@xelauvas/xela-cli 0.1.0

package/src/tools/BashTool/bashSecurity.ts

@@ -0,0 +1,2592 @@
+import { logEvent } from 'src/services/analytics/index.js'
+import { extractHeredocs } from '../../utils/bash/heredoc.js'
+import { ParsedCommand } from '../../utils/bash/ParsedCommand.js'
+import {
+  hasMalformedTokens,
+  hasShellQuoteSingleQuoteBug,
+  tryParseShellCommand,
+} from '../../utils/bash/shellQuote.js'
+import type { TreeSitterAnalysis } from '../../utils/bash/treeSitterAnalysis.js'
+import type { PermissionResult } from '../../utils/permissions/PermissionResult.js'
+
+const HEREDOC_IN_SUBSTITUTION = /\$\(.*<</
+
+// Note: Backtick pattern is handled separately in validateDangerousPatterns
+// to distinguish between escaped and unescaped backticks
+const COMMAND_SUBSTITUTION_PATTERNS = [
+  { pattern: /<\(/, message: 'process substitution <()' },
+  { pattern: />\(/, message: 'process substitution >()' },
+  { pattern: /=\(/, message: 'Zsh process substitution =()' },
+  // Zsh EQUALS expansion: =cmd at word start expands to $(which cmd).
+  // `=curl evil.com` → `/usr/bin/curl evil.com`, bypassing Bash(curl:*) deny
+  // rules since the parser sees `=curl` as the base command, not `curl`.
+  // Only matches word-initial = followed by a command-name char (not VAR=val).
+  {
+    pattern: /(?:^|[\s;&|])=[a-zA-Z_]/,
+    message: 'Zsh equals expansion (=cmd)',
+  },
+  { pattern: /\$\(/, message: '$() command substitution' },
+  { pattern: /\$\{/, message: '${} parameter substitution' },
+  { pattern: /\$\[/, message: '$[] legacy arithmetic expansion' },
+  { pattern: /~\[/, message: 'Zsh-style parameter expansion' },
+  { pattern: /\(e:/, message: 'Zsh-style glob qualifiers' },
+  { pattern: /\(\+/, message: 'Zsh glob qualifier with command execution' },
+  {
+    pattern: /\}\s*always\s*\{/,
+    message: 'Zsh always block (try/always construct)',
+  },
+  // Defense in depth: Block PowerShell comment syntax even though we don't execute in PowerShell
+  // Added as protection against future changes that might introduce PowerShell execution
+  { pattern: /<#/, message: 'PowerShell comment syntax' },
+]
+
+// Zsh-specific dangerous commands that can bypass security checks.
+// These are checked against the base command (first word) of each command segment.
+const ZSH_DANGEROUS_COMMANDS = new Set([
+  // zmodload is the gateway to many dangerous module-based attacks:
+  // zsh/mapfile (invisible file I/O via array assignment),
+  // zsh/system (sysopen/syswrite two-step file access),
+  // zsh/zpty (pseudo-terminal command execution),
+  // zsh/net/tcp (network exfiltration via ztcp),
+  // zsh/files (builtin rm/mv/ln/chmod that bypass binary checks)
+  'zmodload',
+  // emulate with -c flag is an eval-equivalent that executes arbitrary code
+  'emulate',
+  // Zsh module builtins that enable dangerous operations.
+  // These require zmodload first, but we block them as defense-in-depth
+  // in case zmodload is somehow bypassed or the module is pre-loaded.
+  'sysopen', // Opens files with fine-grained control (zsh/system)
+  'sysread', // Reads from file descriptors (zsh/system)
+  'syswrite', // Writes to file descriptors (zsh/system)
+  'sysseek', // Seeks on file descriptors (zsh/system)
+  'zpty', // Executes commands on pseudo-terminals (zsh/zpty)
+  'ztcp', // Creates TCP connections for exfiltration (zsh/net/tcp)
+  'zsocket', // Creates Unix/TCP sockets (zsh/net/socket)
+  'mapfile', // Not actually a command, but the associative array is set via zmodload
+  'zf_rm', // Builtin rm from zsh/files
+  'zf_mv', // Builtin mv from zsh/files
+  'zf_ln', // Builtin ln from zsh/files
+  'zf_chmod', // Builtin chmod from zsh/files
+  'zf_chown', // Builtin chown from zsh/files
+  'zf_mkdir', // Builtin mkdir from zsh/files
+  'zf_rmdir', // Builtin rmdir from zsh/files
+  'zf_chgrp', // Builtin chgrp from zsh/files
+])
+
+// Numeric identifiers for bash security checks (to avoid logging strings)
+const BASH_SECURITY_CHECK_IDS = {
+  INCOMPLETE_COMMANDS: 1,
+  JQ_SYSTEM_FUNCTION: 2,
+  JQ_FILE_ARGUMENTS: 3,
+  OBFUSCATED_FLAGS: 4,
+  SHELL_METACHARACTERS: 5,
+  DANGEROUS_VARIABLES: 6,
+  NEWLINES: 7,
+  DANGEROUS_PATTERNS_COMMAND_SUBSTITUTION: 8,
+  DANGEROUS_PATTERNS_INPUT_REDIRECTION: 9,
+  DANGEROUS_PATTERNS_OUTPUT_REDIRECTION: 10,
+  IFS_INJECTION: 11,
+  GIT_COMMIT_SUBSTITUTION: 12,
+  PROC_ENVIRON_ACCESS: 13,
+  MALFORMED_TOKEN_INJECTION: 14,
+  BACKSLASH_ESCAPED_WHITESPACE: 15,
+  BRACE_EXPANSION: 16,
+  CONTROL_CHARACTERS: 17,
+  UNICODE_WHITESPACE: 18,
+  MID_WORD_HASH: 19,
+  ZSH_DANGEROUS_COMMANDS: 20,
+  BACKSLASH_ESCAPED_OPERATORS: 21,
+  COMMENT_QUOTE_DESYNC: 22,
+  QUOTED_NEWLINE: 23,
+} as const
+
+type ValidationContext = {
+  originalCommand: string
+  baseCommand: string
+  unquotedContent: string
+  fullyUnquotedContent: string
+  /** fullyUnquoted before stripSafeRedirections — used by validateBraceExpansion
+   * to avoid false negatives from redirection stripping creating backslash adjacencies */
+  fullyUnquotedPreStrip: string
+  /** Like fullyUnquotedPreStrip but preserves quote characters ('/"): e.g.,
+   * echo 'x'# → echo ''# (the quote chars remain, revealing adjacency to #) */
+  unquotedKeepQuoteChars: string
+  /** Tree-sitter analysis data, if available. Validators can use this for
+   * more accurate analysis when present, falling back to regex otherwise. */
+  treeSitter?: TreeSitterAnalysis | null
+}
+
+type QuoteExtraction = {
+  withDoubleQuotes: string
+  fullyUnquoted: string
+  /** Like fullyUnquoted but preserves quote characters ('/"): strips quoted
+   * content while keeping the delimiters. Used by validateMidWordHash to detect
+   * quote-adjacent # (e.g., 'x'# where quote stripping would hide adjacency). */
+  unquotedKeepQuoteChars: string
+}
+
+function extractQuotedContent(command: string, isJq = false): QuoteExtraction {
+  let withDoubleQuotes = ''
+  let fullyUnquoted = ''
+  let unquotedKeepQuoteChars = ''
+  let inSingleQuote = false
+  let inDoubleQuote = false
+  let escaped = false
+
+  for (let i = 0; i < command.length; i++) {
+    const char = command[i]
+
+    if (escaped) {
+      escaped = false
+      if (!inSingleQuote) withDoubleQuotes += char
+      if (!inSingleQuote && !inDoubleQuote) fullyUnquoted += char
+      if (!inSingleQuote && !inDoubleQuote) unquotedKeepQuoteChars += char
+      continue
+    }
+
+    if (char === '\\' && !inSingleQuote) {
+      escaped = true
+      if (!inSingleQuote) withDoubleQuotes += char
+      if (!inSingleQuote && !inDoubleQuote) fullyUnquoted += char
+      if (!inSingleQuote && !inDoubleQuote) unquotedKeepQuoteChars += char
+      continue
+    }
+
+    if (char === "'" && !inDoubleQuote) {
+      inSingleQuote = !inSingleQuote
+      unquotedKeepQuoteChars += char
+      continue
+    }
+
+    if (char === '"' && !inSingleQuote) {
+      inDoubleQuote = !inDoubleQuote
+      unquotedKeepQuoteChars += char
+      // For jq, include quotes in extraction to ensure content is properly analyzed
+      if (!isJq) continue
+    }
+
+    if (!inSingleQuote) withDoubleQuotes += char
+    if (!inSingleQuote && !inDoubleQuote) fullyUnquoted += char
+    if (!inSingleQuote && !inDoubleQuote) unquotedKeepQuoteChars += char
+  }
+
+  return { withDoubleQuotes, fullyUnquoted, unquotedKeepQuoteChars }
+}
+
+function stripSafeRedirections(content: string): string {
+  // SECURITY: All three patterns MUST have a trailing boundary (?=\s|$).
+  // Without it, `> /dev/nullo` matches `/dev/null` as a PREFIX, strips
+  // `> /dev/null` leaving `o`, so `echo hi > /dev/nullo` becomes `echo hi o`.
+  // validateRedirections then sees no `>` and passes. The file write to
+  // /dev/nullo is auto-allowed via the read-only path (checkReadOnlyConstraints).
+  // Main bashPermissions flow is protected (checkPathConstraints validates the
+  // original command), but speculation.ts uses checkReadOnlyConstraints alone.
+  return content
+    .replace(/\s+2\s*>&\s*1(?=\s|$)/g, '')
+    .replace(/[012]?\s*>\s*\/dev\/null(?=\s|$)/g, '')
+    .replace(/\s*<\s*\/dev\/null(?=\s|$)/g, '')
+}
+
+/**
+ * Checks if content contains an unescaped occurrence of a single character.
+ * Handles bash escape sequences correctly where a backslash escapes the following character.
+ *
+ * IMPORTANT: This function only handles single characters, not strings. If you need to extend
+ * this to handle multi-character strings, be EXTREMELY CAREFUL about shell ANSI-C quoting
+ * (e.g., $'\n', $'\x41', $'\u0041') which can encode arbitrary characters and strings in ways
+ * that are very difficult to parse correctly. Incorrect handling could introduce security
+ * vulnerabilities by allowing attackers to bypass security checks.
+ *
+ * @param content - The string to search (typically from extractQuotedContent)
+ * @param char - Single character to search for (e.g., '`')
+ * @returns true if unescaped occurrence found, false otherwise
+ *
+ * Examples:
+ *   hasUnescapedChar("test \`safe\`", '`') → false (escaped backticks)
+ *   hasUnescapedChar("test `dangerous`", '`') → true (unescaped backticks)
+ *   hasUnescapedChar("test\\`date`", '`') → true (escaped backslash + unescaped backtick)
+ */
+function hasUnescapedChar(content: string, char: string): boolean {
+  if (char.length !== 1) {
+    throw new Error('hasUnescapedChar only works with single characters')
+  }
+
+  let i = 0
+  while (i < content.length) {
+    // If we see a backslash, skip it and the next character (they form an escape sequence)
+    if (content[i] === '\\' && i + 1 < content.length) {
+      i += 2 // Skip backslash and escaped character
+      continue
+    }
+
+    // Check if current character matches
+    if (content[i] === char) {
+      return true // Found unescaped occurrence
+    }
+
+    i++
+  }
+
+  return false // No unescaped occurrences found
+}
+
+function validateEmpty(context: ValidationContext): PermissionResult {
+  if (!context.originalCommand.trim()) {
+    return {
+      behavior: 'allow',
+      updatedInput: { command: context.originalCommand },
+      decisionReason: { type: 'other', reason: 'Empty command is safe' },
+    }
+  }
+  return { behavior: 'passthrough', message: 'Command is not empty' }
+}
+
+function validateIncompleteCommands(
+  context: ValidationContext,
+): PermissionResult {
+  const { originalCommand } = context
+  const trimmed = originalCommand.trim()
+
+  if (/^\s*\t/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.INCOMPLETE_COMMANDS,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message: 'Command appears to be an incomplete fragment (starts with tab)',
+    }
+  }
+
+  if (trimmed.startsWith('-')) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.INCOMPLETE_COMMANDS,
+      subId: 2,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command appears to be an incomplete fragment (starts with flags)',
+    }
+  }
+
+  if (/^\s*(&&|\|\||;|>>?|<)/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.INCOMPLETE_COMMANDS,
+      subId: 3,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command appears to be a continuation line (starts with operator)',
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'Command appears complete' }
+}
+
+/**
+ * Checks if a command is a "safe" heredoc-in-substitution pattern that can
+ * bypass the generic $() validator.
+ *
+ * This is an EARLY-ALLOW path: returning `true` causes bashCommandIsSafe to
+ * return `passthrough`, bypassing ALL subsequent validators. Given this
+ * authority, the check must be PROVABLY safe, not "probably safe".
+ *
+ * The only pattern we allow is:
+ *   [prefix] $(cat <<'DELIM'\n
+ *   [body lines]\n
+ *   DELIM\n
+ *   ) [suffix]
+ *
+ * Where:
+ * - The delimiter must be single-quoted ('DELIM') or escaped (\DELIM) so the
+ *   body is literal text with no expansion
+ * - The closing delimiter must be on a line BY ITSELF (or with only trailing
+ *   whitespace + `)` for the $(cat <<'EOF'\n...\nEOF)` inline form)
+ * - The closing delimiter must be the FIRST such line — matching bash's
+ *   behavior exactly (no skipping past early delimiters to find EOF))
+ * - There must be non-whitespace text BEFORE the $( (i.e., the substitution
+ *   is used in argument position, not as a command name). Otherwise the
+ *   heredoc body becomes an arbitrary command name with [suffix] as args.
+ * - The remaining text (with the heredoc stripped) must pass all validators
+ *
+ * This implementation uses LINE-BASED matching, not regex [\s\S]*?, to
+ * precisely replicate bash's heredoc-closing behavior.
+ */
+function isSafeHeredoc(command: string): boolean {
+  if (!HEREDOC_IN_SUBSTITUTION.test(command)) return false
+
+  // SECURITY: Use [ \t] (not \s) between << and the delimiter. \s matches
+  // newlines, but bash requires the delimiter word on the same line as <<.
+  // Matching across newlines could accept malformed syntax that bash rejects.
+  // Handle quote variations: 'EOF', ''EOF'' (splitCommand may mangle quotes).
+  const heredocPattern =
+    /\$\(cat[ \t]*<<(-?)[ \t]*(?:'+([A-Za-z_]\w*)'+|\\([A-Za-z_]\w*))/g
+  let match
+  type HeredocMatch = {
+    start: number
+    operatorEnd: number
+    delimiter: string
+    isDash: boolean
+  }
+  const safeHeredocs: HeredocMatch[] = []
+
+  while ((match = heredocPattern.exec(command)) !== null) {
+    const delimiter = match[2] || match[3]
+    if (delimiter) {
+      safeHeredocs.push({
+        start: match.index,
+        operatorEnd: match.index + match[0].length,
+        delimiter,
+        isDash: match[1] === '-',
+      })
+    }
+  }
+
+  // If no safe heredoc patterns found, it's not safe
+  if (safeHeredocs.length === 0) return false
+
+  // SECURITY: For each heredoc, find the closing delimiter using LINE-BASED
+  // matching that exactly replicates bash's behavior. Bash closes a heredoc
+  // at the FIRST line that exactly matches the delimiter. Any subsequent
+  // occurrence of the delimiter is just content (or a new command). Regex
+  // [\s\S]*? can skip past the first delimiter to find a later `DELIM)`
+  // pattern, hiding injected commands between the two delimiters.
+  type VerifiedHeredoc = { start: number; end: number }
+  const verified: VerifiedHeredoc[] = []
+
+  for (const { start, operatorEnd, delimiter, isDash } of safeHeredocs) {
+    // The opening line must end immediately after the delimiter (only
+    // horizontal whitespace allowed before the newline). If there's other
+    // content (like `; rm -rf /`), this is not a simple safe heredoc.
+    const afterOperator = command.slice(operatorEnd)
+    const openLineEnd = afterOperator.indexOf('\n')
+    if (openLineEnd === -1) return false // No content at all
+    const openLineTail = afterOperator.slice(0, openLineEnd)
+    if (!/^[ \t]*$/.test(openLineTail)) return false // Extra content on open line
+
+    // Body starts after the newline
+    const bodyStart = operatorEnd + openLineEnd + 1
+    const body = command.slice(bodyStart)
+    const bodyLines = body.split('\n')
+
+    // Find the FIRST line that closes the heredoc. There are two valid forms:
+    //   1. `DELIM` alone on a line (bash-standard), followed by `)` on the
+    //      next line (with only whitespace before it)
+    //   2. `DELIM)` on a line (the inline $(cat <<'EOF'\n...\nEOF) form,
+    //      where bash's PST_EOFTOKEN closes both heredoc and substitution)
+    // For <<-, leading tabs are stripped before matching.
+    let closingLineIdx = -1
+    let closeParenLineIdx = -1 // Line index where `)` appears
+    let closeParenColIdx = -1 // Column index of `)` on that line
+
+    for (let i = 0; i < bodyLines.length; i++) {
+      const rawLine = bodyLines[i]!
+      const line = isDash ? rawLine.replace(/^\t*/, '') : rawLine
+
+      // Form 1: delimiter alone on a line
+      if (line === delimiter) {
+        closingLineIdx = i
+        // The `)` must be on the NEXT line with only whitespace before it
+        const nextLine = bodyLines[i + 1]
+        if (nextLine === undefined) return false // No closing `)`
+        const parenMatch = nextLine.match(/^([ \t]*)\)/)
+        if (!parenMatch) return false // `)` not at start of next line
+        closeParenLineIdx = i + 1
+        closeParenColIdx = parenMatch[1]!.length // Position of `)`
+        break
+      }
+
+      // Form 2: delimiter immediately followed by `)` (PST_EOFTOKEN form)
+      // Only whitespace allowed between delimiter and `)`.
+      if (line.startsWith(delimiter)) {
+        const afterDelim = line.slice(delimiter.length)
+        const parenMatch = afterDelim.match(/^([ \t]*)\)/)
+        if (parenMatch) {
+          closingLineIdx = i
+          closeParenLineIdx = i
+          // Column is in rawLine (pre-tab-strip), so recompute
+          const tabPrefix = isDash ? (rawLine.match(/^\t*/)?.[0] ?? '') : ''
+          closeParenColIdx =
+            tabPrefix.length + delimiter.length + parenMatch[1]!.length
+          break
+        }
+        // Line starts with delimiter but has other trailing content —
+        // this is NOT the closing line (bash requires exact match or EOF`)`).
+        // But it's also a red flag: if this were inside $(), bash might
+        // close early via PST_EOFTOKEN with other shell metacharacters.
+        // We already handle that case in extractHeredocs — here we just
+        // reject it as not matching our safe pattern.
+        if (/^[)}`|&;(<>]/.test(afterDelim)) {
+          return false // Ambiguous early-closure pattern
+        }
+      }
+    }
+
+    if (closingLineIdx === -1) return false // No closing delimiter found
+
+    // Compute the absolute end position (one past the `)` character)
+    let endPos = bodyStart
+    for (let i = 0; i < closeParenLineIdx; i++) {
+      endPos += bodyLines[i]!.length + 1 // +1 for newline
+    }
+    endPos += closeParenColIdx + 1 // +1 to include the `)` itself
+
+    verified.push({ start, end: endPos })
+  }
+
+  // SECURITY: Reject nested matches. The regex finds $(cat <<'X' patterns
+  // in RAW TEXT without understanding quoted-heredoc semantics. When the
+  // outer heredoc has a quoted delimiter (<<'A'), its body is LITERAL text
+  // in bash — any inner $(cat <<'B' is just characters, not a real heredoc.
+  // But our regex matches both, producing NESTED ranges. Stripping nested
+  // ranges corrupts indices: after stripping the inner range, the outer
+  // range's `end` is stale (points past the shrunken string), causing
+  // `remaining.slice(end)` to return '' and silently drop any suffix
+  // (e.g., `; rm -rf /`). Since all our matched heredocs have quoted/escaped
+  // delimiters, a nested match inside the body is ALWAYS literal text —
+  // no legitimate user writes this pattern. Bail to safe fallback.
+  for (const outer of verified) {
+    for (const inner of verified) {
+      if (inner === outer) continue
+      if (inner.start > outer.start && inner.start < outer.end) {
+        return false
+      }
+    }
+  }
+
+  // Strip all verified heredocs from the command, building `remaining`.
+  // Process in reverse order so earlier indices stay valid.
+  const sortedVerified = [...verified].sort((a, b) => b.start - a.start)
+  let remaining = command
+  for (const { start, end } of sortedVerified) {
+    remaining = remaining.slice(0, start) + remaining.slice(end)
+  }
+
+  // SECURITY: The remaining text must NOT start with only whitespace before
+  // the (now-stripped) heredoc position IF there's non-whitespace after it.
+  // If the $() is in COMMAND-NAME position (no prefix), its output becomes
+  // the command to execute, with any suffix text as arguments:
+  //   $(cat <<'EOF'\nchmod\nEOF\n) 777 /etc/shadow
+  //   → runs `chmod 777 /etc/shadow`
+  // We only allow the substitution in ARGUMENT position: there must be a
+  // command word before the $(.
+  // After stripping, `remaining` should look like `cmd args... [more args]`.
+  // If remaining starts with only whitespace (or is empty), the $() WAS the
+  // command — that's only safe if there are no trailing arguments.
+  const trimmedRemaining = remaining.trim()
+  if (trimmedRemaining.length > 0) {
+    // There's a prefix command — good. But verify the original command
+    // also had a non-whitespace prefix before the FIRST $( (the heredoc
+    // could be one of several; we need the first one's prefix).
+    const firstHeredocStart = Math.min(...verified.map(v => v.start))
+    const prefix = command.slice(0, firstHeredocStart)
+    if (prefix.trim().length === 0) {
+      // $() is in command-name position but there's trailing text — UNSAFE.
+      // The heredoc body becomes the command name, trailing text becomes args.
+      return false
+    }
+  }
+
+  // Check that remaining text contains only safe characters.
+  // After stripping safe heredocs, the remaining text should only be command
+  // names, arguments, quotes, and whitespace. Reject ANY shell metacharacter
+  // to prevent operators (|, &, &&, ||, ;) or expansions ($, `, {, <, >) from
+  // being used to chain dangerous commands after a safe heredoc.
+  // SECURITY: Use explicit ASCII space/tab only — \s matches unicode whitespace
+  // like \u00A0 which can be used to hide content. Newlines are also blocked
+  // (they would indicate multi-line commands outside the heredoc body).
+  if (!/^[a-zA-Z0-9 \t"'.\-/_@=,:+~]*$/.test(remaining)) return false
+
+  // SECURITY: The remaining text (command with heredocs stripped) must also
+  // pass all security validators. Without this, appending a safe heredoc to a
+  // dangerous command (e.g., `zmodload zsh/system $(cat <<'EOF'\nx\nEOF\n)`)
+  // causes this early-allow path to return passthrough, bypassing
+  // validateZshDangerousCommands, validateProcEnvironAccess, and any other
+  // main validator that checks allowlist-safe character patterns.
+  // No recursion risk: `remaining` has no `$(... <<` pattern, so the recursive
+  // call's validateSafeCommandSubstitution returns passthrough immediately.
+  if (bashCommandIsSafe_DEPRECATED(remaining).behavior !== 'passthrough')
+    return false
+
+  return true
+}
+
+/**
+ * Detects well-formed $(cat <<'DELIM'...DELIM) heredoc substitution patterns.
+ * Returns the command with matched heredocs stripped, or null if none found.
+ * Used by the pre-split gate to strip safe heredocs and re-check the remainder.
+ */
+export function stripSafeHeredocSubstitutions(command: string): string | null {
+  if (!HEREDOC_IN_SUBSTITUTION.test(command)) return null
+
+  const heredocPattern =
+    /\$\(cat[ \t]*<<(-?)[ \t]*(?:'+([A-Za-z_]\w*)'+|\\([A-Za-z_]\w*))/g
+  let result = command
+  let found = false
+  let match
+  const ranges: Array<{ start: number; end: number }> = []
+  while ((match = heredocPattern.exec(command)) !== null) {
+    if (match.index > 0 && command[match.index - 1] === '\\') continue
+    const delimiter = match[2] || match[3]
+    if (!delimiter) continue
+    const isDash = match[1] === '-'
+    const operatorEnd = match.index + match[0].length
+
+    const afterOperator = command.slice(operatorEnd)
+    const openLineEnd = afterOperator.indexOf('\n')
+    if (openLineEnd === -1) continue
+    if (!/^[ \t]*$/.test(afterOperator.slice(0, openLineEnd))) continue
+
+    const bodyStart = operatorEnd + openLineEnd + 1
+    const bodyLines = command.slice(bodyStart).split('\n')
+    for (let i = 0; i < bodyLines.length; i++) {
+      const rawLine = bodyLines[i]!
+      const line = isDash ? rawLine.replace(/^\t*/, '') : rawLine
+      if (line.startsWith(delimiter)) {
+        const after = line.slice(delimiter.length)
+        let closePos = -1
+        if (/^[ \t]*\)/.test(after)) {
+          const lineStart =
+            bodyStart +
+            bodyLines.slice(0, i).join('\n').length +
+            (i > 0 ? 1 : 0)
+          closePos = command.indexOf(')', lineStart)
+        } else if (after === '') {
+          const nextLine = bodyLines[i + 1]
+          if (nextLine !== undefined && /^[ \t]*\)/.test(nextLine)) {
+            const nextLineStart =
+              bodyStart + bodyLines.slice(0, i + 1).join('\n').length + 1
+            closePos = command.indexOf(')', nextLineStart)
+          }
+        }
+        if (closePos !== -1) {
+          ranges.push({ start: match.index, end: closePos + 1 })
+          found = true
+        }
+        break
+      }
+    }
+  }
+  if (!found) return null
+  for (let i = ranges.length - 1; i >= 0; i--) {
+    const r = ranges[i]!
+    result = result.slice(0, r.start) + result.slice(r.end)
+  }
+  return result
+}
+
+/** Detection-only check: does the command contain a safe heredoc substitution? */
+export function hasSafeHeredocSubstitution(command: string): boolean {
+  return stripSafeHeredocSubstitutions(command) !== null
+}
+
+function validateSafeCommandSubstitution(
+  context: ValidationContext,
+): PermissionResult {
+  const { originalCommand } = context
+
+  if (!HEREDOC_IN_SUBSTITUTION.test(originalCommand)) {
+    return { behavior: 'passthrough', message: 'No heredoc in substitution' }
+  }
+
+  if (isSafeHeredoc(originalCommand)) {
+    return {
+      behavior: 'allow',
+      updatedInput: { command: originalCommand },
+      decisionReason: {
+        type: 'other',
+        reason:
+          'Safe command substitution: cat with quoted/escaped heredoc delimiter',
+      },
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'Command substitution needs validation',
+  }
+}
+
+function validateGitCommit(context: ValidationContext): PermissionResult {
+  const { originalCommand, baseCommand } = context
+
+  if (baseCommand !== 'git' || !/^git\s+commit\s+/.test(originalCommand)) {
+    return { behavior: 'passthrough', message: 'Not a git commit' }
+  }
+
+  // SECURITY: Backslashes can cause our regex to mis-identify quote boundaries
+  // (e.g., `git commit -m "test\"msg" && evil`). Legitimate commit messages
+  // virtually never contain backslashes, so bail to the full validator chain.
+  if (originalCommand.includes('\\')) {
+    return {
+      behavior: 'passthrough',
+      message: 'Git commit contains backslash, needs full validation',
+    }
+  }
+
+  // SECURITY: The `.*?` before `-m` must NOT match shell operators. Previously
+  // `.*?` matched anything except `\n`, including `;`, `&`, `|`, `` ` ``, `$(`.
+  // For `git commit ; curl evil.com -m 'x'`, `.*?` swallowed `; curl evil.com `
+  // leaving remainder=`` (falsy → remainder check skipped) → returned `allow`
+  // for a compound command. Early-allow skips ALL main validators (line ~1908),
+  // nullifying validateQuotedNewline, validateBackslashEscapedOperators, etc.
+  // While splitCommand currently catches this downstream, early-allow is a
+  // POSITIVE ASSERTION that the FULL command is safe — which it is NOT.
+  //
+  // Also: `\s+` between `git` and `commit` must NOT match `\n`/`\r` (command
+  // separators in bash). Use `[ \t]+` for horizontal-only whitespace.
+  //
+  // The `[^;&|`$<>()\n\r]*?` class excludes shell metacharacters. We also
+  // exclude `<` and `>` here (redirects) — they're allowed in the REMAINDER
+  // for `--author="Name <email>"` but must not appear BEFORE `-m`.
+  const messageMatch = originalCommand.match(
+    /^git[ \t]+commit[ \t]+[^;&|`$<>()\n\r]*?-m[ \t]+(["'])([\s\S]*?)\1(.*)$/,
+  )
+
+  if (messageMatch) {
+    const [, quote, messageContent, remainder] = messageMatch
+
+    if (quote === '"' && messageContent && /\$\(|`|\$\{/.test(messageContent)) {
+      logEvent('tengu_bash_security_check_triggered', {
+        checkId: BASH_SECURITY_CHECK_IDS.GIT_COMMIT_SUBSTITUTION,
+        subId: 1,
+      })
+      return {
+        behavior: 'ask',
+        message: 'Git commit message contains command substitution patterns',
+      }
+    }
+
+    // SECURITY: Check remainder for shell operators that could chain commands
+    // or redirect output. The `.*` before `-m` in the regex can swallow flags
+    // like `--amend`, leaving `&& evil` or `> ~/.bashrc` in the remainder.
+    // Previously we only checked for $() / `` / ${} here, missing operators
+    // like ; | & && || < >.
+    //
+    // `<` and `>` can legitimately appear INSIDE quotes in --author values
+    // like `--author="Name <email>"`. An UNQUOTED `>` is a shell redirect
+    // operator. Because validateGitCommit is an EARLY validator, returning
+    // `allow` here short-circuits bashCommandIsSafe and SKIPS
+    // validateRedirections. So we must bail to passthrough on unquoted `<>`
+    // to let the main validators handle it.
+    //
+    // Attack: `git commit --allow-empty -m 'payload' > ~/.bashrc`
+    //   validateGitCommit returns allow → bashCommandIsSafe short-circuits →
+    //   validateRedirections NEVER runs → ~/.bashrc overwritten with git
+    //   stdout containing `payload` → RCE on next shell login.
+    if (remainder && /[;|&()`]|\$\(|\$\{/.test(remainder)) {
+      return {
+        behavior: 'passthrough',
+        message: 'Git commit remainder contains shell metacharacters',
+      }
+    }
+    if (remainder) {
+      // Strip quoted content, then check for `<` or `>`. Quoted `<>` (email
+      // brackets in --author) are safe; unquoted `<>` are shell redirects.
+      // NOTE: This simple quote tracker has NO backslash handling. `\'`/`\"`
+      // outside quotes would desync it (bash: \' = literal ', tracker: toggles
+      // SQ). BUT line 584 already bailed on ANY backslash in originalCommand,
+      // so we never reach here with backslashes. For backslash-free input,
+      // simple quote toggling is correct (no way to escape quotes without \\).
+      let unquoted = ''
+      let inSQ = false
+      let inDQ = false
+      for (let i = 0; i < remainder.length; i++) {
+        const c = remainder[i]
+        if (c === "'" && !inDQ) {
+          inSQ = !inSQ
+          continue
+        }
+        if (c === '"' && !inSQ) {
+          inDQ = !inDQ
+          continue
+        }
+        if (!inSQ && !inDQ) unquoted += c
+      }
+      if (/[<>]/.test(unquoted)) {
+        return {
+          behavior: 'passthrough',
+          message: 'Git commit remainder contains unquoted redirect operator',
+        }
+      }
+    }
+
+    // Security hardening: block messages starting with dash
+    // This catches potential obfuscation patterns like git commit -m "---"
+    if (messageContent && messageContent.startsWith('-')) {
+      logEvent('tengu_bash_security_check_triggered', {
+        checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+        subId: 5,
+      })
+      return {
+        behavior: 'ask',
+        message: 'Command contains quoted characters in flag names',
+      }
+    }
+
+    return {
+      behavior: 'allow',
+      updatedInput: { command: originalCommand },
+      decisionReason: {
+        type: 'other',
+        reason: 'Git commit with simple quoted message is allowed',
+      },
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'Git commit needs validation' }
+}
+
+function validateJqCommand(context: ValidationContext): PermissionResult {
+  const { originalCommand, baseCommand } = context
+
+  if (baseCommand !== 'jq') {
+    return { behavior: 'passthrough', message: 'Not jq' }
+  }
+
+  if (/\bsystem\s*\(/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.JQ_SYSTEM_FUNCTION,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'jq command contains system() function which executes arbitrary commands',
+    }
+  }
+
+  // File arguments are now allowed - they will be validated by path validation in readOnlyValidation.ts
+  // Only block dangerous flags that could read files into jq variables
+  const afterJq = originalCommand.substring(3).trim()
+  if (
+    /(?:^|\s)(?:-f\b|--from-file|--rawfile|--slurpfile|-L\b|--library-path)/.test(
+      afterJq,
+    )
+  ) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.JQ_FILE_ARGUMENTS,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'jq command contains dangerous flags that could execute code or read arbitrary files',
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'jq command is safe' }
+}
+
+function validateShellMetacharacters(
+  context: ValidationContext,
+): PermissionResult {
+  const { unquotedContent } = context
+  const message =
+    'Command contains shell metacharacters (;, |, or &) in arguments'
+
+  if (/(?:^|\s)["'][^"']*[;&][^"']*["'](?:\s|$)/.test(unquotedContent)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.SHELL_METACHARACTERS,
+      subId: 1,
+    })
+    return { behavior: 'ask', message }
+  }
+
+  const globPatterns = [
+    /-name\s+["'][^"']*[;|&][^"']*["']/,
+    /-path\s+["'][^"']*[;|&][^"']*["']/,
+    /-iname\s+["'][^"']*[;|&][^"']*["']/,
+  ]
+
+  if (globPatterns.some(p => p.test(unquotedContent))) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.SHELL_METACHARACTERS,
+      subId: 2,
+    })
+    return { behavior: 'ask', message }
+  }
+
+  if (/-regex\s+["'][^"']*[;&][^"']*["']/.test(unquotedContent)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.SHELL_METACHARACTERS,
+      subId: 3,
+    })
+    return { behavior: 'ask', message }
+  }
+
+  return { behavior: 'passthrough', message: 'No metacharacters' }
+}
+
+function validateDangerousVariables(
+  context: ValidationContext,
+): PermissionResult {
+  const { fullyUnquotedContent } = context
+
+  if (
+    /[<>|]\s*\$[A-Za-z_]/.test(fullyUnquotedContent) ||
+    /\$[A-Za-z_][A-Za-z0-9_]*\s*[|<>]/.test(fullyUnquotedContent)
+  ) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.DANGEROUS_VARIABLES,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains variables in dangerous contexts (redirections or pipes)',
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No dangerous variables' }
+}
+
+function validateDangerousPatterns(
+  context: ValidationContext,
+): PermissionResult {
+  const { unquotedContent } = context
+
+  // Special handling for backticks - check for UNESCAPED backticks only
+  // Escaped backticks (e.g., \`) are safe and commonly used in SQL commands
+  if (hasUnescapedChar(unquotedContent, '`')) {
+    return {
+      behavior: 'ask',
+      message: 'Command contains backticks (`) for command substitution',
+    }
+  }
+
+  // Other command substitution checks (include double-quoted content)
+  for (const { pattern, message } of COMMAND_SUBSTITUTION_PATTERNS) {
+    if (pattern.test(unquotedContent)) {
+      logEvent('tengu_bash_security_check_triggered', {
+        checkId:
+          BASH_SECURITY_CHECK_IDS.DANGEROUS_PATTERNS_COMMAND_SUBSTITUTION,
+        subId: 1,
+      })
+      return { behavior: 'ask', message: `Command contains ${message}` }
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No dangerous patterns' }
+}
+
+function validateRedirections(context: ValidationContext): PermissionResult {
+  const { fullyUnquotedContent } = context
+
+  if (/</.test(fullyUnquotedContent)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.DANGEROUS_PATTERNS_INPUT_REDIRECTION,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains input redirection (<) which could read sensitive files',
+    }
+  }
+
+  if (/>/.test(fullyUnquotedContent)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.DANGEROUS_PATTERNS_OUTPUT_REDIRECTION,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains output redirection (>) which could write to arbitrary files',
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No redirections' }
+}
+
+function validateNewlines(context: ValidationContext): PermissionResult {
+  // Use fullyUnquotedPreStrip (before stripSafeRedirections) to prevent bypasses
+  // where stripping `>/dev/null` creates a phantom backslash-newline continuation.
+  // E.g., `cmd \>/dev/null\nwhoami` → after stripping becomes `cmd \\nwhoami`
+  // which looks like a safe continuation but actually hides a second command.
+  const { fullyUnquotedPreStrip } = context
+
+  // Check for newlines in unquoted content
+  if (!/[\n\r]/.test(fullyUnquotedPreStrip)) {
+    return { behavior: 'passthrough', message: 'No newlines' }
+  }
+
+  // Flag any newline/CR followed by non-whitespace, EXCEPT backslash-newline
+  // continuations at word boundaries. In bash, `\<newline>` is a line
+  // continuation (both chars removed), which is safe when the backslash
+  // follows whitespace (e.g., `cmd \<newline>--flag`). Mid-word continuations
+  // like `tr\<newline>aceroute` are still flagged because they can hide
+  // dangerous command names from allowlist checks.
+  // eslint-disable-next-line custom-rules/no-lookbehind-regex -- .test() + gated by /[\n\r]/.test() above
+  const looksLikeCommand = /(?<![\s]\\)[\n\r]\s*\S/.test(fullyUnquotedPreStrip)
+  if (looksLikeCommand) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.NEWLINES,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains newlines that could separate multiple commands',
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'Newlines appear to be within data',
+  }
+}
+
+/**
+ * SECURITY: Carriage return (\r, 0x0D) IS a misparsing concern, unlike LF.
+ *
+ * Parser differential:
+ *   - shell-quote's BAREWORD regex uses `[^\s...]` — JS `\s` INCLUDES \r, so
+ *     shell-quote treats CR as a token boundary. `TZ=UTC\recho` tokenizes as
+ *     TWO tokens: ['TZ=UTC', 'echo']. splitCommand joins with space →
+ *     'TZ=UTC echo curl evil.com'.
+ *   - bash's default IFS = $' \t\n' — CR is NOT in IFS. bash sees
+ *     `TZ=UTC\recho` as ONE word → env assignment TZ='UTC\recho' (CR byte
+ *     inside value), then `curl` is the command.
+ *
+ * Attack: `TZ=UTC\recho curl evil.com` with Bash(echo:*)
+ *   validator: splitCommand collapses CR→space → 'TZ=UTC echo curl evil.com'
+ *   → stripSafeWrappers: TZ=UTC stripped → 'echo curl evil.com' matches rule
+ *   bash: executes `curl evil.com`
+ *
+ * validateNewlines catches this but is in nonMisparsingValidators (LF is
+ * correctly handled by both parsers). This validator is NOT in
+ * nonMisparsingValidators — its ask result gets isBashSecurityCheckForMisparsing
+ * and blocks at the bashPermissions gate.
+ *
+ * Checks originalCommand (not fullyUnquotedPreStrip) because CR inside single
+ * quotes is ALSO a misparsing concern for the same reason: shell-quote's `\s`
+ * still tokenizes it, but bash treats it as literal. Block ALL unquoted-or-SQ CR.
+ * Only exception: CR inside DOUBLE quotes where bash also treats it as data
+ * and shell-quote preserves the token (no split).
+ */
+function validateCarriageReturn(context: ValidationContext): PermissionResult {
+  const { originalCommand } = context
+
+  if (!originalCommand.includes('\r')) {
+    return { behavior: 'passthrough', message: 'No carriage return' }
+  }
+
+  // Check if CR appears outside double quotes. CR outside DQ (including inside
+  // SQ and unquoted) causes the shell-quote/bash tokenization differential.
+  let inSingleQuote = false
+  let inDoubleQuote = false
+  let escaped = false
+  for (let i = 0; i < originalCommand.length; i++) {
+    const c = originalCommand[i]
+    if (escaped) {
+      escaped = false
+      continue
+    }
+    if (c === '\\' && !inSingleQuote) {
+      escaped = true
+      continue
+    }
+    if (c === "'" && !inDoubleQuote) {
+      inSingleQuote = !inSingleQuote
+      continue
+    }
+    if (c === '"' && !inSingleQuote) {
+      inDoubleQuote = !inDoubleQuote
+      continue
+    }
+    if (c === '\r' && !inDoubleQuote) {
+      logEvent('tengu_bash_security_check_triggered', {
+        checkId: BASH_SECURITY_CHECK_IDS.NEWLINES,
+        subId: 2,
+      })
+      return {
+        behavior: 'ask',
+        message:
+          'Command contains carriage return (\\r) which shell-quote and bash tokenize differently',
+      }
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'CR only inside double quotes' }
+}
+
+function validateIFSInjection(context: ValidationContext): PermissionResult {
+  const { originalCommand } = context
+
+  // Detect any usage of IFS variable which could be used to bypass regex validation
+  // Check for $IFS and ${...IFS...} patterns (including parameter expansions like ${IFS:0:1}, ${#IFS}, etc.)
+  // Using ${[^}]*IFS to catch all parameter expansion variations with IFS
+  if (/\$IFS|\$\{[^}]*IFS/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.IFS_INJECTION,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains IFS variable usage which could bypass security validation',
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No IFS injection detected' }
+}
+
+// Additional hardening against reading environment variables via /proc filesystem.
+// Path validation typically blocks /proc access, but this provides defense-in-depth.
+// Environment files in /proc can expose sensitive data like API keys and secrets.
+function validateProcEnvironAccess(
+  context: ValidationContext,
+): PermissionResult {
+  const { originalCommand } = context
+
+  // Check for /proc paths that could expose environment variables
+  // This catches patterns like:
+  // - /proc/self/environ
+  // - /proc/1/environ
+  // - /proc/*/environ (with any PID)
+  if (/\/proc\/.*\/environ/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.PROC_ENVIRON_ACCESS,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command accesses /proc/*/environ which could expose sensitive environment variables',
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'No /proc/environ access detected',
+  }
+}
+
+/**
+ * Detects commands with malformed tokens (unbalanced delimiters) combined with
+ * command separators. This catches potential injection patterns where ambiguous
+ * shell syntax could be exploited.
+ *
+ * Security: This check catches the eval bypass discovered in HackerOne review.
+ * When shell-quote parses ambiguous patterns like `echo {"hi":"hi;evil"}`,
+ * it may produce unbalanced tokens (e.g., `{hi:"hi`). Combined with command
+ * separators, this can lead to unintended command execution via eval re-parsing.
+ *
+ * By forcing user approval for these patterns, we ensure the user sees exactly
+ * what will be executed before approving.
+ */
+function validateMalformedTokenInjection(
+  context: ValidationContext,
+): PermissionResult {
+  const { originalCommand } = context
+
+  const parseResult = tryParseShellCommand(originalCommand)
+  if (!parseResult.success) {
+    // Parse failed - this is handled elsewhere (bashToolHasPermission checks this)
+    return {
+      behavior: 'passthrough',
+      message: 'Parse failed, handled elsewhere',
+    }
+  }
+
+  const parsed = parseResult.tokens
+
+  // Check for command separators (;, &&, ||)
+  const hasCommandSeparator = parsed.some(
+    entry =>
+      typeof entry === 'object' &&
+      entry !== null &&
+      'op' in entry &&
+      (entry.op === ';' || entry.op === '&&' || entry.op === '||'),
+  )
+
+  if (!hasCommandSeparator) {
+    return { behavior: 'passthrough', message: 'No command separators' }
+  }
+
+  // Check for malformed tokens (unbalanced delimiters)
+  if (hasMalformedTokens(originalCommand, parsed)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.MALFORMED_TOKEN_INJECTION,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains ambiguous syntax with command separators that could be misinterpreted',
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'No malformed token injection detected',
+  }
+}
+
+function validateObfuscatedFlags(context: ValidationContext): PermissionResult {
+  // Block shell quoting bypass patterns used to circumvent negative lookaheads we use in our regexes to block known dangerous flags
+
+  const { originalCommand, baseCommand } = context
+
+  // Echo is safe for obfuscated flags, BUT only for simple echo commands.
+  // For compound commands (with |, &, ;), we need to check the whole command
+  // because the dangerous ANSI-C quoting might be after the operator.
+  const hasShellOperators = /[|&;]/.test(originalCommand)
+  if (baseCommand === 'echo' && !hasShellOperators) {
+    return {
+      behavior: 'passthrough',
+      message: 'echo command is safe and has no dangerous flags',
+    }
+  }
+
+  // COMPREHENSIVE OBFUSCATION DETECTION
+  // These checks catch various ways to hide flags using shell quoting
+
+  // 1. Block ANSI-C quoting ($'...') - can encode any character via escape sequences
+  // Simple pattern that matches $'...' anywhere. This correctly handles:
+  // - grep '$' file => no match ($ is regex anchor inside quotes, no $'...' structure)
+  // - 'test'$'-exec' => match (quote concatenation with ANSI-C)
+  // - Zero-width space and other invisible chars => match
+  // The pattern requires $' followed by content (can be empty) followed by closing '
+  if (/\$'[^']*'/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 5,
+    })
+    return {
+      behavior: 'ask',
+      message: 'Command contains ANSI-C quoting which can hide characters',
+    }
+  }
+
+  // 2. Block locale quoting ($"...")  - can also use escape sequences
+  // Same simple pattern as ANSI-C quoting above
+  if (/\$"[^"]*"/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 6,
+    })
+    return {
+      behavior: 'ask',
+      message: 'Command contains locale quoting which can hide characters',
+    }
+  }
+
+  // 3. Block empty ANSI-C or locale quotes followed by dash
+  // $''-exec or $""-exec
+  if (/\$['"]{2}\s*-/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 9,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains empty special quotes before dash (potential bypass)',
+    }
+  }
+
+  // 4. Block ANY sequence of empty quotes followed by dash
+  // This catches: ''-  ""-  ''""-  ""''-  ''""''-  etc.
+  // The pattern looks for one or more empty quote pairs followed by optional whitespace and dash
+  if (/(?:^|\s)(?:''|"")+\s*-/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 7,
+    })
+    return {
+      behavior: 'ask',
+      message: 'Command contains empty quotes before dash (potential bypass)',
+    }
+  }
+
+  // 4b. SECURITY: Block homogeneous empty quote pair(s) immediately adjacent
+  // to a quoted dash. Patterns like `"""-f"` (empty `""` + quoted `"-f"`)
+  // concatenate in bash to `-f` but slip past all the above checks:
+  //   - Regex (4) above: `(?:''|"")+\s*-` matches `""` pair, then expects
+  //     optional space and dash — but finds a third `"` instead. No match.
+  //   - Quote-content scanner (below): Sees the first `""` pair with empty
+  //     content (doesn't start with dash). The third `"` opens a new quoted
+  //     region handled by the main quote-state tracker.
+  //   - Quote-state tracker: `""` toggles inDoubleQuote on/off; third `"`
+  //     opens it again. The `-` inside `"-f"` is INSIDE quotes → skipped.
+  //   - Flag scanner: Looks for `\s` before `-`. The `-` is preceded by `"`.
+  //   - fullyUnquotedContent: Both `""` and `"-f"` get stripped.
+  //
+  // In bash, `"""-f"` = empty string + string "-f" = `-f`. This bypass works
+  // for ANY dangerous-flag check (jq -f, find -exec, fc -e) with a matching
+  // prefix permission (Bash(jq:*), Bash(find:*)).
+  //
+  // The regex `(?:""|'')+['"]-` matches:
+  //   - One or more HOMOGENEOUS empty pairs (`""` or `''`) — the concatenation
+  //     point where bash joins the empty string to the flag.
+  //   - Immediately followed by ANY quote char — opens the flag-quoted region.
+  //   - Immediately followed by `-` — the obfuscated flag.
+  //
+  // POSITION-AGNOSTIC: We do NOT require word-start (`(?:^|\s)`) because
+  // prefixes like `$x"""-f"` (unset/empty variable) concatenate the same way.
+  // The homogeneous-empty-pair requirement filters out the `'"'"'` idiom
+  // (no homogeneous empty pair — it's close, double-quoted-content, open).
+  //
+  // FALSE POSITIVE: Matches `echo '"""-f" text'` (pattern inside single-quoted
+  // string). Extremely rare (requires echoing the literal attack). Acceptable.
+  if (/(?:""|'')+['"]-/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 10,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains empty quote pair adjacent to quoted dash (potential flag obfuscation)',
+    }
+  }
+
+  // 4c. SECURITY: Also block 3+ consecutive quotes at word start even without
+  // an immediate dash. Broader safety net for multi-quote obfuscation patterns
+  // not enumerated above (e.g., `"""x"-f` where content between quotes shifts
+  // the dash position). Legitimate commands never need `"""x"` when `"x"` works.
+  if (/(?:^|\s)['"]{3,}/.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 11,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains consecutive quote characters at word start (potential obfuscation)',
+    }
+  }
+
+  // Track quote state to avoid false positives for flags inside quoted strings
+  let inSingleQuote = false
+  let inDoubleQuote = false
+  let escaped = false
+
+  for (let i = 0; i < originalCommand.length - 1; i++) {
+    const currentChar = originalCommand[i]
+    const nextChar = originalCommand[i + 1]
+
+    // Update quote state
+    if (escaped) {
+      escaped = false
+      continue
+    }
+
+    // SECURITY: Only treat backslash as escape OUTSIDE single quotes. In bash,
+    // `\` inside `'...'` is LITERAL. Without this guard, `'\'` desyncs the
+    // quote tracker: `\` sets escaped=true, closing `'` is consumed by the
+    // escaped-skip above instead of toggling inSingleQuote. Parser stays in
+    // single-quote mode, and the `if (inSingleQuote || inDoubleQuote) continue`
+    // at line ~1121 skips ALL subsequent flag detection for the rest of the
+    // command. Example: `jq '\' "-f" evil` — bash gets `-f` arg, but desynced
+    // parser thinks ` "-f" evil` is inside quotes → flag detection bypassed.
+    // Defense-in-depth: hasShellQuoteSingleQuoteBug catches `'\'` patterns at
+    // line ~1856 before this runs. But we fix the tracker for consistency with
+    // the CORRECT implementations elsewhere in this file (hasBackslashEscaped*,
+    // extractQuotedContent) which all guard with `!inSingleQuote`.
+    if (currentChar === '\\' && !inSingleQuote) {
+      escaped = true
+      continue
+    }
+
+    if (currentChar === "'" && !inDoubleQuote) {
+      inSingleQuote = !inSingleQuote
+      continue
+    }
+
+    if (currentChar === '"' && !inSingleQuote) {
+      inDoubleQuote = !inDoubleQuote
+      continue
+    }
+
+    // Only look for flags when not inside quoted strings
+    // This prevents false positives like: make test TEST="file.py -v"
+    if (inSingleQuote || inDoubleQuote) {
+      continue
+    }
+
+    // Look for whitespace followed by quote that contains a dash (potential flag obfuscation)
+    // SECURITY: Block ANY quoted content starting with dash - err on side of safety
+    // Catches: "-"exec, "-file", "--flag", '-'output, etc.
+    // Users can approve manually if legitimate (e.g., find . -name "-file")
+    if (
+      currentChar &&
+      nextChar &&
+      /\s/.test(currentChar) &&
+      /['"`]/.test(nextChar)
+    ) {
+      const quoteChar = nextChar
+      let j = i + 2 // Start after the opening quote
+      let insideQuote = ''
+
+      // Collect content inside the quote
+      while (j < originalCommand.length && originalCommand[j] !== quoteChar) {
+        insideQuote += originalCommand[j]!
+        j++
+      }
+
+      // If we found a closing quote and the content looks like an obfuscated flag, block it.
+      // Three attack patterns to catch:
+      //   1. Flag name inside quotes: "--flag", "-exec", "-X" (dashes + letters inside)
+      //   2. Split-quote flag: "-"exec, "--"output (dashes inside, letters continue after quote)
+      //   3. Chained quotes: "-""exec" (dashes in first quote, second quote contains letters)
+      // Pure-dash strings like "---" or "--" followed by whitespace/separator are separators,
+      // not flags, and should not trigger this check.
+      const charAfterQuote = originalCommand[j + 1]
+      // Inside double quotes, $VAR and `cmd` expand at runtime, so "-$VAR" can
+      // become -exec. Blocking $ and ` here over-blocks single-quoted literals
+      // like grep '-$' (where $ is literal), but main's startsWith('-') already
+      // blocked those — this restores status quo, not a new false positive.
+      // Brace expansion ({) does NOT happen inside quotes, so { is not needed here.
+      const hasFlagCharsInside = /^-+[a-zA-Z0-9$`]/.test(insideQuote)
+      // Characters that can continue a flag after a closing quote. This catches:
+      //   a-zA-Z0-9: "-"exec → -exec (direct concatenation)
+      //   \\:        "-"\exec → -exec (backslash escape is stripped)
+      //   -:         "-"-output → --output (extra dashes)
+      //   {:         "-"{exec,delete} → -exec -delete (brace expansion)
+      //   $:         "-"$VAR → -exec when VAR=exec (variable expansion)
+      //   `:         "-"`echo exec` → -exec (command substitution)
+      // Note: glob chars (*?[) are omitted — they require attacker-controlled
+      // filenames in CWD to exploit, and blocking them would break patterns
+      // like `ls -- "-"*` for listing files that start with dash.
+      const FLAG_CONTINUATION_CHARS = /[a-zA-Z0-9\\${`-]/
+      const hasFlagCharsContinuing =
+        /^-+$/.test(insideQuote) &&
+        charAfterQuote !== undefined &&
+        FLAG_CONTINUATION_CHARS.test(charAfterQuote)
+      // Handle adjacent quote chaining: "-""exec" or "-""-"exec or """-"exec concatenates
+      // to -exec in shell. Follow the chain of adjacent quoted segments until
+      // we find one containing an alphanumeric char or hit a non-quote boundary.
+      // Also handles empty prefix quotes: """-"exec where "" is followed by "-"exec
+      // The combined segments form a flag if they contain dash(es) followed by alphanumerics.
+      const hasFlagCharsInNextQuote =
+        // Trigger when: first segment is only dashes OR empty (could be prefix for flag)
+        (insideQuote === '' || /^-+$/.test(insideQuote)) &&
+        charAfterQuote !== undefined &&
+        /['"`]/.test(charAfterQuote) &&
+        (() => {
+          let pos = j + 1 // Start at charAfterQuote (an opening quote)
+          let combinedContent = insideQuote // Track what the shell will see
+          while (
+            pos < originalCommand.length &&
+            /['"`]/.test(originalCommand[pos]!)
+          ) {
+            const segQuote = originalCommand[pos]!
+            let end = pos + 1
+            while (
+              end < originalCommand.length &&
+              originalCommand[end] !== segQuote
+            ) {
+              end++
+            }
+            const segment = originalCommand.slice(pos + 1, end)
+            combinedContent += segment
+
+            // Check if combined content so far forms a flag pattern.
+            // Include $ and ` for in-quote expansion: "-""$VAR" → -exec
+            if (/^-+[a-zA-Z0-9$`]/.test(combinedContent)) return true
+
+            // If this segment has alphanumeric/expansion and we already have dashes,
+            // it's a flag. Catches "-""$*" where segment='$*' has no alnum but
+            // expands to positional params at runtime.
+            // Guard against segment.length === 0: slice(0, -0) → slice(0, 0) → ''.
+            const priorContent =
+              segment.length > 0
+                ? combinedContent.slice(0, -segment.length)
+                : combinedContent
+            if (/^-+$/.test(priorContent)) {
+              if (/[a-zA-Z0-9$`]/.test(segment)) return true
+            }
+
+            if (end >= originalCommand.length) break // Unclosed quote
+            pos = end + 1 // Move past closing quote to check next segment
+          }
+          // Also check the unquoted char at the end of the chain
+          if (
+            pos < originalCommand.length &&
+            FLAG_CONTINUATION_CHARS.test(originalCommand[pos]!)
+          ) {
+            // If we have dashes in combined content, the trailing char completes a flag
+            if (/^-+$/.test(combinedContent) || combinedContent === '') {
+              // Check if we're about to form a flag with the following content
+              const nextChar = originalCommand[pos]!
+              if (nextChar === '-') {
+                // More dashes, could still form a flag
+                return true
+              }
+              if (/[a-zA-Z0-9\\${`]/.test(nextChar) && combinedContent !== '') {
+                // We have dashes and now alphanumeric/expansion follows
+                return true
+              }
+            }
+            // Original check for dashes followed by alphanumeric
+            if (/^-/.test(combinedContent)) {
+              return true
+            }
+          }
+          return false
+        })()
+      if (
+        j < originalCommand.length &&
+        originalCommand[j] === quoteChar &&
+        (hasFlagCharsInside ||
+          hasFlagCharsContinuing ||
+          hasFlagCharsInNextQuote)
+      ) {
+        logEvent('tengu_bash_security_check_triggered', {
+          checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+          subId: 4,
+        })
+        return {
+          behavior: 'ask',
+          message: 'Command contains quoted characters in flag names',
+        }
+      }
+    }
+
+    // Look for whitespace followed by dash - this starts a flag
+    if (currentChar && nextChar && /\s/.test(currentChar) && nextChar === '-') {
+      let j = i + 1 // Start at the dash
+      let flagContent = ''
+
+      // Collect flag content
+      while (j < originalCommand.length) {
+        const flagChar = originalCommand[j]
+        if (!flagChar) break
+
+        // End flag content once we hit whitespace or an equals sign
+        if (/[\s=]/.test(flagChar)) {
+          break
+        }
+        // End flag collection if we hit quote followed by non-flag character. This is needed to handle cases like -d"," which should be parsed as just -d
+        if (/['"`]/.test(flagChar)) {
+          // Special case for cut -d flag: the delimiter value can be quoted
+          // Example: cut -d'"' should parse as flag name: -d, value: '"'
+          // Note: We only apply this exception to cut -d specifically to avoid bypasses.
+          // Without this restriction, a command like `find -e"xec"` could be parsed as
+          // flag name: -e, bypassing our blocklist for -exec. By restricting to cut -d,
+          // we allow the legitimate use case while preventing obfuscation attacks on other
+          // commands where quoted flag values could hide dangerous flag names.
+          if (
+            baseCommand === 'cut' &&
+            flagContent === '-d' &&
+            /['"`]/.test(flagChar)
+          ) {
+            // This is cut -d followed by a quoted delimiter - flagContent is already '-d'
+            break
+          }
+
+          // Look ahead to see what follows the quote
+          if (j + 1 < originalCommand.length) {
+            const nextFlagChar = originalCommand[j + 1]
+            if (nextFlagChar && !/[a-zA-Z0-9_'"-]/.test(nextFlagChar)) {
+              // Quote followed by something that is clearly not part of a flag, end the parsing
+              break
+            }
+          }
+        }
+        flagContent += flagChar
+        j++
+      }
+
+      if (flagContent.includes('"') || flagContent.includes("'")) {
+        logEvent('tengu_bash_security_check_triggered', {
+          checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+          subId: 1,
+        })
+        return {
+          behavior: 'ask',
+          message: 'Command contains quoted characters in flag names',
+        }
+      }
+    }
+  }
+
+  // Also handle flags that start with quotes: "--"output, '-'-output, etc.
+  // Use fullyUnquotedContent to avoid false positives from legitimate quoted content like echo "---"
+  if (/\s['"`]-/.test(context.fullyUnquotedContent)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 2,
+    })
+    return {
+      behavior: 'ask',
+      message: 'Command contains quoted characters in flag names',
+    }
+  }
+
+  // Also handles cases like ""--output
+  // Use fullyUnquotedContent to avoid false positives from legitimate quoted content
+  if (/['"`]{2}-/.test(context.fullyUnquotedContent)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.OBFUSCATED_FLAGS,
+      subId: 3,
+    })
+    return {
+      behavior: 'ask',
+      message: 'Command contains quoted characters in flag names',
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No obfuscated flags detected' }
+}
+
+/**
+ * Detects backslash-escaped whitespace characters (space, tab) outside of quotes.
+ *
+ * In bash, `echo\ test` is a single token (command named "echo test"), but
+ * shell-quote decodes the escape and produces `echo test` (two separate tokens).
+ * This discrepancy allows path traversal attacks like:
+ *   echo\ test/../../../usr/bin/touch /tmp/file
+ * which the parser sees as `echo test/.../touch /tmp/file` (an echo command)
+ * but bash resolves as `/usr/bin/touch /tmp/file` (via directory "echo test").
+ */
+function hasBackslashEscapedWhitespace(command: string): boolean {
+  let inSingleQuote = false
+  let inDoubleQuote = false
+
+  for (let i = 0; i < command.length; i++) {
+    const char = command[i]
+
+    if (char === '\\' && !inSingleQuote) {
+      if (!inDoubleQuote) {
+        const nextChar = command[i + 1]
+        if (nextChar === ' ' || nextChar === '\t') {
+          return true
+        }
+      }
+      // Skip the escaped character (both outside quotes and inside double quotes,
+      // where \\, \", \$, \` are valid escape sequences)
+      i++
+      continue
+    }
+
+    if (char === '"' && !inSingleQuote) {
+      inDoubleQuote = !inDoubleQuote
+      continue
+    }
+
+    if (char === "'" && !inDoubleQuote) {
+      inSingleQuote = !inSingleQuote
+      continue
+    }
+  }
+
+  return false
+}
+
+function validateBackslashEscapedWhitespace(
+  context: ValidationContext,
+): PermissionResult {
+  if (hasBackslashEscapedWhitespace(context.originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.BACKSLASH_ESCAPED_WHITESPACE,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains backslash-escaped whitespace that could alter command parsing',
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'No backslash-escaped whitespace',
+  }
+}
+
+/**
+ * Detects a backslash immediately preceding a shell operator outside of quotes.
+ *
+ * SECURITY: splitCommand normalizes `\;` to a bare `;` in its output string.
+ * When downstream code (checkReadOnlyConstraints, checkPathConstraints, etc.)
+ * re-parses that normalized string, the bare `;` is seen as an operator and
+ * causes a false split. This enables arbitrary file read bypassing path checks:
+ *
+ *   cat safe.txt \; echo ~/.ssh/id_rsa
+ *
+ * In bash: ONE cat command reading safe.txt, ;, echo, ~/.ssh/id_rsa as files.
+ * After splitCommand normalizes: "cat safe.txt ; echo ~/.ssh/id_rsa"
+ * Nested re-parse: ["cat safe.txt", "echo ~/.ssh/id_rsa"] — both segments
+ * pass isCommandReadOnly, sensitive path hidden in echo segment is never
+ * validated by path constraints. Auto-allowed. Private key leaked.
+ *
+ * This check flags any \<operator> regardless of backslash parity. Even counts
+ * (\\;) are dangerous in bash (\\ → \, ; separates). Odd counts (\;) are safe
+ * in bash but trigger the double-parse bug above. Both must be flagged.
+ *
+ * Known false positive: `find . -exec cmd {} \;` — users will be prompted once.
+ *
+ * Note: `(` and `)` are NOT in this set — splitCommand preserves `\(` and `\)`
+ * in its output (round-trip safe), so they don't trigger the double-parse bug.
+ * This allows `find . \( -name x -o -name y \)` to pass without false positives.
+ */
+const SHELL_OPERATORS = new Set([';', '|', '&', '<', '>'])
+
+function hasBackslashEscapedOperator(command: string): boolean {
+  let inSingleQuote = false
+  let inDoubleQuote = false
+
+  for (let i = 0; i < command.length; i++) {
+    const char = command[i]
+
+    // SECURITY: Handle backslash FIRST, before quote toggles. In bash, inside
+    // double quotes, `\"` is an escape sequence producing a literal `"` — it
+    // does NOT close the quote. If we process quote toggles first, `\"` inside
+    // `"..."` desyncs the tracker:
+    //   - `\` is ignored (gated by !inDoubleQuote)
+    //   - `"` toggles inDoubleQuote to FALSE (wrong — bash says still inside)
+    //   - next `"` (the real closing quote) toggles BACK to TRUE — locked desync
+    //   - subsequent `\;` is missed because !inDoubleQuote is false
+    // Exploit: `tac "x\"y" \; echo ~/.ssh/id_rsa` — bash runs ONE tac reading
+    // all args as files (leaking id_rsa), but desynced tracker misses `\;` and
+    // splitCommand's double-parse normalization "sees" two safe commands.
+    //
+    // Fix structure matches hasBackslashEscapedWhitespace (which was correctly
+    // fixed for this in commit prior to d000dfe84e): backslash check first,
+    // gated only by !inSingleQuote (since backslash IS literal inside '...'),
+    // unconditional i++ to skip the escaped char even inside double quotes.
+    if (char === '\\' && !inSingleQuote) {
+      // Only flag \<operator> when OUTSIDE double quotes (inside double quotes,
+      // operators like ;|&<> are already not special, so \; is harmless there).
+      if (!inDoubleQuote) {
+        const nextChar = command[i + 1]
+        if (nextChar && SHELL_OPERATORS.has(nextChar)) {
+          return true
+        }
+      }
+      // Skip the escaped character unconditionally. Inside double quotes, this
+      // correctly consumes backslash pairs: `"x\\"` → pos 6 (`\`) skips pos 7
+      // (`\`), then pos 8 (`"`) toggles inDoubleQuote off correctly. Without
+      // unconditional skip, pos 7 would see `\`, see pos 8 (`"`) as nextChar,
+      // skip it, and the closing quote would NEVER toggle inDoubleQuote —
+      // permanently desyncing and missing subsequent `\;` outside quotes.
+      // Exploit: `cat "x\\" \; echo /etc/passwd` — bash reads /etc/passwd.
+      //
+      // This correctly handles backslash parity: odd-count `\;` (1, 3, 5...)
+      // is flagged (the unpaired `\` before `;` is detected). Even-count `\\;`
+      // (2, 4...) is NOT flagged, which is CORRECT — bash treats `\\` as
+      // literal `\` and `;` as a separator, so splitCommand handles it
+      // normally (no double-parse bug). This matches
+      // hasBackslashEscapedWhitespace line ~1340.
+      i++
+      continue
+    }
+
+    // Quote toggles come AFTER backslash handling (backslash already skipped
+    // any escaped quote char, so these toggles only fire on unescaped quotes).
+    if (char === "'" && !inDoubleQuote) {
+      inSingleQuote = !inSingleQuote
+      continue
+    }
+    if (char === '"' && !inSingleQuote) {
+      inDoubleQuote = !inDoubleQuote
+      continue
+    }
+  }
+
+  return false
+}
+
+function validateBackslashEscapedOperators(
+  context: ValidationContext,
+): PermissionResult {
+  // Tree-sitter path: if tree-sitter confirms no actual operator nodes exist
+  // in the AST, then any \; is just an escaped character in a word argument
+  // (e.g., `find . -exec cmd {} \;`). Skip the expensive regex check.
+  if (context.treeSitter && !context.treeSitter.hasActualOperatorNodes) {
+    return { behavior: 'passthrough', message: 'No operator nodes in AST' }
+  }
+
+  if (hasBackslashEscapedOperator(context.originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.BACKSLASH_ESCAPED_OPERATORS,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains a backslash before a shell operator (;, |, &, <, >) which can hide command structure',
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'No backslash-escaped operators',
+  }
+}
+
+/**
+ * Checks if a character at position `pos` in `content` is escaped by counting
+ * consecutive backslashes before it. An odd number means it's escaped.
+ */
+function isEscapedAtPosition(content: string, pos: number): boolean {
+  let backslashCount = 0
+  let i = pos - 1
+  while (i >= 0 && content[i] === '\\') {
+    backslashCount++
+    i--
+  }
+  return backslashCount % 2 === 1
+}
+
+/**
+ * Detects unquoted brace expansion syntax that Bash expands but shell-quote/tree-sitter
+ * treat as literal strings. This parsing discrepancy allows permission bypass:
+ *   git ls-remote {--upload-pack="touch /tmp/test",test}
+ * Parser sees one literal arg, but Bash expands to: --upload-pack="touch /tmp/test" test
+ *
+ * Brace expansion has two forms:
+ *   1. Comma-separated: {a,b,c} → a b c
+ *   2. Sequence: {1..5} → 1 2 3 4 5
+ *
+ * Both single and double quotes suppress brace expansion in Bash, so we use
+ * fullyUnquotedContent which has both quote types stripped.
+ * Backslash-escaped braces (\{, \}) also suppress expansion.
+ */
+function validateBraceExpansion(context: ValidationContext): PermissionResult {
+  // Use pre-strip content to avoid false negatives from stripSafeRedirections
+  // creating backslash adjacencies (e.g., `\>/dev/null{a,b}` → `\{a,b}` after
+  // stripping, making isEscapedAtPosition think the brace is escaped).
+  const content = context.fullyUnquotedPreStrip
+
+  // SECURITY: Check for MISMATCHED brace counts in fullyUnquoted content.
+  // A mismatch indicates that quoted braces (e.g., `'{'` or `"{"`) were
+  // stripped by extractQuotedContent, leaving unbalanced braces in the content
+  // we analyze. Our depth-matching algorithm below assumes balanced braces —
+  // with a mismatch, it closes at the WRONG position, missing commas that
+  // bash's algorithm WOULD find.
+  //
+  // Exploit: `git diff {@'{'0},--output=/tmp/pwned}`
+  //   - Original: 2 `{`, 2 `}` (quoted `'{'` counts as content, not operator)
+  //   - fullyUnquoted: `git diff {@0},--output=/tmp/pwned}` — 1 `{`, 2 `}`!
+  //   - Our depth-matcher: closes at first `}` (after `0`), inner=`@0`, no `,`
+  //   - Bash (on original): quoted `{` is content; first unquoted `}` has no
+  //     `,` yet → bash treats as literal content, keeps scanning → finds `,`
+  //     → final `}` closes → expands to `@{0} --output=/tmp/pwned`
+  //   - git writes diff to /tmp/pwned. ARBITRARY FILE WRITE, ZERO PERMISSIONS.
+  //
+  // We count ONLY unescaped braces (backslash-escaped braces are literal in
+  // bash). If counts mismatch AND at least one unescaped `{` exists, block —
+  // our depth-matching cannot be trusted on this content.
+  let unescapedOpenBraces = 0
+  let unescapedCloseBraces = 0
+  for (let i = 0; i < content.length; i++) {
+    if (content[i] === '{' && !isEscapedAtPosition(content, i)) {
+      unescapedOpenBraces++
+    } else if (content[i] === '}' && !isEscapedAtPosition(content, i)) {
+      unescapedCloseBraces++
+    }
+  }
+  // Only block when CLOSE count EXCEEDS open count — this is the specific
+  // attack signature. More `}` than `{` means a quoted `{` was stripped
+  // (bash saw it as content, we see extra `}` unaccounted for). The inverse
+  // (more `{` than `}`) is usually legitimate unclosed/escaped braces like
+  // `{foo` or `{a,b\}` where bash doesn't expand anyway.
+  if (unescapedOpenBraces > 0 && unescapedCloseBraces > unescapedOpenBraces) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.BRACE_EXPANSION,
+      subId: 2,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command has excess closing braces after quote stripping, indicating possible brace expansion obfuscation',
+    }
+  }
+
+  // SECURITY: Additionally, check the ORIGINAL command (before quote stripping)
+  // for `'{'` or `"{"` INSIDE an unquoted brace context — this is the specific
+  // attack primitive. A quoted brace inside an outer unquoted `{...}` is
+  // essentially always an obfuscation attempt; legitimate commands don't nest
+  // quoted braces inside brace expansion (awk/find patterns are fully quoted,
+  // like `awk '{print $1}'` where the OUTER brace is inside quotes too).
+  //
+  // This catches the attack even if an attacker crafts a payload with balanced
+  // stripped braces (defense-in-depth). We use a simple heuristic: if the
+  // original command has `'{'` or `'}'` or `"{"` or `"}"` (quoted single brace)
+  // AND also has an unquoted `{`, that's suspicious.
+  if (unescapedOpenBraces > 0) {
+    const orig = context.originalCommand
+    // Look for quoted single-brace patterns: '{', '}', "{",  "}"
+    // These are the attack primitive — a brace char wrapped in quotes.
+    if (/['"][{}]['"]/.test(orig)) {
+      logEvent('tengu_bash_security_check_triggered', {
+        checkId: BASH_SECURITY_CHECK_IDS.BRACE_EXPANSION,
+        subId: 3,
+      })
+      return {
+        behavior: 'ask',
+        message:
+          'Command contains quoted brace character inside brace context (potential brace expansion obfuscation)',
+      }
+    }
+  }
+
+  // Scan for unescaped `{` characters, then check if they form brace expansion.
+  // We use a manual scan rather than a simple regex lookbehind because
+  // lookbehinds can't handle double-escaped backslashes (\\{ is unescaped `{`).
+  for (let i = 0; i < content.length; i++) {
+    if (content[i] !== '{') continue
+    if (isEscapedAtPosition(content, i)) continue
+
+    // Find matching unescaped `}` by tracking nesting depth.
+    // Previous approach broke on nested `{`, missing commas between the outer
+    // `{` and the nested one (e.g., `{--upload-pack="evil",{test}}`).
+    let depth = 1
+    let matchingClose = -1
+    for (let j = i + 1; j < content.length; j++) {
+      const ch = content[j]
+      if (ch === '{' && !isEscapedAtPosition(content, j)) {
+        depth++
+      } else if (ch === '}' && !isEscapedAtPosition(content, j)) {
+        depth--
+        if (depth === 0) {
+          matchingClose = j
+          break
+        }
+      }
+    }
+
+    if (matchingClose === -1) continue
+
+    // Check for `,` or `..` at the outermost nesting level between this
+    // `{` and its matching `}`. Only depth-0 triggers matter — bash splits
+    // brace expansion at outer-level commas/sequences.
+    let innerDepth = 0
+    for (let k = i + 1; k < matchingClose; k++) {
+      const ch = content[k]
+      if (ch === '{' && !isEscapedAtPosition(content, k)) {
+        innerDepth++
+      } else if (ch === '}' && !isEscapedAtPosition(content, k)) {
+        innerDepth--
+      } else if (innerDepth === 0) {
+        if (
+          ch === ',' ||
+          (ch === '.' && k + 1 < matchingClose && content[k + 1] === '.')
+        ) {
+          logEvent('tengu_bash_security_check_triggered', {
+            checkId: BASH_SECURITY_CHECK_IDS.BRACE_EXPANSION,
+            subId: 1,
+          })
+          return {
+            behavior: 'ask',
+            message:
+              'Command contains brace expansion that could alter command parsing',
+          }
+        }
+      }
+    }
+    // No expansion at this level — don't skip past; inner pairs will be
+    // caught by subsequent iterations of the outer loop.
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'No brace expansion detected',
+  }
+}
+
+// Matches Unicode whitespace characters that shell-quote treats as word
+// separators but bash treats as literal word content. While this differential
+// is defense-favorable (shell-quote over-splits), blocking these proactively
+// prevents future edge cases.
+// eslint-disable-next-line no-misleading-character-class
+const UNICODE_WS_RE =
+  /[\u00A0\u1680\u2000-\u200A\u2028\u2029\u202F\u205F\u3000\uFEFF]/
+
+function validateUnicodeWhitespace(
+  context: ValidationContext,
+): PermissionResult {
+  const { originalCommand } = context
+  if (UNICODE_WS_RE.test(originalCommand)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.UNICODE_WHITESPACE,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains Unicode whitespace characters that could cause parsing inconsistencies',
+    }
+  }
+  return { behavior: 'passthrough', message: 'No Unicode whitespace' }
+}
+
+function validateMidWordHash(context: ValidationContext): PermissionResult {
+  const { unquotedKeepQuoteChars } = context
+  // Match # preceded by a non-whitespace character (mid-word hash).
+  // shell-quote treats mid-word # as comment-start but bash treats it as a
+  // literal character, creating a parser differential.
+  //
+  // Uses unquotedKeepQuoteChars (which preserves quote delimiters but strips
+  // quoted content) to catch quote-adjacent # like 'x'# — fullyUnquotedPreStrip
+  // would strip both quotes and content, turning 'x'# into just # (word-start).
+  //
+  // SECURITY: Also check the CONTINUATION-JOINED version. The context is built
+  // from the original command (pre-continuation-join). For `foo\<NL>#bar`,
+  // pre-join the `#` is preceded by `\n` (whitespace → `/\S#/` doesn't match),
+  // but post-join it's preceded by `o` (non-whitespace → matches). shell-quote
+  // operates on the post-join text (line continuations are joined in
+  // splitCommand), so the parser differential manifests on the joined text.
+  // While not directly exploitable (the `#...` fragment still prompts as its
+  // own subcommand), this is a defense-in-depth gap — shell-quote would drop
+  // post-`#` content from path extraction.
+  //
+  // Exclude ${# which is bash string-length syntax (e.g., ${#var}).
+  // Note: the lookbehind must be placed immediately before # (not before \S)
+  // so that it checks the correct 2-char window.
+  const joined = unquotedKeepQuoteChars.replace(/\\+\n/g, match => {
+    const backslashCount = match.length - 1
+    return backslashCount % 2 === 1 ? '\\'.repeat(backslashCount - 1) : match
+  })
+  if (
+    // eslint-disable-next-line custom-rules/no-lookbehind-regex -- .test() with atom search: fast when # absent
+    /\S(?<!\$\{)#/.test(unquotedKeepQuoteChars) ||
+    // eslint-disable-next-line custom-rules/no-lookbehind-regex -- same as above
+    /\S(?<!\$\{)#/.test(joined)
+  ) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.MID_WORD_HASH,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains mid-word # which is parsed differently by shell-quote vs bash',
+    }
+  }
+  return { behavior: 'passthrough', message: 'No mid-word hash' }
+}
+
+/**
+ * Detects when a `#` comment contains quote characters that would desync
+ * downstream quote trackers (like extractQuotedContent).
+ *
+ * In bash, everything after an unquoted `#` on a line is a comment — quote
+ * characters inside the comment are literal text, not quote toggles. But our
+ * quote-tracking functions don't handle comments, so a `'` or `"` after `#`
+ * toggles their quote state. Attackers can craft `# ' "` sequences that
+ * precisely desync the tracker, causing subsequent content (on following
+ * lines) to appear "inside quotes" when it's actually unquoted in bash.
+ *
+ * Example attack:
+ *   echo "it's" # ' " <<'MARKER'\n
+ *   rm -rf /\n
+ *   MARKER
+ * In bash: `#` starts a comment, `rm -rf /` executes on line 2.
+ * In extractQuotedContent: the `'` at position 14 (after #) opens a single
+ * quote, and the `'` before MARKER closes it. But the `'` after MARKER opens
+ * ANOTHER single quote, swallowing the newline and `rm -rf /`, so
+ * validateNewlines sees no unquoted newlines.
+ *
+ * Defense: If we see an unquoted `#` followed by any quote character on the
+ * same line, treat it as a misparsing concern. Legitimate commands rarely
+ * have quote characters in their comments (and if they do, the user can
+ * approve manually).
+ */
+function validateCommentQuoteDesync(
+  context: ValidationContext,
+): PermissionResult {
+  // Tree-sitter path: tree-sitter correctly identifies comment nodes and
+  // quoted content. The desync concern is about regex quote tracking being
+  // confused by quote characters inside comments. When tree-sitter provides
+  // the quote context, this desync cannot happen — the AST is authoritative
+  // regardless of whether the command contains a comment.
+  if (context.treeSitter) {
+    return {
+      behavior: 'passthrough',
+      message: 'Tree-sitter quote context is authoritative',
+    }
+  }
+
+  const { originalCommand } = context
+
+  // Track quote state character-by-character using the same (correct) logic
+  // as extractQuotedContent: single quotes don't toggle inside double quotes.
+  // When we encounter an unquoted `#`, check if the rest of the line (until
+  // newline) contains any quote characters.
+  let inSingleQuote = false
+  let inDoubleQuote = false
+  let escaped = false
+
+  for (let i = 0; i < originalCommand.length; i++) {
+    const char = originalCommand[i]
+
+    if (escaped) {
+      escaped = false
+      continue
+    }
+
+    if (inSingleQuote) {
+      if (char === "'") inSingleQuote = false
+      continue
+    }
+
+    if (char === '\\') {
+      escaped = true
+      continue
+    }
+
+    if (inDoubleQuote) {
+      if (char === '"') inDoubleQuote = false
+      // Single quotes inside double quotes are literal — no toggle
+      continue
+    }
+
+    if (char === "'") {
+      inSingleQuote = true
+      continue
+    }
+
+    if (char === '"') {
+      inDoubleQuote = true
+      continue
+    }
+
+    // Unquoted `#` — in bash, this starts a comment. Check if the rest of
+    // the line contains quote characters that would desync other trackers.
+    if (char === '#') {
+      const lineEnd = originalCommand.indexOf('\n', i)
+      const commentText = originalCommand.slice(
+        i + 1,
+        lineEnd === -1 ? originalCommand.length : lineEnd,
+      )
+      if (/['"]/.test(commentText)) {
+        logEvent('tengu_bash_security_check_triggered', {
+          checkId: BASH_SECURITY_CHECK_IDS.COMMENT_QUOTE_DESYNC,
+        })
+        return {
+          behavior: 'ask',
+          message:
+            'Command contains quote characters inside a # comment which can desync quote tracking',
+        }
+      }
+      // Skip to end of line (rest is comment)
+      if (lineEnd === -1) break
+      i = lineEnd // Loop increment will move past newline
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No comment quote desync' }
+}
+
+/**
+ * Detects a newline inside a quoted string where the NEXT line would be
+ * stripped by stripCommentLines (trimmed line starts with `#`).
+ *
+ * In bash, `\n` inside quotes is a literal character and part of the argument.
+ * But stripCommentLines (called by stripSafeWrappers in bashPermissions before
+ * path validation and rule matching) processes commands LINE-BY-LINE via
+ * `command.split('\n')` without tracking quote state. A quoted newline lets an
+ * attacker position the next line to start with `#` (after trim), causing
+ * stripCommentLines to drop that line entirely — hiding sensitive paths or
+ * arguments from path validation and permission rule matching.
+ *
+ * Example attack (auto-allowed in acceptEdits mode without any Bash rules):
+ *   mv ./decoy '<\n>#' ~/.ssh/id_rsa ./exfil_dir
+ * Bash: moves ./decoy AND ~/.ssh/id_rsa into ./exfil_dir/ (errors on `\n#`).
+ * stripSafeWrappers: line 2 starts with `#` → stripped → "mv ./decoy '".
+ * shell-quote: drops unbalanced trailing quote → ["mv", "./decoy"].
+ * checkPathConstraints: only sees ./decoy (in cwd) → passthrough.
+ * acceptEdits mode: mv with all-cwd paths → ALLOW. Zero clicks, no warning.
+ *
+ * Also works with cp (exfil), rm/rm -rf (delete arbitrary files/dirs).
+ *
+ * Defense: block ONLY the specific stripCommentLines trigger — a newline inside
+ * quotes where the next line starts with `#` after trim. This is the minimal
+ * check that catches the parser differential while preserving legitimate
+ * multi-line quoted arguments (echo 'line1\nline2', grep patterns, etc.).
+ * Safe heredocs ($(cat <<'EOF'...)) and git commit -m "..." are handled by
+ * early validators and never reach this check.
+ *
+ * This validator is NOT in nonMisparsingValidators — its ask result gets
+ * isBashSecurityCheckForMisparsing: true, causing an early block in the
+ * permission flow at bashPermissions.ts before any line-based processing runs.
+ */
+function validateQuotedNewline(context: ValidationContext): PermissionResult {
+  const { originalCommand } = context
+
+  // Fast path: must have both a newline byte AND a # character somewhere.
+  // stripCommentLines only strips lines where trim().startsWith('#'), so
+  // no # means no possible trigger.
+  if (!originalCommand.includes('\n') || !originalCommand.includes('#')) {
+    return { behavior: 'passthrough', message: 'No newline or no hash' }
+  }
+
+  // Track quote state. Mirrors extractQuotedContent / validateCommentQuoteDesync:
+  // - single quotes don't toggle inside double quotes
+  // - backslash escapes the next char (but not inside single quotes)
+  // stripCommentLines splits on '\n' (not \r), so we only treat \n as a line
+  // separator. \r inside a line is removed by trim() and doesn't change the
+  // trimmed-starts-with-# check.
+  let inSingleQuote = false
+  let inDoubleQuote = false
+  let escaped = false
+
+  for (let i = 0; i < originalCommand.length; i++) {
+    const char = originalCommand[i]
+
+    if (escaped) {
+      escaped = false
+      continue
+    }
+
+    if (char === '\\' && !inSingleQuote) {
+      escaped = true
+      continue
+    }
+
+    if (char === "'" && !inDoubleQuote) {
+      inSingleQuote = !inSingleQuote
+      continue
+    }
+
+    if (char === '"' && !inSingleQuote) {
+      inDoubleQuote = !inDoubleQuote
+      continue
+    }
+
+    // A newline inside quotes: the NEXT line (from bash's perspective) starts
+    // inside a quoted string. Check if that line would be stripped by
+    // stripCommentLines — i.e., after trim(), does it start with `#`?
+    // This exactly mirrors: lines.filter(l => !l.trim().startsWith('#'))
+    if (char === '\n' && (inSingleQuote || inDoubleQuote)) {
+      const lineStart = i + 1
+      const nextNewline = originalCommand.indexOf('\n', lineStart)
+      const lineEnd = nextNewline === -1 ? originalCommand.length : nextNewline
+      const nextLine = originalCommand.slice(lineStart, lineEnd)
+      if (nextLine.trim().startsWith('#')) {
+        logEvent('tengu_bash_security_check_triggered', {
+          checkId: BASH_SECURITY_CHECK_IDS.QUOTED_NEWLINE,
+        })
+        return {
+          behavior: 'ask',
+          message:
+            'Command contains a quoted newline followed by a #-prefixed line, which can hide arguments from line-based permission checks',
+        }
+      }
+    }
+  }
+
+  return { behavior: 'passthrough', message: 'No quoted newline-hash pattern' }
+}
+
+/**
+ * Validates that the command doesn't use Zsh-specific dangerous commands that
+ * can bypass security checks. These commands provide capabilities like loading
+ * kernel modules, raw file I/O, network access, and pseudo-terminal execution
+ * that circumvent normal permission checks.
+ *
+ * Also catches `fc -e` which can execute arbitrary editors on command history,
+ * and `emulate` which with `-c` is an eval-equivalent.
+ */
+function validateZshDangerousCommands(
+  context: ValidationContext,
+): PermissionResult {
+  const { originalCommand } = context
+
+  // Extract the base command from the original command, stripping leading
+  // whitespace, env var assignments, and Zsh precommand modifiers.
+  // e.g., "FOO=bar command builtin zmodload" -> "zmodload"
+  const ZSH_PRECOMMAND_MODIFIERS = new Set([
+    'command',
+    'builtin',
+    'noglob',
+    'nocorrect',
+  ])
+  const trimmed = originalCommand.trim()
+  const tokens = trimmed.split(/\s+/)
+  let baseCmd = ''
+  for (const token of tokens) {
+    // Skip env var assignments (VAR=value)
+    if (/^[A-Za-z_]\w*=/.test(token)) continue
+    // Skip Zsh precommand modifiers (they don't change what command runs)
+    if (ZSH_PRECOMMAND_MODIFIERS.has(token)) continue
+    baseCmd = token
+    break
+  }
+
+  if (ZSH_DANGEROUS_COMMANDS.has(baseCmd)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.ZSH_DANGEROUS_COMMANDS,
+      subId: 1,
+    })
+    return {
+      behavior: 'ask',
+      message: `Command uses Zsh-specific '${baseCmd}' which can bypass security checks`,
+    }
+  }
+
+  // Check for `fc -e` which allows executing arbitrary commands via editor
+  // fc without -e is safe (just lists history), but -e specifies an editor
+  // to run on the command, effectively an eval
+  if (baseCmd === 'fc' && /\s-\S*e/.test(trimmed)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.ZSH_DANGEROUS_COMMANDS,
+      subId: 2,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        "Command uses 'fc -e' which can execute arbitrary commands via editor",
+    }
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'No Zsh dangerous commands',
+  }
+}
+
+// Matches non-printable control characters that have no legitimate use in shell
+// commands: 0x00-0x08, 0x0B-0x0C, 0x0E-0x1F, 0x7F. Excludes tab (0x09),
+// newline (0x0A), and carriage return (0x0D) which are handled by other
+// validators. Bash silently drops null bytes and ignores most control chars,
+// so an attacker can use them to slip metacharacters past our checks while
+// bash still executes them (e.g., "echo safe\x00; rm -rf /").
+// eslint-disable-next-line no-control-regex
+const CONTROL_CHAR_RE = /[\x00-\x08\x0B\x0C\x0E-\x1F\x7F]/
+
+/**
+ * @deprecated Legacy regex/shell-quote path. Only used when tree-sitter is
+ * unavailable. The primary gate is parseForSecurity (ast.ts).
+ */
+export function bashCommandIsSafe_DEPRECATED(
+  command: string,
+): PermissionResult {
+  // SECURITY: Block control characters before any other processing. Null bytes
+  // and other non-printable chars are silently dropped by bash but confuse our
+  // validators, allowing metacharacters adjacent to them to slip through.
+  if (CONTROL_CHAR_RE.test(command)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.CONTROL_CHARACTERS,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains non-printable control characters that could be used to bypass security checks',
+      isBashSecurityCheckForMisparsing: true,
+    }
+  }
+
+  // SECURITY: Detect '\' patterns that exploit shell-quote's incorrect handling
+  // of backslashes inside single quotes. Must run before shell-quote parsing.
+  if (hasShellQuoteSingleQuoteBug(command)) {
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains single-quoted backslash pattern that could bypass security checks',
+      isBashSecurityCheckForMisparsing: true,
+    }
+  }
+
+  // SECURITY: Strip heredoc bodies before running security validators.
+  // Only strip bodies for quoted/escaped delimiters (<<'EOF', <<\EOF) where
+  // the body is literal text — $(), backticks, and ${} are NOT expanded.
+  // Unquoted heredocs (<<EOF) undergo full shell expansion, so their bodies
+  // may contain executable command substitutions that validators must see.
+  // When extractHeredocs bails out (can't parse safely), the raw command
+  // goes through all validators — which is the safe direction.
+  const { processedCommand } = extractHeredocs(command, { quotedOnly: true })
+
+  const baseCommand = command.split(' ')[0] || ''
+  const { withDoubleQuotes, fullyUnquoted, unquotedKeepQuoteChars } =
+    extractQuotedContent(processedCommand, baseCommand === 'jq')
+
+  const context: ValidationContext = {
+    originalCommand: command,
+    baseCommand,
+    unquotedContent: withDoubleQuotes,
+    fullyUnquotedContent: stripSafeRedirections(fullyUnquoted),
+    fullyUnquotedPreStrip: fullyUnquoted,
+    unquotedKeepQuoteChars,
+  }
+
+  const earlyValidators = [
+    validateEmpty,
+    validateIncompleteCommands,
+    validateSafeCommandSubstitution,
+    validateGitCommit,
+  ]
+
+  for (const validator of earlyValidators) {
+    const result = validator(context)
+    if (result.behavior === 'allow') {
+      return {
+        behavior: 'passthrough',
+        message:
+          result.decisionReason?.type === 'other' ||
+          result.decisionReason?.type === 'safetyCheck'
+            ? result.decisionReason.reason
+            : 'Command allowed',
+      }
+    }
+    if (result.behavior !== 'passthrough') {
+      return result.behavior === 'ask'
+        ? { ...result, isBashSecurityCheckForMisparsing: true as const }
+        : result
+    }
+  }
+
+  // Validators that don't set isBashSecurityCheckForMisparsing — their ask
+  // results go through the standard permission flow rather than being blocked
+  // early. LF newlines and redirections are normal patterns that splitCommand
+  // handles correctly, not misparsing concerns.
+  //
+  // NOTE: validateCarriageReturn is NOT here — CR IS a misparsing concern.
+  // shell-quote's `[^\s]` treats CR as a word separator (JS `\s` ⊃ \r), but
+  // bash IFS does NOT include CR. splitCommand collapses CR→space, which IS
+  // misparsing. See validateCarriageReturn for the full attack trace.
+  const nonMisparsingValidators = new Set([
+    validateNewlines,
+    validateRedirections,
+  ])
+
+  const validators = [
+    validateJqCommand,
+    validateObfuscatedFlags,
+    validateShellMetacharacters,
+    validateDangerousVariables,
+    // Run comment-quote-desync BEFORE validateNewlines: it detects cases where
+    // the quote tracker would miss newlines due to # comment desync.
+    validateCommentQuoteDesync,
+    // Run quoted-newline BEFORE validateNewlines: it detects the INVERSE case
+    // (newlines INSIDE quotes, which validateNewlines ignores by design). Quoted
+    // newlines let attackers split commands across lines so that line-based
+    // processing (stripCommentLines) drops sensitive content.
+    validateQuotedNewline,
+    // CR check runs BEFORE validateNewlines — CR is a MISPARSING concern
+    // (shell-quote/bash tokenization differential), LF is not.
+    validateCarriageReturn,
+    validateNewlines,
+    validateIFSInjection,
+    validateProcEnvironAccess,
+    validateDangerousPatterns,
+    validateRedirections,
+    validateBackslashEscapedWhitespace,
+    validateBackslashEscapedOperators,
+    validateUnicodeWhitespace,
+    validateMidWordHash,
+    validateBraceExpansion,
+    validateZshDangerousCommands,
+    // Run malformed token check last - other validators should catch specific patterns first
+    // (e.g., $() substitution, backticks, etc.) since they have more precise error messages
+    validateMalformedTokenInjection,
+  ]
+
+  // SECURITY: We must NOT short-circuit when a non-misparsing validator
+  // returns 'ask' if there are still misparsing validators later in the list.
+  // Non-misparsing ask results are discarded at bashPermissions.ts:~1301-1303
+  // (the gate only blocks when isBashSecurityCheckForMisparsing is set). If
+  // validateRedirections (index 10, non-misparsing) fires first on `>`, it
+  // returns ask-without-flag — but validateBackslashEscapedOperators (index 12,
+  // misparsing) would have caught `\;` WITH the flag. Short-circuiting lets a
+  // payload like `cat safe.txt \; echo /etc/passwd > ./out` slip through.
+  //
+  // Fix: defer non-misparsing ask results. Continue running validators; if any
+  // misparsing validator fires, return THAT (with the flag). Only if we reach
+  // the end without a misparsing ask, return the deferred non-misparsing ask.
+  let deferredNonMisparsingResult: PermissionResult | null = null
+  for (const validator of validators) {
+    const result = validator(context)
+    if (result.behavior === 'ask') {
+      if (nonMisparsingValidators.has(validator)) {
+        if (deferredNonMisparsingResult === null) {
+          deferredNonMisparsingResult = result
+        }
+        continue
+      }
+      return { ...result, isBashSecurityCheckForMisparsing: true as const }
+    }
+  }
+  if (deferredNonMisparsingResult !== null) {
+    return deferredNonMisparsingResult
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'Command passed all security checks',
+  }
+}
+
+/**
+ * @deprecated Legacy regex/shell-quote path. Only used when tree-sitter is
+ * unavailable. The primary gate is parseForSecurity (ast.ts).
+ *
+ * Async version of bashCommandIsSafe that uses tree-sitter when available
+ * for more accurate parsing. Falls back to the sync regex version when
+ * tree-sitter is not available.
+ *
+ * This should be used by async callers (bashPermissions.ts, bashCommandHelpers.ts).
+ * Sync callers (readOnlyValidation.ts) should continue using bashCommandIsSafe().
+ */
+export async function bashCommandIsSafeAsync_DEPRECATED(
+  command: string,
+  onDivergence?: () => void,
+): Promise<PermissionResult> {
+  // Try to get tree-sitter analysis
+  const parsed = await ParsedCommand.parse(command)
+  const tsAnalysis = parsed?.getTreeSitterAnalysis() ?? null
+
+  // If no tree-sitter, fall back to sync version
+  if (!tsAnalysis) {
+    return bashCommandIsSafe_DEPRECATED(command)
+  }
+
+  // Run the same security checks but with tree-sitter enriched context.
+  // The early checks (control chars, shell-quote bug) don't benefit from
+  // tree-sitter, so we run them identically.
+  if (CONTROL_CHAR_RE.test(command)) {
+    logEvent('tengu_bash_security_check_triggered', {
+      checkId: BASH_SECURITY_CHECK_IDS.CONTROL_CHARACTERS,
+    })
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains non-printable control characters that could be used to bypass security checks',
+      isBashSecurityCheckForMisparsing: true,
+    }
+  }
+
+  if (hasShellQuoteSingleQuoteBug(command)) {
+    return {
+      behavior: 'ask',
+      message:
+        'Command contains single-quoted backslash pattern that could bypass security checks',
+      isBashSecurityCheckForMisparsing: true,
+    }
+  }
+
+  const { processedCommand } = extractHeredocs(command, { quotedOnly: true })
+
+  const baseCommand = command.split(' ')[0] || ''
+
+  // Use tree-sitter quote context for more accurate analysis
+  const tsQuote = tsAnalysis.quoteContext
+  const regexQuote = extractQuotedContent(
+    processedCommand,
+    baseCommand === 'jq',
+  )
+
+  // Use tree-sitter quote context as primary, but keep regex as reference
+  // for divergence logging
+  const withDoubleQuotes = tsQuote.withDoubleQuotes
+  const fullyUnquoted = tsQuote.fullyUnquoted
+  const unquotedKeepQuoteChars = tsQuote.unquotedKeepQuoteChars
+
+  const context: ValidationContext = {
+    originalCommand: command,
+    baseCommand,
+    unquotedContent: withDoubleQuotes,
+    fullyUnquotedContent: stripSafeRedirections(fullyUnquoted),
+    fullyUnquotedPreStrip: fullyUnquoted,
+    unquotedKeepQuoteChars,
+    treeSitter: tsAnalysis,
+  }
+
+  // Log divergence between tree-sitter and regex quote extraction.
+  // Skip for heredoc commands: tree-sitter strips (quoted) heredoc bodies
+  // to nothing while the regex path replaces them with placeholder strings
+  // (via extractHeredocs), so the two outputs can never match. Logging
+  // divergence for every heredoc command would poison the signal.
+  //
+  // onDivergence callback: when called in a fanout loop (bashPermissions.ts
+  // Promise.all over subcommands), the caller batches divergences into a
+  // single logEvent instead of N separate calls. Each logEvent triggers
+  // getEventMetadata() → buildProcessMetrics() → process.memoryUsage() →
+  // /proc/self/stat read; with memoized metadata these resolve as microtasks
+  // and starve the event loop (CC-643). Single-command callers omit the
+  // callback and get the original per-call logEvent behavior.
+  if (!tsAnalysis.dangerousPatterns.hasHeredoc) {
+    const hasDivergence =
+      tsQuote.fullyUnquoted !== regexQuote.fullyUnquoted ||
+      tsQuote.withDoubleQuotes !== regexQuote.withDoubleQuotes
+    if (hasDivergence) {
+      if (onDivergence) {
+        onDivergence()
+      } else {
+        logEvent('tengu_tree_sitter_security_divergence', {
+          quoteContextDivergence: true,
+        })
+      }
+    }
+  }
+
+  const earlyValidators = [
+    validateEmpty,
+    validateIncompleteCommands,
+    validateSafeCommandSubstitution,
+    validateGitCommit,
+  ]
+
+  for (const validator of earlyValidators) {
+    const result = validator(context)
+    if (result.behavior === 'allow') {
+      return {
+        behavior: 'passthrough',
+        message:
+          result.decisionReason?.type === 'other' ||
+          result.decisionReason?.type === 'safetyCheck'
+            ? result.decisionReason.reason
+            : 'Command allowed',
+      }
+    }
+    if (result.behavior !== 'passthrough') {
+      return result.behavior === 'ask'
+        ? { ...result, isBashSecurityCheckForMisparsing: true as const }
+        : result
+    }
+  }
+
+  const nonMisparsingValidators = new Set([
+    validateNewlines,
+    validateRedirections,
+  ])
+
+  const validators = [
+    validateJqCommand,
+    validateObfuscatedFlags,
+    validateShellMetacharacters,
+    validateDangerousVariables,
+    validateCommentQuoteDesync,
+    validateQuotedNewline,
+    validateCarriageReturn,
+    validateNewlines,
+    validateIFSInjection,
+    validateProcEnvironAccess,
+    validateDangerousPatterns,
+    validateRedirections,
+    validateBackslashEscapedWhitespace,
+    validateBackslashEscapedOperators,
+    validateUnicodeWhitespace,
+    validateMidWordHash,
+    validateBraceExpansion,
+    validateZshDangerousCommands,
+    validateMalformedTokenInjection,
+  ]
+
+  let deferredNonMisparsingResult: PermissionResult | null = null
+  for (const validator of validators) {
+    const result = validator(context)
+    if (result.behavior === 'ask') {
+      if (nonMisparsingValidators.has(validator)) {
+        if (deferredNonMisparsingResult === null) {
+          deferredNonMisparsingResult = result
+        }
+        continue
+      }
+      return { ...result, isBashSecurityCheckForMisparsing: true as const }
+    }
+  }
+  if (deferredNonMisparsingResult !== null) {
+    return deferredNonMisparsingResult
+  }
+
+  return {
+    behavior: 'passthrough',
+    message: 'Command passed all security checks',
+  }
+}