Compare commits

...
Author SHA1 Message Date
yilsemandClaude Opus 5.5 7e4cc0915c AITURK IDE 1.0.0-beta.18: gorunmeyen onay beklemesi yerine aninda "onay gerekli"
A one-shot run (hermes chat --query-file ... -Q, stdin closed, no window)
still installs the prompt_toolkit approval callback but never starts the
prompt UI. The protected instruction-file gate (AGENTS.md / CLAUDE.md /
SOUL.md ...) took the registered callback for a reachable human and waited
out approvals.timeout on a panel painted nowhere. With the profile value of
300000 s the stage job hung for good.

- Gate: in a one-shot run, or when only an automatic approver
  (subagent_auto_approve) is attached, return BLOCKED / APPROVAL REQUIRED
  immediately instead of prompting. Human consent is still required every
  time; yolo and approvals.single_query_mode still do not open this gate.
- One-shot exit prints an "approval_required: {json}" line on stderr
  (session_id, tool_call_id, request_id, targets, reason, applied=false)
  so the launcher can tell "blocked, needs a human" from "still working".
- execute_code: an approval that arrives after the calling script stopped
  waiting (300 s reply timeout, script run timeout, retired kernel cell)
  is not applied.
- Same invisible wait removed from the computer_use and memory-write
  prompts in a one-shot run. The one-shot marker is no longer inherited by
  child processes.
- BLOCKED text no longer breaks on a target path containing braces.

Long-running tools are untouched: no general tool timeout was shortened.

Kart: turkservis-20261007-1421-2

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EwSpve9HuXMy7dHf4RDMzX
2026-10-07 17:17:05 +03:00
yilsemandClaude Opus 5.5 58dbb4523b AITURK IDE 1.0.0-beta.17: telefon kodu yayini yeniden denenir
Uzun sure acik kalan IDE'de "Telefon ekle" ilk istegi olu bir keep-alive
baglantisinda kaybedebiliyordu; bulusma kaydi yayinlanamayinca kod yolu bes
dakika boyunca "Kodla baglanma su an kullanilamiyor" kaliyordu (QR calisiyordu).

- publishPairOffer: en cok 4 deneme (0,5 / 1,5 / 3 sn); ag hatasi, 5xx ve 429
  yeniden denenir, 422 kalicidir; yaniti kaybolan denemeden sonraki 409 kaydin
  yerinde oldugu anlamina gelir (iptal jetonu yoktur, kayit kendiliginden biter).
- Pencerede "Yeniden dene": ayni kod ve QR ile kaydi yeniden yayinlar.
- Role oturumu kaydi ve "Giris yap" baslangici icin tek ek deneme (zaman asimi
  yeniden denenmez).
- Gunluge yalniz "pair offer published / failed after N tries" duser; kod,
  lookup ve jeton yazilmaz.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B8ppceaR5s6cvMnm7uGRnf
2026-10-05 14:27:59 +03:00
24 changed files with 1613 additions and 90 deletions

No files matched your search

@@ -11,6 +11,7 @@ import { describe, expect, it, vi } from 'vitest'
import {
AITURK_CLI_KEY_HOLD_MS,
AITURK_LOGIN_START_RETRY_DELAY_MS,
type AiturkCliExit,
AiturkLoginError,
createAiturkCliRunner,
@@ -38,6 +39,8 @@ const approved = { state: 'approved', key: IDE_KEY, cliKey: CLI_KEY, account: 'a
function harness(
options: {
start?: { status: number; body?: unknown }
/** Thrown by successive `device-auth/start` calls before any answer is given. */
startErrors?: Error[]
polls?: ({ status: number; body?: unknown } | Error)[]
cliExits?: AiturkCliExit[]
validate?: unknown
@@ -47,6 +50,7 @@ function harness(
) {
const calls: { url: string; init: any }[] = []
const polls = [...(options.polls ?? [])]
const startErrors = [...(options.startErrors ?? [])]
const cliExits = [...(options.cliExits ?? [0])]
const clock = { now: 1_000_000 }
const timers: { fn: () => void; ms: number; cleared: boolean }[] = []
@@ -55,6 +59,12 @@ function harness(
calls.push({ url, init })
if (url.endsWith('/start')) {
const failure = startErrors.shift()
if (failure) {
throw failure
}
const r = options.start ?? { status: 201, body: startBody() }
return { status: r.status, json: async () => r.body }
@@ -154,6 +164,47 @@ describe('aiturk login: start', () => {
await expect(harness({ start: { status: 500 } }).login.start()).rejects.toThrow('aiturk-login:unavailable')
})
it('tries start once more after a thrown network error (first request of a long-idle process)', async () => {
const h = harness({ startErrors: [new TypeError('fetch failed')] })
const pending = h.login.start()
await vi.waitFor(() => expect(h.timers.some(timer => timer.ms === AITURK_LOGIN_START_RETRY_DELAY_MS)).toBe(true))
expect(AITURK_LOGIN_START_RETRY_DELAY_MS).toBe(500)
expect(h.calls.filter(call => call.url.endsWith('/start'))).toHaveLength(1)
expect(h.openExternal).not.toHaveBeenCalled()
h.timers.find(timer => timer.ms === AITURK_LOGIN_START_RETRY_DELAY_MS)!.fn()
expect(await pending).toEqual({ userCode: 'ABCD2345', verificationUri: URI, expiresAt: 1_000_000 + 600_000 })
expect(h.calls.filter(call => call.url.endsWith('/start'))).toHaveLength(2)
// One code only: the approval page is opened once.
expect(h.openExternal).toHaveBeenCalledExactlyOnceWith(URI)
})
it('gives up after the second network error, and never retries an HTTP answer or a timeout', async () => {
const twice = harness({ startErrors: [new TypeError('fetch failed'), new TypeError('fetch failed')] })
const failing = twice.login.start().catch(e => e)
await vi.waitFor(() => expect(twice.timers.some(timer => timer.ms === AITURK_LOGIN_START_RETRY_DELAY_MS)).toBe(true))
twice.timers.find(timer => timer.ms === AITURK_LOGIN_START_RETRY_DELAY_MS)!.fn()
expect((await failing).message).toBe('aiturk-login:unavailable')
expect(twice.calls.filter(call => call.url.endsWith('/start'))).toHaveLength(2)
expect(twice.timers.filter(timer => timer.ms === AITURK_LOGIN_START_RETRY_DELAY_MS)).toHaveLength(1)
for (const status of [500, 503, 429, 422]) {
const answered = harness({ start: { status } })
await expect(answered.login.start()).rejects.toThrow(/^aiturk-login:(?:unavailable|rate-limited)$/)
expect(answered.calls.filter(call => call.url.endsWith('/start'))).toHaveLength(1)
expect(answered.timers).toHaveLength(0)
}
const timedOut = harness({ startErrors: [new DOMException('The operation was aborted due to timeout', 'TimeoutError')] })
await expect(timedOut.login.start()).rejects.toThrow('aiturk-login:unavailable')
expect(timedOut.calls.filter(call => call.url.endsWith('/start'))).toHaveLength(1)
expect(timedOut.timers).toHaveLength(0)
})
it('rejects malformed codes', async () => {
await expect(
harness({ start: { status: 201, body: startBody({ userCode: 'abcd2345' }) } }).login.start()
+24 -9
View File
@@ -22,6 +22,8 @@
export const AITURK_LOGIN_ORIGIN = 'https://turkservis.online'
export const AITURK_LOGIN_FETCH_TIMEOUT_MS = 20_000
/** Wait before the single second try of `device-auth/start` after a thrown network error. */
export const AITURK_LOGIN_START_RETRY_DELAY_MS = 500
export const AITURK_CLI_TIMEOUT_MS = 60_000
export const AITURK_CLI_KEY_HOLD_MS = 5 * 60_000
@@ -237,16 +239,29 @@ export function createAiturkLogin(deps: AiturkLoginDeps) {
await cancelOnServer(previous.deviceCode)
}
let response: AiturkFetchResponse
let response: AiturkFetchResponse | undefined
try {
response = await post('/v1/aiturk/device-auth/start', null, {
client: 'ide',
device_name: deviceName(deps.hostname()),
also_cli: true
})
} catch {
throw new AiturkLoginError('unavailable')
// A long-idle process can fail its first request on a dead pooled
// connection. Only a THROWN network error (no HTTP answer, so no code was
// handed to us) is tried once more; an HTTP answer is never retried, so
// one start cannot knowingly leave two codes behind. A timeout already
// used its whole budget and is not retried.
for (let attempt = 1; response === undefined; attempt++) {
try {
response = await post('/v1/aiturk/device-auth/start', null, {
client: 'ide',
device_name: deviceName(deps.hostname()),
also_cli: true
})
} catch (error) {
const name = (error as null | { name?: unknown })?.name
if (attempt >= 2 || name === 'TimeoutError' || name === 'AbortError') {
throw new AiturkLoginError('unavailable')
}
await new Promise<void>(resolve => void setTimer(resolve, AITURK_LOGIN_START_RETRY_DELAY_MS))
}
}
if (response.status === 429) {
+1 -1
View File
@@ -17893,7 +17893,7 @@ function getMobileLink(): MobileLink {
// API traffic is not logged; pairing and refusals are, for diagnosis. Relay
// lines are status changes only ("relay connected", "relay failed:<code>").
log: line => {
if (/^relay /.test(line) || / \/aiturk\//.test(line) || / [45]\d\d$/.test(line)) {
if (/^(?:relay|pair) /.test(line) || / \/aiturk\//.test(line) || / [45]\d\d$/.test(line)) {
rememberLog(`[mobile-lan] ${line}`)
}
}
+6 -3
View File
@@ -20,7 +20,7 @@ const TEXT = {
scan: 'Telefonda AITURK Mobil → “Bilgisayara bağlan” → bu QR kodunu okutun.',
orCode: 'ya da telefonda şu kodu yazın:',
codePending: 'Kodla bağlanma hazırlanıyor…', codeReady: 'Kodla bağlanma hazır.',
codeUnavailable: 'Kodla bağlanma şu an kullanılamıyor; QR kodunu okutun.',
codeUnavailable: 'Kodla bağlanma şu an kullanılamıyor; QR kodunu okutun.', codeRetry: 'Yeniden dene',
securityCode: code => `Güvenlik kodu: ${code} (telefonda da aynı kod görünmeli)`,
anyNetwork: 'Telefon ve bilgisayar aynı ağda olmak zorunda değil.', sameWifi: 'Telefon ve bilgisayar aynı Wi-Fi ağında olmalı.',
countdown: (m, s) => `Kod ${m}:${s} sonra geçersiz olur`, expired: 'Kodun süresi doldu. Yeniden “Telefon ekle” seçin.',
@@ -56,7 +56,7 @@ const TEXT = {
scan: 'On the phone: AITURK Mobile → “Connect to computer” → scan this QR code.',
orCode: 'or type this code on the phone:',
codePending: 'Preparing code sign-in…', codeReady: 'Connecting with the code is ready.',
codeUnavailable: 'Connecting with the code is unavailable right now; scan the QR code.',
codeUnavailable: 'Connecting with the code is unavailable right now; scan the QR code.', codeRetry: 'Try again',
securityCode: code => `Security code: ${code} (the phone must show the same code)`,
anyNetwork: 'The phone and the computer do not have to be on the same network.', sameWifi: 'The phone and the computer must be on the same Wi-Fi network.',
countdown: (m, s) => `The code expires in ${m}:${s}`, expired: 'The code expired. Choose “Add a phone” again.',
@@ -84,7 +84,7 @@ document.documentElement.lang = lang; document.title = `AITURK · ${t.title}`
for (const [id, key] of [['t-eyebrow', 'eyebrow'], ['t-title', 'title'], ['t-intro', 'intro'], ['t-remote', 'remote'], ['t-scan', 'scan'],
['t-or-code', 'orCode'], ['t-devices', 'devices'], ['devices-empty', 'noDevices'], ['t-enabled', 'toggle'], ['t-relay-label', 'relayLabel'],
['t-lan-label', 'lanLabel'], ['t-reissued', 'reissued'], ['lan-disallow', 'lanDisallow'], ['relay-renew', 'renew'],
['pair-begin', 'addPhone'], ['pair-cancel', 'cancel'], ['firewall-remove', 'removeRule']]) $(id).textContent = t[key]
['pair-begin', 'addPhone'], ['pair-cancel', 'cancel'], ['pair-republish', 'codeRetry'], ['firewall-remove', 'removeRule']]) $(id).textContent = t[key]
let current = null, drawnQr = '', busy = false
function notice(text) { $('notice').textContent = text || ''; $('notice').hidden = !text }
@@ -121,6 +121,8 @@ function render(state) {
$('pair-qr').setAttribute('aria-label', t.scan)
$('pair-code').textContent = `${state.pairing.code.slice(0, 4)}-${state.pairing.code.slice(4)}`
$('pair-code-path').textContent = { pending: t.codePending, ready: t.codeReady, unavailable: t.codeUnavailable }[state.pairing.codePath]
// Offered only once every background try is over; while they run the line above says "hazırlanıyor…".
$('pair-republish').hidden = state.pairing.codePath !== 'unavailable'
$('pair-security').textContent = state.pairing.securityCode ? t.securityCode(state.pairing.securityCode) : ''
$('t-same-wifi').textContent = /[?&]r=/.test(state.pairing.qr) ? t.anyNetwork : t.sameWifi
tick()
@@ -155,6 +157,7 @@ async function refresh() { try { render(await host.call('state')) } catch (error
$('enabled').onchange = () => { const on = $('enabled').checked; void run(async () => { notice(''); render(await host.call(on ? 'enable' : 'disable')) }) }
$('pair-begin').onclick = () => run(async () => { notice(''); render(await host.call('pair.begin')) })
$('pair-cancel').onclick = () => run(async () => { render(await host.call('pair.cancel')) })
$('pair-republish').onclick = () => run(async () => { render(await host.call('pair.republish')) })
$('lan-allow').onclick = () => run(async () => { notice(''); render(await host.call('lan.allow')) })
$('lan-disallow').onclick = () => run(async () => { notice(''); render(await host.call('lan.disallow')) })
$('relay-renew').onclick = () => { if (window.confirm(t.renewConfirm)) void run(async () => { render(await host.call('relay.renew')); notice(t.renewed) }) }
+1
View File
@@ -24,3 +24,4 @@ button.link { background:none; color:var(--accent); min-height:0; padding:0; mar
.security { font-family:ui-monospace,Consolas,monospace; color:var(--text); overflow-wrap:anywhere }
footer { display:flex; flex-wrap:wrap; gap:8px 20px }
footer button.link { min-height:44px }
#pair-republish { display:block; min-height:44px; margin:0 auto; padding:0 12px }
+1
View File
@@ -20,6 +20,7 @@
<p id="t-or-code"></p>
<p id="pair-code" class="code" aria-live="off"></p>
<p id="pair-code-path"></p>
<button id="pair-republish" class="link" hidden></button>
<p id="pair-countdown" class="countdown"></p>
<p id="pair-security" class="security"></p>
<p id="t-same-wifi" class="hint"></p>
+2
View File
@@ -20,6 +20,8 @@ export function createMobileWindow({ assets, link, show = true }) {
disable: () => link.disable(),
'pair.begin': () => link.pairBegin(),
'pair.cancel': () => link.pairCancel(),
// Publishes the running pairing's rendezvous record again; the code and the QR stay as they are.
'pair.republish': () => link.pairRepublish(),
devices: () => link.devices(),
'device.revoke': params => link.deviceRevoke(params?.deviceId),
// The only actions that may raise a UAC prompt; both are explicit button presses.
+340 -4
View File
@@ -19,6 +19,7 @@ import { WebSocket, WebSocketServer } from 'ws'
import { createSelfSignedCertificate } from './mobile-lan-cert'
import {
buildPairQrPayload,
cancelPairOffer,
createLanProxy,
createMobileLink,
createWindowsFirewall,
@@ -34,6 +35,8 @@ import {
type MobileLinkEvent,
normalizePairCode,
PAIR_CODE_ALPHABET,
PAIR_OFFER_MAX_TRIES,
PAIR_OFFER_RETRY_DELAYS_MS,
pairLookup,
parseFirewallRuleOutput,
parseMobileLinkStored,
@@ -745,10 +748,19 @@ function linkHarness(over: Record<string, any> = {}) {
const events: MobileLinkEvent[] = []
const logs: string[] = []
const hostSockets: FakeHostSocket[] = []
/** Outcomes of successive `POST /offers` calls (status or 'throw'); afterwards `state.offerStatus` applies. */
const offerPlan: (number | 'throw')[] = [...(over.offerPlan ?? [])]
/** Waits the controller asked for between tries (real values); they are shortened to 1 ms here. */
const retrySleeps: number[] = []
/** With `holdRetries` a wait ends only when the test calls the function. */
const heldRetries: (() => void)[] = []
const RETRY_WAITS = new Set<number>([...PAIR_OFFER_RETRY_DELAYS_MS])
const state = {
remote: false,
offerStatus: over.offerStatus ?? 201,
/** While set, `POST /offers` answers only after it resolves. */
offerGate: null as null | Promise<void>,
/** false = `POST /sessions` fails (no internet). */
sessions: over.sessions ?? true,
sessionCount: 0,
@@ -809,13 +821,20 @@ function linkHarness(over: Record<string, any> = {}) {
return { status: 201, json: async () => ({ relayId: RELAY_IDS[index], hostToken: HOST_TOKENS[index] }) }
}
if (state.offerStatus === 'throw') {
const isOffer = url.endsWith('/offers')
const offerOutcome = isOffer && offerPlan.length ? offerPlan.shift() : state.offerStatus
if (isOffer && state.offerGate) {
await state.offerGate
}
if (offerOutcome === 'throw') {
throw new Error('network down')
}
if (url.endsWith('/offers')) {
if (isOffer) {
return {
status: state.offerStatus,
status: offerOutcome,
json: async () => ({ offerId: '0f8fad5b-d9cb-469f-a165-70867728950e', cancelToken: `tsmp_${'c'.repeat(43)}`, expiresIn: 300 })
}
}
@@ -835,6 +854,23 @@ function linkHarness(over: Record<string, any> = {}) {
platform: 'win32',
listenHost: LAN_HOST,
ports: [0],
// Real timers, except the waits between tries of a request: those are recorded and cut to 1 ms.
setTimer: (fn: () => void, ms: number) => {
if (!RETRY_WAITS.has(ms)) {
return setTimeout(fn, ms)
}
retrySleeps.push(ms)
if (over.holdRetries) {
heldRetries.push(fn)
return null
}
return setTimeout(fn, 1)
},
clearTimer: (handle: unknown) => clearTimeout(handle as NodeJS.Timeout),
log: (line: string) => logs.push(line)
}
@@ -850,7 +886,7 @@ function linkHarness(over: Record<string, any> = {}) {
const firewallWrites = () => firewallCalls.filter(entry => entry !== 'query')
const offers = () => fetchCalls.filter(entry => entry.url.endsWith('/offers'))
return { link: make(), make, vault, store, firewallCalls, firewallWrites, fetchCalls, offers, events, logs, state, hostSockets }
return { link: make(), make, vault, store, firewallCalls, firewallWrites, fetchCalls, offers, events, logs, state, hostSockets, retrySleeps, heldRetries }
}
describe('mobile link controller', () => {
@@ -1278,6 +1314,306 @@ describe('mobile link controller', () => {
})
})
describe('mobile link controller: rendezvous record retries', () => {
const lookupOf = (h: ReturnType<typeof linkHarness>, index: number) => JSON.parse(h.offers()[index].init.body).lookup
const cancels = (h: ReturnType<typeof linkHarness>) => h.fetchCalls.filter(entry => entry.url.endsWith('/cancel'))
const noSecretInLogs = (h: ReturnType<typeof linkHarness>, code: string) => {
const text = h.logs.join('\n')
expect(text).not.toContain(code)
expect(text).not.toContain(pairLookup(code))
expect(text).not.toContain('tsmp_')
}
it('first request rejects, the second answers 201: the code path becomes ready (one lost request does not close it)', async () => {
const h = linkHarness({ offerPlan: ['throw', 201] })
await h.link.enable()
const state = await h.link.pairBegin()
expect(state.pairing!.codePath).toBe('pending')
await until(() => h.events.some(event => event.state.pairing?.codePath === 'ready'))
expect(h.offers()).toHaveLength(2)
expect(h.retrySleeps).toEqual([500])
// Same record both times, and each try has its own timeout signal.
expect(lookupOf(h, 1)).toBe(lookupOf(h, 0))
expect(h.offers()[1].init.signal).not.toBe(h.offers()[0].init.signal)
expect(h.events.some(event => event.state.pairing?.codePath === 'unavailable')).toBe(false)
expect(h.logs.filter(line => line.startsWith('pair '))).toEqual(['pair offer published'])
noSecretInLogs(h, state.pairing!.code)
await h.link.stop()
})
it('two 503 answers, then 201: ready on the third try', async () => {
const h = linkHarness({ offerPlan: [503, 503, 201] })
await h.link.enable()
await h.link.pairBegin()
await until(() => h.events.some(event => event.state.pairing?.codePath === 'ready'))
expect(h.offers()).toHaveLength(3)
expect(h.retrySleeps).toEqual([500, 1500])
expect(h.events.some(event => event.state.pairing?.codePath === 'unavailable')).toBe(false)
await h.link.stop()
})
it('a lost answer followed by 409 means the first request landed: ready, and no cancel is sent without a token', async () => {
const h = linkHarness({ offerPlan: ['throw', 409] })
await h.link.enable()
await h.link.pairBegin()
await until(() => h.events.some(event => event.state.pairing?.codePath === 'ready'))
expect(h.offers()).toHaveLength(2)
expect((await h.link.pairCancel()).pairing).toBeNull()
await new Promise(resolve => setTimeout(resolve, 50))
expect(cancels(h)).toEqual([])
await h.link.stop()
expect(cancels(h)).toEqual([])
})
it('422 is final: one try, unavailable', async () => {
const h = linkHarness({ offerPlan: [422, 201] })
await h.link.enable()
const state = await h.link.pairBegin()
await until(() => h.events.some(event => event.state.pairing?.codePath === 'unavailable'))
await new Promise(resolve => setTimeout(resolve, 50))
expect(h.offers()).toHaveLength(1)
expect(h.retrySleeps).toEqual([])
expect(h.logs.filter(line => line.startsWith('pair '))).toEqual(['pair offer failed after 1 try'])
noSecretInLogs(h, state.pairing!.code)
await h.link.stop()
})
it('four failing tries: unavailable only after the fourth, never earlier', async () => {
const h = linkHarness({ offerPlan: ['throw', 503, 429, 'throw', 201], holdRetries: true })
await h.link.enable()
const state = await h.link.pairBegin()
for (let tries = 1; tries <= 3; tries++) {
await until(() => h.offers().length === tries && h.heldRetries.length === 1)
// Still being tried: the window keeps saying "hazırlanıyor…".
expect((await h.link.state()).pairing!.codePath).toBe('pending')
expect(h.events.some(event => event.state.pairing?.codePath === 'unavailable')).toBe(false)
h.heldRetries.shift()!()
}
await until(() => h.events.some(event => event.state.pairing?.codePath === 'unavailable'))
expect(h.offers()).toHaveLength(4)
expect(h.retrySleeps).toEqual([500, 1500, 3000])
expect(h.heldRetries).toEqual([])
expect(h.logs.filter(line => line.startsWith('pair '))).toEqual(['pair offer failed after 4 tries'])
noSecretInLogs(h, state.pairing!.code)
await h.link.stop()
})
it('cancelling the pairing stops the tries; a record that lands after the cancel is cancelled on the gateway', async () => {
const h = linkHarness({ offerPlan: ['throw', 201], holdRetries: true })
await h.link.enable()
await h.link.pairBegin()
await until(() => h.offers().length === 1 && h.heldRetries.length === 1)
expect((await h.link.pairCancel()).pairing).toBeNull()
// The wait ends after the pairing is gone: no further request.
h.heldRetries.shift()!()
await new Promise(resolve => setTimeout(resolve, 50))
expect(h.offers()).toHaveLength(1)
expect(cancels(h)).toEqual([])
expect(h.logs.filter(line => line.startsWith('pair '))).toEqual([])
// A NEW pairing is not confused with the old one, and a request in flight when pairing ends is cleaned up.
let release: () => void = () => undefined
h.state.offerGate = new Promise<void>(resolve => {
release = resolve
})
await h.link.pairBegin()
await until(() => h.offers().length === 2)
await h.link.pairCancel()
release()
await until(() => cancels(h).length === 1)
expect(cancels(h)[0].init.headers.Authorization).toBe(`Bearer tsmp_${'c'.repeat(43)}`)
await h.link.stop()
})
it('pair.republish publishes the running pairing again: new request, same code, same QR, same expiry', async () => {
const h = linkHarness({ offerPlan: [422] })
await h.link.enable()
const begun = (await h.link.pairBegin()).pairing!
await until(() => h.events.some(event => event.state.pairing?.codePath === 'unavailable'))
expect(h.offers()).toHaveLength(1)
const again = (await h.link.pairRepublish()).pairing!
expect(again).toMatchObject({ code: begun.code, qr: begun.qr, expiresAt: begun.expiresAt, securityCode: begun.securityCode, codePath: 'pending' })
await until(() => h.events.at(-1)?.state.pairing?.codePath === 'ready')
expect(h.offers()).toHaveLength(2)
expect(lookupOf(h, 1)).toBe(lookupOf(h, 0))
expect(JSON.parse(h.offers()[1].init.body)).toEqual(JSON.parse(h.offers()[0].init.body))
// Nothing to do while ready (or pending), and nothing without a pairing.
expect((await h.link.pairRepublish()).pairing).toMatchObject({ code: begun.code, codePath: 'ready' })
expect(h.offers()).toHaveLength(2)
// The code on screen still pairs the phone.
const claim = await call((await h.link.state()).port as number, 'POST', '/aiturk/pair/claim', {
ca: h.vault.value.lanTls.cert,
body: { code: begun.code, device_name: 'Pixel' }
})
expect(claim.status).toBe(200)
await until(() => h.events.some(event => event.type === 'paired'))
expect((await h.link.pairRepublish()).pairing).toBeNull()
expect(h.offers()).toHaveLength(2)
await h.link.stop()
})
it('pair.republish after four lost requests: a 409 on its first try is our own record (ready); after a 422 it is not', async () => {
const lost = linkHarness({ offerPlan: ['throw', 'throw', 'throw', 'throw', 409] })
await lost.link.enable()
await lost.link.pairBegin()
await until(() => lost.events.some(event => event.state.pairing?.codePath === 'unavailable'))
expect(lost.offers()).toHaveLength(4)
await lost.link.pairRepublish()
await until(() => lost.events.at(-1)?.state.pairing?.codePath === 'ready')
expect(lost.offers()).toHaveLength(5)
await lost.link.stop()
expect(cancels(lost)).toEqual([])
const refused = linkHarness({ offerPlan: [422, 409] })
await refused.link.enable()
await refused.link.pairBegin()
await until(() => refused.events.some(event => event.state.pairing?.codePath === 'unavailable'))
await refused.link.pairRepublish()
await until(() => refused.offers().length === 2 && refused.events.at(-1)?.state.pairing?.codePath === 'unavailable')
await refused.link.stop()
})
})
describe('publishPairOffer / cancelPairOffer', () => {
const input = { code: 'ABCDEFGH', addresses: [] as string[], port: 47321, fingerprint: 'f'.repeat(43), relay: RELAY_IDS[0] }
const good = { offerId: '0f8fad5b-d9cb-469f-a165-70867728950e', cancelToken: `tsmp_${'c'.repeat(43)}` }
function plan(outcomes: (number | 'throw')[]) {
const sleeps: number[] = []
const logs: string[] = []
let calls = 0
const fetcher = async () => {
const outcome = outcomes[calls++] ?? 201
if (outcome === 'throw') {
throw new Error('fetch failed')
}
return { status: outcome, json: async () => good }
}
const options = {
sleep: async (ms: number) => {
sleeps.push(ms)
},
log: (line: string) => logs.push(line)
}
return { fetcher, options, sleeps, logs, calls: () => calls }
}
it('waits 0.5 s, 1.5 s and 3 s between at most four tries', async () => {
const p = plan(['throw', 500, 429, 'throw', 201])
expect(await publishPairOffer(p.fetcher, input, p.options)).toBeNull()
expect(p.calls()).toBe(PAIR_OFFER_MAX_TRIES)
expect(p.calls()).toBe(4)
expect(p.sleeps).toEqual([500, 1500, 3000])
expect(p.sleeps).toEqual([...PAIR_OFFER_RETRY_DELAYS_MS])
expect(p.logs).toEqual(['offer failed after 4 tries'])
})
it('retries a thrown error, 5xx and 429; 201 ends it', async () => {
for (const first of ['throw', 500, 502, 503, 504, 429] as const) {
const p = plan([first, 201])
expect(await publishPairOffer(p.fetcher, input, p.options)).toEqual(good)
expect(p.calls()).toBe(2)
expect(p.logs).toEqual(['offer published'])
}
})
it('does not retry 422 or any other final answer; a first-try 409 is somebody else’s record', async () => {
for (const status of [422, 400, 401, 403, 404, 409, 200, 302]) {
const p = plan([status, 201])
expect(await publishPairOffer(p.fetcher, input, p.options)).toBeNull()
expect(p.calls()).toBe(1)
expect(p.sleeps).toEqual([])
expect(p.logs).toEqual(['offer failed after 1 try'])
}
})
it('409 after a lost answer or a 5xx is our own record (no token); 409 after 429 is not', async () => {
for (const first of ['throw', 502] as const) {
const p = plan([first, 409])
expect(await publishPairOffer(p.fetcher, input, p.options)).toEqual({ offerId: null, cancelToken: null })
expect(p.calls()).toBe(2)
expect(p.logs).toEqual(['offer published'])
}
const limited = plan([429, 409])
expect(await publishPairOffer(limited.fetcher, input, limited.options)).toBeNull()
expect(limited.calls()).toBe(2)
const told = plan([409])
expect(await publishPairOffer(told.fetcher, input, { ...told.options, earlierTryMayHaveLanded: true })).toEqual({ offerId: null, cancelToken: null })
})
it('stops as soon as the pairing is over, without a log line', async () => {
const p = plan(['throw', 'throw', 201])
let over = false
const result = await publishPairOffer(p.fetcher, input, {
...p.options,
cancelled: () => over,
sleep: async ms => {
p.sleeps.push(ms)
over = true
}
})
expect(result).toBeNull()
expect(p.calls()).toBe(1)
expect(p.logs).toEqual([])
})
it('cancelPairOffer sends nothing when there is no token', async () => {
const urls: string[] = []
const fetcher = async (url: string) => {
urls.push(url)
return { status: 200, json: async () => ({}) }
}
await cancelPairOffer(fetcher, { offerId: null, cancelToken: null })
expect(urls).toEqual([])
await cancelPairOffer(fetcher, good)
expect(urls).toEqual(['https://turkservis.online/v1/aiturk/mobile-pair/cancel'])
})
})
describe('pure helpers', () => {
it('pairing code: alphabet, normalisation, display form, lookup digest', () => {
for (let i = 0; i < 200; i++) {
+178 -39
View File
@@ -247,9 +247,37 @@ export interface PairFetchResponse {
export type PairFetch = (url: string, init: Record<string, unknown>) => Promise<PairFetchResponse>
/**
* A published rendezvous record. Both fields are null when the record is known
* to exist but its answer never reached us (an earlier try failed on the wire
* and the next one got 409): the code path works, there is just nothing to
* cancel — the record expires on its own after five minutes.
*/
export interface PairOffer {
offerId: string
cancelToken: string
offerId: null | string
cancelToken: null | string
}
export const PAIR_OFFER_MAX_TRIES = 4
/** Waits between the tries: ~0.5 s, 1.5 s, 3 s. */
export const PAIR_OFFER_RETRY_DELAYS_MS: readonly number[] = [500, 1500, 3000]
export interface PairOfferOptions {
/** Injected wait (the controller passes its own timer). Default: a real timer. */
sleep?: (ms: number) => Promise<void>
delaysMs?: readonly number[]
maxTries?: number
/** true = the pairing this record belongs to is over: stop trying. */
cancelled?: () => boolean
/** "offer published" / "offer failed after N tries" only — never the code, the lookup or a token. */
log?: (line: string) => void
/**
* An earlier publish for the SAME code may have reached the gateway (its
* answer was lost), so a 409 on the very first try also means "it is there".
*/
earlierTryMayHaveLanded?: boolean
/** Called when a try ended without a usable answer (thrown error / 5xx), i.e. it may have been stored. */
onMayHaveLanded?: () => void
}
function pairFetchInit(body: unknown, headers: Record<string, string> = {}) {
@@ -263,12 +291,22 @@ function pairFetchInit(body: unknown, headers: Record<string, string> = {}) {
}
/**
* Publish the rendezvous record. Returns null on any failure — the QR path
* keeps working and the UI shows "kodla bağlanma kullanılamıyor".
* Publish the rendezvous record. Returns null when it could not be published —
* the QR path keeps working and the UI shows "kodla bağlanma kullanılamıyor".
*
* Up to four tries. A long-idle process can lose its first request on a dead
* pooled keep-alive connection, and one lost request must not close the code
* path for the whole five minutes:
* - a thrown network error, 5xx and 429 are tried again;
* - 422 and every other answer are final (the request itself is wrong);
* - 409 after a try that failed on the wire / with 5xx means that try DID
* land and only its answer was lost: the record exists, the code path is
* ready (no cancel token is known for it).
*/
export async function publishPairOffer(
fetcher: PairFetch,
input: { code: string; addresses: string[]; port: number; fingerprint: string; relay?: null | string }
input: { code: string; addresses: string[]; port: number; fingerprint: string; relay?: null | string },
options: PairOfferOptions = {}
): Promise<null | PairOffer> {
const relay = isRelayId(input.relay) ? input.relay : null
@@ -282,34 +320,93 @@ export async function publishPairOffer(
return null
}
try {
const response = await fetcher(
`${MOBILE_PAIR_ORIGIN}/v1/aiturk/mobile-pair/offers`,
pairFetchInit({ lookup: pairLookup(input.code), urls, fp: input.fingerprint, ...(relay ? { relay } : {}) })
)
const sleep = options.sleep ?? (ms => new Promise<void>(resolve => setTimeout(resolve, ms)))
const delays = options.delaysMs ?? PAIR_OFFER_RETRY_DELAYS_MS
const maxTries = Math.max(1, options.maxTries ?? PAIR_OFFER_MAX_TRIES)
const cancelled = options.cancelled ?? (() => false)
const body = { lookup: pairLookup(input.code), urls, fp: input.fingerprint, ...(relay ? { relay } : {}) }
let mayHaveLanded = options.earlierTryMayHaveLanded === true
let tries = 0
if (response.status !== 201) {
return null
}
const failed = () => {
options.log?.(`offer failed after ${tries} ${tries === 1 ? 'try' : 'tries'}`)
const data = (await response.json()) as null | Record<string, unknown>
if (!data || typeof data.offerId !== 'string' || !UUID_RE.test(data.offerId)) {
return null
}
if (typeof data.cancelToken !== 'string' || !CANCEL_TOKEN_RE.test(data.cancelToken)) {
return null
}
return { offerId: data.offerId, cancelToken: data.cancelToken }
} catch {
return null
}
const published = (offer: PairOffer) => {
options.log?.('offer published')
return offer
}
while (tries < maxTries) {
if (cancelled()) {
return null
}
tries++
let status: number
let data: null | Record<string, unknown> = null
try {
// A fresh init per try: the timeout signal must not be shared between tries.
const response = await fetcher(`${MOBILE_PAIR_ORIGIN}/v1/aiturk/mobile-pair/offers`, pairFetchInit(body))
status = response.status
if (status === 201) {
data = (await response.json()) as null | Record<string, unknown>
}
} catch {
// No (complete) answer: the request may or may not have been stored.
status = 0
}
if (status === 201) {
if (
!data ||
typeof data.offerId !== 'string' ||
!UUID_RE.test(data.offerId) ||
typeof data.cancelToken !== 'string' ||
!CANCEL_TOKEN_RE.test(data.cancelToken)
) {
return failed()
}
return published({ offerId: data.offerId, cancelToken: data.cancelToken })
}
if (status === 409 && mayHaveLanded) {
return published({ offerId: null, cancelToken: null })
}
const transient = status === 0 || status === 429 || (status >= 500 && status <= 599)
if (!transient) {
return failed()
}
if (status !== 429) {
mayHaveLanded = true
options.onMayHaveLanded?.()
}
if (tries < maxTries) {
await sleep(delays[Math.min(tries - 1, delays.length - 1)] ?? 0)
}
}
return cancelled() ? null : failed()
}
/** Best effort; the record expires on its own after five minutes anyway. */
/** Best effort; the record expires on its own after five minutes anyway. No token = nothing to cancel. */
export async function cancelPairOffer(fetcher: PairFetch, offer: PairOffer): Promise<void> {
if (!offer.cancelToken) {
return
}
try {
await fetcher(
`${MOBILE_PAIR_ORIGIN}/v1/aiturk/mobile-pair/cancel`,
@@ -1380,6 +1477,8 @@ interface PairSession {
addresses: string[]
port: number
fingerprint: string
/** A publish got no answer at least once: a later 409 for this code is our own record. */
offerMayHaveLanded: boolean
}
export function createMobileLink(deps: MobileLinkDeps) {
@@ -1526,6 +1625,42 @@ export function createMobileLink(deps: MobileLinkDeps) {
}
}
const sleep = (ms: number) => new Promise<void>(resolve => void setTimer(resolve, ms))
/**
* Publishes (or publishes again) the rendezvous record of `current` in the
* background; the code path stays `pending` until every try is over. The
* code and the QR are never touched here.
*/
function publishOffer(current: PairSession, relayId: null | string) {
current.codePath = 'pending'
const input = { code: current.code, addresses: current.addresses, port: current.port, fingerprint: current.fingerprint, relay: relayId }
void publishPairOffer(deps.fetch, input, {
sleep,
cancelled: () => session !== current,
log: line => deps.log?.(`pair ${line}`),
earlierTryMayHaveLanded: current.offerMayHaveLanded,
onMayHaveLanded: () => {
current.offerMayHaveLanded = true
}
}).then(offer => {
if (session !== current) {
// Pairing ended meanwhile; do not leave the record behind.
if (offer) {
void cancelPairOffer(deps.fetch, offer)
}
return
}
current.offer = offer
current.codePath = offer ? 'ready' : 'unavailable'
emit()
})
}
/** The relay registration changed while a QR is on screen: show the current address. */
function refreshSessionQr(data: MobileLinkStored) {
if (!session) {
@@ -1892,27 +2027,31 @@ export function createMobileLink(deps: MobileLinkDeps) {
}, Math.max(0, expiresAt - now()) + 50),
addresses,
port,
fingerprint
fingerprint,
offerMayHaveLanded: false
}
session = current
ensureRelay(data)?.start()
// Not awaited: the QR must not wait for the gateway.
void publishPairOffer(deps.fetch, { code, addresses, port, fingerprint, relay: relayUrl }).then(offer => {
if (session !== current) {
// Pairing ended meanwhile; do not leave the record behind.
if (offer) {
void cancelPairOffer(deps.fetch, offer)
}
publishOffer(current, relayUrl)
return
}
return snapshot(data)
}),
current.offer = offer
current.codePath = offer ? 'ready' : 'unavailable'
emit()
})
/**
* "Yeniden dene" next to "Kodla bağlanma şu an kullanılamıyor": publishes
* the rendezvous record of the RUNNING pairing again. Same code, same QR,
* same expiry; nothing happens unless the code path is `unavailable`.
*/
pairRepublish: () =>
serial(async () => {
const data = await load()
if (session?.codePath === 'unavailable' && proxy?.pairingOpen() && session) {
publishOffer(session, data.relayId)
}
return snapshot(data)
}),
@@ -28,6 +28,7 @@ import {
createRelayClient,
encodeRelayFrame,
isRelayId,
MOBILE_NETWORK_RETRY_DELAY_MS,
parseRelaySession,
registerRelaySession,
RELAY_BACKOFF_MS,
@@ -925,10 +926,15 @@ describe('relay client', () => {
})
it('a failing registration is a fixed code (no upstream text) and is retried; renew() swaps the registration only when the new one is stored', async () => {
const h = clientHarness({ sessions: ['throw', 429, 500, 201, 500, 201] })
const h = clientHarness({ sessions: ['throw', 'throw', 429, 500, 201, 500, 201] })
h.client.start()
await h.settle()
// A thrown network error gets one more try 0.5 s later; until then nothing is reported as failed.
expect(h.client.state()).toEqual({ status: 'connecting', code: null })
h.fire(500)
await h.settle()
expect(h.fetches).toHaveLength(2)
expect(h.client.state()).toEqual({ status: 'failed', code: 'register_failed' })
expect(h.sockets).toHaveLength(0)
h.fire(2000)
@@ -982,6 +988,60 @@ describe('relay client', () => {
broken.stop()
})
it('registration: a thrown network error is tried once more after 0.5 s; an HTTP answer or a timeout is not', async () => {
const h = clientHarness({ sessions: ['throw', 201] })
const pending = h.client.ensureRegistered()
await h.settle()
expect(h.fetches).toHaveLength(1)
expect(h.timers.map(timer => timer.ms)).toEqual([MOBILE_NETWORK_RETRY_DELAY_MS])
expect(MOBILE_NETWORK_RETRY_DELAY_MS).toBe(500)
h.fire(500)
expect(await pending).toBe(RELAY_IDS[0])
expect(h.fetches).toHaveLength(2)
expect(h.fetches[1].init.signal).not.toBe(h.fetches[0].init.signal)
expect(h.saved).toEqual([{ credentials: { relayId: RELAY_IDS[0], hostToken: HOST_TOKENS[0] }, reissued: false }])
// The function itself: exactly one extra try, only for a thrown error.
const sleeps: number[] = []
const run = async (outcomes: (number | 'throw' | 'timeout')[]) => {
let calls = 0
const result = await registerRelaySession(
async () => {
const outcome = outcomes[calls++] ?? 201
if (outcome === 'throw') {
throw new TypeError('fetch failed')
}
if (outcome === 'timeout') {
throw new DOMException('The operation was aborted due to timeout', 'TimeoutError')
}
return { status: outcome, json: async () => ({ relayId: RELAY_IDS[0], hostToken: HOST_TOKENS[0] }) }
},
{
sleep: async ms => {
sleeps.push(ms)
}
}
)
return { result, calls }
}
expect(await run(['throw', 201])).toEqual({ result: { relayId: RELAY_IDS[0], hostToken: HOST_TOKENS[0] }, calls: 2 })
expect(await run(['throw', 'throw', 201])).toEqual({ result: { error: 'register_failed' }, calls: 2 })
expect(await run(['throw', 429])).toEqual({ result: { error: 'rate_limited' }, calls: 2 })
expect(sleeps).toEqual([500, 500, 500])
expect(await run([500, 201])).toEqual({ result: { error: 'register_failed' }, calls: 1 })
expect(await run([429, 201])).toEqual({ result: { error: 'rate_limited' }, calls: 1 })
expect(await run(['timeout', 201])).toEqual({ result: { error: 'register_failed' }, calls: 1 })
expect(sleeps).toHaveLength(3)
})
it('streams: opened / close, bytes copied untouched in ≤ 64 KiB frames, unknown ids ignored, stream and buffer limits', async () => {
const h = clientHarness({ stored: { relayId: RELAY_IDS[0], hostToken: HOST_TOKENS[0] } })
+62 -17
View File
@@ -92,24 +92,67 @@ export type RelayFailureCode =
| 'busy'
| 'timeout'
/** `POST /sessions`; a fixed failure code instead of any upstream text. */
export async function registerRelaySession(fetcher: RelayFetch): Promise<RelayCredentials | { error: RelayFailureCode }> {
try {
const response = await fetcher(`${MOBILE_RELAY_ORIGIN}${MOBILE_RELAY_PATH}/sessions`, {
method: 'POST',
headers: { 'Content-Type': 'application/json', Accept: 'application/json' },
body: '{}',
redirect: 'error',
signal: AbortSignal.timeout(MOBILE_RELAY_FETCH_TIMEOUT_MS)
})
/** Wait before the single second try of a request whose first try never got an answer. */
export const MOBILE_NETWORK_RETRY_DELAY_MS = 500
if (response.status !== 201) {
return { error: response.status === 429 ? 'rate_limited' : 'register_failed' }
/** An aborted / timed-out request already used its whole budget: it is not tried again. */
export function isTimeoutError(error: unknown): boolean {
const name = (error as null | { name?: unknown })?.name
return name === 'TimeoutError' || name === 'AbortError'
}
export interface NetworkRetryOptions {
/** Injected wait (tests: fake clock). Default: a real timer. */
sleep?: (ms: number) => Promise<void>
retryDelayMs?: number
}
const realSleep = (ms: number) => new Promise<void>(resolve => setTimeout(resolve, ms))
/**
* `POST /sessions`; a fixed failure code instead of any upstream text.
*
* A process that sat idle for a long time can fail its first request on a dead
* pooled keep-alive connection. So a THROWN network error (no HTTP answer at
* all) is tried once more after ~0.5 s. Any HTTP answer is final: no retry.
*/
export async function registerRelaySession(
fetcher: RelayFetch,
options: NetworkRetryOptions = {}
): Promise<RelayCredentials | { error: RelayFailureCode }> {
const sleep = options.sleep ?? realSleep
for (let attempt = 1; ; attempt++) {
let response: RelayFetchResponse
try {
response = await fetcher(`${MOBILE_RELAY_ORIGIN}${MOBILE_RELAY_PATH}/sessions`, {
method: 'POST',
headers: { 'Content-Type': 'application/json', Accept: 'application/json' },
body: '{}',
redirect: 'error',
signal: AbortSignal.timeout(MOBILE_RELAY_FETCH_TIMEOUT_MS)
})
} catch (error) {
if (attempt >= 2 || isTimeoutError(error)) {
return { error: 'register_failed' }
}
await sleep(options.retryDelayMs ?? MOBILE_NETWORK_RETRY_DELAY_MS)
continue
}
return parseRelaySession(await response.json()) ?? { error: 'register_failed' }
} catch {
return { error: 'register_failed' }
try {
if (response.status !== 201) {
return { error: response.status === 429 ? 'rate_limited' : 'register_failed' }
}
return parseRelaySession(await response.json()) ?? { error: 'register_failed' }
} catch {
return { error: 'register_failed' }
}
}
}
@@ -221,6 +264,8 @@ export function createRelayClient(deps: RelayClientDeps) {
let retryTimer: unknown = null
let silenceTimer: unknown = null
const streams = new Map<number, Stream>()
/** Same (injectable) clock as every other wait of this client. */
const sleep = (ms: number) => new Promise<void>(resolve => void setTimer(resolve, ms))
function setState(status: RelayStatus, code: null | RelayFailureCode = null) {
if (state.status === status && state.code === code) {
@@ -270,7 +315,7 @@ export function createRelayClient(deps: RelayClientDeps) {
return stored
}
const fresh = await registerRelaySession(deps.fetch)
const fresh = await registerRelaySession(deps.fetch, { sleep })
if ('error' in fresh) {
return fresh
@@ -626,7 +671,7 @@ export function createRelayClient(deps: RelayClientDeps) {
* only once the new one is stored; false = nothing changed.
*/
async renew(): Promise<boolean> {
const fresh = await registerRelaySession(deps.fetch)
const fresh = await registerRelaySession(deps.fetch, { sleep })
if ('error' in fresh) {
return false
+1 -1
View File
@@ -2,7 +2,7 @@
"name": "aiturk-ide",
"productName": "AITURK IDE",
"private": true,
"version": "1.0.0-beta.16",
"version": "1.0.0-beta.18",
"description": "TurkServis resmi yapay zekâ geliştirme ortamı. Hermes Agent tabanlıdır.",
"author": "AITURK / TurkServis",
"repository": {
+29
View File
@@ -1470,6 +1470,34 @@ def _wait_for_oneshot_background_completions(cli) -> None:
)
def _emit_approval_required(session_id) -> None:
"""Report approvals a one-shot run could not put in front of a human.
One ``approval_required: <json>`` line per blocked request on stderr,
next to ``session_id:``, so the process that launched this run can tell
"blocked, needs a human" from "still working" without parsing prose.
Nothing was applied for these requests; resuming the same session in an
interactive surface is the way to get them approved.
"""
try:
from tools.approval import drain_human_approval_required
records = drain_human_approval_required()
except Exception:
return
for record in records:
try:
if not record.get("session_id"):
record["session_id"] = session_id or ""
record["resume"] = "interactive"
print(
"approval_required: "
+ json.dumps(record, ensure_ascii=True, sort_keys=True),
file=sys.stderr,
)
except Exception:
continue
def _finalize_single_query(cli) -> None:
"""Close one-shot CLI resources before releasing the active session lease."""
try:
@@ -22360,6 +22388,7 @@ def main(
# Session ID goes to stderr so piped stdout is clean.
print(f"\nsession_id: {cli.session_id}", file=sys.stderr)
_emit_approval_required(cli.session_id)
# Ensure proper exit code for automation wrappers.
#
+1 -1
View File
@@ -65,7 +65,7 @@
},
"apps/desktop": {
"name": "aiturk-ide",
"version": "1.0.0-beta.16",
"version": "1.0.0-beta.18",
"license": "MIT",
"dependencies": {
"@assistant-ui/core": "0.2.23",
+1
View File
@@ -272,6 +272,7 @@ _HERMES_BEHAVIORAL_VARS = frozenset({
"HERMES_REAL_HOME",
"TERMINAL_HOME_MODE",
"HERMES_INTERACTIVE",
"HERMES_SINGLE_QUERY_SESSION",
"HERMES_QUIET",
"HERMES_TOOL_PROGRESS",
"HERMES_TOOL_PROGRESS_MODE",
@@ -0,0 +1,578 @@
"""Protected instruction-file gate when no human can be reached.
A one-shot run (``chat -q`` / ``--query-file``) installs the CLI approval
callback but never starts the prompt UI. The gate used to treat the
registered callback as a reachable human and wait out ``approvals.timeout``
on a panel nobody could see. These tests pin the fixed contract:
* no reachable human -> BLOCKED at once, nothing written, the outcome is
recorded for the launching process;
* the gate itself is unchanged: still asks a real human every time, still
ignores yolo and ``approvals.single_query_mode``;
* an approval that arrives after the calling execute_code script stopped
waiting is not applied.
"""
import json
import time
import pytest
@pytest.fixture(autouse=True)
def _gate_on(monkeypatch):
import tools.approval as A
import tools.file_tools as ft
monkeypatch.setattr(ft, "_protected_instruction_config", lambda: (True, []))
monkeypatch.delenv("HERMES_SINGLE_QUERY_SESSION", raising=False)
A.drain_human_approval_required()
A.set_approval_reply_deadline(None)
yield
A.drain_human_approval_required()
A.set_approval_reply_deadline(None)
@pytest.fixture
def approvals():
"""CLI approval callback that records calls and gives a scripted answer."""
from tools.terminal_tool import set_approval_callback
state = {"calls": [], "answer": "deny"}
def cb(command, description, **kwargs):
state["calls"].append({"command": command, **kwargs})
answer = state["answer"]
if callable(answer):
return answer()
return answer
set_approval_callback(cb)
yield state
set_approval_callback(None)
@pytest.fixture
def one_shot(monkeypatch):
monkeypatch.setenv("HERMES_SINGLE_QUERY_SESSION", "1")
def _write(path, content="injected"):
from tools.file_tools import write_file_tool
return json.loads(write_file_tool(str(path), content))
def _records():
import tools.approval as A
return A.drain_human_approval_required()
# ---- one-shot run: callback registered, nobody to answer ------------------
class TestOneShotRun:
def test_blocks_at_once_without_prompting(self, tmp_path, approvals, one_shot):
target = tmp_path / "AGENTS.md"
target.write_text("rules\n", encoding="utf-8")
approvals["answer"] = "once" # would be honored if anyone were asked
started = time.monotonic()
res = _write(target)
elapsed = time.monotonic() - started
assert res.get("error") and "BLOCKED" in res["error"]
assert "APPROVAL REQUIRED" in res["error"]
assert target.read_text(encoding="utf-8") == "rules\n"
assert approvals["calls"] == [], "prompted on a surface nobody can see"
assert elapsed < 5
def test_outcome_is_recorded_for_the_launcher(self, tmp_path, approvals, one_shot):
import tools.approval as A
target = tmp_path / "CLAUDE.md"
sid = A._approval_session_id.set("sess-123")
tcid = A._approval_tool_call_id.set("call-9")
try:
_write(target)
finally:
A._approval_tool_call_id.reset(tcid)
A._approval_session_id.reset(sid)
records = _records()
assert len(records) == 1
rec = records[0]
assert rec["kind"] == "protected_instruction_file"
assert rec["reason"] == "no_interactive_user"
assert rec["applied"] is False
assert rec["session_id"] == "sess-123"
assert rec["tool_call_id"] == "call-9"
assert rec["request_id"]
assert any("CLAUDE.md" in t for t in rec["targets"])
assert _records() == [], "drain must clear"
def test_yolo_does_not_open_the_gate(self, tmp_path, approvals, one_shot, monkeypatch):
import tools.approval as A
monkeypatch.setattr(A, "_YOLO_MODE_FROZEN", True)
target = tmp_path / "AGENTS.md"
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
def test_single_query_mode_approve_does_not_open_the_gate(
self, tmp_path, approvals, one_shot, monkeypatch
):
import tools.approval as A
real_cfg_get = A.cfg_get
def fake_cfg_get(config, *keys, default=None):
if keys == ("approvals", "single_query_mode"):
return "approve"
return real_cfg_get(config, *keys, default=default)
monkeypatch.setattr(A, "cfg_get", fake_cfg_get)
target = tmp_path / "SOUL.md"
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
def test_regular_file_still_writes(self, tmp_path, approvals, one_shot):
target = tmp_path / "notes.md"
res = _write(target, "hello")
assert not res.get("error"), res
assert target.read_text(encoding="utf-8") == "hello"
assert _records() == []
def test_multi_file_patch_applies_nothing(self, tmp_path, approvals, one_shot):
from tools.file_tools import patch_tool
agents = tmp_path / "AGENTS.md"
agents.write_text("rules\n", encoding="utf-8")
plain = tmp_path / "plain.txt"
plain.write_text("hello\n", encoding="utf-8")
patch = (
"*** Begin Patch\n"
f"*** Update File: {plain}\n"
"@@\n"
"-hello\n"
"+world\n"
f"*** Update File: {agents}\n"
"@@\n"
"-rules\n"
"+injected\n"
"*** End Patch"
)
res = json.loads(patch_tool(mode="patch", patch=patch))
assert res.get("error") and "BLOCKED" in res["error"]
assert plain.read_text(encoding="utf-8") == "hello\n"
assert agents.read_text(encoding="utf-8") == "rules\n"
assert approvals["calls"] == []
def test_retry_in_same_run_blocks_again(self, tmp_path, approvals, one_shot):
target = tmp_path / "AGENTS.md"
_write(target)
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
assert len(_records()) == 2
# ---- no channel at all ----------------------------------------------------
def test_no_channel_fails_closed_and_is_recorded(tmp_path):
target = tmp_path / "AGENTS.md"
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
records = _records()
assert [r["reason"] for r in records] == ["no_approval_channel"]
def test_undeliverable_gateway_request_is_recorded(tmp_path):
import tools.approval as A
session_key = "protected-unattended-undelivered"
token = A.set_current_session_key(session_key)
try:
def notify(approval_data):
raise RuntimeError("client went away")
A.register_gateway_notify(session_key, notify)
try:
target = tmp_path / "AGENTS.md"
res = _write(target)
finally:
A.unregister_gateway_notify(session_key)
finally:
A.reset_current_session_key(token)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
assert [r["reason"] for r in _records()] == ["approval_not_delivered"]
# ---- a reachable human: the gate behaves exactly as before -----------------
class TestReachableHuman:
def test_allow_writes(self, tmp_path, approvals):
target = tmp_path / "AGENTS.md"
approvals["answer"] = "once"
res = _write(target, "approved")
assert not res.get("error"), res
assert target.read_text(encoding="utf-8") == "approved"
assert len(approvals["calls"]) == 1
assert _records() == []
@pytest.mark.parametrize("answer", ["deny", "timeout"])
def test_deny_and_silence_block(self, tmp_path, approvals, answer):
target = tmp_path / "AGENTS.md"
approvals["answer"] = answer
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
assert len(approvals["calls"]) == 1
# A human was asked and did not consent: not an "unattended" outcome.
assert _records() == []
def test_prompt_that_dies_blocks(self, tmp_path, approvals):
"""Cancel / lost surface: the prompt raising is not consent."""
def boom():
raise RuntimeError("prompt surface lost")
target = tmp_path / "AGENTS.md"
approvals["answer"] = boom
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
# ---- approval that arrives after the caller stopped waiting ----------------
class TestLateApproval:
def test_late_cli_approval_is_not_applied(self, tmp_path, approvals):
import tools.approval as A
target = tmp_path / "AGENTS.md"
target.write_text("rules\n", encoding="utf-8")
def answer_after_deadline():
# The human says yes, but only once the execute_code script that
# asked for this write has already timed out.
A.set_approval_reply_deadline(time.monotonic() - 1)
return "once"
A.set_approval_reply_deadline(time.monotonic() + 300)
approvals["answer"] = answer_after_deadline
res = _write(target)
assert res.get("error") and "BLOCKED" in res["error"]
assert "NOT applied" in res["error"]
assert target.read_text(encoding="utf-8") == "rules\n"
def test_late_approval_does_not_carry_to_the_next_request(self, tmp_path, approvals):
import tools.approval as A
target = tmp_path / "AGENTS.md"
A.set_approval_reply_deadline(time.monotonic() - 1)
approvals["answer"] = "once"
assert "BLOCKED" in _write(target).get("error", "")
assert not target.exists()
# A fresh call (new deadline) asks again; the old yes is not reused.
A.set_approval_reply_deadline(time.monotonic() + 300)
approvals["answer"] = "deny"
assert "BLOCKED" in _write(target).get("error", "")
assert not target.exists()
assert len(approvals["calls"]) == 2
def test_in_time_approval_is_applied(self, tmp_path, approvals):
import tools.approval as A
target = tmp_path / "AGENTS.md"
A.set_approval_reply_deadline(time.monotonic() + 300)
approvals["answer"] = "once"
res = _write(target, "approved")
assert not res.get("error"), res
assert target.read_text(encoding="utf-8") == "approved"
def test_late_gateway_approval_is_not_applied(self, tmp_path):
import tools.approval as A
session_key = "protected-unattended-late"
token = A.set_current_session_key(session_key)
try:
def notify(approval_data):
A.resolve_gateway_approval(session_key, "once")
A.register_gateway_notify(session_key, notify)
try:
A.set_approval_reply_deadline(time.monotonic() - 1)
target = tmp_path / "AGENTS.md"
res = _write(target)
finally:
A.unregister_gateway_notify(session_key)
finally:
A.reset_current_session_key(token)
assert res.get("error") and "NOT applied" in res["error"]
assert not target.exists()
def test_no_deadline_outside_execute_code(self):
import tools.approval as A
assert A.approval_reply_window_remaining() is None
assert A.approval_arrived_too_late() is False
def test_rpc_dispatch_binds_and_restores_the_deadline(self):
import tools.approval as A
import tools.code_execution_tool as cet
start = time.monotonic()
previous = cet._bind_rpc_reply_deadline(start)
try:
remaining = A.approval_reply_window_remaining()
assert remaining is not None
assert 0 < remaining <= cet._RPC_REPLY_TIMEOUT_SECONDS
finally:
cet._restore_rpc_reply_deadline(previous)
assert A.approval_reply_window_remaining() is None
def test_deadline_is_per_thread(self):
import threading
import tools.approval as A
A.set_approval_reply_deadline(time.monotonic() - 1)
seen = {}
def other():
seen["late"] = A.approval_arrived_too_late()
worker = threading.Thread(target=other)
worker.start()
worker.join()
assert seen["late"] is False
assert A.approval_arrived_too_late() is True
# ---- what the launching process sees ---------------------------------------
def test_recorded_outcomes_are_bounded():
import tools.approval as A
for i in range(A._HUMAN_APPROVAL_REQUIRED_MAX + 20):
A.note_human_approval_required("protected_instruction_file", [f"f{i}"], "x")
records = A.drain_human_approval_required()
assert len(records) == A._HUMAN_APPROVAL_REQUIRED_MAX
assert records[-1]["targets"] == [f"f{A._HUMAN_APPROVAL_REQUIRED_MAX + 19}"]
def test_one_shot_exit_reports_approval_required_on_stderr(capsys):
import cli as cli_module
import tools.approval as A
A.note_human_approval_required(
"protected_instruction_file", ["AGENTS.md"], "no_interactive_user"
)
cli_module._emit_approval_required("sess-42")
err = capsys.readouterr().err
lines = [l for l in err.splitlines() if l.startswith("approval_required: ")]
assert len(lines) == 1
payload = json.loads(lines[0].split(": ", 1)[1])
assert payload["session_id"] == "sess-42"
assert payload["applied"] is False
assert payload["reason"] == "no_interactive_user"
assert payload["resume"] == "interactive"
assert payload["targets"] == ["AGENTS.md"]
cli_module._emit_approval_required("sess-42")
assert "approval_required" not in capsys.readouterr().err
# ---- execute_code RPC: the dispatcher really binds the window ---------------
def _drive_rpc_once(dispatch, stop_event=None):
"""Send one tool call through the real RPC server loop; return its reply."""
import socket
import threading
import tools.code_execution_tool as cet
server = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
server.bind(("127.0.0.1", 0))
server.listen(1)
stop_event = stop_event or threading.Event()
worker = threading.Thread(
target=cet._rpc_server_loop,
args=(server, "task", [], [0], 5, frozenset({"patch"}), stop_event, "tok", dispatch),
daemon=True,
)
worker.start()
client = socket.create_connection(server.getsockname(), timeout=10)
try:
client.sendall(
(json.dumps({"tool": "patch", "args": {}, "token": "tok"}) + "\n").encode()
)
buf = b""
while not buf.endswith(b"\n"):
chunk = client.recv(65536)
if not chunk:
break
buf += chunk
finally:
client.close()
stop_event.set()
worker.join(timeout=10)
server.close()
return buf.decode().strip()
def test_rpc_server_loop_binds_reply_window_around_dispatch():
import tools.approval as A
import tools.code_execution_tool as cet
seen = {}
def dispatch(tool_name, tool_args):
seen["remaining"] = A.approval_reply_window_remaining()
seen["late"] = A.approval_arrived_too_late()
return json.dumps({"ok": True})
reply = _drive_rpc_once(dispatch)
assert json.loads(reply) == {"ok": True}
assert seen["late"] is False
assert 0 < seen["remaining"] <= cet._RPC_REPLY_TIMEOUT_SECONDS
def test_rpc_approval_after_script_ended_is_late():
"""Script killed by its own run timeout while the prompt is still open."""
import threading
import tools.approval as A
stop_event = threading.Event()
seen = {}
def dispatch(tool_name, tool_args):
seen["before"] = A.approval_arrived_too_late()
stop_event.set() # the script ended while a human was deciding
seen["after"] = A.approval_arrived_too_late()
return json.dumps({"ok": True})
_drive_rpc_once(dispatch, stop_event)
assert seen == {"before": False, "after": True}
def test_protected_write_through_rpc_is_dropped_when_script_ended(tmp_path, approvals):
"""End to end through the RPC loop: late yes -> BLOCKED, file untouched."""
import threading
from tools.file_tools import write_file_tool
from tools.terminal_tool import _get_approval_callback, set_approval_callback
target = tmp_path / "AGENTS.md"
target.write_text("rules\n", encoding="utf-8")
stop_event = threading.Event()
callback = _get_approval_callback()
def answer_after_script_died():
stop_event.set()
return "once"
approvals["answer"] = answer_after_script_died
def dispatch(tool_name, tool_args):
set_approval_callback(callback) # the RPC thread carries the cell's callback
try:
return write_file_tool(str(target), "injected")
finally:
set_approval_callback(None)
reply = json.loads(_drive_rpc_once(dispatch, stop_event))
assert "NOT applied" in reply.get("error", "")
assert target.read_text(encoding="utf-8") == "rules\n"
assert len(approvals["calls"]) == 1
def test_retired_kernel_cell_counts_as_gone():
import tools.approval as A
cell = {"active": True}
previous = A.also_gone_when(lambda: not cell["active"])
try:
assert A.approval_arrived_too_late() is False
cell["active"] = False
assert A.approval_arrived_too_late() is True
finally:
A.restore_approval_reply_deadline(previous)
assert A.approval_arrived_too_late() is False
# ---- an automatic approver is not a person ----------------------------------
def test_subagent_auto_approve_cannot_open_the_gate(tmp_path):
from tools.delegate_tool import _subagent_auto_approve
from tools.terminal_tool import set_approval_callback
target = tmp_path / "AGENTS.md"
set_approval_callback(_subagent_auto_approve)
try:
res = _write(target)
finally:
set_approval_callback(None)
assert res.get("error") and "BLOCKED" in res["error"]
assert not target.exists()
assert [r["reason"] for r in _records()] == ["no_interactive_user"]
def test_blocked_message_survives_braces_in_target(tmp_path):
import tools.file_tools as ft
res = ft._request_protected_instruction_approval(["{be}/AGENTS.md"])
assert res.startswith("BLOCKED") and "{be}/AGENTS.md" in res
# ---- sibling gates in a one-shot run ----------------------------------------
def test_computer_use_does_not_wait_in_one_shot_run(one_shot, monkeypatch):
import tools.computer_use.tool as cu
calls = []
monkeypatch.setattr(cu, "_approval_callback", lambda *a, **k: calls.append(a) or "approve_once")
res = cu._request_approval("click", {}, session_id="unattended-test")
assert res is not None and "one-shot run" in json.loads(res)["error"]
assert calls == []
def test_computer_use_still_asks_when_interactive(monkeypatch):
import tools.computer_use.tool as cu
calls = []
monkeypatch.setattr(cu, "_approval_callback", lambda *a, **k: calls.append(a) or "approve_once")
assert cu._request_approval("click", {}, session_id="attended-test") is None
assert len(calls) == 1
def test_memory_write_prompt_unavailable_in_one_shot_run(approvals, monkeypatch):
import tools.write_approval as wa
assert wa._interactive_approval_available() is True
monkeypatch.setenv("HERMES_SINGLE_QUERY_SESSION", "1")
assert wa._interactive_approval_available() is False
def test_one_shot_marker_is_not_inherited_by_child_processes(one_shot):
from tools.environments.local import _make_run_env
assert "HERMES_SINGLE_QUERY_SESSION" not in _make_run_env({})
+109
View File
@@ -78,6 +78,115 @@ _hermes_interactive_ctx: contextvars.ContextVar[Optional[str]] = contextvars.Con
)
# Absolute ``time.monotonic()`` instant after which whoever issued the
# current tool call has stopped waiting for its result. Set by the
# execute_code RPC dispatchers, whose sandbox client gives up after a fixed
# reply timeout. An approval that resolves after this instant must NOT be
# acted on: the caller already reported the call as failed, so applying it
# would be a write the agent believes never happened (and may redo).
# The reply timeout is not the only way a caller goes away: the script can
# hit its own run timeout or be interrupted first. ``gone`` is a probe the
# dispatcher supplies for that ("has my script/cell already ended?").
# Thread-local on purpose: the dispatcher and the tool handler it calls run
# on the same thread, whatever context the handler is re-bound to.
_approval_reply_deadline = threading.local()
def set_approval_reply_deadline(deadline: Optional[float], gone=None):
"""Bind the caller's reply window for this thread; return the prior binding.
``deadline`` is monotonic seconds (None = unbounded); ``gone`` is an
optional zero-argument callable that turns true once the caller has
ended. Pass the returned value to :func:`restore_approval_reply_deadline`.
"""
previous = (
getattr(_approval_reply_deadline, "value", None),
getattr(_approval_reply_deadline, "gone", None),
)
_approval_reply_deadline.value = deadline
_approval_reply_deadline.gone = gone
return previous
def restore_approval_reply_deadline(previous) -> None:
"""Undo :func:`set_approval_reply_deadline` / :func:`also_gone_when`."""
deadline, gone = previous if previous else (None, None)
_approval_reply_deadline.value = deadline
_approval_reply_deadline.gone = gone
def also_gone_when(probe):
"""Add another "caller has ended" probe, keeping the current window."""
deadline = getattr(_approval_reply_deadline, "value", None)
current = getattr(_approval_reply_deadline, "gone", None)
def gone() -> bool:
return bool(probe()) or bool(current and current())
return set_approval_reply_deadline(deadline, gone)
def approval_reply_window_remaining() -> Optional[float]:
"""Seconds left before the caller stops waiting, or None when unbounded."""
deadline = getattr(_approval_reply_deadline, "value", None)
if deadline is None:
return None
return deadline - time.monotonic()
def approval_arrived_too_late() -> bool:
"""True when the caller of the current tool call has stopped waiting."""
gone = getattr(_approval_reply_deadline, "gone", None)
if gone is not None:
try:
if gone():
return True
except Exception:
pass
remaining = approval_reply_window_remaining()
return remaining is not None and remaining <= 0
# Approvals that could not be put in front of a human because the run has
# no interactive channel (single-query ``chat -Q`` stage runs, scripts).
# The single-query CLI reports them on stderr as ``approval_required:``
# lines so the coordinating process gets a machine-readable outcome
# instead of inferring one from prose. Bounded: long-lived processes that
# never drain it must not grow without limit.
_human_approval_required: list[dict] = []
_HUMAN_APPROVAL_REQUIRED_MAX = 50
_human_approval_required_lock = threading.Lock()
def note_human_approval_required(kind: str, targets: list[str], reason: str) -> dict:
"""Record an approval the run could not present to a human.
Captures the active session/tool-call identity so the coordinator can
tie the outcome to the exact job and resume the same session.
"""
record = {
"kind": kind,
"targets": list(dict.fromkeys(str(t) for t in targets)),
"reason": reason,
"request_id": uuid.uuid4().hex,
"tool_call_id": _approval_tool_call_id.get() or "",
"session_id": _approval_session_id.get() or os.getenv("HERMES_SESSION_ID", ""),
"applied": False,
}
with _human_approval_required_lock:
_human_approval_required.append(record)
del _human_approval_required[:-_HUMAN_APPROVAL_REQUIRED_MAX]
return dict(record)
def drain_human_approval_required() -> list[dict]:
"""Return and clear the recorded approval-required outcomes."""
with _human_approval_required_lock:
records = list(_human_approval_required)
_human_approval_required.clear()
return records
def set_hermes_interactive_context(interactive: bool) -> contextvars.Token:
"""Bind interactive mode for the current context (thread or asyncio task).
+39
View File
@@ -716,6 +716,36 @@ def _call(tool_name, args):
# RPC server (runs in a thread inside the parent process)
# ---------------------------------------------------------------------------
# How long the generated sandbox client waits for one tool-call reply
# (``_sock.settimeout(300)`` / the file transport's 300s poll deadline).
_RPC_REPLY_TIMEOUT_SECONDS = 300
def _bind_rpc_reply_deadline(call_start: float, stop_event=None):
"""Tell the approval gates when this call's requester stops waiting.
The requester is gone at the reply timeout, or earlier when the script
itself ended (run timeout, interrupt): ``stop_event`` is set then.
"""
try:
from tools.approval import set_approval_reply_deadline
gone = stop_event.is_set if stop_event is not None else None
return (True, set_approval_reply_deadline(
call_start + _RPC_REPLY_TIMEOUT_SECONDS, gone))
except Exception:
return (False, None)
def _restore_rpc_reply_deadline(previous) -> None:
if not previous or not previous[0]:
return
try:
from tools.approval import restore_approval_reply_deadline
restore_approval_reply_deadline(previous[1])
except Exception:
pass
# Terminal parameters that must not be used from ephemeral sandbox scripts
_TERMINAL_BLOCKED_PARAMS = {"background", "pty", "notify", "notify_on_complete", "watch_patterns"}
@@ -826,12 +856,18 @@ def _rpc_server_loop(
# Dispatch through the standard tool handler.
# Suppress stdout/stderr from internal tool handlers so
# their status prints don't leak into the CLI spinner.
# The sandbox client stops waiting for this reply after
# _RPC_REPLY_TIMEOUT_SECONDS; an approval granted later
# must not apply a write the script already saw fail.
_prev_deadline = _bind_rpc_reply_deadline(call_start, stop_event)
try:
with thread_scoped_silence():
result = dispatch(tool_name, tool_args)
except Exception as exc:
logger.error("Tool call failed in sandbox: %s", exc, exc_info=True)
result = tool_error(str(exc))
finally:
_restore_rpc_reply_deadline(_prev_deadline)
tool_call_counter[0] += 1
call_duration = time.monotonic() - call_start
@@ -1101,6 +1137,7 @@ def _rpc_poll_loop(
tool_args.pop(param, None)
# Dispatch through the standard tool handler
_prev_deadline = _bind_rpc_reply_deadline(call_start, stop_event)
try:
with thread_scoped_silence():
tool_result = handle_function_call(
@@ -1110,6 +1147,8 @@ def _rpc_poll_loop(
logger.error("Tool call failed in remote sandbox: %s",
exc, exc_info=True)
tool_result = tool_error(str(exc))
finally:
_restore_rpc_reply_deadline(_prev_deadline)
tool_call_counter[0] += 1
call_duration = time.monotonic() - call_start
+16
View File
@@ -322,9 +322,25 @@ class CellAuthority:
set_sudo(self._sudo_cb)
except Exception:
previous = None
# A cell that settled (timeout, interrupt, error) no longer waits for
# this call: an approval granted afterwards must not be applied.
reply_window = None
try:
from tools.approval import also_gone_when
reply_window = (also_gone_when(lambda: not self.active),)
except Exception:
reply_window = None
try:
return handle_function_call(tool_name, tool_args, task_id=self.task_id)
finally:
if reply_window is not None:
try:
from tools.approval import restore_approval_reply_deadline
restore_approval_reply_deadline(reply_window[0])
except Exception:
pass
if previous is not None and self._callback_setters is not None:
set_approval, set_sudo = self._callback_setters
try:
+16
View File
@@ -613,6 +613,22 @@ def _request_approval(action: str, args: Dict[str, Any],
# No CLI approval wired — default allow. Gateway approval is handled
# one layer out via the normal tool-approval infra.
return None
try:
from tools.approval import _is_single_query_approval_context
unattended = _is_single_query_approval_context()
except Exception:
unattended = False
if unattended:
# One-shot run: the prompt callback is installed but its panel is
# never shown, so asking would only wait out approvals.timeout.
return json.dumps({
"error": (
"approval required but this is a one-shot run with no "
"interactive user to approve it. Silence is not consent; "
"do not retry without the user."
),
"action": action,
})
summary = _summarize_action(action, args)
try:
verdict = cb(action, args, summary)
+5
View File
@@ -102,6 +102,11 @@ def _subagent_auto_approve(command: str, description: str, **kwargs) -> str:
return "once"
# Not a person: gates that require real human consent every time (protected
# instruction files) must not accept this callback's answer.
_subagent_auto_approve.unattended = True
def _get_subagent_approval_callback():
"""Return the callback to install into subagent worker threads.
+4
View File
@@ -1546,6 +1546,10 @@ def _make_run_env(env: dict) -> dict:
run_env[real_key] = v
elif _is_hermes_internal_secret(k):
continue
elif k == "HERMES_SINGLE_QUERY_SESSION":
# Describes this process only. A child (an interactive hermes
# started from a one-shot run's terminal) sets its own.
continue
else:
first_party = _is_terminal_first_party_env(k)
passthrough = _is_passthrough(k)
+79 -14
View File
@@ -861,12 +861,18 @@ def _request_protected_instruction_approval(
"required (not bypassed by auto-approve)."
)
display = f"<write to {targets}>"
blocked = (
f"BLOCKED: write to protected agent-instruction file(s) ({targets}) "
"{why} The user has NOT consented to this write. Do NOT retry it or "
"attempt the same edit via another path (terminal, execute_code, "
"etc.)."
)
class _Blocked:
# Not str.format: a target path may itself contain braces.
@staticmethod
def format(why: str) -> str:
return (
"BLOCKED: write to protected agent-instruction file(s) "
f"({targets}) {why} The user has NOT consented to this "
"write. Do NOT retry it or attempt the same edit via "
"another path (terminal, execute_code, etc.)."
)
blocked = _Blocked
try:
import tools.approval as _approval
@@ -874,6 +880,41 @@ def _request_protected_instruction_approval(
return blocked.format(why="requires approval but the approval "
"subsystem is unavailable.")
def _needs_human(reason: str, why: str) -> str:
# No human could be shown this request. Record it so a one-shot
# run can hand its coordinator a machine-readable outcome, and tell
# the agent where a human CAN approve it instead of waiting on a
# prompt nobody sees.
try:
_approval.note_human_approval_required(
"protected_instruction_file", list(reasons), reason)
except Exception:
pass
return blocked.format(
why=why + " APPROVAL REQUIRED: a human must approve this from "
"an interactive session (resume this same session in "
"the desktop app or an interactive terminal and "
"request the write again there).")
def _too_late() -> str | None:
# The approval came back after whoever issued this call stopped
# waiting (execute_code RPC reply timeout). The caller already saw
# a failure, so applying the write now would be an edit the agent
# believes never happened.
try:
late = _approval.approval_arrived_too_late()
except Exception:
late = False
if not late:
return None
logger.warning(
"Approval for protected instruction file(s) %s arrived after "
"the caller stopped waiting; write not applied.", targets)
return blocked.format(
why="was approved only after the calling script had stopped "
"waiting for the result, so the write was NOT applied. "
"The approval is spent; a new request needs a new approval.")
# Gateway surface: block on the button round-trip when a notify callback
# is registered for this session (Telegram/Discord/Slack). One-operation
# only — no session/permanent buttons are offered.
@@ -898,14 +939,15 @@ def _request_protected_instruction_approval(
session_key, notify_cb, approval_data, surface="gateway",
)
if decision.get("notify_failed"):
return blocked.format(
why="requires approval but the approval request could not "
"be delivered.")
return _needs_human(
"approval_not_delivered",
"requires approval but the approval request could not "
"be delivered.")
choice = decision.get("choice")
if decision.get("resolved") and choice in {"once", "session", "always"}:
# One-operation grant regardless of the tapped scope — nothing
# is persisted for this gate.
return None
return _too_late()
if not decision.get("resolved"):
return blocked.format(
why="approval prompt timed out without a user response. "
@@ -921,6 +963,28 @@ def _request_protected_instruction_approval(
callback = None
if callback is not None:
# A one-shot run (``chat -q`` / ``--query-file``) installs the
# prompt callback but never starts the prompt UI: the panel would
# be painted nowhere and the wait would last the whole
# approvals.timeout with nobody able to answer. A registered
# callback is not a reachable human. Fail closed NOW;
# approvals.single_query_mode does not apply to this gate.
try:
unattended = _approval._is_single_query_approval_context()
except Exception:
unattended = False
if unattended:
return _needs_human(
"no_interactive_user",
"requires approval but this is a one-shot run with no "
"interactive user to approve it.")
if getattr(callback, "unattended", False):
# An auto-answering callback (subagent auto-approve) is not a
# person; this gate takes consent from a human only.
return _needs_human(
"no_interactive_user",
"requires approval but only an automatic approver is "
"attached to this run, not a person.")
choice = _approval.prompt_dangerous_approval(
display, description,
allow_permanent=False,
@@ -929,7 +993,7 @@ def _request_protected_instruction_approval(
)
if choice in {"once", "session", "always"}:
# One-operation grant; never persisted (see docstring).
return None
return _too_late()
if choice == "timeout":
return blocked.format(
why="approval prompt timed out without a user response. "
@@ -938,9 +1002,10 @@ def _request_protected_instruction_approval(
# No human channel at all (script, cron, background thread): fail
# closed. Auto-approving here would recreate the persistence vector.
return blocked.format(
why="requires approval but no interactive user or gateway is "
"present to approve it.")
return _needs_human(
"no_approval_channel",
"requires approval but no interactive user or gateway is "
"present to approve it.")
def _check_protected_instruction_write(paths: list[str],
+8
View File
@@ -327,6 +327,14 @@ def _interactive_approval_available() -> bool:
review affordance (``/memory pending``) instead.
* Scripts, cron, and background threads — no user present.
"""
try:
from tools.approval import _is_single_query_approval_context
if _is_single_query_approval_context():
# One-shot run: the callback exists but its panel is never
# shown, so an inline prompt would wait with nobody to answer.
return False
except Exception:
pass
try:
from tools.terminal_tool import _get_approval_callback
return _get_approval_callback() is not None