fix(careful): close three check-careful bypasses via real JSON extraction

The grep-based command extractor stopped at the first escaped quote, so any
quoted argument truncated the command before the pattern checks ran —
`git commit -m "wip" && rm -rf /` was silently allowed. Replace it with a
python3/node JSON parse that fails CLOSED on unreadable payloads, add an
IFS/base64-to-shell obfuscation tripwire, and stop multi-line commands from
riding the single-line safe-exception whitelist (line-based grep would have
approved `rm -rf /` when a later line matched node_modules — a hazard the
real newline decoding exposed).

Contributed by @wtamminga (PR #2426; the -R hunk was dropped — it landed in
v1.61.0.0 — and output shapes updated to the nested hookSpecificOutput form).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

careful/bin/check-careful.sh

@@ -9,16 +9,47 @@ set -euo pipefail
 # Read stdin (JSON with tool_input)
 INPUT=$(cat)
 
-# Extract the "command" field value from tool_input
+# Extract the "command" field value from tool_input with a real JSON parser.
-# Try grep/sed first (handles 99% of cases), fall back to Python for escaped quotes
+#
-CMD=$(printf '%s' "$INPUT" | grep -o '"command"[[:space:]]*:[[:space:]]*"[^"]*"' | head -1 | sed 's/.*:[[:space:]]*"//;s/"$//' || true)
+# The previous extractor was
+#   grep -o '"command"[[:space:]]*:[[:space:]]*"[^"]*"'
+# whose [^"]* stops at the first escaped quote in the JSON string value. Any
+# destructive command preceded by a quoted argument was therefore truncated
+# away before the pattern checks ever ran:
+#
+#   git commit -m "wip" && rm -rf /   ->  CMD='git commit -m \'   -> allowed
+#   bash -c "rm -rf /"                ->  CMD='bash -c \'         -> allowed
+#   echo "x"; rm -rf ~                ->  CMD='echo \'            -> allowed
+#
+# The python3 fallback never rescued these because CMD was non-empty, so the
+# `[ -z "$CMD" ]` guard did not fire. Parse the payload properly instead, and
+# fail CLOSED when it cannot be parsed at all — a hook that gates destructive
+# commands must not allow-by-default on unreadable input.
+#
+# python3 is tried first because it ships with macOS and most Linux distros and
+# is reliably on PATH in a hook environment; node is the fallback.
+extract_cmd() {
+  if command -v python3 >/dev/null 2>&1; then
+    printf '%s' "$INPUT" | python3 -c 'import sys,json; d=json.loads(sys.stdin.read()); c=d.get("tool_input",{}).get("command",""); sys.stdout.write(c if isinstance(c,str) else "")' 2>/dev/null && return 0
+  fi
+  if command -v node >/dev/null 2>&1; then
+    printf '%s' "$INPUT" | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>{try{const j=JSON.parse(s);const c=(j&&j.tool_input&&j.tool_input.command)||"";process.stdout.write(typeof c==="string"?c:"")}catch(e){process.exit(3)}})' 2>/dev/null && return 0
+  fi
+  return 1
+}
 
-# Python fallback if grep returned empty (e.g., escaped quotes in command)
+set +e
-if [ -z "$CMD" ]; then
+CMD=$(extract_cmd)
-  CMD=$(printf '%s' "$INPUT" | python3 -c 'import sys,json; print(json.loads(sys.stdin.read()).get("tool_input",{}).get("command",""))' 2>/dev/null || true)
+EXTRACT_RC=$?
+set -e
+
+# No parser available, or the payload is not parseable JSON. Fail closed.
+if [ "$EXTRACT_RC" -ne 0 ] && [ -n "$INPUT" ]; then
+  printf '{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"ask","permissionDecisionReason":"[careful] Could not parse the tool payload to safety-check this command. Approve only if you know what it does."}}\n'
+  exit 0
 fi
 
-# If we still couldn't extract a command, allow
+# Parsed fine, but there is genuinely no command field (non-Bash payload) — allow.
 if [ -z "$CMD" ]; then
   echo '{}'
   exit 0
@@ -27,6 +58,23 @@ fi
 # Normalize: lowercase for case-insensitive SQL matching
 CMD_LOWER=$(printf '%s' "$CMD" | tr '[:upper:]' '[:lower:]')
 
+# --- Shell-obfuscation tripwire ---
+# Every check below inspects the command as a STRING, but bash executes what the
+# string MEANS after expansion. ${IFS} holds the default field separator and
+# contains no literal whitespace, so
+#
+#   rm${IFS}-rf${IFS}/
+#
+# matches none of the `rm\s+` patterns while executing as a full recursive
+# delete. The same holds for a command assembled by a base64 decode piped to a
+# shell. Rather than try to out-parse bash, treat these splitting/decoding
+# primitives as a reason to ask: they are vanishingly rare in commands a human
+# actually means to run unattended.
+if printf '%s' "$CMD" | grep -qE '\$\{IFS\}|\$IFS|\$\(echo[^)]*base64[^)]*\)|base64[[:space:]]+(-d|--decode)[^|]*\|[[:space:]]*(sh|bash)' 2>/dev/null; then
+  printf '{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"ask","permissionDecisionReason":"[careful] Shell obfuscation detected (IFS word-splitting or base64-to-shell). Read the command carefully before approving."}}\n'
+  exit 0
+fi
+
 # --- Check for safe exceptions (one standalone rm of build artifacts) ---
 # Match the complete command. Parsing only the last rm is unsafe because shell
 # syntax or comments can hide an earlier destructive command, for example:
@@ -39,10 +87,20 @@ CMD_LOWER=$(printf '%s' "$CMD" | tr '[:upper:]' '[:lower:]')
 #     ENDS in a whitelisted suffix (`rm -rf $(./wipe-all)/node_modules`)
 #     cannot ride the whitelist. Plain $VAR expansion (no parenthesis) is
 #     still allowed.
-if printf '%s' "$CMD" | grep -qE '^[[:space:]]*rm[[:space:]]+(-[a-zA-Z]*[rR][a-zA-Z]*[[:space:]]+|--recursive[[:space:]]+)(([^[:space:];&|#(`]*/)?(node_modules|\.next|dist|__pycache__|\.cache|build|\.turbo|coverage)[[:space:]]*)+$' 2>/dev/null; then
+#   - multi-line commands never ride the whitelist: grep matches the anchored
-  echo '{}'
+#     shape against EACH line, so `rm -rf /\nrm -rf node_modules` would be
-  exit 0
+#     allowed by its second line. With the JSON-parser extraction the \n in
-fi
+#     the payload is a real newline (the old grep extractor kept it as two
+#     literal characters, which broke the anchored match by accident).
+case "$CMD" in
+  *$'\n'*) : ;; # multi-line: fall through to the destructive checks
+  *)
+    if printf '%s' "$CMD" | grep -qE '^[[:space:]]*rm[[:space:]]+(-[a-zA-Z]*[rR][a-zA-Z]*[[:space:]]+|--recursive[[:space:]]+)(([^[:space:];&|#(`]*/)?(node_modules|\.next|dist|__pycache__|\.cache|build|\.turbo|coverage)[[:space:]]*)+$' 2>/dev/null; then
+      echo '{}'
+      exit 0
+    fi
+    ;;
+esac
 
 # --- Destructive pattern checks ---
 WARN=""

test/hook-scripts.test.ts

@@ -161,6 +161,21 @@ describe('check-careful.sh', () => {
       expect(output.hookSpecificOutput?.permissionDecision).toBeUndefined();
     });
 
+    // The old grep extractor stopped at the first escaped quote in the JSON
+    // string, so any quoted argument truncated the command BEFORE the pattern
+    // checks ran — hiding everything after it. (#2426)
+    test.each([
+      'git commit -m "wip" && rm -rf /',
+      'bash -c "rm -rf /"',
+      'echo "x"; rm -rf ~',
+      'npm run build --msg "done" && rm -rf /',
+    ])('a quoted argument cannot hide a later destructive command: %s', (command) => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, carefulInput(command));
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBe('ask');
+      expect(output.hookSpecificOutput?.permissionDecisionReason).toContain('recursive delete');
+    });
+
     // JSON-escaped newline (literal two-char \n surviving the grep extraction
     // path) breaks the anchored whitelist shape → falls through to the warn.
     test('newline-chained rm warns (escaped-newline separator branch)', () => {
@@ -197,11 +212,62 @@ describe('check-careful.sh', () => {
     });
   });
 
+  // --- Shell obfuscation ---
+
+  describe('shell obfuscation', () => {
+    test.each([
+      'rm${IFS}-rf${IFS}/',
+      'rm$IFS-rf$IFS/',
+      'echo cm0gLXJmIC8= | base64 -d | sh',
+    ])('asks when the command hides its shape behind expansion: %s', (command) => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, carefulInput(command));
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBe('ask');
+      expect(output.hookSpecificOutput?.permissionDecisionReason).toContain('obfuscation');
+    });
+
+    test('ordinary commands are unaffected', () => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, carefulInput('cat file.b64 | base64 -d > out.bin'));
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBeUndefined();
+    });
+  });
+
+  // --- JSON payload extraction ---
+
+  describe('command extraction', () => {
+    test('fails closed when the payload is not valid JSON', () => {
+      const { exitCode, output } = runHookRaw(CAREFUL_SCRIPT, 'this is not json');
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBe('ask');
+      expect(output.hookSpecificOutput?.permissionDecisionReason).toContain('parse');
+    });
+
+    test('allows a well-formed payload with no command field', () => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, { tool_input: { file_path: '/tmp/x' } });
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBeUndefined();
+    });
+
+    test('allows when command is present but not a string', () => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, { tool_input: { command: 42 } });
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBeUndefined();
+    });
+
+    test('preserves escaped quotes in the extracted command', () => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, carefulInput('echo "hello world"'));
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBeUndefined();
+    });
+  });
+
   // --- SQL destructive commands ---
-  // Note: SQL commands that contain embedded double quotes (e.g., psql -c "DROP TABLE")
+  // Embedded double quotes are now safe to use here. They previously truncated the
-  // get their command value truncated by the grep-based JSON extractor because \"
+  // extracted command (the grep-based extractor stopped at the first \"), which hid
-  // terminates the [^"]* match. We use commands WITHOUT embedded quotes so the grep
+  // the SQL keyword from the pattern matcher — so the older tests had to be written
-  // extraction works and the SQL keywords are visible to the pattern matcher.
+  // without quotes, in a shape no one actually types. The JSON-parser extraction
+  // fixed that, and the quoted forms below are the realistic ones.
 
   describe('SQL destructive commands', () => {
     test('psql DROP TABLE warns with DROP in message', () => {
@@ -211,6 +277,16 @@ describe('check-careful.sh', () => {
       expect(output.hookSpecificOutput?.permissionDecisionReason).toContain('DROP');
     });
 
+    test.each([
+      'psql -c "DROP TABLE users"',
+      'psql -c "TRUNCATE orders"',
+      'mysql -e "DROP DATABASE prod"',
+    ])('a quoted SQL statement is still inspected: %s', (command) => {
+      const { exitCode, output } = runHook(CAREFUL_SCRIPT, carefulInput(command));
+      expect(exitCode).toBe(0);
+      expect(output.hookSpecificOutput?.permissionDecision).toBe('ask');
+    });
+
     test('mysql drop database warns (case insensitive)', () => {
       const { exitCode, output } = runHook(CAREFUL_SCRIPT, carefulInput('mysql -e drop database mydb'));
       expect(exitCode).toBe(0);
@@ -325,10 +401,13 @@ describe('check-careful.sh', () => {
       expect(output.hookSpecificOutput?.permissionDecision).toBeUndefined();
     });
 
-    test('malformed JSON input allows gracefully (exit 0, output {})', () => {
+    test('malformed JSON input fails CLOSED (asks instead of allowing)', () => {
-      const { exitCode, raw } = runHookRaw(CAREFUL_SCRIPT, 'this is not json at all{{{{');
+      // Pre-#2426 this allowed (`{}`) — a hook that gates destructive commands
+      // must not allow-by-default on input it cannot read.
+      const { exitCode, output } = runHookRaw(CAREFUL_SCRIPT, 'this is not json at all{{{{');
       expect(exitCode).toBe(0);
-      expect(raw).toBe('{}');
+      expect(output.hookSpecificOutput?.permissionDecision).toBe('ask');
+      expect(output.hookSpecificOutput?.permissionDecisionReason).toContain('parse');
     });
 
     test('Python fallback: grep fails on multiline JSON, Python parses it', () => {