vskill 1.2.6 → 1.2.7

package/dist/scanner/patterns.js

@@ -1,8 +1,29 @@
 // ---------------------------------------------------------------------------
 // Tier 1 Security Scan Patterns
-// 52 regex-based patterns for static analysis of skill content.
+// 58 regex-based patterns for static analysis of skill content, plus the
+// hidden-content checks (HC-*) in hidden-content.ts.
 // ---------------------------------------------------------------------------
-// ---- Patterns (52 total) --------------------------------------------------
+import { HIDDEN_CONTENT_RULES, scanHiddenContent } from "./hidden-content.js";
+import { DOT_SEGMENT_URL_SOURCE, OFFICIAL_INSTALLER_URL_SOURCE } from "./installer-hosts.js";
+// ---- Shared fragments ------------------------------------------------------
+/**
+ * netcat in command position: at the start of a line or string or after a
+ * list marker or shell prompt there, after a shell separator, colon, quote,
+ * backtick or `$(`, or after a wrapper or verb such as sudo, nohup,
+ * timeout N, run or Run; an absolute path is allowed. No fragment repeats and
+ * no whitespace run crosses a newline, so a long line or a run of blank lines
+ * is scanned in one pass.
+ */
+const NETCAT_COMMAND = String.raw `(?:^|^[^\S\n]*(?:[-*+>$#]|\d+[.)])(?=\s)|[\n|;&(\x60'":]|\$\(|\b(?:sudo|doas|busybox|nohup|exec|env|command|time|then|do|[Rr]un|[Ee]xecute)(?=\s)|\btimeout[^\S\n]+[\w.]+(?=\s))[^\S\n]*(?:(?:\/[\w.-]+)*\/)?(?:nc|ncat|netcat)\b`;
+/**
+ * An official installer URL start with no dot segment in reach: a dot
+ * segment can climb out of an allowlisted owner's path on
+ * raw.githubusercontent.com, so such a URL is not official.
+ */
+const OFFICIAL_INSTALLER_URL_SOURCE_NO_DOT_SEGMENTS = `(?!${DOT_SEGMENT_URL_SOURCE})(?:${OFFICIAL_INSTALLER_URL_SOURCE})`;
+/** A filesystem API call opening its argument list (FS-003). */
+const FS_CALL = String.raw `(?:readFile|writeFile|appendFile|createReadStream|createWriteStream|open|unlink|stat|lstat|fstat|access|rename|copyFile|chmod|chown|truncate)(?:Sync)?\s*\(`;
+// ---- Patterns (58 total) --------------------------------------------------
 export const SCAN_PATTERNS = [
     // --- Command Injection (1-7) ---------------------------------------------
     {
@@ -50,7 +71,7 @@ export const SCAN_PATTERNS = [
         name: "Shell pipe operator",
         severity: "medium",
         description: "Detects shell piping within exec-like calls",
-        pattern: /exec\s*\([^)]*\|[^)]*\)/g,
+        pattern: /exec\s*\((?:(?!exec\s*\()[^)|\n])*\|(?:(?!exec\s*\()[^)\n])*\)/g,
         category: "command-injection",
     },
     {
@@ -66,7 +87,38 @@ export const SCAN_PATTERNS = [
         name: "Pipe-to-shell execution",
         severity: "critical",
         description: "Detects curl/wget output piped directly to a shell interpreter (download-and-execute)",
-        pattern: /\b(?:curl|wget)\b[^|]*\|\s*(?:(?:\/\w+)*\/)?(?:env\s+)?(?:ba|z|da|k)?sh\b/g,
+        // Each scan stops at the next curl/wget, so a pipe-free run costs one
+        // pass however many times the command repeats, and padding cannot hide the pipe.
+        pattern: /\b(?:curl|wget)\b(?:(?!\b(?:curl|wget)\b)[^|\n])*\|\s*(?:(?:\/\w+)*\/)?(?:env\s+)?(?:ba|z|da|k)?sh\b/g,
+        category: "command-injection",
+    },
+    {
+        id: "CI-009",
+        name: "Python shell execution",
+        severity: "high",
+        description: "Detects os.system/os.popen and subprocess calls with shell=True",
+        // os.system( is already CI-003 ("system("); only the spaced form is added.
+        pattern: /\bos\.popen\s*\(|\bos\.system\s+\(|\bsubprocess\.(?:run|call|Popen|check_output|check_call)\s*\([^)]{0,400}?\bshell\s*=\s*True/g,
+        category: "command-injection",
+    },
+    {
+        id: "CI-010",
+        name: "PowerShell dynamic execution",
+        severity: "high",
+        description: "Detects Invoke-Expression/iex, -EncodedCommand payloads and base64 decoding in PowerShell",
+        pattern: /\bInvoke-Expression\b|\|\s*iex\b|\biex\s*\(\s*(?:\(|\$|\[|['"]|New-Object\b|Invoke-|i(?:wr|rm)\b|Get-Content\b)|\s-(?:EncodedCommand|enc|ec)\s+[A-Za-z0-9+\/=]{20,}|\[(?:System\.)?Convert\]::FromBase64String/gi,
+        category: "command-injection",
+    },
+    {
+        // CLI-only: the PowerShell twin of CI-008. A CI-010 match inside this
+        // span is reported as CI-011 instead, so installer docs in fenced code
+        // (irm bun.sh/install.ps1 | iex) get the same downgrade as curl | sh.
+        // The download target must be literal: a $variable URL stays CI-010.
+        id: "CI-011",
+        name: "PowerShell download-and-execute",
+        severity: "critical",
+        description: "Detects irm/iwr output piped to iex, or iex of a downloaded script (PowerShell installer one-liners)",
+        pattern: /\b(?:irm|iwr|Invoke-RestMethod|Invoke-WebRequest)\b[^|\n$]{1,400}\|\s*(?:iex|Invoke-Expression)\b|\b(?:iex|Invoke-Expression)\s*\(?\s*\(\s*(?:(?:irm|iwr|Invoke-RestMethod|Invoke-WebRequest)\b[^()\n$]{1,400}|New-Object\s+(?:System\.)?Net\.WebClient\s*\)\s*\.DownloadString\s*\(\s*['"]https?:\/\/[^()\n$'"]{1,400}['"]\s*)\)/gi,
         category: "command-injection",
     },
     // --- Data Exfiltration (8-12) --------------------------------------------
@@ -99,7 +151,10 @@ export const SCAN_PATTERNS = [
         name: "DNS exfiltration pattern",
         severity: "medium",
         description: "Detects encoding data into DNS lookups",
-        pattern: /dns\.(?:resolve|lookup)\s*\(.*(?:\+|concat|\$\{)/g,
+        // Scans like this one stop at the next call of the same function, so a
+        // repeated prefix cannot make a long line quadratic, and no length bound
+        // lets padding hide the rest of the call.
+        pattern: /dns\.(?:resolve|lookup)\s*\((?:(?!dns\.(?:resolve|lookup)\b)[^\n])*?(?:\+|concat|\$\{)/g,
         category: "data-exfiltration",
     },
     {
@@ -107,7 +162,7 @@ export const SCAN_PATTERNS = [
         name: "Base64 encode and send",
         severity: "medium",
         description: "Detects base64 encoding followed by network send (exfil pattern)",
-        pattern: /btoa\s*\(|Buffer\.from\([^)]*\)\.toString\s*\(\s*['"]base64['"]/g,
+        pattern: /btoa\s*\(|Buffer\.from\((?:(?!Buffer\.from\()[^)\n])*\)\.toString\s*\(\s*['"]base64['"]/g,
         category: "data-exfiltration",
     },
     // --- Privilege Escalation (13-17) ----------------------------------------
@@ -157,7 +212,7 @@ export const SCAN_PATTERNS = [
         name: "Read .env file",
         severity: "critical",
         description: "Detects reading .env files which contain secrets",
-        pattern: /readFile(?:Sync)?\s*\([^)]*\.env\b/g,
+        pattern: /readFile(?:Sync)?\s*\((?:(?!readFile)[^)\n])*?\.env\b/g,
         category: "credential-theft",
     },
     {
@@ -189,7 +244,7 @@ export const SCAN_PATTERNS = [
         name: "Secrets in environment",
         severity: "high",
         description: "Detects broad process.env access patterns",
-        pattern: /process\.env\[(?:[^'"\]]*\+|[a-zA-Z_]\w*\])/g,
+        pattern: /process\.env\[(?:(?:(?!process\.env\[)[^'"\]\n])*\+|[a-zA-Z_]\w*\])/g,
         category: "credential-theft",
     },
     {
@@ -247,15 +302,15 @@ export const SCAN_PATTERNS = [
         name: "Write to system paths",
         severity: "critical",
         description: "Detects writing to system-critical paths",
-        pattern: /writeFile(?:Sync)?\s*\([^)]*(?:\/etc\/|\/usr\/|\/var\/|\/System\/|C:\\Windows)/g,
+        pattern: /writeFile(?:Sync)?\s*\((?:(?!writeFile)[^)\n])*?(?:\/etc\/|\/usr\/|\/var\/|\/System\/|C:\\Windows)/g,
         category: "filesystem-access",
     },
     {
         id: "FS-003",
         name: "Path traversal",
         severity: "high",
-        description: "Detects directory traversal attempts",
-        pattern: /\.\.[/\\]\.\.[/\\]|\.\.(?:[/\\]){2,}/g,
+        description: "Detects directory traversal in filesystem API call arguments",
+        pattern: new RegExp(String.raw `${FS_CALL}(?:(?!${FS_CALL})[^)\n])*\.\.[/\\]\.\.[/\\]`, "g"),
         category: "filesystem-access",
     },
     {
@@ -280,7 +335,10 @@ export const SCAN_PATTERNS = [
         name: "Reverse shell pattern",
         severity: "critical",
         description: "Detects reverse shell connection patterns",
-        pattern: /(?:\/dev\/tcp\/|nc\s+-[a-z]*e|ncat\s|bash\s+-i\s+>&|mkfifo|0<&\d+)/g,
+        // A flag or shell pipe after the host only counts when netcat is the
+        // command itself, and the argument run stops at the first | ; or &, so
+        // `nc -z db 5432 && psql -c ...` and prose about `wc -c` stay clean.
+        pattern: new RegExp(String.raw `(?:\/dev\/tcp\/|\b(?:nc|netcat)\s+-[a-z]*e|\bncat\b\s+-|${NETCAT_COMMAND}[^\n|;&]{0,160}?\s(?:-e|-c|--exec|--sh-exec)\b|${NETCAT_COMMAND}[^\n|;&]{0,160}\|\s*(?:(?:\/\w+)*\/)?(?:ba|z|da|k)?sh\b|bash\s+-i\s+>&|mkfifo\s+\S|0<&\d+)`, "g"),
         category: "network-access",
     },
     {
@@ -291,6 +349,23 @@ export const SCAN_PATTERNS = [
         pattern: /(?:http|https):\/\/\$\{|(?:http|https):\/\/['"]\s*\+/g,
         category: "network-access",
     },
+    {
+        id: "NA-004",
+        name: "Fake prerequisite download",
+        severity: "high",
+        description: "Detects instructions to download and run an external binary or script as a setup step (ClawHavoc-style)",
+        pattern: new RegExp(String.raw `\b(?:before (?:using|running|installing)|prerequisites?|required?|first)\b[^.\n]{0,120}?\b(?:download|install|run|execute)\b[^\n]{0,160}?(?:https?:\/\/(?!${OFFICIAL_INSTALLER_URL_SOURCE_NO_DOT_SEGMENTS})[^\/\s)'"\x60]{1,253}\/[^\s)'"\x60]{0,400}?\.(?:zip|exe|dmg|pkg|msi|sh|ps1|bat|cmd|command|app|scr)\b|\b(?:glot\.io|pastebin\.com|rentry\.co|paste\.ee|transfer\.sh|hastebin\.com)\/)`, "gi"),
+        category: "network-access",
+    },
+    {
+        id: "NA-005",
+        name: "Password-protected archive",
+        severity: "high",
+        description: "Detects password-protected archives, a common way to hide payloads from scanners",
+        // No /i flag: unzip -P is the password but -p pipes to stdout.
+        pattern: /\b(?:[Zz][Ii][Pp]|[Aa][Rr][Cc][Hh][Ii][Vv][Ee]|[Rr][Aa][Rr]|7[Zz])\b[^\n]{0,80}?\b[Pp][Aa][Ss][Ss](?:[Ww][Oo][Rr][Dd]|[Cc][Oo][Dd][Ee])\s*(?:[Ii][Ss]\s*)?[:=]\s*\S+|\bunzip\b[^|&;\n]{0,60}?\s-[a-zA-Z]*P\s*\S+|\b(?:7z|7za|7zz|unrar)\s+[ex]\b[^|&;\n]{0,60}?\s-[pP]\S+/g,
+        category: "network-access",
+    },
     // --- Code Execution (35-37) ----------------------------------------------
     {
         id: "CE-001",
@@ -316,15 +391,21 @@ export const SCAN_PATTERNS = [
         pattern: /import\s*\(\s*(?:`[^`]*\$\{|[a-zA-Z_]\w*\s*[\+)])/g,
         category: "code-execution",
     },
+    {
+        id: "CE-004",
+        name: "Python dynamic execution or deserialization",
+        severity: "high",
+        description: "Detects pickle/marshal loads and exec() of decoded or compiled code",
+        pattern: /\b(?:pickle|cPickle|_pickle|marshal|dill)\.loads?\s*\(|\bexec\s*\(\s*(?:compile|base64|codecs|zlib|bytes\.fromhex|marshal|__import__)\b|\b__import__\s*\(\s*['"](?:os|subprocess|socket|pty)['"]\s*\)/g,
+        category: "code-execution",
+    },
     // --- DCI Block Abuse (38-51) -----------------------------------------------
-    // DCI blocks are shell commands in SKILL.md executed via ! prefix.
-    // These patterns detect malicious use within DCI contexts.
     {
         id: "DCI-001",
         name: "DCI credential file read",
         severity: "critical",
         description: "DCI block reads credential files (~/.ssh/, ~/.aws/, .env)",
-        pattern: /^\s*!\s*`[^`]*(?:~\/\.ssh\/|~\/\.aws\/|\.env\b|\.gnupg\/)/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*(?:~\/\.ssh\/|~\/\.aws\/|\.env\b|\.gnupg\/)/gm,
         category: "dci-abuse",
     },
     {
@@ -332,7 +413,7 @@ export const SCAN_PATTERNS = [
         name: "DCI network exfiltration",
         severity: "critical",
         description: "DCI block uses curl/wget for network access",
-        pattern: /^\s*!\s*`[^`]*\b(?:curl|wget)\b/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*\b(?:curl|wget)\b/gm,
         category: "dci-abuse",
     },
     {
@@ -340,7 +421,9 @@ export const SCAN_PATTERNS = [
         name: "DCI fetch/nc network call",
         severity: "critical",
         description: "DCI block uses fetch or netcat for network access",
-        pattern: /^\s*!\s*`[^`]*\b(?:fetch|nc|ncat|netcat)\b/gm,
+        // `git fetch` and other package-manager fetch subcommands are not network
+        // calls the skill makes. The lookbehind runs only where `fetch` matched.
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*(?:\bfetch\b(?<!\b(?:git|npm|pnpm|yarn)\s+fetch)|\b(?:nc|ncat|netcat)\b)/gm,
         category: "dci-abuse",
     },
     {
@@ -348,7 +431,9 @@ export const SCAN_PATTERNS = [
         name: "DCI agent config write",
         severity: "critical",
         description: "DCI block writes to agent config files (CLAUDE.md, AGENTS.md, .claude/)",
-        pattern: /^\s*!\s*`[^`]*(?:>\s*.*(?:CLAUDE\.md|AGENTS\.md|\.claude\/|\.specweave\/))/gm,
+        // The prefix cannot pass a `>`, so only the first redirect is tried: a
+        // later target also follows it, and the line is scanned once.
+        pattern: /^[^\S\n]*![^\S\n]*`[^`>\n]*>[^`\n]*(?:CLAUDE\.md|AGENTS\.md|\.claude\/|\.specweave\/)/gm,
         category: "dci-abuse",
     },
     {
@@ -356,7 +441,9 @@ export const SCAN_PATTERNS = [
         name: "DCI agent config modify",
         severity: "critical",
         description: "DCI block modifies agent config via tee/sed/echo append",
-        pattern: /^\s*!\s*`[^`]*(?:tee|sed\s+-i|echo\s+.*>>)\s*.*(?:CLAUDE\.md|AGENTS\.md|\.claude\/)/gm,
+        // Each tee/sed -i scan stops at the next one, each echo scan at the next
+        // echo, and each `>>` scan at the next `>>`: linear with no length bound.
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*?(?:(?:\btee\b|\bsed\s+-i)(?:(?!\btee\b|\bsed\s+-i)[^`\n])*|\becho\s(?:(?!\becho\s)[^`\n])*?>>(?:(?!>>)[^`\n])*)(?:CLAUDE\.md|AGENTS\.md|\.claude\/)/gm,
         category: "dci-abuse",
     },
     {
@@ -364,7 +451,7 @@ export const SCAN_PATTERNS = [
         name: "DCI base64 decode",
         severity: "critical",
         description: "DCI block contains base64 decoding (obfuscation)",
-        pattern: /^\s*!\s*`[^`]*\b(?:base64\s+(?:-[dD]|--decode)|atob\s*\()/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*\b(?:base64\s+(?:-[dD]|--decode)|atob\s*\()/gm,
         category: "dci-abuse",
     },
     {
@@ -372,7 +459,7 @@ export const SCAN_PATTERNS = [
         name: "DCI hex escape obfuscation",
         severity: "critical",
         description: "DCI block contains hex escape sequences (obfuscation)",
-        pattern: /^\s*!\s*`[^`]*\\x[0-9a-fA-F]{2}(?:\\x[0-9a-fA-F]{2}){3,}/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*\\x[0-9a-fA-F]{2}(?:\\x[0-9a-fA-F]{2}){3,}/gm,
         category: "dci-abuse",
     },
     {
@@ -380,7 +467,7 @@ export const SCAN_PATTERNS = [
         name: "DCI eval execution",
         severity: "critical",
         description: "DCI block uses eval for code execution",
-        pattern: /^\s*!\s*`[^`]*\beval\b/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*\beval\b/gm,
         category: "dci-abuse",
     },
     {
@@ -388,7 +475,9 @@ export const SCAN_PATTERNS = [
         name: "DCI download and execute",
         severity: "critical",
         description: "DCI block pipes downloaded content to shell (download-and-execute)",
-        pattern: /^\s*!\s*`[^`]*\b(?:curl|wget)\b[^`]*\|\s*(?:ba|z|da|k)?sh\b/gm,
+        // The prefix stops at the first curl/wget; a later shell pipe also
+        // follows it, so one pass decides the line.
+        pattern: /^[^\S\n]*![^\S\n]*`(?:(?!\b(?:curl|wget)\b)[^`\n])*\b(?:curl|wget)\b[^`\n]*\|\s*(?:ba|z|da|k)?sh\b/gm,
         category: "dci-abuse",
     },
     {
@@ -396,7 +485,7 @@ export const SCAN_PATTERNS = [
         name: "DCI reverse shell",
         severity: "critical",
         description: "DCI block establishes a reverse shell connection",
-        pattern: /^\s*!\s*`[^`]*(?:\/dev\/tcp\/|bash\s+-i\s+>&|mkfifo|nc\s+-[a-z]*e)/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*(?:\/dev\/tcp\/|bash\s+-i\s+>&|mkfifo|nc\s+-[a-z]*e)/gm,
         category: "dci-abuse",
     },
     {
@@ -404,7 +493,7 @@ export const SCAN_PATTERNS = [
         name: "DCI sudo escalation",
         severity: "critical",
         description: "DCI block uses sudo for privilege escalation",
-        pattern: /^\s*!\s*`[^`]*\bsudo\b/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*\bsudo\b/gm,
         category: "dci-abuse",
     },
     {
@@ -412,7 +501,7 @@ export const SCAN_PATTERNS = [
         name: "DCI rm destructive command",
         severity: "critical",
         description: "DCI block executes destructive rm -rf command",
-        pattern: /^\s*!\s*`[^`]*\brm\s+-[a-zA-Z]*r[a-zA-Z]*f/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`[^`\n]*\brm\s+-[a-zA-Z]*r[a-zA-Z]*f/gm,
         category: "dci-abuse",
     },
     {
@@ -420,7 +509,9 @@ export const SCAN_PATTERNS = [
         name: "DCI home dir exfiltration",
         severity: "critical",
         description: "DCI block reads from home directory sensitive paths",
-        pattern: /^\s*!\s*`[^`]*(?:cat|less|head|tail|strings)\s+[^`]*(?:~\/\.|\/home\/[^`]*\.)/gm,
+        // Independent lookaheads with no length bound, so padding cannot push the
+        // path out of reach; each /home/ scan stops at the next /home/.
+        pattern: /^[^\S\n]*![^\S\n]*`(?=[^`\n]*\b(?:cat|less|head|tail|strings)\s)(?=[^`\n]*(?:~\/\.|\/home\/(?:(?!\/home\/)[^`\s])*\.))/gm,
         category: "dci-abuse",
     },
     {
@@ -428,7 +519,7 @@ export const SCAN_PATTERNS = [
         name: "DCI data pipe to network",
         severity: "critical",
         description: "DCI block pipes local data to a network command",
-        pattern: /^\s*!\s*`[^`]*(?:cat|tar|zip)\s+[^|`]*\|\s*(?:curl|wget|nc)\b/gm,
+        pattern: /^[^\S\n]*![^\S\n]*`(?=[^`\n]*\b(?:cat|tar|zip)\s)(?=[^`\n]*\|\s*(?:curl|wget|nc)\b)/gm,
         category: "dci-abuse",
     },
 ];
@@ -437,7 +528,7 @@ export const SCAN_PATTERNS = [
 // Suppress DCI-abuse findings when the line matches this pattern AND
 // no malicious DCI pattern also matches (prevents appended-command bypass).
 const SAFE_DCI_PATTERNS = [
-    /^\s*!\s*`for\s+d\s+in\s+\.specweave\/skill-memories/,
+    /^[^\S\n]*![^\S\n]*`for\s+d\s+in\s+\.specweave\/skill-memories/,
 ];
 /** All malicious DCI patterns extracted for the two-pass safe check */
 const MALICIOUS_DCI_PATTERNS = SCAN_PATTERNS
@@ -547,71 +638,46 @@ function computeFencedCodeBlockLines(lines) {
     }
     return fenced;
 }
-// ---- HTML comment line detection -------------------------------------------
-// Pre-compute which lines are inside HTML comments (<!-- ... -->).
-// Findings inside HTML comments are suppressed entirely.
-function computeHtmlCommentLines(lines) {
-    const commented = new Set();
-    let inComment = false;
-    for (let i = 0; i < lines.length; i++) {
-        const line = lines[i];
-        if (inComment) {
-            if (line.includes("-->")) {
-                // Closing line — don't mark it (content may follow -->)
-                inComment = false;
-            }
-            else {
-                // Interior line — fully inside comment
-                commented.add(i);
-            }
-        }
-        else if (line.includes("<!--") && !line.includes("-->")) {
-            // Multi-line comment starts — don't mark opening line
-            // (content may precede <!-- or follow on same line)
-            inComment = true;
-        }
-        // Single-line <!-- ... --> comments are NOT marked to prevent
-        // bypass via: <!-- --> eval("malicious")
-    }
-    return commented;
-}
 // ---- Documentation-safe patterns -------------------------------------------
 // These patterns commonly fire on installation instructions and code examples
 // inside fenced code blocks. Downgrade to "info" when in documentation context.
-const DOCUMENTATION_SAFE_PATTERNS = new Set(["PE-001", "PE-002", "PE-003", "FS-001", "CI-008", "CT-002", "CT-004", "NA-001", "PI-003"]);
-// ---- Markdown-link safe context for FS-003 ---------------------------------
-/**
- * Returns true if the match at `matchIndex` falls inside a markdown link
- * target, i.e. between `](` and `)`.  Standard relative links like
- * `[text](../../tools/file.md)` are not path-traversal attacks.
- */
-function isInsideMarkdownLink(line, matchIndex) {
-    const before = line.slice(0, matchIndex);
-    const linkOpen = before.lastIndexOf("](");
-    if (linkOpen === -1)
-        return false;
-    // Ensure the link isn't already closed before our match
-    if (before.slice(linkOpen + 2).includes(")"))
-        return false;
-    // Ensure there's a closing ) after the match
-    return line.indexOf(")", matchIndex) !== -1;
+// NA-004/NA-005 are deliberately absent: fake-prerequisite and password-archive
+// lures are written as prose, so documentation context must not hide them.
+const DOCUMENTATION_SAFE_PATTERNS = new Set([
+    "PE-001", "PE-002", "PE-003", "FS-001", "CI-008", "CI-009", "CI-010", "CI-011",
+    "CE-004", "CT-002", "CT-004", "NA-001", "PI-003",
+]);
+// CI-010 forms that run hidden or computed code keep their severity even in a
+// fenced example: an -EncodedCommand payload, base64 decoded on a line that
+// also executes, and Invoke-Expression/iex of a variable or a download.
+const CI010_ENCODED_RE = /^\s-(?:EncodedCommand|enc|ec)\s/i;
+const CI010_EXEC_SINK_RE = /\b(?:iex|Invoke-Expression|Invoke-Command|Start-Process|powershell|pwsh)\b|\[ScriptBlock\]::Create|\.Invoke\s*\(|&\s*\(/i;
+const CI010_COMPUTED_RE = /\$[\w{(:]|DownloadString|FromBase64String|GetString|\b(?:irm|iwr|Invoke-RestMethod|Invoke-WebRequest)\b/i;
+function ci010KeepsSeverity(match, line) {
+    if (CI010_ENCODED_RE.test(match))
+        return true;
+    if (/FromBase64String/i.test(match))
+        return CI010_EXEC_SINK_RE.test(line);
+    return CI010_COMPUTED_RE.test(line);
 }
 // ---- Markdown inline code detection ----------------------------------------
 // Detect if a regex match falls inside markdown inline code (backticks).
 // Inline code in SKILL.md is referential — it names functions/commands, not
 // invokes them. Example: `eval()` in "Watch for: `eval()`, `exec()`".
 /**
- * Returns true if the match at [matchStart, matchEnd) falls inside a
- * markdown inline code span (single or double backtick delimiters).
- * Fenced code block markers (```) are ignored — handled separately.
+ * The [contentStart, contentEnd) ranges of markdown inline code spans (single
+ * or double backtick delimiters) on a line, in order. Fenced code block
+ * markers (```) are ignored — handled separately. Computed once per line so
+ * many matches on one long line stay linear.
  */
-function isInsideInlineCode(line, matchStart, matchEnd) {
+function inlineCodeSpans(line) {
+    const spans = [];
     let i = 0;
     while (i < line.length) {
-        if (line[i] !== "`") {
-            i++;
-            continue;
-        }
+        const tick = line.indexOf("`", i);
+        if (tick < 0)
+            break;
+        i = tick;
         // Count opening backticks
         let openLen = 0;
         while (i + openLen < line.length && line[i + openLen] === "`")
@@ -629,12 +695,91 @@ function isInsideInlineCode(line, matchStart, matchEnd) {
             i = contentStart;
             continue;
         }
-        // Check if match falls within this inline code span
-        if (matchStart >= contentStart && matchEnd <= closeIdx)
-            return true;
+        spans.push([contentStart, closeIdx]);
         i = closeIdx + openLen;
     }
-    return false;
+    return spans;
+}
+// ---- Finding context ---------------------------------------------------------
+// Each context line is cut to a window around the match column, so a match on
+// a huge line costs O(window), not O(line).
+const CONTEXT_WINDOW = 300;
+function contextWindow(text, column) {
+    if (text.length <= CONTEXT_WINDOW)
+        return text;
+    let start = Math.max(0, Math.min(column - CONTEXT_WINDOW / 2, text.length - CONTEXT_WINDOW));
+    let end = start + CONTEXT_WINDOW;
+    // Never split a surrogate pair.
+    const high = (c) => c >= 0xd800 && c <= 0xdbff;
+    if (start > 0 && high(text.charCodeAt(start - 1)))
+        start++;
+    if (end < text.length && high(text.charCodeAt(end - 1)))
+        end--;
+    return text.slice(start, end);
+}
+// ---- CI-010 vs CI-011 -------------------------------------------------------
+// CI-011 (PowerShell download-and-execute) is a more specific form of CI-010.
+// A CI-010 match inside a CI-011 span is dropped so the installer one-liner is
+// reported once, under the id that gets the documentation-context downgrade.
+/** Longest rescanned-payload match kept on a finding. */
+const MAX_PAYLOAD_MATCH = 200;
+const INSTALLER_SOURCE = SCAN_PATTERNS.find((p) => p.id === "CI-011").pattern;
+function installerSpans(text) {
+    const spans = [];
+    const re = new RegExp(INSTALLER_SOURCE.source, INSTALLER_SOURCE.flags);
+    let m;
+    while ((m = re.exec(text)) !== null) {
+        if (m[0].length === 0) {
+            re.lastIndex++;
+            continue;
+        }
+        spans.push([m.index, m.index + m[0].length]);
+    }
+    return spans;
+}
+/** True if [start, end) lies inside one of the sorted, non-overlapping spans. */
+function insideSpan(spans, start, end) {
+    let lo = 0;
+    let hi = spans.length - 1;
+    while (lo <= hi) {
+        const mid = (lo + hi) >> 1;
+        if (spans[mid][0] <= start)
+            lo = mid + 1;
+        else
+            hi = mid - 1;
+    }
+    return hi >= 0 && end <= spans[hi][1];
+}
+// ---- Finding caps -------------------------------------------------------------
+// A 2MB file of one repeated match must not become hundreds of thousands of
+// findings. Past the caps a finding folds into the latest one of the same
+// pattern and severity; one is still listed when no such finding exists, so
+// a new severity is never lost.
+const MAX_FINDINGS_PER_PATTERN = 100;
+const MAX_FINDINGS = 2000;
+// Decoded payloads are rescanned up to these limits; the rest are reported
+// as one HC-010 marker so the judge knows they were not rescanned.
+const MAX_RESCAN_PAYLOADS = 50;
+const MAX_RESCAN_CHARS = 64 * 1024;
+const UNSCANNED_RULE = HIDDEN_CONTENT_RULES.find((r) => r.id === "HC-010");
+/** Adds findings to `out` under the caps; `add` returns the finding that holds the new one. */
+function cappedFindings(out) {
+    const listed = new Map();
+    const latest = new Map();
+    return (finding) => {
+        const key = `${finding.patternId}:${finding.severity}`;
+        const count = listed.get(finding.patternId) ?? 0;
+        const holder = latest.get(key);
+        if (holder && (count >= MAX_FINDINGS_PER_PATTERN || out.length >= MAX_FINDINGS)) {
+            holder.occurrences = (holder.occurrences ?? 1) + (finding.occurrences ?? 1);
+            holder.foldedFindings = (holder.foldedFindings ?? 0) + 1 + (finding.foldedFindings ?? 0);
+            return holder;
+        }
+        out.push(finding);
+        listed.set(finding.patternId, count + 1);
+        latest.set(key, finding);
+        return finding;
+    };
 }
 // ---- Scanner function -----------------------------------------------------
 /**
@@ -643,67 +788,162 @@ function isInsideInlineCode(line, matchStart, matchEnd) {
 export function scanContent(content) {
     const lines = content.split("\n");
     const fencedLines = computeFencedCodeBlockLines(lines);
-    const commentLines = computeHtmlCommentLines(lines);
     const findings = [];
+    const add = cappedFindings(findings);
+    const installerSpanCache = new Map();
+    const inlineCodeCache = new Map();
+    const safeDciCache = new Map();
+    const rmSeverityCache = new Map();
     for (const pattern of SCAN_PATTERNS) {
         // Reset the regex lastIndex for each pattern (they use /g flag)
         const regex = new RegExp(pattern.pattern.source, pattern.pattern.flags);
         for (let lineIdx = 0; lineIdx < lines.length; lineIdx++) {
             const line = lines[lineIdx];
             let match;
+            // One finding per severity for this pattern on this line; further
+            // matches only raise its occurrence count.
+            let folded = null;
             while ((match = regex.exec(line)) !== null) {
-                // Suppress DCI-abuse findings for known-safe DCI patterns
-                if (pattern.category === "dci-abuse" && isSafeDciBlock(line)) {
+                // Guard against zero-length matches looping forever on a /g regex.
+                if (match[0].length === 0) {
+                    regex.lastIndex++;
                     continue;
                 }
-                // Suppress all findings inside HTML comments
-                if (commentLines.has(lineIdx)) {
-                    continue;
+                // Suppress DCI-abuse findings for known-safe DCI patterns
+                if (pattern.category === "dci-abuse") {
+                    let safe = safeDciCache.get(lineIdx);
+                    if (safe === undefined)
+                        safeDciCache.set(lineIdx, (safe = isSafeDciBlock(line)));
+                    if (safe)
+                        continue;
                 }
-                // Suppress FS-003 path traversal inside markdown links
-                if (pattern.id === "FS-003" && isInsideMarkdownLink(line, match.index)) {
-                    continue;
+                // HTML comments are NOT suppressed: they render invisibly to a human
+                // reviewer but an agent reads them verbatim, so they are where hidden
+                // instructions go. hidden-content.ts adds HC-005 for directive bodies.
+                if (pattern.id === "CI-010") {
+                    let spans = installerSpanCache.get(lineIdx);
+                    if (!spans)
+                        installerSpanCache.set(lineIdx, (spans = installerSpans(line)));
+                    if (insideSpan(spans, match.index, match.index + match[0].length))
+                        continue;
                 }
-                // Build context: up to 1 line before and after
-                const contextLines = [];
-                if (lineIdx > 0)
-                    contextLines.push(lines[lineIdx - 1]);
-                contextLines.push(line);
-                if (lineIdx < lines.length - 1)
-                    contextLines.push(lines[lineIdx + 1]);
                 // Downgrade FS-001 severity based on context:
                 // safe targets → info, system paths → keep high, other → low
                 // Only for `rm` commands — rimraf() keeps base severity ("high")
                 let severity = pattern.severity;
-                if (pattern.id === "FS-001" && /\brm\s+-/.test(line)) {
-                    if (isRmTargetingSafePaths(line)) {
-                        severity = "info";
-                    }
-                    else if (!isRmTargetingSystemPaths(line)) {
-                        severity = "low";
+                if (pattern.id === "FS-001") {
+                    let rmSeverity = rmSeverityCache.get(lineIdx);
+                    if (rmSeverity === undefined) {
+                        rmSeverity = !/\brm\s+-/.test(line)
+                            ? null
+                            : isRmTargetingSafePaths(line)
+                                ? "info"
+                                : isRmTargetingSystemPaths(line)
+                                    ? null
+                                    : "low";
+                        rmSeverityCache.set(lineIdx, rmSeverity);
                     }
+                    if (rmSeverity)
+                        severity = rmSeverity;
                 }
                 // Downgrade documentation-safe patterns inside fenced code blocks
-                if (DOCUMENTATION_SAFE_PATTERNS.has(pattern.id) && fencedLines.has(lineIdx)) {
+                if (DOCUMENTATION_SAFE_PATTERNS.has(pattern.id) &&
+                    fencedLines.has(lineIdx) &&
+                    !(pattern.id === "CI-010" && ci010KeepsSeverity(match[0], line))) {
                     severity = "info";
                 }
                 // Downgrade non-DCI patterns inside markdown inline code to info.
                 // Inline code references (e.g., `eval()`) are educational, not invocations.
-                if (pattern.category !== "dci-abuse" && isInsideInlineCode(line, match.index, match.index + match[0].length)) {
-                    severity = "info";
+                if (pattern.category !== "dci-abuse") {
+                    let spans = inlineCodeCache.get(lineIdx);
+                    if (!spans)
+                        inlineCodeCache.set(lineIdx, (spans = inlineCodeSpans(line)));
+                    if (insideSpan(spans, match.index, match.index + match[0].length))
+                        severity = "info";
                 }
-                findings.push({
+                const prior = folded?.get(severity);
+                if (prior) {
+                    prior.occurrences = (prior.occurrences ?? 1) + 1;
+                    continue;
+                }
+                // Build context: up to 1 line before and after, each windowed
+                // around the match column.
+                const contextLines = [];
+                if (lineIdx > 0)
+                    contextLines.push(contextWindow(lines[lineIdx - 1], match.index));
+                contextLines.push(contextWindow(line, match.index));
+                if (lineIdx < lines.length - 1)
+                    contextLines.push(contextWindow(lines[lineIdx + 1], match.index));
+                const finding = {
                     patternId: pattern.id,
                     patternName: pattern.name,
                     severity,
+                    ...(severity !== pattern.severity ? { downgradedFrom: pattern.severity } : {}),
                     category: pattern.category,
                     match: match[0],
                     lineNumber: lineIdx + 1,
                     context: contextLines.join("\n"),
-                });
+                };
+                // Later matches on this line go to whichever finding now holds it.
+                (folded ??= new Map()).set(severity, add(finding));
+            }
+        }
+    }
+    // Hidden channels: invisible Unicode, comment directives, encoded text.
+    // Decoded payloads are rescanned at full severity: hidden text gets no
+    // documentation-context downgrade.
+    const hidden = scanHiddenContent(content);
+    const extra = [...hidden.findings];
+    let budget = MAX_RESCAN_CHARS;
+    for (const [index, payload] of hidden.payloads.entries()) {
+        if (index >= MAX_RESCAN_PAYLOADS || budget <= 0) {
+            const skipped = hidden.payloads.length - index;
+            extra.push({
+                patternId: UNSCANNED_RULE.id,
+                patternName: UNSCANNED_RULE.name,
+                severity: UNSCANNED_RULE.severity,
+                category: "prompt-injection",
+                match: `[vskill: ${skipped} decoded payload${skipped === 1 ? "" : "s"} not rescanned]`,
+                lineNumber: payload.lineNumber,
+                context: "",
+            });
+            break;
+        }
+        // The whole decoded text is rescanned (within the budget), not the
+        // display copy.
+        const text = payload.text.length > budget ? payload.text.slice(0, budget) : payload.text;
+        budget -= text.length;
+        const spans = installerSpans(text);
+        for (const pattern of SCAN_PATTERNS) {
+            const regex = new RegExp(pattern.pattern.source, pattern.pattern.flags);
+            let match;
+            let first = null;
+            while ((match = regex.exec(text)) !== null) {
+                if (match[0].length === 0) {
+                    regex.lastIndex++;
+                    continue;
+                }
+                if (pattern.id === "CI-010" && insideSpan(spans, match.index, match.index + match[0].length))
+                    continue;
+                if (first) {
+                    first.occurrences = (first.occurrences ?? 1) + 1;
+                    continue;
+                }
+                first = {
+                    patternId: pattern.id,
+                    patternName: pattern.name,
+                    severity: pattern.severity,
+                    category: pattern.category,
+                    match: `[hidden ${payload.kind}] ${match[0].slice(0, MAX_PAYLOAD_MATCH)}`,
+                    lineNumber: payload.lineNumber,
+                    context: contextWindow(text, match.index),
+                };
+                extra.push(first);
             }
         }
     }
+    for (const finding of extra)
+        add(finding);
     return findings;
 }
 //# sourceMappingURL=patterns.js.map