@poncho-ai/harness 0.34.0 → 0.34.1

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.34.0 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.34.1 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
@@ -8,8 +8,8 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js 336.31 KB
-ESM ⚡️ Build success in 164ms
+ESM dist/index.js 336.65 KB
+ESM ⚡️ Build success in 155ms
 DTS Build start
-DTS ⚡️ Build success in 7760ms
-DTS dist/index.d.ts 33.99 KB
+DTS ⚡️ Build success in 7364ms
+DTS dist/index.d.ts 34.28 KB

package/CHANGELOG.md

@@ -1,5 +1,14 @@
 # @poncho-ai/harness
 
+## 0.34.1
+
+### Patch Changes
+
+- [`59a88cc`](https://github.com/cesr/poncho-ai/commit/59a88cc52b5c3aa7432b820424bb8067174233e5) Thanks [@cesr](https://github.com/cesr)! - fix: improve token estimation accuracy and handle missing attachments
+  - Use a JSON-specific token ratio for tool definitions to avoid inflating counts with many MCP tools.
+  - Track actual context size from model responses for compaction triggers instead of cumulative input tokens.
+  - Gracefully degrade when file attachments are missing or expired instead of crashing.
+
 ## 0.34.0
 
 ### Minor Changes

package/dist/index.d.ts

@@ -83,6 +83,11 @@ declare const resolveCompactionConfig: (explicit?: Partial<CompactionConfig>) =>
 declare const estimateTokens: (text: string) => number;
 /**
  * Estimate the total token count of a system prompt + messages + tool defs.
+ *
+ * Tool definitions are structured JSON (property names, braces, enum values)
+ * which tokenizes more efficiently than natural language — roughly 5-6
+ * chars/token vs ~4 chars/token for prose.  We estimate them separately to
+ * avoid inflating the count when there are many MCP tools (100+).
  */
 declare const estimateTotalTokens: (systemPrompt: string, messages: Message[], toolDefinitionsJson?: string) => number;
 /**

package/dist/index.js

@@ -397,10 +397,11 @@ var estimateTotalTokens = (systemPrompt, messages, toolDefinitionsJson) => {
       return sum + 200;
     }, 0);
   }
+  let tokens = Math.ceil(chars / 4 * OVERHEAD_MULTIPLIER);
   if (toolDefinitionsJson) {
-    chars += toolDefinitionsJson.length;
+    tokens += Math.ceil(toolDefinitionsJson.length / 6);
   }
-  return Math.ceil(chars / 4 * OVERHEAD_MULTIPLIER);
+  return tokens;
 };
 var findSafeSplitPoint = (messages, keepRecentMessages) => {
   const candidateIdx = messages.length - keepRecentMessages;
@@ -7289,19 +7290,24 @@ ${textContent}` };
                   };
                 }
                 let resolvedData;
-                if (part.data.startsWith(PONCHO_UPLOAD_SCHEME) && this.uploadStore) {
-                  const buf = await this.uploadStore.get(part.data);
-                  resolvedData = buf.toString("base64");
-                } else if (part.data.startsWith("https://") || part.data.startsWith("http://")) {
-                  if (this.uploadStore) {
+                try {
+                  if (part.data.startsWith(PONCHO_UPLOAD_SCHEME) && this.uploadStore) {
                     const buf = await this.uploadStore.get(part.data);
                     resolvedData = buf.toString("base64");
+                  } else if (part.data.startsWith("https://") || part.data.startsWith("http://")) {
+                    if (this.uploadStore) {
+                      const buf = await this.uploadStore.get(part.data);
+                      resolvedData = buf.toString("base64");
+                    } else {
+                      const resp = await fetch(part.data);
+                      resolvedData = Buffer.from(await resp.arrayBuffer()).toString("base64");
+                    }
                   } else {
-                    const resp = await fetch(part.data);
-                    resolvedData = Buffer.from(await resp.arrayBuffer()).toString("base64");
+                    resolvedData = part.data;
                   }
-                } else {
-                  resolvedData = part.data;
+                } catch {
+                  const label = part.filename ?? part.mediaType;
+                  return { type: "text", text: `[Attached file: ${label} \u2014 file is no longer available]` };
                 }
                 if (isSupportedImage) {
                   return {
@@ -7330,8 +7336,8 @@ ${textContent}` };
         const compactionConfig = resolveCompactionConfig(agent.frontmatter.compaction);
         if (compactionConfig.enabled && (step === 1 || step % COMPACTION_CHECK_INTERVAL_STEPS === 0)) {
           const estimated = estimateTotalTokens(systemPrompt, messages, toolDefsJsonForEstimate);
-          const lastReportedInput = totalInputTokens > 0 ? totalInputTokens : 0;
-          const effectiveTokens = Math.max(estimated, lastReportedInput);
+          const lastReportedContext = latestContextTokens > 0 ? latestContextTokens + toolOutputEstimateSinceModel : 0;
+          const effectiveTokens = Math.max(estimated, lastReportedContext);
           if (effectiveTokens > compactionConfig.trigger * contextWindow) {
             yield pushEvent({ type: "compaction:started", estimatedTokens: effectiveTokens });
             const compactResult = await compactMessages(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.34.0",
+  "version": "0.34.1",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/compaction.ts

@@ -49,6 +49,11 @@ export const estimateTokens = (text: string): number =>
 
 /**
  * Estimate the total token count of a system prompt + messages + tool defs.
+ *
+ * Tool definitions are structured JSON (property names, braces, enum values)
+ * which tokenizes more efficiently than natural language — roughly 5-6
+ * chars/token vs ~4 chars/token for prose.  We estimate them separately to
+ * avoid inflating the count when there are many MCP tools (100+).
  */
 export const estimateTotalTokens = (
   systemPrompt: string,
@@ -64,10 +69,13 @@ export const estimateTotalTokens = (
           return sum + 200; // rough estimate for file/image references
         }, 0);
   }
+  let tokens = Math.ceil((chars / 4) * OVERHEAD_MULTIPLIER);
   if (toolDefinitionsJson) {
-    chars += toolDefinitionsJson.length;
+    // JSON-specific ratio — no overhead multiplier (structural tokens are
+    // already accounted for by the higher chars-per-token ratio).
+    tokens += Math.ceil(toolDefinitionsJson.length / 6);
   }
-  return Math.ceil((chars / 4) * OVERHEAD_MULTIPLIER);
+  return tokens;
 };
 
 /**

package/src/harness.ts

@@ -2154,19 +2154,24 @@ ${boundedMainMemory.trim()}`
                 // Always resolve to base64 so the model doesn't need to
                 // fetch URLs itself (which fails for private blob stores).
                 let resolvedData: string;
-                if (part.data.startsWith(PONCHO_UPLOAD_SCHEME) && this.uploadStore) {
-                  const buf = await this.uploadStore.get(part.data);
-                  resolvedData = buf.toString("base64");
-                } else if (part.data.startsWith("https://") || part.data.startsWith("http://")) {
-                  if (this.uploadStore) {
+                try {
+                  if (part.data.startsWith(PONCHO_UPLOAD_SCHEME) && this.uploadStore) {
                     const buf = await this.uploadStore.get(part.data);
                     resolvedData = buf.toString("base64");
+                  } else if (part.data.startsWith("https://") || part.data.startsWith("http://")) {
+                    if (this.uploadStore) {
+                      const buf = await this.uploadStore.get(part.data);
+                      resolvedData = buf.toString("base64");
+                    } else {
+                      const resp = await fetch(part.data);
+                      resolvedData = Buffer.from(await resp.arrayBuffer()).toString("base64");
+                    }
                   } else {
-                    const resp = await fetch(part.data);
-                    resolvedData = Buffer.from(await resp.arrayBuffer()).toString("base64");
+                    resolvedData = part.data;
                   }
-                } else {
-                  resolvedData = part.data;
+                } catch {
+                  const label = part.filename ?? part.mediaType;
+                  return { type: "text" as const, text: `[Attached file: ${label} — file is no longer available]` };
                 }
                 if (isSupportedImage) {
                   return {
@@ -2200,8 +2205,13 @@ ${boundedMainMemory.trim()}`
         const compactionConfig = resolveCompactionConfig(agent.frontmatter.compaction);
         if (compactionConfig.enabled && (step === 1 || step % COMPACTION_CHECK_INTERVAL_STEPS === 0)) {
           const estimated = estimateTotalTokens(systemPrompt, messages, toolDefsJsonForEstimate);
-          const lastReportedInput = totalInputTokens > 0 ? totalInputTokens : 0;
-          const effectiveTokens = Math.max(estimated, lastReportedInput);
+          // Use the actual context size from the last model response (input tokens
+          // + tool output accumulated since), not totalInputTokens which is a
+          // cumulative sum across all steps and would wildly overcount.
+          const lastReportedContext = latestContextTokens > 0
+            ? latestContextTokens + toolOutputEstimateSinceModel
+            : 0;
+          const effectiveTokens = Math.max(estimated, lastReportedContext);
 
           if (effectiveTokens > compactionConfig.trigger * contextWindow) {
             yield pushEvent({ type: "compaction:started", estimatedTokens: effectiveTokens });