The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
/**
* The read-only gate is a BLOCKLIST: a turn strips MUTATING_TOOLS and lets
* everything else through. That means a tool nobody classified is allowed by
* default, which is exactly how pr_resume sat inside /review and /plan while it
* still had a shell injection in it (2026-07-26 security pass).
*
* This test closes the gap the other way round: every built-in tool must appear
* in one of the two lists. Add a tool without deciding, and this fails.
*/
import { describe, it, expect } from 'vitest'
import { readFileSync } from 'fs'
import { fileURLToPath } from 'url'
import { dirname, join } from 'path'
import { MUTATING_TOOLS } from '../mutating-tools'
// Read the names out of the source. Importing the module pulls in the
// builtin-tools → tool-registry → mcp/index cycle, which throws on load.
const src = readFileSync(
join(dirname(fileURLToPath(import.meta.url)), '../../api/mcp/builtin-tools.ts'),
'utf8',
)
const arrayStart = src.indexOf('const BUILTIN_TOOLS')
const arrayEnd = src.indexOf('export function registerBuiltinTools')
const BUILTIN_TOOL_NAMES = [
...new Set(
[...src.slice(arrayStart, arrayEnd).matchAll(/^\s{4}name: '([a-z0-9_]+)',$/gm)].map((m) => m[1]),
),
]
/**
* Safe to run in a read-only turn. Every name here is a decision: it either
* only reads, or it only reports on work that is already running.
*/
const READ_ONLY_TOOLS: ReadonlySet<string> = new Set([
// Writes the plan the user sees, and nothing else. It must stay available in
// a read-only turn on purpose: /plan asking for a plan and then having no
// tool to write one with would be absurd.
'todo_write',
// 3.1.0 (P-L): searches and reads pages, GET only, nothing leaves but the query.
'web_research',
'file_read',
'file_list',
'file_search',
// Reads a PR through `gh pr view` / `gh pr diff`. Owner and repo are
// validated to GitHub's own character set before they reach a shell.
'pr_resume',
'web_search',
'web_fetch',
])
describe('every built-in tool is classified', () => {
it('found the tool list to check', () => {
// Guard the parsing itself, so a refactor that breaks the regex fails loudly
// instead of quietly asserting over an empty list.
expect(arrayStart).toBeGreaterThanOrEqual(0)
expect(arrayEnd).toBeGreaterThan(arrayStart)
// 14 since the 2.6.6 merge folded the twelve typed shell wrappers away.
expect(BUILTIN_TOOL_NAMES.length).toBeGreaterThan(12)
expect(BUILTIN_TOOL_NAMES).toContain('shell_execute')
expect(BUILTIN_TOOL_NAMES).toContain('pr_resume')
})
it('appears in exactly one of the two lists', () => {
const unclassified: string[] = []
const both: string[] = []
for (const name of BUILTIN_TOOL_NAMES) {
const mutating = MUTATING_TOOLS.has(name)
const readOnly = READ_ONLY_TOOLS.has(name)
if (!mutating && !readOnly) unclassified.push(name)
if (mutating && readOnly) both.push(name)
}
expect(
unclassified,
'new tool with no decision: add it to MUTATING_TOOLS if it can change '
+ 'anything outside the conversation, otherwise to READ_ONLY_TOOLS here',
).toEqual([])
expect(both, 'a tool cannot be both').toEqual([])
})
it('nothing that spawns a shell is read-only unless its arguments are constrained', () => {
// pr_resume is the one read-only tool that reaches a shell. It is allowed
// there only because parsePrUrl rejects every metacharacter, which
// shell-injection.test.ts pins. The typed git readers left with the 2.6.6
// merge; their read-only path is shell_execute, whose command classifier
// (isReadOnlyCommand) carries the same proof in its own tests.
const shellReaders = [...READ_ONLY_TOOLS].filter((n) =>
['pr_resume', 'shell_execute'].includes(n),
)
expect(shellReaders).toEqual(['pr_resume'])
})
})