claude-code-cache-fix 2.0.1 → 2.0.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-cache-fix",
-  "version": "2.0.1",
+  "version": "2.0.3",
   "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

@@ -752,9 +752,10 @@ function normalizeToolUseInputsInBody(body) {
 //
 // This sticky extension maintains per-session state tracking where markers
 // have appeared in prior turns, and reinstates them on future turns as
-// additive preservation. Up to 3 historical message-level markers are
+// additive preservation. Up to 2 historical message-level markers are
 // tracked (Anthropic's hard limit is 4 cache_control markers total — 1 for
-// system[2] + 3 for message-level breakpoints). When a historical position
+// system[2] + 1 canonical from cache_control_normalize + 2 historical from
+// sticky = 4). When a historical position
 // would exceed the cap, the oldest tracked entry is dropped (LRU).
 //
 // Messages are identified by a stable hash so that compaction rewrites /
@@ -772,7 +773,10 @@ function normalizeToolUseInputsInBody(body) {
 // --------------------------------------------------------------------------
 
 const CACHE_CONTROL_STICKY_DIR = join(homedir(), ".claude", "cache-fix-state");
-const CACHE_CONTROL_STICKY_MAX_POSITIONS = 3;
+// Anthropic hard limit: 4 cache_control markers total per request.
+// CC uses 1 on system[2] + cache_control_normalize places 1 on last user msg = 2 reserved.
+// Sticky can use at most 2 historical positions to stay within the 4-marker cap.
+const CACHE_CONTROL_STICKY_MAX_POSITIONS = 2;
 const CACHE_CONTROL_STICKY_DEFAULT_MARKER = { type: "ephemeral", ttl: "1h" };
 
 /**
@@ -938,10 +942,30 @@ function updateCacheControlStickyState(body, priorState) {
     capped = capped.slice(capped.length - CACHE_CONTROL_STICKY_MAX_POSITIONS);
   }
 
+  // Count existing cache_control markers across the entire body (system +
+  // messages) so sticky never pushes the total past Anthropic's hard limit
+  // of 4. CC may use 2 or 3 of those slots itself depending on version.
+  const ANTHROPIC_MARKER_LIMIT = 4;
+  let existingMarkers = 0;
+  if (Array.isArray(body.system)) {
+    for (const b of body.system) {
+      if (b && typeof b === "object" && b.cache_control) existingMarkers++;
+    }
+  }
+  for (const msg of body.messages) {
+    if (!msg || !Array.isArray(msg.content)) continue;
+    for (const b of msg.content) {
+      if (b && typeof b === "object" && b.cache_control) existingMarkers++;
+    }
+  }
+  const stickyBudget = Math.max(0, ANTHROPIC_MARKER_LIMIT - existingMarkers);
+
   // Compute mutations: for each tracked hash present in this body, if the
   // message doesn't already have any marker, add one at its last block.
+  // Stop once the sticky budget is exhausted.
   const mutations = [];
   for (const pos of capped) {
+    if (mutations.length >= stickyBudget) break;
     const msgIdx = hashToMsgIdx.get(pos.msg_hash);
     if (msgIdx === undefined) continue;
     const msg = body.messages[msgIdx];
@@ -2354,7 +2378,7 @@ globalThis.fetch = async function (url, options) {
       // the tail of each new user turn, the previous position loses the ~43
       // bytes of cache_control framing — a tail-of-message byte drift that
       // breaks every downstream cached block. This extension tracks marker
-      // positions by stable message-hash across turns (up to 3) and re-adds
+      // positions by stable message-hash across turns (up to 2) and re-adds
       // them on future bodies. Runs AFTER cache_control_normalize (when
       // present) so normalize pins the canonical tail-marker first and
       // sticky re-adds the historical ones. State file is per-project at