@poncho-ai/harness 0.59.6 → 0.59.7

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.59.6 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.59.7 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
@@ -9,8 +9,8 @@
 CLI Target: es2022
 ESM Build start
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM dist/index.js            557.73 KB
-ESM ⚡️ Build success in 256ms
+ESM dist/index.js            558.06 KB
+ESM ⚡️ Build success in 257ms
 DTS Build start
-DTS ⚡️ Build success in 8276ms
+DTS ⚡️ Build success in 7680ms
 DTS dist/index.d.ts 101.66 KB

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # @poncho-ai/harness
 
+## 0.59.7
+
+### Patch Changes
+
+- [`c73cb19`](https://github.com/cesr/poncho-ai/commit/c73cb19ec8bf61fe0598262ae4d050fb84c939b5) Thanks [@cesr](https://github.com/cesr)! - Auto-compaction never fired on cached conversations: the per-step context
+  measure (`latestContextTokens`) used `usage.inputTokens`, which with
+  Anthropic prompt caching is only the NON-cached slice — a real 190k+
+  conversation reported ~12k of "context", so the trigger comparison never
+  tripped and transcripts grew past the model's window. Context now counts
+  input + cache-read + cache-write tokens (everything the model read). Also
+  pins claude-fable-5 / opus-4-8 / opus-4-7 in the context-window registry
+  (previously relying on the silent 200k default).
+
 ## 0.59.6
 
 ### Patch Changes

package/dist/index.js

@@ -7490,6 +7490,12 @@ var completeOpenAICodexDeviceAuth = async (request) => {
 
 // src/model-factory.ts
 var MODEL_CONTEXT_WINDOWS = {
+  // Pinned conservatively at 200k. The API has accepted >204k for fable-5
+  // (its real window is larger), but compacting at trigger×200k keeps
+  // long-conversation cost bounded; raise deliberately, not by omission.
+  "claude-fable-5": 2e5,
+  "claude-opus-4-8": 2e5,
+  "claude-opus-4-7": 2e5,
   "claude-opus-4-6": 2e5,
   "claude-sonnet-4-6": 2e5,
   "claude-opus-4-5": 2e5,
@@ -11362,7 +11368,7 @@ ${textContent}` };
           totalOutputTokens += usage.outputTokens ?? 0;
           totalCachedTokens += stepCachedTokens;
           totalCacheWriteTokens += stepCacheWriteTokens;
-          latestContextTokens = stepInputTokens;
+          latestContextTokens = stepInputTokens + stepCachedTokens + stepCacheWriteTokens;
           toolOutputEstimateSinceModel = 0;
           yield pushEvent({
             type: "model:response",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.59.6",
+  "version": "0.59.7",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/harness.ts

@@ -3208,7 +3208,14 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
       totalOutputTokens += usage.outputTokens ?? 0;
       totalCachedTokens += stepCachedTokens;
       totalCacheWriteTokens += stepCacheWriteTokens;
-      latestContextTokens = stepInputTokens;
+      // Context size = EVERYTHING the model read this step. With prompt
+      // caching, Anthropic's `usage.input_tokens` is only the non-cached
+      // slice — the bulk of a long conversation arrives as cache reads.
+      // Counting input alone made the auto-compaction check see ~12k of
+      // "context" on a real 190k+ conversation, so compaction never fired
+      // and the transcript grew unbounded (observed 2026-06-12: 205k real
+      // context, trigger at 190k, no compaction).
+      latestContextTokens = stepInputTokens + stepCachedTokens + stepCacheWriteTokens;
       toolOutputEstimateSinceModel = 0;
 
       yield pushEvent({

package/src/model-factory.ts

@@ -9,6 +9,12 @@ import {
 export type ModelProviderFactory = (modelName: string) => LanguageModel;
 
 const MODEL_CONTEXT_WINDOWS: Record<string, number> = {
+  // Pinned conservatively at 200k. The API has accepted >204k for fable-5
+  // (its real window is larger), but compacting at trigger×200k keeps
+  // long-conversation cost bounded; raise deliberately, not by omission.
+  "claude-fable-5": 200_000,
+  "claude-opus-4-8": 200_000,
+  "claude-opus-4-7": 200_000,
   "claude-opus-4-6": 200_000,
   "claude-sonnet-4-6": 200_000,
   "claude-opus-4-5": 200_000,