@poncho-ai/harness 0.59.5 → 0.59.7

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.59.5 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
@@ -8,9 +8,9 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js            557.00 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 252ms
+ESM dist/index.js            558.06 KB
+ESM ⚡️ Build success in 257ms
 DTS Build start
-DTS ⚡️ Build success in 7698ms
+DTS ⚡️ Build success in 7680ms
 DTS dist/index.d.ts 101.66 KB

package/CHANGELOG.md

@@ -1,5 +1,33 @@
 # @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
+
+- [`e573f72`](https://github.com/cesr/poncho-ai/commit/e573f72ca31627e48dbdbf296946a72c59a488db) Thanks [@cesr](https://github.com/cesr)! - Preserve the LLM transcript when a turn dies. The errored branch of
+  runConversationTurn persisted only the display draft — `_harnessMessages`
+  was never updated, so the model's next turn had no memory of the entire
+  failed interaction (its user message included), even though the user could
+  see it on screen. Both the errored branch and the cancelled-without-
+  `run:cancelled.messages` fallback now append a faithful plain-text
+  reconstruction (user message + assistant text-so-far + tool activity + an
+  interruption note) to the transcript. Plain text on purpose: replaying real
+  tool_use blocks would need paired results or the next API call rejects the
+  dangling pair.
+
 ## 0.59.5
 
 ### 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 @@ ${textContent}` };
           totalOutputTokens += usage.outputTokens ?? 0;
           totalCachedTokens += stepCachedTokens;
           totalCacheWriteTokens += stepCacheWriteTokens;
-          latestContextTokens = stepInputTokens;
+          latestContextTokens = stepInputTokens + stepCachedTokens + stepCacheWriteTokens;
           toolOutputEstimateSinceModel = 0;
           yield pushEvent({
             type: "model:response",
@@ -14498,6 +14504,20 @@ var runConversationTurn = async (opts) => {
     };
   } catch (error) {
     flushTurnDraft(draft);
+    const reconstructTranscriptTail = (reason) => {
+      const parts = [];
+      if (draft.assistantResponse.length > 0) parts.push(draft.assistantResponse);
+      if (draft.toolTimeline.length > 0) {
+        parts.push(`Tool activity before interruption:
+${draft.toolTimeline.join("\n")}`);
+      }
+      parts.push(`[This turn was interrupted: ${reason}. The work above may be incomplete.]`);
+      return [
+        ...conversation._harnessMessages ?? [],
+        userMessage,
+        { role: "assistant", content: parts.join("\n\n") }
+      ];
+    };
     const aborted = opts.abortSignal?.aborted === true;
     if (aborted || runCancelled) {
       if (draft.assistantResponse.length > 0 || draft.toolTimeline.length > 0 || draft.sections.length > 0) {
@@ -14508,7 +14528,7 @@ var runConversationTurn = async (opts) => {
             latestRunId,
             contextTokens: 0,
             contextWindow: 0,
-            harnessMessages: cancelHarnessMessages,
+            harnessMessages: cancelHarnessMessages ?? reconstructTranscriptTail("cancelled"),
             toolResultArchive: opts.harness.getToolResultArchive(opts.conversationId)
           },
           { shouldRebuildCanonical: true }
@@ -14551,6 +14571,9 @@ var runConversationTurn = async (opts) => {
     }
     if (draft.assistantResponse.length > 0 || draft.toolTimeline.length > 0 || draft.sections.length > 0) {
       conversation.messages = buildMessages(false);
+      conversation._harnessMessages = reconstructTranscriptTail(
+        error instanceof Error ? `error \u2014 ${error.message}` : "error"
+      );
       conversation.updatedAt = Date.now();
       await opts.conversationStore.update(conversation);
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.59.5",
+  "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,

package/src/orchestrator/run-conversation-turn.ts

@@ -420,6 +420,31 @@ export const runConversationTurn = async (
     };
   } catch (error) {
     flushTurnDraft(draft);
+
+    // The LLM transcript (`_harnessMessages`) is normally only written at a
+    // clean finalize / a cancel that delivered `run:cancelled.messages`. A
+    // turn that dies any other way (in-process error, abort that never
+    // surfaced the cancel event) would leave the transcript WITHOUT this
+    // turn at all — the display shows the partial work but the model has
+    // amnesia about the whole interaction on the next turn. Reconstruct a
+    // faithful plain-text record from the draft instead: the user message
+    // plus an assistant message carrying the text-so-far + tool activity.
+    // Plain text on purpose — replaying real tool_use blocks would need
+    // paired results or the next API call 400s on the dangling pair.
+    const reconstructTranscriptTail = (reason: string): Message[] => {
+      const parts: string[] = [];
+      if (draft.assistantResponse.length > 0) parts.push(draft.assistantResponse);
+      if (draft.toolTimeline.length > 0) {
+        parts.push(`Tool activity before interruption:\n${draft.toolTimeline.join("\n")}`);
+      }
+      parts.push(`[This turn was interrupted: ${reason}. The work above may be incomplete.]`);
+      return [
+        ...(conversation._harnessMessages ?? []),
+        userMessage,
+        { role: "assistant" as const, content: parts.join("\n\n") },
+      ];
+    };
+
     const aborted = opts.abortSignal?.aborted === true;
     if (aborted || runCancelled) {
       if (
@@ -434,7 +459,8 @@ export const runConversationTurn = async (
             latestRunId,
             contextTokens: 0,
             contextWindow: 0,
-            harnessMessages: cancelHarnessMessages,
+            harnessMessages:
+              cancelHarnessMessages ?? reconstructTranscriptTail("cancelled"),
             toolResultArchive: opts.harness.getToolResultArchive(opts.conversationId),
           },
           { shouldRebuildCanonical: true },
@@ -484,6 +510,12 @@ export const runConversationTurn = async (
       draft.sections.length > 0
     ) {
       conversation.messages = buildMessages(false); // terminal: errored
+      // Keep the LLM transcript faithful too (see reconstructTranscriptTail
+      // above) — without this, the next turn's model context skipped the
+      // whole errored interaction.
+      conversation._harnessMessages = reconstructTranscriptTail(
+        error instanceof Error ? `error — ${error.message}` : "error",
+      );
       conversation.updatedAt = Date.now();
       await opts.conversationStore.update(conversation);
     }