Topic: macos-locked-use
Status: implemented experimental first version; bounded one-display Tart VM evidence and SIP-enabled physical idle-lock, successive-task and expiry evidence on macOS 26.6.2. Operator-confirmed physical takeover/recovery is live-tested with a scripted timing caveat. Multiple displays, additional OS revisions, Intel execution and distribution qualification remain open.
Current (2026-11-19), Tactical 165: agent-facing locked behavior, reproduced in a disposable Tart workspace (no physical qualification):
- With Browser while locked on, new one-shot Browser owners work on a locked console, with no unlock, approval or notice wait.
browserWhileLocked.activenow reports that even when prepared Desktop is also ready.requiresUnlockedApprovalis true only for Ask first.- Host doctor's
locked_consolecheck names the usable routes and says not to ask the person to unlock. The person is needed when access is Off (as after any resident restart) or no locked route is ready. - A held prepared owner leaves
IOConsoleLockedfalse and renews indefinitely within the claim. It relocks within about 3 seconds when it closes or its process dies. control keep-unlocked --for Dis the bounded way to hold that for a program such as a build.- Takeover by physical input is not VM-qualifiable here: neither Screen Sharing nor kernel-posted HID events count as physical. It remains a physical cell.
Current (2026-11-06), Tactical 159:
- A plain CLI/SDK owner may request
preparedConsole: trueand receives the same covered owner as a flagless one. Disposable-workspace evidence: 18/18 covered checks, plus a live plain--prepared-consolecall that unlocked, returned and relocked. - Views advertise
renewalonly when renewal can succeed. - Workstation capabilities omit
session.unlockandauthorization.*.
Current (2026-11-05), bounded signed readiness qualification: a cached helper setup observation becoming unavailable must not retire an already-admitted finite protected owner. Desktop readiness now retains that owner's route only while its original locked-console identity, consent, heartbeat, phase, safety latch and positive-helper silence bound remain valid. New tasks still need fresh setup readiness. A deterministic channel regression fails without this rule and passes with it; all 231 Swift tests pass. The full signed campaign exposed an intermittent first-read refusal with unavailable cached helper facts, but did not retain its exact shared blocking reason. This repair is consistent with that observation; exact incident attribution remains open. Closed owner diagnostics remain available for future failures. Tactical 156 records exact signed-package acceptance of the repaired shared-policy route: 900-second expiry, sustained same-owner renewal, protected faults, native Stop and boot-bound consent all pass in a single-display SIP-disabled ARM64 Tart workspace. This does not extend the earlier physical evidence to the new shared-policy implementation or identify the original intermittent trigger.
Current (2026-11-04), bounded shared-policy adoption: the unreleased native guard binds Desktop surface consent and one live shared owner, while Swift/root retain their independent native safety ceiling. Clean bbb53fd passes 18 local ad-hoc ARM64 appliance checks, including two isolated effects, concealment, clean completion/relock with consent retained, native Stop and final relock. Swift obtains the actual shared terminal reason before releasing a removed owner, and the native Stop chord follows the active Command layout. Tactical 154 owns exact source, cleanup and limits, including a preparation timeout whose cause remains open. This does not replace the physical/signed evidence or close the sustained editing/renewal investigation below.
Current (2026-11-02), investigation:
Tactical 148
isolates a diagnostic defect that can invalidate its own signed candidate:
the direct SDK adapter bypassed the installed CLI's no-bytecode entry point.
Both Python routes are now guarded, with candidate seal/manifest checks and
stronger independent save/renewal oracles. A separate successive covered
startup fails on an intact candidate before activation; its exact stage remains
unknown. A matched SDK trial separately ends desktop_locked_browser_only
before activation despite prior readiness. Neither corrected trial qualifies
sustained or renewed editing. These findings do not close the original physical
guardian/editing reports; successive-task readiness is the next investigation.
Current (2026-10-25), unreleased repair: the remote lockout incident exposed consent deletion after an ambiguous two-display startup. Operational failure now ends the finite session and fences desktop dispatch while retaining saved operator consent and independently permitted locked-browser access. Root recovery requires fresh same-console lock, healthy installation and quiet evidence, with a 30-second fault cooldown increasing to a 300-second cap after repeated failures. Clean covered completion resets fault history. A fresh task is required; no failed action is replayed. Admission readiness is separate from ongoing direct OS authorization and live guardian enforcement. Locked-origin tasks still require fresh positive current-helper liveness within 30 seconds; a missing cached sample alone does not masquerade as operator disable. A fresh current-root fault ends active control immediately. The root's own relock uses a pinned lock-only runner in the original user's GUI bootstrap; direct system-bootstrap locking was measured to have no effect. Historical composed relock evidence did not isolate root independence. The platform report records that finding, and Tactical 138 requires a timed root-only cell with resident and guardian both stopped. Desktop-scope restoration after a resident restart also requires positive root cleanup completion, even before its fault pause has been published; locked-browser scopes remain independent. Login-window discovery searches all bounded candidate windows and refuses ambiguous fields. Physical multiple-display qualification remains open. Tactical 138 owns implementation and signed VM qualification. The final signed development assembly passes composed locked-browser and covered-fault acceptance, including root-only relock with both user-session control processes stopped. This qualifies the bounded VM route and leaves physical multiple-display and SIP-enabled fault execution separate.
Current (2026-10-16): a reported physical task ended with
unlock_password_panel_unavailable. A bounded one-display VM completion on
signed 0.6.1 passes covered activation, independent native fixture effects,
filtered capture, independently observed relock and retained approval. An
earlier VM trial ended with the distinct disabled reason. Neither result
explains the physical failure. Source diagnostics now distinguish login-window
discovery, AX read errors, traversal limits and field readiness without reading
password values; identity, focused-field and targeted-input requirements remain.
The physical diagnostic build is not installed or qualified. See
Tactical 129 and the
platform evidence.
Current (2026-10-13): desktop 0.6.0 has bounded physical evidence of covered startup ending immediately despite approved access and a healthy helper. The asynchronous session cache can overwrite newer native evidence. The source repair fences older polls and binds activation to the guardian's unlocked observation of the original console. Deterministic regressions pass; the same-publisher signed ARM64 local candidate is installed with permissions and healthy helper retained. After explicit operator opt-in, its one-display physical completion cell passes covered unlock, independent native effects, filtered capture, completion relock and retention of the original approval. Expiry, takeover, successive tasks and broader protected-session qualification remain open for this repaired candidate. The platform report owns the evidence and Tactical 115 owns the bounded diagnosis and repair. The fix is now published in desktop 0.6.1, with both Mac packages independently authenticated and notarized. The physical effects above remain evidence of the signed local candidate; the public packages have not been installed for another physical completion test.
Add an initially disabled, persistent local setting named Allow Machine Control while screen is locked. The person completes native helper approval in Permissions, then enables the preference, approves ordinary access, and locks the Mac normally. The checkbox never requests a grant or installs software. An approved, bounded control session may temporarily unlock the existing console session behind opaque covers on every display. The setting grants no ordinary access by itself. No account password is stored by this feature.
Decision rationale: the static Sky lock-screen review found a password-free authorization plug-in peer check based on Team ID and signing identifier without an observed strict live-code validity or exact artifact check. That is a security-boundary gap, not a demonstrated exploit. Machine Control therefore retains strict dynamic signature validation plus the installer-pinned resident CDHash at its root broker. Publisher identity, process names, claimed session IDs and distant ancestry cannot arm unlock.
This follows the documented OpenAI locked-use flow. Product wording and task-end relock are Machine Control decisions, rather than claims about undocumented OpenAI UI or implementation details.
Decision, current source: cleanup may request lock only when the task's
native starting observation was locked. An unlocked-origin legacy control
lease ends when a later idle lock occurs; it cannot enter covered unlock or
inherit relock authority; that expected ending preserves valid standing
approval for a fresh locked-origin task. Ordinary completion leaves an unlocked console
unlocked. The new regression is in LockedUseTests; native Mac execution of
this adjustment remains pending. Tactical 102
owns the cross-platform correction and Windows cover slice.
Current source: physical keyboard/pointer activity cancels control and requests immediate relock while preserving otherwise valid ordinary access. Manual Pause gates observations/actions and ends covered control without revoking that access. Root automatic-unlock inhibition uses fresh lock/health/quiet recovery; physical acceptance of this revision remains pending. Display changes, sleep, session changes, expired access, Stop, disconnect, process failure, and disabling the setting end temporary unlock. Covers remain until lock is independently observed. A root watchdog owns failure relock; ordinary provider calls remain grant checked and generation bound.
Current policy: prepared locked use retains existing ordinary approval across idle locks and clean task completion, within the approving console and approval lifetime. Every subsequent task still acquires a finite owner connection. Verified console replacement, explicit operator Stop and timed approval expiry invalidate consent. Unknown state, safety failure and missing preparation make affected scopes unavailable without deleting saved consent. Ordinary manual/physical pauses retain valid access.
Decision — next direction: ordinary physical takeover should pause access and preserve otherwise valid standing authorization. Fresh covered control may resume when locked and physically quiet; an owner who unlocks to work remains protected by a local-use pause. Explicit Pause, Resume and Stop have distinct effects. The cross-platform access admission and pause topic owns this successor contract and its proposed implementation plan. These semantics are partly implemented in source; unattended resumption and physical acceptance remain open. Earlier physical evidence retains the behavior of its tested revision rather than inheriting these new semantics.
Open (2026-10-29), interactive evidence: covered browser edits sometimes appeared absent from native reads and became visible after clean completion and re-admission; replaying an apparently ineffective edit produced duplicate text. Separate 900-second owner requests ended with guardian failure at very different elapsed times. Their source enforced a 900-second owner bound, separate from target claims, but normal expiry and guardian faults have different end reasons; deadline expiry alone does not explain the observations. Read-only status also raced normal lifecycle transitions. The sanitized gap record separates observations, caller mistakes and diagnostics. The original intermittent guardian failures and delayed real-site edits remain unexplained.
Current, unreleased implementation: Tactical 141 separates guardian diagnosis, input/observation consistency, automated Tart acceptance and duration policy. 143 adds closed guardian health/timing predicates and independent model/save/reload, hover and presentation oracles. Its ordinary multiline Chrome regression identified dropped LF characters and now passes with native Return delivery. This is a specific input repair, not an explanation of every reported edit.
Decision, bounded live VM regression passed: cover windows use alpha compositing with full opaque alpha-1 content paint so the compositor's opaque window optimization does not suspend covered browser semantics. Full-frame paint is explicitly flushed before unlock; geometry, paint, visibility and alpha remain fail-closed health checks. Native browser model/Save/reload and independent unfiltered concealment pass together in the documented Tart profile. The sustained VM campaign passes; physical qualification remains open in Tactical 143.
Decision, unreleased startup handoff: wait for guardian activation off the resident main loop while continuing owner heartbeat/cancellation. Adopt a late result only for the exact live lease and phase after fresh authority/EOF checks. An originally locked task may prepare an asleep display only while a fresh independent observation still reports locked during waiting/unlocking. All active-control display-sleep endings remain. Tactical 143 owns qualification.
Decision, source implemented: 144 permits negotiated finite renewal of the same live owner under its original approval. Current retained SDK/CLI clients negotiate automatic renewal by default, so a supported task has no fifteen-minute total cutoff and needs no repeated ownership acquisition. Every protected extension still requires resident/guardian/root agreement; the original timed grant ceiling, Stop, Pause, takeover, connection loss and watchdogs still win. Claim renewal confers no desktop authority, and legacy finite owner commands remain bounded. Sustained exact-package qualification passes in the documented VM profile in the tactical, including 960 seconds under one owner and independent completion/disconnect relock. This does not qualify physical hardware or change unsupported providers' finite contracts.
Decision (2026-11-01), planned: support approved target-native desktop control on an awake Mac with its lid closed, including tasks that start with the existing console locked. This is requested product scope, not implemented support. The platform finding records the current explicit lid guards and bounded refusal evidence.
Implementation and acceptance must:
- Replace the blanket lid refusal in admission, display preparation and the guardian with explicit readiness checks for qualified display configurations. Cover both an external display and an awake closed-lid configuration without one; report unavailable configurations honestly.
- Preserve ordinary approval, finite owner connections, native capture/input, takeover, watchdogs and completion/disconnect relock. Do not make persistent sleep prevention or an outer control route a prerequisite.
- Independently verify semantic, keyboard and pointer effects and useful capture while the lid remains closed. Browser-only success does not qualify desktop control.
- Qualify lid opening/closing, display attachment/removal and sleep/wake on physical hardware, including concealment on newly visible displays, safe termination/relock and fresh admission after a transition. Preserve truthful limits on exposure guarantees.
- Report the specific blocking prerequisite in doctor and admission refusals; helper readiness alone must not imply closed-lid desktop readiness.
Open: which hardware, power and display configurations can meet those conditions without changing system sleep policy. System-sleep wake remains a separate capability; this work first targets a Mac that is already awake.
Current, unreleased source: Browser-only locked access has a separate default-off preference that leaves the OS console locked and requires no unlock helper. It retains existing approval for browser operations without conferring desktop authority. The platform evidence and Tactical 130 distinguish source-native validation from signed installed qualification.
- First version: awake Mac, open lid, existing console session only.
- No sleep prevention, closed-lid control, fresh login, FileVault/preboot, general remote unlock, or arbitrary privileged command API.
- First implementation uses native desktop input/capture; unsupported routes must refuse rather than produce covered screenshots or bypass takeover.
- Existing explicit appliance unlock retains its separate authority/profile.
- Same-user shell access is not contained. The authorization plug-in consumes a short session-wide grant; it cannot authenticate the original agent.
- Local enable/disable belongs to the native operator, never the agent socket.
Permissions owns SMAppService registration, native System Settings helper approval, typed preparation, and the capture-consent preflight. The Settings checkbox only persists the local preference after preparation is ready. After a signed update, native maintenance refreshes an already approved managed helper on the same unlocked console, once per launch and only while idle and unpaused. Preparation preserves the checkbox choice; initial setup remains off. Ordinary same-build launches leave a healthy helper alone. Initial installation, revoked OS approval and actual failures remain visible in Permissions. Disabling immediately persists local revocation; installed idle helpers have no standing grant. The existing signed executable runs a separate cover/input guardian, while the root broker owns the maximum deadline, heartbeat watchdog, and durable restart marker. Operator windows are hidden during coverage to preserve self-interface protection without blocking underlying pointer targets.
The same typed session.control / exact-ID session.control.end contract works
locally and remotely. The calling integration must own that bounded connection
for the real task; the resident does not infer completion from unrelated calls.
Native capture, AX/key/pointer effects, physical keyboard/pointer takeover,
completion, expiry, disconnect, resident stall/crash, persistent pause, and
manual recovery have bounded VM evidence. See the
current guide for usage and removal.
Current: integrated physical control passed programmatic lock, automatic covered unlock, native fixture AX/pointer/keyboard effects, filtered capture, and independent completion relock with no remaining covers or manual-unlock pause. The trigger verifies loginwindow's kernel identity and posts directly to it; screen notifications check actual layout and cover health. Native helper repair also succeeded after fixing re-registration ordering. Tactical 067 owns the bounded evidence and the earlier negative/diagnostic trials.
Current: the installed revision retains prepared, unpaused approval across expected locks and binds each grant synchronously to the approving console. Physical idle lock, two consecutive already-locked tasks and bounded task expiry passed with an unchanged approval and independent OS relock. Each task verified exact single-fixture PID/oracle attribution and native AX/pointer/keyboard effects. Agent capture showed the fixture beneath excluded covers. None of those cells required a manual unlock, external wake or new approval between tasks. Physical takeover also recorded immediate cancellation, OS lock, revoked Access and a manual-unlock pause; observed recovery cleared the pause while a new task still refused without fresh approval. The operator confirmed the takeover worked. It raced the runner's startup assertion, and the independent probe ran after recovery, so the full scripted takeover cell is not passed. Tactical 070 owns those precise live-tested boundaries and the timing caveat.
Current: all 91 Swift tests, ARM64/Intel framework builds and strict signed
candidate verification pass.
Automatic maintenance of the existing approved helper and same-build relaunch
passed on the physical Mac without a Repair click. App restart still clears
ordinary Access; maintenance does not regrant it.
A bounded native display wake in guardian startup passed task activation from
an inactive display while the OS remained awake. The temporary assertion is
released on readiness or bounded failure; no standing sleep-prevention policy
is installed. That signed update again maintained the helper automatically and
preserved the enabled checkbox. Tactical 073 records the earlier readiness gap
and bounded physical fix.
Decision, current source: system.login.screensaver is shared with
other authorization plug-ins, such as Codex Computer Use's locked-use entry.
Setup, disable and removal change only Machine Control's entry in that
one-of-n rule, before the password fallback, and never restore a snapshot.
Health checks only our entry; another writer's removal reports
unlock_policy_entry_missing for operator repair, and status lists the other
entries. Helper refusal codes now reach Permissions.
Tactical 106 owns
this slice; signed coexistence acceptance is open.
Open: a full scripted physical takeover cell with independent lock readback before recovery, multiple displays, additional OS revisions, Intel execution and distribution qualification.
Tactical 064 owns implementation, validation, and the precise final acceptance scope.
Tactical 065 replaces the initial command/password setup with the native permission flow. Its live acceptance remains separate from Tactical 064's earlier VM evidence.
Tactical 066 records the failed physical trial and safe manual recovery, rather than physical acceptance.
Tactical 067 owns the programmatic-lock and targeted-trigger iteration and its remaining acceptance.
Tactical 070 owns console-bound access retention and successive unattended-task acceptance.
Tactical 072 owns approved-helper update maintenance, choice preservation and ordinary relaunch.
Tactical 073 owns bounded display wake for authorized task startup after idle lock.
Current (source/policy fixtures): Tactical 077 replaces the physical takeover root latch with a durable physical/local-use pause. The same locked console may resume after 30 seconds of trustworthy HID idle and pause age; owner-unlocked work stays paused. Helper faults retain separate recovery. This has not yet replaced the prior signed physical result.