Where agent teams gather. Cross-platform team chat where AI agents are first-class teammates — with cloud or bring-your-own (Claude Code / Codex) brains.
TypeScript
/**
* The standing prompt must expose sandboxed development commands while keeping
* Cumora actions on the fixed MCP bridge (#145).
*
* Run: node --import tsx --test server/src/__tests__/agents-computer-prompt-surface.test.ts
*/
import { test } from 'node:test'
import assert from 'node:assert/strict'
import {
actionSurfaceFor, actionSurfaceText, calendarExampleText, postingMechanicsText,
} from '../agents/computer/prompt-surface.js'
test('secure mode gets the MCP tool, compatibility mode keeps the CLI', () => {
// The switch is the unsandboxed opt-in itself, because that is exactly what
// gates writing the shim into the model-visible <home>/bin.
assert.equal(actionSurfaceFor(false), 'mcp')
assert.equal(actionSurfaceFor(true), 'cli')
})
test('the secure prompt allows local development and routes Cumora actions through MCP', () => {
const text = actionSurfaceText('mcp')
assert.match(text, /mcp__cumora__cli/)
assert.match(text, /argv/)
assert.match(text, /no `cumora` on\s+PATH/)
assert.match(text, /no runtime URL or token/)
assert.match(text, /sandboxed shell/)
assert.match(text, /scripts, tests and builds/)
assert.match(text, /`workspace\/`/)
assert.match(text, /build caches there/)
assert.match(text, /Cumora actions still go through the MCP tool/)
assert.doesNotMatch(text, /NO shell|no `node`|try to run a command line|whole surface/)
assert.doesNotMatch(text, /CLI on your PATH/)
})
test('the compatibility prompt is unchanged', () => {
// Operators who opted into unsandboxed BYOA do have the shim on PATH, and
// nothing about their instructions should shift.
assert.equal(actionSurfaceText('cli'), 'You act on Cumora through the `cumora` CLI on your PATH.\n\n')
})
test('secure posting advice does not point at the flag the bridge refuses', () => {
// The sharper half of the bug. The MCP shim rejects these outright:
// "local file/stdin flags are unavailable; pass body text directly in argv"
// so an agent that DID find the tool and followed the old prompt still failed.
const text = postingMechanicsText('mcp')
assert.match(text, /directly in argv/)
assert.match(text, /Do NOT use/)
assert.match(text, /`--file` or `--stdin`/)
assert.doesNotMatch(text, /notes\/reply\.md/)
assert.doesNotMatch(text, /SINGLE quotes/)
})
test('compatibility posting advice keeps the file route, which is right there', () => {
const text = postingMechanicsText('cli')
assert.match(text, /--file notes\/reply\.md/)
assert.match(text, /the shell mangles/)
})
test('the secure calendar example is argv, not a quoted command line', () => {
// Single quotes are load-bearing in the shell form and would arrive as
// literal characters inside a calendar title through the bridge.
const text = calendarExampleText('mcp', 'agent-ada')
assert.match(text, /"argv":\["calendar","create"/)
assert.match(text, /"agent-ada"/)
assert.doesNotMatch(text, /'<chase>'/)
})
test('the compatibility calendar example still shell-quotes', () => {
const text = calendarExampleText('cli', 'agent-ada')
assert.match(text, /cumora calendar create '<chase>'/)
assert.match(text, /--assignee agent-ada/)
})
test('neither surface leaves an unbalanced backtick in the prompt', () => {
// These strings are stitched into one prompt full of `code spans`; an odd
// count means a span swallows the rest of the paragraph in whatever renders
// it, which is easy to do and invisible until someone reads the output.
for (const text of [
actionSurfaceText('cli'), actionSurfaceText('mcp'),
postingMechanicsText('cli'), postingMechanicsText('mcp'),
calendarExampleText('cli', 'a'), calendarExampleText('mcp', 'a'),
]) {
const ticks = (text.match(/`/g) ?? []).length
assert.equal(ticks % 2, 0, `unbalanced backticks (${ticks}) in: ${text}`)
}
})