
+var IdentityAttributeSpanProcessor = class {
+  onStart(span, parentContext) {
+    const sessionId = parentContext.getValue(TELEMETRY_SESSION_ID_KEY);
+    if (typeof sessionId === "string" && sessionId) {
+      span.setAttribute("session.id", sessionId);
+    }
+    const userId = parentContext.getValue(TELEMETRY_USER_ID_KEY);
+    if (typeof userId === "string" && userId) {
+      span.setAttribute("user.id", userId);
+    }
+  }
+  onEnd() {
+  }
+  forceFlush() {
+    return Promise.resolve();
+  }
+  shutdown() {
+    return Promise.resolve();
+  }
+};
 function formatOtlpError(err) {
   if (!(err instanceof Error)) return String(err);
   const parts = [];
@@ -10339,7 +10361,9 @@ var AgentHarness = class _AgentHarness {
       const processor = new BatchSpanProcessor(exporter);
       this.otlpSpanProcessor = processor;
       const provider2 = new NodeTracerProvider({
-        spanProcessors: [processor]
+        // Identity injector FIRST so every span (root, LLM steps, tool
+        // executions) carries session.id/user.id before batching/export.
+        spanProcessors: [new IdentityAttributeSpanProcessor(), processor]
       });
       provider2.register();
       this.otlpTracerProvider = provider2;
@@ -10506,7 +10530,13 @@ var AgentHarness = class _AgentHarness {
           ...input.tenantId ? { "tenant.id": input.tenantId } : {}
         }
       });
-      const spanContext = trace.setSpan(otelContext.active(), rootSpan);
+      let spanContext = trace.setSpan(otelContext.active(), rootSpan);
+      if (input.conversationId) {
+        spanContext = spanContext.setValue(TELEMETRY_SESSION_ID_KEY, input.conversationId);
+      }
+      if (this.telemetryUserId) {
+        spanContext = spanContext.setValue(TELEMETRY_USER_ID_KEY, this.telemetryUserId);
+      }
       try {
         const gen = this.run(input);
         let next;
@@ -10690,11 +10720,11 @@ ${skillContextWindow}${browserContext}${fsContext}${isolateContext}` : `${agentP
       const timeContext = `
 
 Current UTC time (hour precision): ${hourlyTime}`;
-      const dynamicPart = `${memoryContext}${todoContext}${timeContext}`;
-      return { staticPart, dynamicPart };
+      const dynamicPart = `${todoContext}${timeContext}`;
+      return { staticPart, memoryPart: memoryContext, dynamicPart };
     };
-    let { staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts();
-    let systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+    let { staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts();
+    let systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
     let lastPromptFingerprint = `${this.agentFileFingerprint}
 ${this.skillFingerprint}`;
     const pushEvent = (event) => {
@@ -11173,6 +11203,18 @@ ${textContent}` };
                 anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } }
               }
             },
+            // Memory: per-user, conversation-independent, changes only on
+            // an explicit memory write — its own 1h breakpoint means a
+            // memory edit busts THIS block forward but a normal turn reads
+            // it (plus everything before it) from cache. Breakpoint budget:
+            // Anthropic allows 4; this is #2 of 3 (static, memory, tail).
+            ...memorySystemPart.length > 0 ? [{
+              role: "system",
+              content: memorySystemPart,
+              providerOptions: {
+                anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } }
+              }
+            }] : [],
             ...dynamicSystemPart.length > 0 ? [{ role: "system", content: dynamicSystemPart }] : [],
             ...cachedMessages
           ] : cachedMessages;
@@ -11814,8 +11856,8 @@ ${textContent}` };
               const currentFingerprint = `${this.agentFileFingerprint}
 ${this.skillFingerprint}`;
               if (currentFingerprint !== lastPromptFingerprint) {
-                ({ staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts());
-                systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+                ({ staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts());
+                systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
                 lastPromptFingerprint = currentFingerprint;
               }
             }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.59.9",
+  "version": "0.59.11",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/harness.ts

@@ -67,9 +67,41 @@ import { createSkillTools, normalizeScriptPolicyPath } from "./skill-tools.js";
 import { createSearchTools } from "./search-tools.js";
 import { createSubagentTools } from "./subagent-tools.js";
 import type { SubagentManager } from "./subagent-manager.js";
-import { trace, context as otelContext, SpanStatusCode, SpanKind, diag, DiagConsoleLogger, DiagLogLevel } from "@opentelemetry/api";
+import { trace, context as otelContext, createContextKey, type Context as OtelContextType, SpanStatusCode, SpanKind, diag, DiagConsoleLogger, DiagLogLevel } from "@opentelemetry/api";
+import type { Span as OtelSdkSpan, SpanProcessor } from "@opentelemetry/sdk-trace-node";
 import { NodeTracerProvider, BatchSpanProcessor } from "@opentelemetry/sdk-trace-node";
 import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-http";
+
+// ── Telemetry identity propagation ──────────────────────────────────────────
+// Observability backends (Latitude) resolve a span's session/user from the
+// span's OWN attributes — the console's session & conversation views key on
+// the LLM generation spans, not the root span. So stamping `session.id` /
+// `user.id` only on the invoke_agent root groups the API-level trace but
+// leaves the console treating every turn as its own session. The fix is the
+// same one vendor SDKs use: carry the identity in the OTel Context and have
+// a SpanProcessor stamp it onto EVERY span at start.
+const TELEMETRY_SESSION_ID_KEY = createContextKey("poncho.telemetry.session_id");
+const TELEMETRY_USER_ID_KEY = createContextKey("poncho.telemetry.user_id");
+
+class IdentityAttributeSpanProcessor implements SpanProcessor {
+  onStart(span: OtelSdkSpan, parentContext: OtelContextType): void {
+    const sessionId = parentContext.getValue(TELEMETRY_SESSION_ID_KEY);
+    if (typeof sessionId === "string" && sessionId) {
+      span.setAttribute("session.id", sessionId);
+    }
+    const userId = parentContext.getValue(TELEMETRY_USER_ID_KEY);
+    if (typeof userId === "string" && userId) {
+      span.setAttribute("user.id", userId);
+    }
+  }
+  onEnd(): void {}
+  forceFlush(): Promise<void> {
+    return Promise.resolve();
+  }
+  shutdown(): Promise<void> {
+    return Promise.resolve();
+  }
+}
 import { normalizeOtlp } from "./telemetry.js";
 
 /** Extract useful details from OTLPExporterError (has .code + .data) or plain Error. */
@@ -1883,7 +1915,9 @@ export class AgentHarness {
       const processor = new BatchSpanProcessor(exporter);
       this.otlpSpanProcessor = processor;
       const provider = new NodeTracerProvider({
-        spanProcessors: [processor],
+        // Identity injector FIRST so every span (root, LLM steps, tool
+        // executions) carries session.id/user.id before batching/export.
+        spanProcessors: [new IdentityAttributeSpanProcessor(), processor],
       });
       provider.register();
       this.otlpTracerProvider = provider;
@@ -2074,7 +2108,15 @@ export class AgentHarness {
         },
       });
 
-      const spanContext = trace.setSpan(otelContext.active(), rootSpan);
+      let spanContext = trace.setSpan(otelContext.active(), rootSpan);
+      // Identity rides the context so IdentityAttributeSpanProcessor stamps
+      // session.id/user.id on every descendant span (see processor docs).
+      if (input.conversationId) {
+        spanContext = spanContext.setValue(TELEMETRY_SESSION_ID_KEY, input.conversationId);
+      }
+      if (this.telemetryUserId) {
+        spanContext = spanContext.setValue(TELEMETRY_USER_ID_KEY, this.telemetryUserId);
+      }
 
       try {
         const gen = this.run(input);
@@ -2289,13 +2331,17 @@ ${typeStubs}
 Code is wrapped in an async IIFE — use \`return\` to return a value to the tool result.`;
     }
 
-    // Split the system prompt into a static portion (stable across turns
-    // and jobs within an hour, modulo MCP connect/skill author/memory edit)
-    // and a dynamic tail (memory, todos, time). The static portion gets a
-    // 1-hour Anthropic cache breakpoint downstream; the tail rides the
-    // existing 5-min message-level breakpoint. See the streamText site for
-    // the breakpoint wiring.
-    const buildSystemPromptParts = async (): Promise<{ staticPart: string; dynamicPart: string }> => {
+    // Split the system prompt into THREE cacheability tiers (see the
+    // streamText site for the breakpoint wiring):
+    //   1. staticPart — agent body + skills + runtime context. Stable across
+    //      turns, conversations, and jobs within an hour. 1h breakpoint.
+    //   2. memoryPart — the user's memory file. Per-user, shared by every
+    //      conversation, and only changes on an explicit memory write — so
+    //      it gets its own 1h breakpoint instead of riding the volatile
+    //      tail (where it busted the message-history cache for no reason).
+    //   3. dynamicPart — todos + hour-quantized time. Genuinely volatile
+    //      within a conversation; uncached, kept as small as possible.
+    const buildSystemPromptParts = async (): Promise<{ staticPart: string; memoryPart: string; dynamicPart: string }> => {
       const agentPrompt = renderCurrentAgentPrompt();
       const tenantSkills = await this.getSkillsForTenant(input.tenantId);
       const skillContextWindow = buildSkillContextWindow(tenantSkills);
@@ -2316,13 +2362,13 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
         return `${weekday} ${d.toISOString().slice(0, 13)}Z`;
       })();
       const timeContext = `\n\nCurrent UTC time (hour precision): ${hourlyTime}`;
-      const dynamicPart = `${memoryContext}${todoContext}${timeContext}`;
-      return { staticPart, dynamicPart };
+      const dynamicPart = `${todoContext}${timeContext}`;
+      return { staticPart, memoryPart: memoryContext, dynamicPart };
     };
-    let { staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } =
+    let { staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } =
       await buildSystemPromptParts();
     // Concatenated form for legacy consumers (token estimation, telemetry).
-    let systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+    let systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
     let lastPromptFingerprint = `${this.agentFileFingerprint}\n${this.skillFingerprint}`;
 
     const pushEvent = (event: AgentEvent): AgentEvent => {
@@ -2966,6 +3012,20 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
                   anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } },
                 },
               },
+              // Memory: per-user, conversation-independent, changes only on
+              // an explicit memory write — its own 1h breakpoint means a
+              // memory edit busts THIS block forward but a normal turn reads
+              // it (plus everything before it) from cache. Breakpoint budget:
+              // Anthropic allows 4; this is #2 of 3 (static, memory, tail).
+              ...(memorySystemPart.length > 0
+                ? [{
+                    role: "system" as const,
+                    content: memorySystemPart,
+                    providerOptions: {
+                      anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } },
+                    },
+                  }]
+                : []),
               ...(dynamicSystemPart.length > 0
                 ? [{ role: "system" as const, content: dynamicSystemPart }]
                 : []),
@@ -3752,9 +3812,9 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
             agent = this.parsedAgent as ParsedAgent;
             const currentFingerprint = `${this.agentFileFingerprint}\n${this.skillFingerprint}`;
             if (currentFingerprint !== lastPromptFingerprint) {
-              ({ staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } =
+              ({ staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } =
                 await buildSystemPromptParts());
-              systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+              systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
               lastPromptFingerprint = currentFingerprint;
             }
           }