claude-code-cache-fix 2.0.0-beta.4 → 2.0.0

package/README.md

@@ -4,7 +4,7 @@
 
 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 v2.1.111. Opus 4.7 compatible.
+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 v2.1.112. 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.
 
@@ -607,6 +607,7 @@ measurable signature of cache-efficiency degradation.
 - **[@JEONG-JIWOO](https://github.com/JEONG-JIWOO)** — VS Code extension investigation: discovered `claudeCode.claudeProcessWrapper` as the working integration path, wrote the C wrapper for Windows (#16)
 - **[@X-15](https://github.com/X-15)** — VS Code extension validation, per-fix health status analysis confirming safety check behavior on v2.1.105 (#16)
 - **[@ArkNill](https://github.com/ArkNill)** — Fingerprint verification fix for CC v2.1.108+ (`isMeta` filter change, PR #21), Korean README (PR #22), original [claude-code-hidden-problem-analysis](https://github.com/ArkNill/claude-code-hidden-problem-analysis) research
+- **[@deafsquad](https://github.com/deafsquad)** — Universal smoosh_split un-smoosh fix (PR #26), source-level function attribution of resume scatter bug (anthropics/claude-code#43657), OTEL telemetry discovery
 
 If you contributed to the community effort on these issues and aren't listed here, please open an issue or PR — we want to credit everyone properly.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-cache-fix",
-  "version": "2.0.0-beta.4",
+  "version": "2.0.0",
   "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

@@ -330,6 +330,403 @@ function stripSessionKnowledge(text) {
   );
 }
 
+// --------------------------------------------------------------------------
+// SessionStart:resume → :startup rewrite (Bug: anthropics/claude-code#43657)
+// --------------------------------------------------------------------------
+//
+// On `claude --continue`, CC fires processSessionStartHooks('resume', …) at
+// src/utils/sessionStart.ts:35. The resulting attachment text wraps the
+// hook's stdout in `<system-reminder>\nSessionStart:resume hook success: …`.
+// The original (pre-resume) session sent the same block as
+// `SessionStart:startup hook success: …`. Byte difference at msg[0] content[N]
+// → whole message prefix re-caches → full-session-cost miss.
+//
+// Some SessionStart hooks additionally embed `<session-id>` tags or
+// `Last active: <timestamp>` lines inside the reminder body, both of which
+// carry UUID/date volatility on top of the event-name flip.
+//
+// This helper rewrites the outbound text to match the originally-cached
+// form. Runs on both standalone text blocks and tool_result.content strings
+// (covers the case where the SessionStart reminder got smooshed by CC's
+// smooshSystemReminderSiblings pass before we see it).
+//
+// Agent behavior is unaffected — CC does not condition behavior on the
+// event-name text, and session-id / timestamps are ephemeral runtime
+// metadata, not semantic inputs.
+// --------------------------------------------------------------------------
+
+const SESSION_START_RESUME_MARKER = /SessionStart:resume hook success:/g;
+const SESSION_START_ID_TAG = /\n?<session-id>[^<]*<\/session-id>/g;
+const SESSION_START_LAST_ACTIVE_LINE = /\nLast active:[^\n]*/g;
+
+/**
+ * Normalize a single text payload (a text block's .text or a tool_result's
+ * string .content) to remove SessionStart-resume volatility. Returns
+ * [newText, mutationCount]. Callers only need the text, but the count is
+ * exposed for stats. The function is a pure string-to-string transform
+ * (idempotent: running twice produces the same output as running once).
+ */
+function normalizeSessionStartText(text) {
+  if (typeof text !== "string" || !text.includes("SessionStart:")) return [text, 0];
+  let count = 0;
+  let out = text;
+  if (SESSION_START_RESUME_MARKER.test(out)) {
+    SESSION_START_RESUME_MARKER.lastIndex = 0;
+    out = out.replace(SESSION_START_RESUME_MARKER, "SessionStart:startup hook success:");
+    count++;
+  }
+  if (SESSION_START_ID_TAG.test(out)) {
+    SESSION_START_ID_TAG.lastIndex = 0;
+    out = out.replace(SESSION_START_ID_TAG, "");
+    count++;
+  }
+  if (SESSION_START_LAST_ACTIVE_LINE.test(out)) {
+    SESSION_START_LAST_ACTIVE_LINE.lastIndex = 0;
+    out = out.replace(SESSION_START_LAST_ACTIVE_LINE, "");
+    count++;
+  }
+  return [out, count];
+}
+
+// --------------------------------------------------------------------------
+// Continue-trailer strip (Bug: anthropics/claude-code#12 / resume UX)
+// --------------------------------------------------------------------------
+//
+// On `claude --continue`, CC appends a text block whose text is EXACTLY
+// "Continue from where you left off." to the last user message before
+// firing the first post-resume request. The pre-exit body did not carry
+// that block, so its presence in the resumed body creates a tail-of-last-
+// user-message drift (~40 bytes plus JSON framing) that breaks cache at
+// that position.
+//
+// The trailer is a semantic no-op — the agent already has the full prior
+// conversation as context. Removing it makes the post-resume body byte-
+// match what the pre-exit body cached at the tail.
+//
+// Match is intentionally narrow (exact string equality on the block's
+// .text) so mentions of the phrase inside a longer user sentence don't
+// get caught.
+// --------------------------------------------------------------------------
+
+const CONTINUE_TRAILER_TEXT = "Continue from where you left off.";
+
+/**
+ * Returns true iff the block is an exact-match Continue-trailer text block
+ * (a `{type: "text", text: "Continue from where you left off."}` shape —
+ * cache_control field on the same block is allowed and ignored). Pure
+ * predicate; exported for unit tests.
+ */
+function isContinueTrailerBlock(block) {
+  return (
+    !!block &&
+    typeof block === "object" &&
+    block.type === "text" &&
+    block.text === CONTINUE_TRAILER_TEXT
+  );
+}
+
+// --------------------------------------------------------------------------
+// Deferred-tools restore (MCP reconnect race)
+// --------------------------------------------------------------------------
+//
+// Observed empirically: on `claude --continue`, if MCP servers haven't
+// finished reconnecting by the time CC fires the first post-resume
+// request, the `<system-reminder>The following deferred tools are now
+// available via ToolSearch…` block at msg[0] (or wherever the attachment
+// lands post-compaction) shrinks dramatically. A full list of ~40 tools
+// collapses to a handful of CC built-ins (AskUserQuestion, EnterPlanMode,
+// ExitPlanMode, PushNotification) and CC injects a trailing
+// `The following deferred tools are no longer available (their MCP server
+// disconnected). Do not search for them — ToolSearch will return no match:`
+// notice.
+//
+// That block change at the root of the message array breaks cache at the
+// very top — the entire ~940K prompt re-caches. By the time the second
+// post-resume request fires, MCPs are usually reconnected and the block is
+// full again, but the cache is already committed to the shrunk version
+// for this session.
+//
+// This extension snapshots the block to
+// `~/.claude/cache-fix-state/deferred-tools-<sha1(key)>.txt` every time
+// it's sent in its full form (no UNAVAILABLE marker), keyed by a caller-
+// supplied project key (default: cwd). On a subsequent request where the
+// block is shorter AND contains the UNAVAILABLE marker, the persisted
+// full bytes are substituted so the on-wire body matches the server's
+// cached prefix.
+//
+// Trade-off: the restored block may reference MCP tools that haven't
+// actually reconnected yet. Agent calls ToolSearch → no match → one retry.
+// Tiny cost versus a full-prompt cache miss on every resume.
+// --------------------------------------------------------------------------
+
+const DEFERRED_TOOLS_AVAILABLE_MARKER =
+  "The following deferred tools are now available via ToolSearch";
+const DEFERRED_TOOLS_UNAVAILABLE_MARKER =
+  "The following deferred tools are no longer available";
+const DEFERRED_TOOLS_SNAPSHOT_DIR = join(homedir(), ".claude", "cache-fix-state");
+
+/**
+ * Build the absolute snapshot path for a given key. Exported for tests so
+ * they can assert on path derivation without duplicating the hash logic.
+ */
+function deferredToolsSnapshotPath(key) {
+  const hash = createHash("sha1").update(String(key)).digest("hex").slice(0, 16);
+  return join(DEFERRED_TOOLS_SNAPSHOT_DIR, `deferred-tools-${hash}.txt`);
+}
+
+/**
+ * Locate the deferred-tools reminder block anywhere in `body.messages`.
+ * The block's position varies by session shape (pre-compaction it often
+ * sits at `msg[0].content[0]`; post-compaction it can land at
+ * `msg[1].content[N]` next to other attachments). Returns
+ * `{ msgIdx, blockIdx, text } | null`.
+ *
+ * Assistant messages are skipped so that if the agent happens to mention
+ * the AVAILABLE_MARKER phrase verbatim in its own output, we don't
+ * misidentify it as a real deferred-tools block.
+ */
+function findDeferredToolsBlockInBody(body) {
+  if (!body || !Array.isArray(body.messages)) return null;
+  for (let m = 0; m < body.messages.length; m++) {
+    const msg = body.messages[m];
+    if (msg?.role !== "user" || !Array.isArray(msg.content)) continue;
+    for (let i = 0; i < msg.content.length; i++) {
+      const b = msg.content[i];
+      if (
+        b?.type === "text" &&
+        typeof b.text === "string" &&
+        b.text.includes(DEFERRED_TOOLS_AVAILABLE_MARKER)
+      ) {
+        return { msgIdx: m, blockIdx: i, text: b.text };
+      }
+    }
+  }
+  return null;
+}
+
+// --------------------------------------------------------------------------
+// Bookkeeping-reminder strip
+// --------------------------------------------------------------------------
+//
+// Complements `smoosh_normalize` / `smoosh_split`: where normalize stabilizes
+// bytes in-place and split peels smooshed reminders back into standalone
+// text blocks, this pass REMOVES purely-bookkeeping reminder blocks entirely
+// from the outbound body. Zero model visibility, zero drift.
+//
+// Targeted patterns (all CC-internal, per-turn values the agent doesn't need
+// to condition behavior on):
+//   - `Token usage: <N>/<M>; <K> remaining`
+//   - `Output tokens — turn: <X> · session: <Y>`
+//   - `USD budget: $<X>/$<Y>; $<Z> remaining`
+//   - `The task tools haven't been used recently. …`
+//   - `The TodoWrite tool hasn't been used recently. …`
+//   - `Remaining conversation turns: <N>`
+//   - `Messages until auto-compact: <N>`
+//
+// Hook-injected reminders (thinking-enrichment, action-tracker,
+// PreToolUse/PostToolUse blocking errors, UserPromptSubmit additional
+// context, custom user hooks) are deliberately NOT stripped here — the
+// agent needs that feedback visible in the turn it fires, and attempting a
+// history-only filter creates per-turn drift of its own (the "last user
+// message" shifts each turn, so a reminder preserved at turn N gets
+// stripped at N+1 when its host message falls into history). Leaving hook
+// reminders untouched is the safer choice; their residual drift is small
+// compared to bookkeeping churn.
+// --------------------------------------------------------------------------
+
+const REMINDER_WRAP_REGEX =
+  /^<system-reminder>\n([\s\S]*?)\n<\/system-reminder>\s*$/;
+
+const BOOKKEEPING_REMINDER_PATTERNS = [
+  /^Token usage: \d+\/\d+; \d+ remaining\s*$/,
+  /^Output tokens \u2014 turn: [^\n]+ \u00b7 session: [^\n]+\s*$/,
+  /^USD budget: \$[\d.]+\/\$[\d.]+; \$[\d.]+ remaining\s*$/,
+  /^The task tools haven't been used recently\./,
+  /^The TodoWrite tool hasn't been used recently\./,
+  /^Remaining conversation turns: /,
+  /^Messages? until auto-compact: /,
+];
+
+/**
+ * Returns true iff the text is a `<system-reminder>`-wrapped block whose
+ * inner content matches a bookkeeping pattern. Pure predicate, exported
+ * for unit tests.
+ */
+function isBookkeepingReminder(text) {
+  if (typeof text !== "string") return false;
+  const m = text.match(REMINDER_WRAP_REGEX);
+  if (!m) return false;
+  const inner = m[1];
+  for (const rx of BOOKKEEPING_REMINDER_PATTERNS) {
+    if (rx.test(inner)) return true;
+  }
+  return false;
+}
+
+// --------------------------------------------------------------------------
+// cache_control marker position-normalizer
+// --------------------------------------------------------------------------
+//
+// Anthropic's prompt-cache uses `cache_control: {type: "ephemeral", ttl: ...}`
+// markers on content blocks as cache breakpoints. CC places this marker on
+// "the last block of the last user message" each turn — which shifts as new
+// turns arrive. When the marker moves, the PREVIOUS last-block's JSON loses
+// the cache_control field → that block's bytes differ from the server's
+// cached version → partial re-cache on top of the stable system-prompt
+// cache.
+//
+// Enforce a canonical position on every outbound body:
+//   1. Strip every existing cache_control marker from user-message content
+//      blocks.
+//   2. Place a single {type: "ephemeral", ttl: "1h"} marker on the LAST
+//      content block of the LAST user message.
+//
+// Fast path: if the canonical block already has the correct marker AND it's
+// the only user-side marker, the body is left untouched — ensures the pass
+// is a true no-op when nothing changed.
+//
+// System-side markers (e.g., on `system[2]` for the global prompt) are NOT
+// touched — they're CC's stable breakpoint for the system prompt and work
+// correctly.
+// --------------------------------------------------------------------------
+
+const CACHE_CONTROL_CANONICAL_MARKER = { type: "ephemeral", ttl: "1h" };
+
+/**
+ * Strip every cache_control marker from a single user message's content
+ * blocks. Returns the number stripped. Mutates the message's content array
+ * in place.
+ */
+function stripCacheControlMarkers(msg) {
+  if (!msg || msg.role !== "user" || !Array.isArray(msg.content)) return 0;
+  let n = 0;
+  for (let i = 0; i < msg.content.length; i++) {
+    const block = msg.content[i];
+    if (block && typeof block === "object" && block.cache_control) {
+      const { cache_control, ...rest } = block;
+      msg.content[i] = rest;
+      n++;
+    }
+  }
+  return n;
+}
+
+/**
+ * Count cache_control markers across all user-message content blocks.
+ * Exported so the call-site's fast-path check has a tested helper.
+ */
+function countUserCacheControlMarkers(body) {
+  if (!body || !Array.isArray(body.messages)) return 0;
+  let n = 0;
+  for (const msg of body.messages) {
+    if (msg?.role !== "user" || !Array.isArray(msg.content)) continue;
+    for (const block of msg.content) {
+      if (block && typeof block === "object" && block.cache_control) n++;
+    }
+  }
+  return n;
+}
+
+// --------------------------------------------------------------------------
+// tool_use.input field-set normalization
+// --------------------------------------------------------------------------
+//
+// CC's serialization of `tool_use.input` can drift between turns when the
+// caller passes fields not declared in the tool's `input_schema.properties`.
+// Observed case: a SendMessage tool call where the caller passed
+// `{to, summary, message, type, recipient, content}`. Pre-miss body
+// serialized input as `{to, summary, message}` (3 schema-only keys).
+// Post-miss body (same tool_use_id, same turn position) serialized the
+// same block as `{to, summary, message, type, recipient, content}` (6 keys
+// — extras preserved). That byte drift at a mid-history assistant message
+// re-caches every block from that message forward → full-session-cost miss.
+//
+// Concrete instance: 2334-byte drift on ONE assistant-side tool_use block
+// caused a 619,722 `cache_creation_input_tokens` miss at 15:16:52 UTC on
+// msg[844] of a long-running session.
+//
+// This helper walks every assistant-role message's tool_use blocks, looks
+// up the tool's declared `input_schema.properties` from `body.tools`, and
+// rewrites `input` to contain ONLY the schema keys (in schema declaration
+// order). Tools with no schema in `body.tools` are left untouched — we
+// can't determine what's legitimate vs extra.
+//
+// Agent behavior is unaffected — extras weren't declared in the schema so
+// downstream consumers shouldn't rely on them. The point of this pass is
+// to pin the serialization to the schema's field set so CC's own drift
+// between turns can't break cache.
+// --------------------------------------------------------------------------
+
+/**
+ * Mutate `body` in place: for every assistant-role message's tool_use
+ * blocks whose tool name matches an entry in `body.tools` with a known
+ * `input_schema.properties`, replace `input` with a new object containing
+ * ONLY the schema-declared keys, preserved in schema declaration order.
+ * Returns the count of tool_use blocks modified (0 if nothing changed or
+ * preconditions missing). Pure transform: safe to call repeatedly.
+ */
+function normalizeToolUseInputsInBody(body) {
+  if (!body || typeof body !== "object") return 0;
+  if (!Array.isArray(body.messages) || !Array.isArray(body.tools)) return 0;
+
+  // Build toolSchemas: { name: orderedKeys[] } from body.tools entries
+  // that declare input_schema.properties.
+  const toolSchemas = Object.create(null);
+  for (const tool of body.tools) {
+    if (!tool || typeof tool !== "object") continue;
+    const name = tool.name;
+    if (typeof name !== "string") continue;
+    const props = tool.input_schema && tool.input_schema.properties;
+    if (!props || typeof props !== "object") continue;
+    toolSchemas[name] = Object.keys(props);
+  }
+
+  let modified = 0;
+  for (const msg of body.messages) {
+    if (!msg || msg.role !== "assistant") continue;
+    if (!Array.isArray(msg.content)) continue;
+    for (let i = 0; i < msg.content.length; i++) {
+      const block = msg.content[i];
+      if (!block || block.type !== "tool_use") continue;
+      if (!block.input || typeof block.input !== "object" || Array.isArray(block.input)) continue;
+      const schemaKeys = toolSchemas[block.name];
+      if (!schemaKeys) continue; // unknown tool — skip
+      const currentKeys = Object.keys(block.input);
+      // Determine if any non-schema key is present. If all current keys
+      // are in schema AND their order already matches a subset of
+      // schemaKeys order, we could skip — but we always rebuild to also
+      // canonicalize key order, which is what JSON.stringify consumers
+      // depend on for byte stability.
+      const schemaKeySet = new Set(schemaKeys);
+      const hasExtras = currentKeys.some((k) => !schemaKeySet.has(k));
+      // Also rebuild when order differs from schema declaration order,
+      // because extras stripping alone doesn't guarantee a canonical
+      // byte sequence across turns.
+      const presentSchemaKeys = schemaKeys.filter((k) =>
+        Object.prototype.hasOwnProperty.call(block.input, k)
+      );
+      const currentInSchema = currentKeys.filter((k) => schemaKeySet.has(k));
+      let orderDiffers = presentSchemaKeys.length !== currentInSchema.length;
+      if (!orderDiffers) {
+        for (let j = 0; j < presentSchemaKeys.length; j++) {
+          if (presentSchemaKeys[j] !== currentInSchema[j]) {
+            orderDiffers = true;
+            break;
+          }
+        }
+      }
+      if (!hasExtras && !orderDiffers) continue;
+      const newInput = {};
+      for (const k of presentSchemaKeys) {
+        newInput[k] = block.input[k];
+      }
+      msg.content[i] = { ...block, input: newInput };
+      modified++;
+    }
+  }
+  return modified;
+}
+
 /**
  * Core fix: on EVERY call, scan the entire message array for the LATEST
  * relocatable blocks (skills, MCP, deferred tools, hooks) and ensure they
@@ -728,6 +1125,12 @@ const _STATS_SCHEMA = {
   cwd_normalize: { applied: 0, skipped: 0, lastApplied: null },
   smoosh_normalize: { applied: 0, skipped: 0, lastApplied: null },
   smoosh_split: { applied: 0, skipped: 0, lastApplied: null },
+  session_start_normalize: { applied: 0, skipped: 0, lastApplied: null },
+  continue_trailer_strip: { applied: 0, skipped: 0, lastApplied: null },
+  deferred_tools_restore: { applied: 0, skipped: 0, lastApplied: null },
+  reminder_strip: { applied: 0, skipped: 0, lastApplied: null },
+  cache_control_normalize: { applied: 0, skipped: 0, lastApplied: null },
+  tool_use_input_normalize: { applied: 0, skipped: 0, lastApplied: null },
 };
 
 function _createEmptyStats() {
@@ -1349,6 +1752,67 @@ globalThis.fetch = async function (url, options) {
         }
       }
 
+      // Extension: session_start_normalize — SessionStart:resume → :startup rewrite
+      // and ephemeral session-id / Last-active strip. Runs BEFORE smoosh_normalize
+      // so drift at msg[0] content[N] is stabilized before any subsequent pass
+      // reads from the same text. Applies to both standalone text blocks and
+      // tool_result.content strings (in case CC's smooshSystemReminderSiblings
+      // folded the reminder before we see it).
+      // Bug: anthropics/claude-code#43657
+      // Opt-out via CACHE_FIX_SKIP_SESSION_START_NORMALIZE=1 (defaults ON).
+      if (shouldApplyFix("session_start_normalize") && payload.messages) {
+        let ssnApplied = 0;
+        for (const msg of payload.messages) {
+          if (msg?.role !== "user" || !Array.isArray(msg.content)) continue;
+          for (let i = 0; i < msg.content.length; i++) {
+            const block = msg.content[i];
+            if (block?.type === "text" && typeof block.text === "string") {
+              const [t, n] = normalizeSessionStartText(block.text);
+              if (n > 0) {
+                msg.content[i] = { ...block, text: t };
+                ssnApplied += n;
+              }
+            } else if (block?.type === "tool_result" && typeof block.content === "string") {
+              const [c, n] = normalizeSessionStartText(block.content);
+              if (n > 0) {
+                msg.content[i] = { ...block, content: c };
+                ssnApplied += n;
+              }
+            }
+          }
+        }
+        if (ssnApplied > 0) {
+          modified = true;
+          debugLog(`APPLIED: session-start-normalize rewrote ${ssnApplied} marker(s)`);
+          recordFixResult("session_start_normalize", "applied");
+        } else {
+          recordFixResult("session_start_normalize", "skipped");
+        }
+      }
+
+      // Extension: tool_use_input_normalize — strip tool_use.input keys not
+      // declared in body.tools[*].input_schema.properties. CC's serialization
+      // of tool_use.input can drift between turns when the caller passed
+      // extra fields; the pre-miss body may serialize only the schema keys
+      // while the post-miss body serializes the full caller-supplied set
+      // (or vice versa). That byte drift at a mid-history assistant message
+      // re-caches every block from that message forward.
+      //
+      // Runs AFTER session_start_normalize so mid-history drift is pinned
+      // before any downstream pass (smoosh_*, fingerprint, ttl) hashes the
+      // same block. Default ON, opt-out via
+      // CACHE_FIX_SKIP_TOOL_USE_INPUT_NORMALIZE=1.
+      if (shouldApplyFix("tool_use_input_normalize")) {
+        const tuinApplied = normalizeToolUseInputsInBody(payload);
+        if (tuinApplied > 0) {
+          modified = true;
+          debugLog(`APPLIED: tool-use-input-normalize rewrote ${tuinApplied} tool_use block(s)`);
+          recordFixResult("tool_use_input_normalize", "applied");
+        } else {
+          recordFixResult("tool_use_input_normalize", "skipped");
+        }
+      }
+
       // 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,
@@ -1474,6 +1938,159 @@ globalThis.fetch = async function (url, options) {
         }
       }
 
+      // Extension: continue_trailer_strip — remove the "Continue from where
+      // you left off." text block CC appends to the last user message on
+      // --continue. Pre-exit bodies didn't carry it, so its presence in the
+      // resumed body creates tail-of-last-msg drift that breaks cache.
+      // Exact-match string equality on `.text` — user sentences mentioning
+      // the phrase inside longer content are not touched.
+      // Bug: anthropics/claude-code#12 (resume UX), observed empirically.
+      // Opt-out via CACHE_FIX_SKIP_CONTINUE_TRAILER_STRIP=1 (defaults ON).
+      if (shouldApplyFix("continue_trailer_strip") && payload.messages) {
+        let trailerStripped = 0;
+        for (const msg of payload.messages) {
+          if (msg?.role !== "user" || !Array.isArray(msg.content)) continue;
+          const kept = msg.content.filter((block) => {
+            if (isContinueTrailerBlock(block)) {
+              trailerStripped++;
+              return false;
+            }
+            return true;
+          });
+          if (kept.length !== msg.content.length) msg.content = kept;
+        }
+        if (trailerStripped > 0) {
+          modified = true;
+          debugLog(`APPLIED: continue-trailer-strip removed ${trailerStripped} trailer block(s)`);
+          recordFixResult("continue_trailer_strip", "applied");
+        } else {
+          recordFixResult("continue_trailer_strip", "skipped");
+        }
+      }
+
+      // Extension: deferred_tools_restore — persist-and-restore the
+      // deferred-tools attachment block across sessions so MCP reconnect
+      // race at resume-time doesn't shrink msg[0] and bust the whole cache.
+      // Snapshot key defaults to process.cwd() (one snapshot per project).
+      // Opt-out via CACHE_FIX_SKIP_DEFERRED_TOOLS_RESTORE=1 (defaults ON).
+      if (shouldApplyFix("deferred_tools_restore") && payload.messages) {
+        let dtrRestored = 0;
+        const found = findDeferredToolsBlockInBody(payload);
+        if (found) {
+          const hasUnavail = found.text.includes(DEFERRED_TOOLS_UNAVAILABLE_MARKER);
+          const snapshotPath = deferredToolsSnapshotPath(process.cwd());
+          if (!hasUnavail) {
+            // Clean baseline — persist it for future resumes. Silent on
+            // any I/O error; snapshot is best-effort.
+            try {
+              mkdirSync(DEFERRED_TOOLS_SNAPSHOT_DIR, { recursive: true });
+              writeFileSync(snapshotPath, found.text, "utf-8");
+            } catch {}
+          } else {
+            // Shrunk block with explicit "no longer available" signal →
+            // attempt restore. Only substitute if the persisted version is
+            // strictly longer (never downgrade to a stale shorter snapshot).
+            let snapshot = null;
+            try { snapshot = readFileSync(snapshotPath, "utf-8"); } catch {}
+            if (snapshot && snapshot.length > found.text.length) {
+              const targetMsg = payload.messages[found.msgIdx];
+              const newContent = targetMsg.content.slice();
+              newContent[found.blockIdx] = { ...newContent[found.blockIdx], text: snapshot };
+              payload.messages[found.msgIdx] = { ...targetMsg, content: newContent };
+              dtrRestored = 1;
+            }
+          }
+        }
+        if (dtrRestored > 0) {
+          modified = true;
+          debugLog(`APPLIED: deferred-tools-restore substituted full block at msg[${found.msgIdx}].content[${found.blockIdx}]`);
+          recordFixResult("deferred_tools_restore", "applied");
+        } else {
+          recordFixResult("deferred_tools_restore", "skipped");
+        }
+      }
+
+      // Extension: reminder_strip — remove bookkeeping system-reminder blocks
+      // (Token usage / USD budget / Output tokens / TodoWrite nudge / turn
+      // counters) entirely from user messages. Runs AFTER smoosh_split so
+      // blocks peeled out of tool_result.content are visible as standalone
+      // text and can be matched by isBookkeepingReminder.
+      // Zero model visibility, zero drift.
+      // Opt-out via CACHE_FIX_SKIP_REMINDER_STRIP=1 (defaults ON).
+      if (shouldApplyFix("reminder_strip") && payload.messages) {
+        let reminderStripped = 0;
+        for (const msg of payload.messages) {
+          if (msg?.role !== "user" || !Array.isArray(msg.content)) continue;
+          const kept = msg.content.filter((block) => {
+            if (block?.type !== "text") return true;
+            if (isBookkeepingReminder(block.text)) {
+              reminderStripped++;
+              return false;
+            }
+            return true;
+          });
+          if (kept.length !== msg.content.length) msg.content = kept;
+        }
+        if (reminderStripped > 0) {
+          modified = true;
+          debugLog(`APPLIED: reminder-strip removed ${reminderStripped} bookkeeping reminder block(s)`);
+          recordFixResult("reminder_strip", "applied");
+        } else {
+          recordFixResult("reminder_strip", "skipped");
+        }
+      }
+
+      // Extension: cache_control_normalize — pin the cache_control marker at
+      // a canonical position (last block of last user message) on every
+      // outbound body. Prevents marker-shuffle drift between turns from
+      // invalidating the previous-last-block's cached bytes. Runs LAST
+      // (after smoosh_split and any other content-mutating pass) so the
+      // canonical position is calculated against the final content array.
+      // Fast path: if canonical position already holds the correct marker
+      // and it's the only user-side marker, body passes through untouched.
+      // Opt-out via CACHE_FIX_SKIP_CACHE_CONTROL_NORMALIZE=1 (defaults ON).
+      if (shouldApplyFix("cache_control_normalize") && payload.messages && payload.messages.length > 0) {
+        // Locate canonical position: last block of last user message with an
+        // array content. If no valid target, skip.
+        let targetMsgIdx = -1;
+        let targetBlockIdx = -1;
+        for (let i = payload.messages.length - 1; i >= 0; i--) {
+          const m = payload.messages[i];
+          if (m?.role !== "user") continue;
+          if (!Array.isArray(m.content) || m.content.length === 0) break;
+          targetMsgIdx = i;
+          targetBlockIdx = m.content.length - 1;
+          break;
+        }
+
+        let ccMutated = false;
+        if (targetMsgIdx !== -1) {
+          const targetBlock = payload.messages[targetMsgIdx].content[targetBlockIdx];
+          const existingCC = targetBlock?.cache_control;
+          const canonicalAlreadyCorrect =
+            existingCC &&
+            existingCC.type === CACHE_CONTROL_CANONICAL_MARKER.type &&
+            existingCC.ttl === CACHE_CONTROL_CANONICAL_MARKER.ttl;
+
+          if (!(canonicalAlreadyCorrect && countUserCacheControlMarkers(payload) === 1)) {
+            // Strip all markers from user messages, then place canonical.
+            for (const msg of payload.messages) stripCacheControlMarkers(msg);
+            const tm = payload.messages[targetMsgIdx];
+            const newContent = tm.content.slice();
+            newContent[targetBlockIdx] = { ...newContent[targetBlockIdx], cache_control: { ...CACHE_CONTROL_CANONICAL_MARKER } };
+            payload.messages[targetMsgIdx] = { ...tm, content: newContent };
+            ccMutated = true;
+          }
+        }
+        if (ccMutated) {
+          modified = true;
+          debugLog(`APPLIED: cache_control_normalize pinned marker at msg[${targetMsgIdx}].content[${targetBlockIdx}]`);
+          recordFixResult("cache_control_normalize", "applied");
+        } else {
+          recordFixResult("cache_control_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".
@@ -1899,5 +2516,17 @@ export {
   isClearArtifact,
   rewriteOutputEfficiencyInstruction,
   normalizeOutputEfficiencyReplacement,
+  normalizeSessionStartText,
+  isContinueTrailerBlock,
+  CONTINUE_TRAILER_TEXT,
+  findDeferredToolsBlockInBody,
+  deferredToolsSnapshotPath,
+  DEFERRED_TOOLS_AVAILABLE_MARKER,
+  DEFERRED_TOOLS_UNAVAILABLE_MARKER,
+  isBookkeepingReminder,
+  stripCacheControlMarkers,
+  countUserCacheControlMarkers,
+  CACHE_CONTROL_CANONICAL_MARKER,
+  normalizeToolUseInputsInBody,
   _pinnedBlocks,  // exported so tests can reset between runs
 };