mono/packages/kbot/ref/tools/PowerShellTool/modeValidation.ts
2026-04-01 01:05:48 +02:00

405 lines
17 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

/**
* PowerShell permission mode validation.
*
* Checks if commands should be auto-allowed based on the current permission mode.
* In acceptEdits mode, filesystem-modifying PowerShell cmdlets are auto-allowed.
* Follows the same patterns as BashTool/modeValidation.ts.
*/
import type { ToolPermissionContext } from '../../Tool.js'
import type { PermissionResult } from '../../utils/permissions/PermissionResult.js'
import type { ParsedPowerShellCommand } from '../../utils/powershell/parser.js'
import {
deriveSecurityFlags,
getPipelineSegments,
PS_TOKENIZER_DASH_CHARS,
} from '../../utils/powershell/parser.js'
import {
argLeaksValue,
isAllowlistedPipelineTail,
isCwdChangingCmdlet,
isSafeOutputCommand,
resolveToCanonical,
} from './readOnlyValidation.js'
/**
* Filesystem-modifying cmdlets that are auto-allowed in acceptEdits mode.
* Stored as canonical (lowercase) cmdlet names.
*
* Tier 3 cmdlets with complex parameter binding removed — they fall through to
* 'ask'. Only simple write cmdlets (first positional = -Path) are auto-allowed
* here, and they get path validation via CMDLET_PATH_CONFIG in pathValidation.ts.
*/
const ACCEPT_EDITS_ALLOWED_CMDLETS = new Set([
'set-content',
'add-content',
'remove-item',
'clear-content',
])
function isAcceptEditsAllowedCmdlet(name: string): boolean {
// resolveToCanonical handles aliases via COMMON_ALIASES, so e.g. 'rm' → 'remove-item',
// 'ac' → 'add-content'. Any alias that resolves to an allowed cmdlet is automatically
// allowed. Tier 3 cmdlets (new-item, copy-item, move-item, etc.) and their aliases
// (mkdir, ni, cp, mv, etc.) resolve to cmdlets NOT in the set and fall through to 'ask'.
const canonical = resolveToCanonical(name)
return ACCEPT_EDITS_ALLOWED_CMDLETS.has(canonical)
}
/**
* New-Item -ItemType values that create filesystem links (reparse points or
* hard links). All three redirect path resolution at runtime — symbolic links
* and junctions are directory/file reparse points; hard links alias a file's
* inode. Any of these let a later relative-path write land outside the
* validator's view.
*/
const LINK_ITEM_TYPES = new Set(['symboliclink', 'junction', 'hardlink'])
/**
* Check if a lowered, dash-normalized arg (colon-value stripped) is an
* unambiguous PowerShell abbreviation of New-Item's -ItemType or -Type param.
* Min prefixes: `-it` (avoids ambiguity with other New-Item params), `-ty`
* (avoids `-t` colliding with `-Target`).
*/
function isItemTypeParamAbbrev(p: string): boolean {
return (
(p.length >= 3 && '-itemtype'.startsWith(p)) ||
(p.length >= 3 && '-type'.startsWith(p))
)
}
/**
* Detects New-Item creating a filesystem link (-ItemType SymbolicLink /
* Junction / HardLink, or the -Type alias). Links poison subsequent path
* resolution the same way Set-Location/New-PSDrive do: a relative path
* through the link resolves to the link target, not the validator's view.
* Finding #18.
*
* Handles PS parameter abbreviation (`-it`, `-ite`, ... `-itemtype`; `-ty`,
* `-typ`, `-type`), unicode dash prefixes (en-dash/em-dash/horizontal-bar),
* and colon-bound values (`-it:Junction`).
*/
export function isSymlinkCreatingCommand(cmd: {
name: string
args: string[]
}): boolean {
const canonical = resolveToCanonical(cmd.name)
if (canonical !== 'new-item') return false
for (let i = 0; i < cmd.args.length; i++) {
const raw = cmd.args[i] ?? ''
if (raw.length === 0) continue
// Normalize unicode dash prefixes (, —, ―) and forward-slash (PS 5.1
// parameter prefix) → ASCII `-` so prefix comparison works. PS tokenizer
// treats all four dash chars plus `/` as parameter markers. (bug #26)
const normalized =
PS_TOKENIZER_DASH_CHARS.has(raw[0]!) || raw[0] === '/'
? '-' + raw.slice(1)
: raw
const lower = normalized.toLowerCase()
// Split colon-bound value: -it:SymbolicLink → param='-it', val='symboliclink'
const colonIdx = lower.indexOf(':', 1)
const paramRaw = colonIdx > 0 ? lower.slice(0, colonIdx) : lower
// Strip backtick escapes: -Item`Type → -ItemType (bug #22)
const param = paramRaw.replace(/`/g, '')
if (!isItemTypeParamAbbrev(param)) continue
const rawVal =
colonIdx > 0
? lower.slice(colonIdx + 1)
: (cmd.args[i + 1]?.toLowerCase() ?? '')
// Strip backtick escapes from colon-bound value: -it:Sym`bolicLink → symboliclink
// Mirrors the param-name strip at L103. Space-separated args use .value
// (backtick-resolved by .NET parser), but colon-bound uses .text (raw source).
// Strip surrounding quotes: -it:'SymbolicLink' or -it:"Junction" (bug #6)
const val = rawVal.replace(/`/g, '').replace(/^['"]|['"]$/g, '')
if (LINK_ITEM_TYPES.has(val)) return true
}
return false
}
/**
* Checks if commands should be handled differently based on the current permission mode.
*
* In acceptEdits mode, auto-allows filesystem-modifying PowerShell cmdlets.
* Uses the AST to resolve aliases before checking the allowlist.
*
* @param input - The PowerShell command input
* @param parsed - The parsed AST of the command
* @param toolPermissionContext - Context containing mode and permissions
* @returns
* - 'allow' if the current mode permits auto-approval
* - 'passthrough' if no mode-specific handling applies
*/
export function checkPermissionMode(
input: { command: string },
parsed: ParsedPowerShellCommand,
toolPermissionContext: ToolPermissionContext,
): PermissionResult {
// Skip bypass and dontAsk modes (handled elsewhere)
if (
toolPermissionContext.mode === 'bypassPermissions' ||
toolPermissionContext.mode === 'dontAsk'
) {
return {
behavior: 'passthrough',
message: 'Mode is handled in main permission flow',
}
}
if (toolPermissionContext.mode !== 'acceptEdits') {
return {
behavior: 'passthrough',
message: 'No mode-specific validation required',
}
}
// acceptEdits mode: check if all commands are filesystem-modifying cmdlets
if (!parsed.valid) {
return {
behavior: 'passthrough',
message: 'Cannot validate mode for unparsed command',
}
}
// SECURITY: Check for subexpressions, script blocks, or member invocations
// that could be used to smuggle arbitrary code through acceptEdits mode.
const securityFlags = deriveSecurityFlags(parsed)
if (
securityFlags.hasSubExpressions ||
securityFlags.hasScriptBlocks ||
securityFlags.hasMemberInvocations ||
securityFlags.hasSplatting ||
securityFlags.hasAssignments ||
securityFlags.hasStopParsing ||
securityFlags.hasExpandableStrings
) {
return {
behavior: 'passthrough',
message:
'Command contains subexpressions, script blocks, or member invocations that require approval',
}
}
const segments = getPipelineSegments(parsed)
// SECURITY: Empty segments with valid parse = no commands to check, don't auto-allow
if (segments.length === 0) {
return {
behavior: 'passthrough',
message: 'No commands found to validate for acceptEdits mode',
}
}
// SECURITY: Compound cwd desync guard — BashTool parity.
// When any statement in a compound contains Set-Location/Push-Location/Pop-Location
// (or aliases like cd, sl, chdir, pushd, popd), the cwd changes between statements.
// Path validation resolves relative paths against the stale process cwd, so a write
// cmdlet in a later statement targets a different directory than the validator checked.
// Example: `Set-Location ./.claude; Set-Content ./settings.json '...'` — the validator
// sees ./settings.json as /project/settings.json, but PowerShell writes to
// /project/.claude/settings.json. Refuse to auto-allow any write operation in a
// compound that contains a cwd-changing command. This matches BashTool's
// compoundCommandHasCd guard (BashTool/pathValidation.ts:630-655).
const totalCommands = segments.reduce(
(sum, seg) => sum + seg.commands.length,
0,
)
if (totalCommands > 1) {
let hasCdCommand = false
let hasSymlinkCreate = false
let hasWriteCommand = false
for (const seg of segments) {
for (const cmd of seg.commands) {
if (cmd.elementType !== 'CommandAst') continue
if (isCwdChangingCmdlet(cmd.name)) hasCdCommand = true
if (isSymlinkCreatingCommand(cmd)) hasSymlinkCreate = true
if (isAcceptEditsAllowedCmdlet(cmd.name)) hasWriteCommand = true
}
}
if (hasCdCommand && hasWriteCommand) {
return {
behavior: 'passthrough',
message:
'Compound command contains a directory-changing command (Set-Location/Push-Location/Pop-Location) with a write operation — cannot auto-allow because path validation uses stale cwd',
}
}
// SECURITY: Link-create compound guard (finding #18). Mirrors the cd
// guard above. `New-Item -ItemType SymbolicLink -Path ./link -Value /etc;
// Get-Content ./link/passwd` — path validation resolves ./link/passwd
// against cwd (no link there at validation time), but runtime follows
// the just-created link to /etc/passwd. Same TOCTOU shape as cwd desync.
// Applies to SymbolicLink, Junction, and HardLink — all three redirect
// path resolution at runtime.
// No `hasWriteCommand` requirement: read-through-symlink is equally
// dangerous (exfil via Get-Content ./link/etc/shadow), and any other
// command using paths after a just-created link is unvalidatable.
if (hasSymlinkCreate) {
return {
behavior: 'passthrough',
message:
'Compound command creates a filesystem link (New-Item -ItemType SymbolicLink/Junction/HardLink) — cannot auto-allow because path validation cannot follow just-created links',
}
}
}
for (const segment of segments) {
for (const cmd of segment.commands) {
if (cmd.elementType !== 'CommandAst') {
// SECURITY: This guard is load-bearing for THREE cases. Do not narrow it.
//
// 1. Expression pipeline sources (designed): '/etc/passwd' | Remove-Item
// — the string literal is CommandExpressionAst, piped value binds to
// -Path. We cannot statically know what path it represents.
//
// 2. Control-flow statements (accidental but relied upon):
// foreach ($x in ...) { Remove-Item $x }. Non-PipelineAst statements
// produce a synthetic CommandExpressionAst entry in segment.commands
// (parser.ts transformStatement). Without this guard, Remove-Item $x
// in nestedCommands would be checked below and auto-allowed — but $x
// is a loop-bound variable we cannot validate.
//
// 3. Non-PipelineAst redirection coverage (accidental): cmd && cmd2 > /tmp
// also produces a synthetic element here. isReadOnlyCommand relies on
// the same accident (its allowlist rejects the synthetic element's
// full-text name), so both paths fail safe together.
return {
behavior: 'passthrough',
message: `Pipeline contains expression source (${cmd.elementType}) that cannot be statically validated`,
}
}
// SECURITY: nameType is computed from the raw name before stripModulePrefix.
// 'application' = raw name had path chars (. \\ /). scripts\\Remove-Item
// strips to Remove-Item and would match ACCEPT_EDITS_ALLOWED_CMDLETS below,
// but PowerShell runs scripts\\Remove-Item.ps1. Same gate as isAllowlistedCommand.
if (cmd.nameType === 'application') {
return {
behavior: 'passthrough',
message: `Command '${cmd.name}' resolved from a path-like name and requires approval`,
}
}
// SECURITY: elementTypes whitelist — same as isAllowlistedCommand.
// deriveSecurityFlags above checks hasSubExpressions/etc. but does NOT
// flag bare Variable/Other elementTypes. `Remove-Item $env:PATH`:
// elementTypes = ['StringConstant', 'Variable']
// deriveSecurityFlags: no subexpression → passes
// checkPathConstraints: resolves literal text '$env:PATH' as relative
// path → cwd/$env:PATH → inside cwd → allow
// RUNTIME: PowerShell expands $env:PATH → deletes actual env value path
// isAllowlistedCommand rejects non-StringConstant/Parameter; this is the
// acceptEdits parity gate.
//
// Also check colon-bound expression metachars (same as isAllowlistedCommand's
// colon-bound check). `Remove-Item -Path:(1 > /tmp/x)`:
// elementTypes = ['StringConstant', 'Parameter'] — passes whitelist above
// deriveSecurityFlags: ParenExpressionAst in .Argument not detected by
// Get-SecurityPatterns (ParenExpressionAst not in FindAll filter)
// checkPathConstraints: literal text '-Path:(1 > /tmp/x)' not a path
// RUNTIME: paren evaluates, redirection writes /tmp/x → arbitrary write
if (cmd.elementTypes) {
for (let i = 1; i < cmd.elementTypes.length; i++) {
const t = cmd.elementTypes[i]
if (t !== 'StringConstant' && t !== 'Parameter') {
return {
behavior: 'passthrough',
message: `Command argument has unvalidatable type (${t}) — variable paths cannot be statically resolved`,
}
}
if (t === 'Parameter') {
// elementTypes[i] ↔ args[i-1] (elementTypes[0] is the command name).
const arg = cmd.args[i - 1] ?? ''
const colonIdx = arg.indexOf(':')
if (colonIdx > 0 && /[$(@{[]/.test(arg.slice(colonIdx + 1))) {
return {
behavior: 'passthrough',
message:
'Colon-bound parameter contains an expression that cannot be statically validated',
}
}
}
}
}
// Safe output cmdlets (Out-Null, etc.) and allowlisted pipeline-tail
// transformers (Format-*, Measure-Object, Select-Object, etc.) don't
// affect the semantics of the preceding command. Skip them so
// `Remove-Item ./foo | Out-Null` or `Set-Content ./foo hi | Format-Table`
// auto-allows the same as the bare write cmdlet. isAllowlistedPipelineTail
// is the narrow fallback for cmdlets moved from SAFE_OUTPUT_CMDLETS to
// CMDLET_ALLOWLIST (argLeaksValue validates their args).
if (
isSafeOutputCommand(cmd.name) ||
isAllowlistedPipelineTail(cmd, input.command)
) {
continue
}
if (!isAcceptEditsAllowedCmdlet(cmd.name)) {
return {
behavior: 'passthrough',
message: `No mode-specific handling for '${cmd.name}' in acceptEdits mode`,
}
}
// SECURITY: Reject commands with unclassifiable argument types. 'Other'
// covers HashtableAst, ConvertExpressionAst, BinaryExpressionAst — all
// can contain nested redirections or code that the parser cannot fully
// decompose. isAllowlistedCommand (readOnlyValidation.ts) already
// enforces this whitelist via argLeaksValue; this closes the same gap
// in acceptEdits mode. Without this, @{k='payload' > ~/.bashrc} as a
// -Value argument passes because HashtableAst maps to 'Other'.
// argLeaksValue also catches colon-bound variables (-Flag:$env:SECRET).
if (argLeaksValue(cmd.name, cmd)) {
return {
behavior: 'passthrough',
message: `Arguments in '${cmd.name}' cannot be statically validated in acceptEdits mode`,
}
}
}
// Also check nested commands from control flow statements
if (segment.nestedCommands) {
for (const cmd of segment.nestedCommands) {
if (cmd.elementType !== 'CommandAst') {
// SECURITY: Same as above — non-CommandAst element in nested commands
// (control flow bodies) cannot be statically validated as a path source.
return {
behavior: 'passthrough',
message: `Nested expression element (${cmd.elementType}) cannot be statically validated`,
}
}
if (cmd.nameType === 'application') {
return {
behavior: 'passthrough',
message: `Nested command '${cmd.name}' resolved from a path-like name and requires approval`,
}
}
if (
isSafeOutputCommand(cmd.name) ||
isAllowlistedPipelineTail(cmd, input.command)
) {
continue
}
if (!isAcceptEditsAllowedCmdlet(cmd.name)) {
return {
behavior: 'passthrough',
message: `No mode-specific handling for '${cmd.name}' in acceptEdits mode`,
}
}
// SECURITY: Same argLeaksValue check as the main command loop above.
if (argLeaksValue(cmd.name, cmd)) {
return {
behavior: 'passthrough',
message: `Arguments in nested '${cmd.name}' cannot be statically validated in acceptEdits mode`,
}
}
}
}
}
// All commands are filesystem-modifying cmdlets -- auto-allow
return {
behavior: 'allow',
updatedInput: input,
decisionReason: {
type: 'mode',
mode: 'acceptEdits',
},
}
}