Check Cache BugsCAUTION

The most comprehensive Claude Code guide: agentic workflows, hooks, skills, MCP servers, quizzes, and production-ready templates. 430K+ lines.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: check-cache-bugs
description: "Audit Claude Code setup for cache bugs (CC#40524): sentinel, --resume/--continue, attribution header + ArkNill B3/B4/B5"
effort: low
disable-model-invocation: true
---

# Check cache bugs (CC#40524)

Audit your Claude Code setup for cache and cost bugs discovered in March-April 2026.

**Time**: ~30 seconds | **Scope**: version, config files, CLAUDE.md, skills, hooks, shell profiles, all claude binaries

> **Note on cache contamination**: This skill loads content containing `cch=` strings into the current session's message array. For cleanest results, run this command at the very start of a fresh session, or via `claude -p "$(cat .claude/commands/check-cache-bugs.md)"` as a one-shot print-mode invocation.

**Reference**: `anthropics/claude-code#40524` | **Discovered by**: `@jmarianski` + `@whiletrue0x` | **Extended by**: `@ArkNill` (ArkNill/claude-code-cache-analysis, April 2026)

---

## Background

### Fix Status (current as of v2.1.92)

| Bug | Versions affected | Status |
|-----|------------------|--------|
| Bug 1: cch sentinel (standalone binary) | v2.1.36-v2.1.90 | **FIXED in v2.1.91** |
| Bug 2: deferred_tools_delta on --resume | v2.1.69-v2.1.89 | **FIXED in v2.1.90** |
| Bug 3: attribution header per-session hash | v2.1.69+ | **Active** (env var workaround) |
| B4: Microcompact / silent context stripping | all versions through v2.1.92 | **Active** (GrowthBook controlled) |
| B5: Tool result budget cap 200K | all versions through v2.1.92 | **Active** (MCP tools exempted) |

### Original three bugs (CC#40524)

- **Bug 1** (FIXED v2.1.91): Bun's native HTTP layer did a same-length byte replacement of the
  `cch=00000` attestation placeholder after `JSON.stringify` but before TLS. Triggered only if
  `cch=00000` appeared literally in `messages[]` content. Confirmed closed; npm and standalone
  binary are now equivalent on v2.1.91+.

- **Bug 2** (FIXED v2.1.90): The session JSONL writer stripped `deferred_tools_delta` attachment
  records before writing to disk. On `--resume`, those records were absent; the deferred tools
  layer had no prior history and re-announced all tools from scratch, shifting every message
  position and breaking the messages-level cache prefix entirely. Each resume rebuilt 87-118K
  tokens as `cache_creation`. Anthropic tracked internally as inc-4747.

- **Bug 3** (ACTIVE, v2.1.69+): Claude Code injects a billing header as the first block of the
  system prompt. The header contains a 3-character SHA-256 hash derived from characters [4, 7, 20]
  of the first user message + CC version, unique per session, per subagent, per side-query. Since
  the cache is prefix-based, this causes a cold miss on ~12K system prompt tokens on every
  invocation. A session with 5 subagents = 6 cold misses. Empirical: 48% -> 99.98% cache hit with
  env var workaround.

### Additional bugs (ArkNill, April 2026)

- **B4: Microcompact** (all versions): Three mechanisms silently replace tool results with
  `[Old tool result content cleared]` before sending to the API. Controlled by server-side
  GrowthBook flags, bypasses `DISABLE_AUTO_COMPACT`. The `/export` command shows original
  context, NOT what the model received. 327 clearing events measured in one session.

- **B5: Tool result budget cap** (all versions): `applyToolResultBudget()` truncates tool results
  when aggregate chars exceed 200K. Built-in tools (Read, Bash, Grep, Glob, Edit) are affected.
  MCP tools can override via `_meta["anthropic/maxResultSizeChars"]` (v2.1.91+).

Cost impact summary: 2-5x per session start/subagent for Bug 3 in isolation; 10-20x per turn for
Bug 2 with 10+ skills and 3-4 resumes (both estimates are correct in their respective context).

---

## Instructions

You are an auditor. Run all phases in order, collect every result, and produce the final report. Do not skip phases or stop early.

---

### Phase 1: Claude Code version and install method

```bash
# Version check
claude --version

# All installed claude binaries
which -a claude 2>/dev/null

# Check if active binary is standalone or npm
file $(which claude) 2>/dev/null
ls -la $(which claude) 2>/dev/null
```

- Version < 2.1.90 -> flag **BUG 2 RISK** (resume/DTD; fixed in v2.1.90)
- Version < 2.1.91 AND standalone binary -> flag **BUG 1 MECHANISM** (sentinel; fixed in v2.1.91)
- Version >= 2.1.69 -> flag **BUG 3 RISK** (attribution header; no fix yet, check Phase 4)
- Binary is Mach-O / ELF executable -> standalone (relevant for Bug 1 on v2.1.36-v2.1.90)
- Binary is a symlink to `node_modules` or contains `cli.js` -> npm/npx

---

### Phase 2: Sentinel scan (Bug 1)

Search for the literal string `cch=` in all static config files. Exclude `.jsonl` files (ephemeral conversation history) and this command file itself.

```bash
# Global config files
grep -r "cch=" \
  ~/.claude/CLAUDE.md \
  ~/.claude/MEMORY.md \
  ~/.claude/TONE.md \
  ~/.claude/FLAGS.md \
  ~/.claude/RULES.md \
  ~/.claude/RTK.md \
  ~/.claude/ANTI_AI.md \
  2>/dev/null

# Global skills, commands, agents, hooks (excluding this command file)
grep -rl "cch=" ~/.claude/skills/ 2>/dev/null
grep -rl "cch=" ~/.claude/commands/ --exclude="check-cache-bugs.md" 2>/dev/null
grep -rl "cch=" ~/.claude/agents/ 2>/dev/null
grep -rl "cch=" ~/.claude/hooks/ 2>/dev/null

# Project-level config
grep -r "cch=" CLAUDE.md .claude/CLAUDE.md .claude/MEMORY.md 2>/dev/null
grep -rl "cch=" .claude/skills/ 2>/dev/null
grep -rl "cch=" .claude/commands/ --exclude="check-cache-bugs.md" 2>/dev/null
grep -rl "cch=" .claude/agents/ 2>/dev/null
grep -rl "cch=" .claude/hooks/ 2>/dev/null

# Broader scan: all CLAUDE.md files across projects
find ~ -name "CLAUDE.md" \
  -not -path "*/node_modules/*" \
  -not -path "*/.git/*" \
  2>/dev/null | xargs grep -l "cch=" 2>/dev/null
```

Flag as **BUG 1 RISK** if any match found outside `.jsonl` files AND version < v2.1.91.
On v2.1.91+: sentinel mechanism is fixed, no risk regardless of config content.

---

### Phase 3: Resume/conti

CAUTIONgrade B · trust 89/100 Install with care. The audit found things worth knowing before you trust its output.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (2)

whitepapers/recap-cards/en/_extensions

whitepapers/recap-cards/fr/_extensions

Gates applied: no_behavioural_pass.