jeopi-snapcompact 16.2.26 → 16.2.28

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [16.2.27] - 2026-07-07
+
+### Fixed
+
+- `serializeConversation` no longer renders assistant `thinking` blocks as italic prose into the archived conversation text before it gets rasterized into frame images. Those frame images are resent to the model as part of the persisted compaction summary on every later turn, so baking a prior turn's own reasoning into pixels was still a `reasoning_extraction` classifier-refusal trigger on Anthropic targets even though it's not literal text tokens (mirrors the plaintext-drop fix already applied to `packages/agent/src/compaction/utils.ts`). `redactedThinking` blocks were already dropped implicitly and needed no change.
+
 ## [16.2.8] - 2026-06-30
 
 ### Fixed

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "jeopi-snapcompact",
-	"version": "16.2.26",
+	"version": "16.2.28",
 	"description": "Bitmap-frame context compression for vision-capable LLMs",
 	"homepage": "https://github.com/akillness/jeopi",
 	"author": "Can Boluk",
@@ -31,10 +31,10 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"jeopi-ai": "16.2.26",
-		"jeopi-natives": "16.2.26",
-		"jeopi-utils": "16.2.26",
-		"jeopi-wire": "16.2.26"
+		"jeopi-ai": "16.2.28",
+		"jeopi-natives": "16.2.28",
+		"jeopi-utils": "16.2.28",
+		"jeopi-wire": "16.2.28"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/snapcompact.ts

@@ -791,17 +791,21 @@ export function serializeConversation(messages: Message[], options?: SerializeOp
 							.join("");
 			if (content) parts.push(`# User${HEADING_MARKER}\n${stripDimMarkers(content)}`);
 		} else if (msg.role === "assistant") {
-			// Stream blocks in content order: buffer thinking/text, then flush a
-			// `# Assistant` block (thinking as italics above the text) right before
-			// each tool call, so text or thinking after a call stays after it.
-			let pendingThinking: string[] = [];
+			// Stream blocks in content order: buffer text, then flush a
+			// `# Assistant` block right before each tool call, so text after a
+			// call stays after it. Thinking/redactedThinking blocks are dropped
+			// entirely (never buffered, never rasterized into a frame image):
+			// this archive gets rendered into ImageContent frames that re-enter
+			// the conversation history sent back to the model, and an image
+			// depicting a model's own prior natural-language reasoning is a
+			// plausible reasoning_extraction refusal trigger even though it's
+			// pixels, not text tokens (mirrors the plaintext drop in
+			// packages/agent/src/compaction/utils.ts serializeConversation).
 			let pendingText: string[] = [];
 			const flushAssistant = () => {
 				const sections: string[] = [];
-				if (pendingThinking.length > 0) sections.push(`_${pendingThinking.join("\n")}_`);
 				if (pendingText.length > 0) sections.push(pendingText.join("\n"));
 				if (sections.length > 0) parts.push(`# Assistant${HEADING_MARKER}\n${sections.join("\n\n")}`);
-				pendingThinking = [];
 				pendingText = [];
 			};
 
@@ -809,9 +813,6 @@ export function serializeConversation(messages: Message[], options?: SerializeOp
 				if (block.type === "text") {
 					const text = stripDimMarkers(block.text);
 					if (text.trim()) pendingText.push(text);
-				} else if (block.type === "thinking") {
-					const thinking = stripDimMarkers(block.thinking);
-					if (thinking.trim()) pendingThinking.push(thinking);
 				} else if (block.type === "toolCall") {
 					if (uselessCallIds.has(block.id)) continue;
 					flushAssistant();