The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.
TypeScript 3d-modelingagent-frameworkagent-harnessai-agentscadclaude-codecodexcoding-agentsdeveloper-toolsesp32game-developmentmcp-servermulti-agentopen-hardwareopencodeopenharnesspcbroboticsterminaltmux
Continue Harness Codex resumes in restricted mode (#3406)
harness-for-autonomous[bot]e052e399e41c
diff --git a/cli/src/engines/codex/openRestricted.spec.ts b/cli/src/engines/codex/openRestricted.spec.tsindex 592147c7c..a70a4839a 100644--- a/cli/src/engines/codex/openRestricted.spec.ts+++ b/cli/src/engines/codex/openRestricted.spec.ts@@ -2,29 +2,33 @@ import { readFileSync } from 'node:fs'import { describe, expect, it } from 'vitest'import { openRestrictedKeys, screen } from './screen.js'-/** As codex-cli 0.165.0 drew them in a tmux pane (2026-11-05, task #3009): a Team-written start's empty engine home with no- * entry for its checkout, and the same home naming the checkout `untrusted`. */+/** Real tmux captures: 0.165 on 2026-11-05 (#3009), and 0.166 on Linux on 2026-11-06 (#3904). */const fixture = (name: string) => readFileSync(new URL(`../../superTabs/fixtures/screens/${name}.txt`, import.meta.url), 'utf8')-const RESTRICTED = fixture('codex-0165-open-restricted')const TRUST = fixture('codex-0165-trust-fresh-home')// "Open restricted" keeps the folder's config, hooks and exec policies off and saves no trust: the answer an untrusted// Team-written branch gets. "Trust and continue" would trust the branch, so nothing is ever sent on that screen.-describe('opening an untrusted folder restricted in Codex', () => {+describe.each([+ { version: '0165', secondRow: 'Quit', escape: 'quit' },+ { version: '0166', secondRow: 'Back to Agent Command Center', escape: 'back' },+])('opening an untrusted folder restricted in Codex $version', ({ version, secondRow, escape }) => {+ const RESTRICTED = fixture(`codex-${version}-open-restricted`)+ const footer = `enter continue · esc ${escape}`it('answers the restricted screen with Enter on its highlighted first row', () => {expect(openRestrictedKeys(RESTRICTED)).toEqual(['Enter'])})it.each([['the trust screen, whose first row trusts the folder', TRUST],- ['Quit highlighted instead', RESTRICTED.replace('› 1. Open restricted', ' 1. Open restricted').replace(' 2. Quit', '› 2. Quit')],- ['rows in another order', RESTRICTED.replace('› 1. Open restricted\n 2. Quit', '› 1. Quit\n 2. Open restricted')],- ['an extra row', RESTRICTED.replace(' 2. Quit', ' 2. Quit\n 3. Trust and continue')],- ['a missing Quit row', RESTRICTED.replace(' 2. Quit\n', '')],+ ['the second row highlighted instead', RESTRICTED.replace('› 1. Open restricted', ' 1. Open restricted').replace(` 2. ${secondRow}`, `› 2. ${secondRow}`)],+ ['rows in another order', RESTRICTED.replace(`› 1. Open restricted\n 2. ${secondRow}`, `› 1. ${secondRow}\n 2. Open restricted`)],+ ['an extra row', RESTRICTED.replace(` 2. ${secondRow}`, ` 2. ${secondRow}\n 3. Trust and continue`)],+ ['a missing second row', RESTRICTED.replace(` 2. ${secondRow}\n`, '')],['no sentence saying untrusted folders stay disabled', RESTRICTED.replace('Config, hooks, and exec policies from untrusted folders stay disabled.', 'Folder settings will apply.')],['a trust question on the same screen', `${TRUST}\n${RESTRICTED}`],- ['another footer', RESTRICTED.replace('enter continue · esc quit', 'enter continue · esc skip')],- ['no footer', RESTRICTED.replace('\n\n enter continue · esc quit', '')],+ ['another footer', RESTRICTED.replace(footer, 'enter continue · esc skip')],+ ['the other version\'s footer', RESTRICTED.replace(footer, `enter continue · esc ${escape === 'quit' ? 'back' : 'quit'}`)],+ ['no footer', RESTRICTED.replace(`\n\n ${footer}`, '')],['the screen left in scrollback above the composer', `${RESTRICTED}\n\n› Write tests\n gpt-6-astra high · ~/project\n`],])('sends nothing on %s', (_name, capture) => {expect(openRestrictedKeys(capture)).toBeNull()diff --git a/cli/src/engines/codex/screen.ts b/cli/src/engines/codex/screen.tsindex bbf039272..7164ab4cc 100644--- a/cli/src/engines/codex/screen.ts+++ b/cli/src/engines/codex/screen.ts@@ -49,8 +49,8 @@ export function updateDismissKeys(capture: string): string[] | null {}/**- * The key that opens an untrusted folder restricted, or null unless the capture is exactly Codex's screen for it. Only for a- * Team-written Swarm start (superTabs/machines.ts), whose engine home names its checkout untrusted (superTabs/start.ts): in+ * The key that opens an untrusted folder restricted, or null unless the capture is exactly Codex's screen for it. For a+ * Swarm start or resume, whose engine home names its checkout untrusted (superTabs/start.ts): in* a home that names nothing, codex-cli 0.165 asks "Trust this folder?" instead, nobody answers it, and every such start* timed out at 3 minutes (task #3009). "Open restricted" keeps the folder's config, hooks and exec policies off and saves no* trust, which is what that home is for. It is Enter on the highlighted first row, so the screen must say so word for word:@@ -63,8 +63,12 @@ export function openRestrictedKeys(capture: string): string[] | null {const text = lines.join(' ')if (/Trust this folder\?|Do you trust the contents/.test(text) || !text.includes(RESTRICTED_SENTENCE)) return nullconst rows = lines.filter(line => /^[›>]?\s*\d+\./.test(line))- const menu = rows.length === 2 && /^[›>]\s*1\. Open restricted$/.test(rows[0]) && rows[1] === '2. Quit'- return menu && lines[lines.length - 1] === 'enter continue · esc quit' ? ['Enter'] : null+ if (rows.length !== 2 || !/^[›>]\s*1\. Open restricted$/.test(rows[0])) return null+ const footer = lines[lines.length - 1]+ // 0.166 returns to Agent Command Center instead of quitting (#3904); both screens open restricted with Enter.+ const menu = (rows[1] === '2. Quit' && footer === 'enter continue · esc quit')+ || (rows[1] === '2. Back to Agent Command Center' && footer === 'enter continue · esc back')+ return menu ? ['Enter'] : null}const RESTRICTED_SENTENCE = 'Config, hooks, and exec policies from untrusted folders stay disabled.'diff --git a/cli/src/superTabs/fixtures/screens/codex-0166-open-restricted.txt b/cli/src/superTabs/fixtures/screens/codex-0166-open-restricted.txtnew file mode 100644index 000000000..9cf235151--- /dev/null+++ b/cli/src/superTabs/fixtures/screens/codex-0166-open-restricted.txt@@ -0,0 +1,11 @@++ Folder access+ /tmp/throwaway/repo++ Config, hooks, and exec policies from untrusted folders stay disabled. Trusted project folders can still contribute+ settings. Skills still load, and tools follow your permission settings. Opening will not change saved trust.++› 1. Open restricted+ 2. Back to Agent Command Center++ enter continue · esc backdiff --git a/cli/src/superTabs/machines.spec.ts b/cli/src/superTabs/machines.spec.tsindex 64c23e585..20495c24b 100644--- a/cli/src/superTabs/machines.spec.ts+++ b/cli/src/superTabs/machines.spec.ts@@ -2957,11 +2957,12 @@ describe('a launched start whose engine stops on a screen', () => {/** Two passes, 16 s apart: a screen fails a start only when two reads in a row show it the same (security review 1 of #3063). */const twice = async (f: ReturnType<typeof fixture>) => { await f.machine.tick(); f.advance(); await f.machine.tick() }- it('opens a Team-written Codex start\'s untrusted checkout restricted, with Enter on that exact screen alone', async () => {- const { f, screen, inputIfScreen, failures } = launched(RESTRICTED, 'trust_open')+ it.each(['0165', '0166'])('opens a Team-written Codex %s start\'s untrusted checkout restricted, with Enter on that exact screen alone', async version => {+ const restricted = screens(`codex-${version}-open-restricted`)+ const { f, screen, inputIfScreen, failures } = launched(restricted, 'trust_open')await f.machine.tick()expect(screen).toHaveBeenCalledWith('launched', 0)- expect(inputIfScreen).toHaveBeenCalledWith('launched', ['Enter'], RESTRICTED)+ expect(inputIfScreen).toHaveBeenCalledWith('launched', ['Enter'], restricted)expect(failures()).toEqual([])})diff --git a/cli/src/superTabs/terminal.spec.ts b/cli/src/superTabs/terminal.spec.tsindex 2add33c98..b4de85cd6 100644--- a/cli/src/superTabs/terminal.spec.ts+++ b/cli/src/superTabs/terminal.spec.ts@@ -1,4 +1,11 @@-import { readFileSync } from 'node:fs'+import { mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'+import { spawn } from 'node:child_process'+import { createInterface } from 'node:readline'+import { createSuperTabTerminal } from './inCore/terminal.js'+import { screen as codexScreen } from '../engines/codex/screen.js'+import type { RegisteredSession } from '../lib/registry.js'+import { tmpdir } from 'node:os'+import { join } from 'node:path'import { describe, expect, it, vi } from 'vitest'import type { CoreApi, SuperTabScreen } from '../core/api.js'import { fakeCore } from '../testing/fakeCore.js'@@ -651,3 +658,81 @@ describe('reading a model that keeps failing off a screen', () => {expect(modelFailing(`I will make the client say "Unable to connect to API" when the relay is down.\n${'line\n'.repeat(30)}> `)).toBe(false)})})++// #3904: a resumed Swarm Codex already has an identity, so startup settlement never sees its folder prompt.+describe.each([+ { version: '0165', secondRow: 'Quit' },+ { version: '0166', secondRow: 'Back to Agent Command Center' },+])('opening a Swarm Codex folder restricted after start or resume ($version)', ({ version, secondRow }) => {+ const restricted = fixture(`codex-${version}-open-restricted`)+ it.each(['start', 'resume'])('continues a fake Codex %s without granting trust', async (mode) => {+ const home = mkdtempSync(join(tmpdir(), 'codex-restricted-'))+ const config = join(home, 'config.toml')+ const savedConfig = '[projects."/worktree"]\ntrust_level = "untrusted"\n'+ writeFileSync(config, savedConfig)+ // A separate fake engine waits on input on both startup paths, reading only its disposable home.+ const child = spawn(process.execPath, ['--input-type=module', '-e', `+ import { readFileSync } from 'node:fs';+ import { createInterface } from 'node:readline';+ const [mode, config, prompt] = process.argv.slice(1);+ if (!['start', 'resume'].includes(mode) || !readFileSync(config, 'utf8').includes('"untrusted"')) process.exit(1);+ console.log(JSON.stringify(prompt));+ createInterface({ input: process.stdin }).on('line', key => {+ if (key === 'Enter') console.log(JSON.stringify('› Ask Codex to do anything'));+ });+ `, mode, config, restricted], { stdio: ['pipe', 'pipe', 'pipe'], windowsHide: true })+ const lines = createInterface({ input: child.stdout })[Symbol.asyncIterator]()+ const nextScreen = async () => {+ let timer: ReturnType<typeof setTimeout> | undefined+ try {+ const line = await Promise.race([lines.next(), new Promise<never>((_, reject) => {+ timer = setTimeout(() => reject(new Error('Fake Codex did not draw its next screen')), 2_000)+ })])+ if (line.done) throw new Error('Fake Codex exited before drawing its screen')+ return JSON.parse(line.value) as string+ } finally { clearTimeout(timer) }+ }+ try {+ let capture = await nextScreen()+ const t = setup({ engine: 'codex' })+ const terminal = createSuperTabTerminal({+ resolve: id => id === 'h1' ? { agentId: id, engine: 'codex' } as RegisteredSession : undefined,+ capture: async () => ({ state: 'succeeded', value: capture }),+ read: async (_session, text) => codexScreen.inspect(text ?? ''),+ type: async () => false,+ key: async (_id, key) => { child.stdin.write(`${key}\n`); capture = await nextScreen(); return true },+ locked: async (_id, write) => write(),+ })+ Object.assign(t.core.superTabs!, terminal)+ await t.host.tick([row()])+ expect(capture).toBe('› Ask Codex to do anything')+ expect(readFileSync(config, 'utf8')).toBe(savedConfig)+ t.wait(180_000); await t.host.tick([row()])+ expect(t.alerts()).toEqual([])+ } finally {+ child.kill()+ rmSync(home, { recursive: true, force: true })+ }+ })++ it.each([+ ['trust grant', CODEX_TRUST_PROMPT, 'codex', [row()]],+ ['another engine', restricted, 'claude', [row()]],+ ['personal session outside the roster', restricted, 'codex', []],+ ['the second row highlighted', restricted.replace('› 1. Open restricted', ' 1. Open restricted').replace(` 2. ${secondRow}`, `› 2. ${secondRow}`), 'codex', [row()]],+ ])('never types on %s', async (_name, text, engine, rows) => {+ const t = setup({ engine, screen: () => ({ ok: true, text, hold: 'trust_open' }) })+ await t.host.tick(rows)+ expect(t.core.superTabs!.inputIfScreen).not.toHaveBeenCalled()+ })++ it('bounds failed conditional inputs and alerts when the prompt persists', async () => {+ const t = setup({ engine: 'codex', screen: () => ({ ok: true, text: restricted, hold: 'trust_open' }) })+ vi.mocked(t.core.superTabs!.inputIfScreen).mockResolvedValue({ ok: false, error: 'Screen changed.' })+ for (let i = 0; i < 5; i++) { await t.host.tick([row()]); t.wait(15_000) }+ expect(t.core.superTabs!.inputIfScreen).toHaveBeenCalledTimes(3)+ expect(t.core.superTabs!.inputIfScreen).toHaveBeenCalledWith('h1', ['Enter'], restricted)+ t.wait(180_000); await t.host.tick([row()])+ expect(t.alerts()).toContainEqual(expect.objectContaining({ alert: 'prompt', active: true }))+ })+})diff --git a/cli/src/superTabs/terminal.ts b/cli/src/superTabs/terminal.tsindex 826319bd7..8dc0f90d0 100644--- a/cli/src/superTabs/terminal.ts+++ b/cli/src/superTabs/terminal.ts@@ -10,7 +10,7 @@import { CommandRetries, permanentRefusal } from './retries.js'import { createHash } from 'node:crypto'import type { CoreApi, SuperTabScreen } from '../core/api.js'-import { updateDismissKeys } from '../engines/codex/screen.js'+import { openRestrictedKeys, updateDismissKeys } from '../engines/codex/screen.js'import { feedbackDismissKeys } from '../engines/claude/screen.js'import type { MessageHold } from '../engines/facets/screen.js'import { backend } from './host.js'@@ -29,7 +29,7 @@ export const SERVE_EVERY_MS = 3_000/** A screen read per agent; the 3-minute prompt wait does not need it finer. */const WATCH_EVERY_MS = 15_000const PROMPT_WAIT_MS = 3 * 60_000-/** How often Skip is typed for one stuck update screen before the 3-minute alert is left to say it did not work. */+/** How often a safe dismissal is tried before the 3-minute alert is left to say it did not work. */const SKIP_TRIES = 3/** The spec's cap on what a screen read returns. */const MAX_SCREEN_BYTES = 64 * 1024@@ -340,7 +340,7 @@ export class TerminalHost {// A screen that could not be read says nothing either way.const read = await this.core.superTabs!.screen(agent.harnessId, 0).catch(() => null)if (!read?.ok) continue- state.skips = await this.skipUpdate(agent, read, state.skips ?? 0)+ state.skips = await this.dismissScreen(agent, read, state.skips ?? 0)const now = this.now()const digest = screenDigest(read.text), seen = this.progress.get(agent.harnessId)const failing = unanswered(read.text)@@ -364,19 +364,20 @@ export class TerminalHost {}}- /** Codex's "Update available" screen on a Swarm agent's terminal, where no one will answer it and its task waits: Esc, through- * the core's door that types only while the screen is still exactly the one read here, so a screen another writer cleared- * meanwhile gets nothing (superTabs/inCore/terminal.ts `inputIfScreen`). Never a row number and Enter (on a misread row- * that could run the update), and only a screen read as exactly that prompt (engines/codex/screen.ts `updateDismissKeys`).- * After [SKIP_TRIES] the 3-minute alert takes over. Returns the tries made for this screen, none once it is gone. The held- * task goes on by itself: the mailbox types it again once the prompt is gone. */- private async skipUpdate(agent: Own, read: Extract<SuperTabScreen, { ok: true }>, tries: number): Promise<number> {- // Codex's update screen, or one of Claude Code's own feedback popups (engines/claude/screen.ts feedbackDismissKeys).- const keys = read.hold === 'update_prompt_open' ? updateDismissKeys(read.text) : feedbackDismissKeys(read.text)+ /** Safe choices on an unattended Swarm pane: skip an update or feedback popup, or open Codex restricted.+ * The core rechecks the exact screen under its input lock, so another writer clearing it receives no stray key.+ * After [SKIP_TRIES] the existing prompt alert takes over; the mailbox resumes once the screen is gone. */+ private async dismissScreen(agent: Own, read: Extract<SuperTabScreen, { ok: true }>, tries: number): Promise<number> {+ // Resumes already have a session identity, so startScreen cannot settle their folder-access prompt (#3904).+ // Only the exact restricted choice is safe; never answer a trust grant or another engine's screen.+ const restricted = read.hold === 'trust_open' && this.core.agents.byAgent(agent.harnessId)?.engine === 'codex'+ ? openRestrictedKeys(read.text) : null+ const keys = restricted ?? (read.hold === 'update_prompt_open' ? updateDismissKeys(read.text) : feedbackDismissKeys(read.text))if (!keys) return 0if (tries >= SKIP_TRIES) return triesconst dismissed = await this.core.superTabs!.inputIfScreen(agent.harnessId, keys, read.text).catch(error => ({ ok: false as const, error: String(error) }))if (!dismissed.ok) this.log(`could not dismiss a screen for ${agent.agentId}: ${dismissed.error}`)+ else if (restricted) this.log(`Opened Codex restricted for ${agent.agentId}; no folder trust granted`)return tries + 1}diff --git a/docs/plans/2026-11-06-codex-restricted-resume-plan.md b/docs/plans/2026-11-06-codex-restricted-resume-plan.mdnew file mode 100644index 000000000..44b9158a2--- /dev/null+++ b/docs/plans/2026-11-06-codex-restricted-resume-plan.md@@ -0,0 +1,39 @@+# Codex restricted resume implementation plan++> Implement task by task in this Engineer session, test first, with independent pull request reviews.++**Goal:** Harness-run Codex starts and resumes continue past Open restricted without granting folder trust.+**Architecture:** Reuse the exact-screen Codex recognizer and the core's conditional input door in the existing Swarm terminal watcher. The startup handler remains responsible for pre-session starts; the watcher handles resumed sessions.+**Tech Stack:** TypeScript, Vitest, fake Codex terminal.+**Spec:** Harness task #3904, conversation 10f805b0f0832aaf73da073d.++## Global Constraints++- Never mark a worktree trusted or enable untrusted config, hooks or exec policies.+- Only Harness Swarm agents; personal sessions outside Harness are unaffected.+- Two independent security reviewers must review sandbox and trust boundaries.+- Bounded retries, conditional input against the exact captured screen, no new polling.+- Platforms: macOS, Linux, Windows; no process launch behavior changes.++## Design++No new stored or exposed data. Growth, locking, read cost, scale, compatibility, migration, rollback, audience and identity: existing bounded Swarm terminal watch and conditional input contracts remain in use. Failure: after bounded attempts the existing visible prompt alert applies. Rollback: revert the watcher change.++## Review Focus++- A Trust and continue prompt must receive no input.+- Changed or malformed screens must receive no input.+- Other engines and sessions outside the Swarm roster must receive no input.+- A stuck prompt must stop automatic attempts and retain the existing alert.+- Fresh and resumed fake Codex sessions must continue with saved trust unchanged.++## Task 1: Continue restricted Codex sessions in the Swarm watcher++**Files:** Modify `cli/src/superTabs/terminal.ts`; test `cli/src/superTabs/terminal.spec.ts`; reuse `cli/src/engines/codex/screen.ts` and its tests.+**Interfaces:** Consume `openRestrictedKeys(capture: string): string[] | null` and `inputIfScreen(harnessId, keys, capture)`; retain `TerminalHost.tick` API.++- [ ] Add fake Codex start/resume tests that wait for Enter and then expose a working composer; assert no trust/config change. Cover wrong engine, non-roster session, trust prompt, changed screen and bounded attempts.+- [ ] Commit tests; run affected tests and record the expected failure.+- [ ] Extend the existing bounded dismissal path only for a Codex agent and the exact restricted screen.+- [ ] Run affected terminal, engine recognizer and startup tests, typecheck, and compare watcher timing on the same idle workload before/after. Commit implementation.+- [ ] Push draft, record red/green evidence, mark ready and hand off for two security reviews and platform testing.