claude-code-cache-fix 2.0.0-beta.1 → 2.0.0-beta.2

package/README.md

@@ -4,7 +4,9 @@
 
 English | [中文](./README.zh.md) | [한국어](./README.ko.md) | [Português](./docs/guia-pt-br.md)
 
-Fixes prompt cache regressions in [Claude Code](https://github.com/anthropics/claude-code) that cause **up to 20x cost increase** on resumed sessions, plus monitoring for silent context degradation. Confirmed through v2.1.107.
+Fixes prompt cache regressions in [Claude Code](https://github.com/anthropics/claude-code) that cause **up to 20x cost increase** on resumed sessions, plus monitoring for silent context degradation. Confirmed through v2.1.111. Opus 4.7 compatible.
+
+> **Opus 4.7 advisory:** Our metered data shows 4.7 burns Q5h quota at **~2.4x the rate of 4.6** for equivalent visible token counts. Two factors: a new tokenizer (up to 35% more tokens, [documented](https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7)) and adaptive thinking overhead (~105%, not documented in usage response). Workaround: `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1` (may reduce quality). Image stripping (`CACHE_FIX_IMAGE_KEEP_LAST`) is even more important on 4.7 due to high-res image support increasing image token counts. See [Discussion #25](https://github.com/cnighswonger/claude-code-cache-fix/discussions/25) for full analysis.
 
 ## Security model
 
@@ -308,7 +310,7 @@ When the server downgrades your TTL to 5m (Layer 2 — quota-aware downgrade at
 
 ## Image stripping
 
-Images read via the Read tool are encoded as base64 and stored in `tool_result` blocks in conversation history. They ride along on **every subsequent API call** until compaction. A single 500KB image costs ~62,500 tokens per turn in carry-forward.
+Images read via the Read tool are encoded as base64 and stored in `tool_result` blocks in conversation history. They ride along on **every subsequent API call** until compaction. A single 500KB image costs ~62,500 tokens per turn on Opus 4.6, and potentially **~85,000+ tokens on Opus 4.7** due to the new tokenizer (up to 35% inflation) and high-res image support (2576px max, up from 1568px). Image stripping is strongly recommended on 4.7.
 
 Enable image stripping to remove old images from tool results:
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-cache-fix",
-  "version": "2.0.0-beta.1",
+  "version": "2.0.0-beta.2",
   "description": "Fixes prompt cache regression in Claude Code that causes up to 20x cost increase on resumed sessions",
   "type": "module",
   "exports": "./preload.mjs",

package/preload.mjs

@@ -726,6 +726,7 @@ const _STATS_SCHEMA = {
   identity: { applied: 0, skipped: 0, lastApplied: null },
   git_status: { applied: 0, skipped: 0, lastApplied: null },
   cwd_normalize: { applied: 0, skipped: 0, lastApplied: null },
+  smoosh_normalize: { applied: 0, skipped: 0, lastApplied: null },
 };
 
 function _createEmptyStats() {
@@ -1347,6 +1348,62 @@ globalThis.fetch = async function (url, options) {
         }
       }
 
+      // Optimization: normalize smooshed dynamic system-reminders in tool_result content
+      // CC's smooshSystemReminderSiblings (messages.ts:1835) folds <system-reminder> text
+      // blocks into tool_result.content strings. Dynamic values (token_usage, budget_usd,
+      // output_token_usage, todo_reminder) change every turn, causing mid-history cache
+      // busts even without resume or attachment scatter.
+      // Bug: anthropics/claude-code#49585 (deafsquad)
+      // Opt-in via CACHE_FIX_NORMALIZE_SMOOSH=1.
+      if (process.env.CACHE_FIX_NORMALIZE_SMOOSH === "1" && shouldApplyFix("smoosh_normalize") && payload.messages) {
+        let smooshNormalized = 0;
+        const smooshPatterns = [
+          // Token usage: 12345/50000; 37655 remaining
+          /(<system-reminder>\nToken usage: )\d+\/\d+; \d+ remaining/g,
+          // USD budget: $1.23/$10.00; $8.77 remaining
+          /(<system-reminder>\nUSD budget: )\$[\d.]+\/\$[\d.]+; \$[\d.]+ remaining/g,
+          // Output tokens — turn: 1,234 / 5,000 · session: 12,345
+          /(<system-reminder>\nOutput tokens \u2014 turn: )[\d,./\s]+ \u00b7 session: [\d,]+/g,
+          // TodoWrite reminder with variable todo list content
+          /(<system-reminder>\nThe TodoWrite tool hasn't been used recently\..*?)(\n\nHere are the existing contents of your todo list:\n\n\[[\s\S]*?\])?(\n<\/system-reminder>)/g,
+        ];
+        const smooshReplacements = [
+          "$1[normalized]/[normalized]; [normalized] remaining",
+          "$1$[normalized]/$[normalized]; $[normalized] remaining",
+          "$1[normalized] \u00b7 session: [normalized]",
+          "$1$3",  // strip the variable todo list, keep the static reminder text
+        ];
+
+        for (const msg of payload.messages) {
+          if (msg.role !== "user") continue;
+          // Handle both string content (smooshed tool_result) and array content
+          if (Array.isArray(msg.content)) {
+            for (let i = 0; i < msg.content.length; i++) {
+              const block = msg.content[i];
+              // Smooshed tool_result with string content
+              if (block.type === "tool_result" && typeof block.content === "string" && block.content.includes("<system-reminder>")) {
+                let newContent = block.content;
+                for (let p = 0; p < smooshPatterns.length; p++) {
+                  smooshPatterns[p].lastIndex = 0; // reset regex state
+                  newContent = newContent.replace(smooshPatterns[p], smooshReplacements[p]);
+                }
+                if (newContent !== block.content) {
+                  msg.content[i] = { ...block, content: newContent };
+                  smooshNormalized++;
+                }
+              }
+            }
+          }
+        }
+        if (smooshNormalized > 0) {
+          modified = true;
+          debugLog(`APPLIED: smoosh-normalized ${smooshNormalized} tool_result block(s) with dynamic system-reminders`);
+          recordFixResult("smoosh_normalize", "applied");
+        } else {
+          recordFixResult("smoosh_normalize", "skipped");
+        }
+      }
+
       // Bug 5: TTL enforcement (configurable per request type)
       // The client gates 1h cache TTL behind a GrowthBook allowlist that checks
       // querySource against patterns like "repl_main_thread*", "sdk", "auto_mode".