Where does the bug appear (feature/product)?
Grok Bot
Describe the Bug
Grok Bot cannot read an existing local file on my Mac. The desktop setting and synchronized per-machine configuration say Ask every time or Always allow, but local tools are rejected as Never. No approval card appears, and no file contents are returned.
This persists after a complete local cleanup, reinstall/sign-in, a new machine ID, updating to 0.71.2, and re-granting macOS Full Disk Access. Two Bots on the same Mac have encountered the issue. The latest post-reinstall check was on the main Bot.
Actual error
The Bot relayed this as the complete result of the fresh local Read:
Error: Local tools are turned off. The user has set local tool access to "Never", so Shell, Read, and AwaitShell with machineId, plus CopyToBox and CopyFromBox, cannot run on their computer. Do not retry them while this setting remains "Never". Use your own computer instead by omitting machineId, or ask the user to change the setting in Settings → Bot → Execution on Local Computer. If they change it away from "Never", you may try again.
A local Shell attempt also failed before spawning with the same message prefixed by Error: Command failed to spawn:.
Steps to Reproduce
- Sign in to the Grok Bot desktop app on the Mac.
- Open Settings → Computer → Execution on Local Computer. Select Ask every time or Always allow, and wait for synchronization to finish.
- Ask the Bot to refresh its device list. It reports one desktop device with
connected: true. - Ask the Bot to inspect the modification time and only the three permission fields in its own cloud-side
/home/box/agent-data/settings.json, using cloud Shell without amachineId. - Ask the Bot to make a fresh local
Readcall for/Users/<user>/.hermes/profiles/optimus/SOUL.md(the username is redacted here), using the currently connected Mac’smachineId. - The Bot reports the Never error quoted above, including after the clean reinstall with a new machine ID.
Expected Behavior
In Ask mode, show an approval request and permit the file read after approval. In Always allow mode, evaluate the request under the configured policy rather than reject it as a user-selected Never setting. A successful test should return the actual file contents.
Operating System
MacOS
Version Information
Grok Bot
Version: 0.71.2
Release Track: stable
OS: darwin, arm64
macOS: 27.0.1 (26A434)
Also reproduced on 0.71.1 after clean reinstall, before completing the official update.
For AI issues: add Request ID with privacy disabled
These are conversation request IDs from the desktop transcript, not independently captured low-level tool-call IDs. All times: 2026-10-17, Asia/Shanghai (UTC+08:00).
13:35:18 — latest, clean reinstall/new device, 0.71.2, Full Disk Access, Always allow:
c79423f6-cde0-4089-a1fa-b0f796244350
13:29:50 — clean reinstall/new device, 0.71.1:
7ed32366-5e7e-47cc-acaf-e3753c0ec45e
13:08:41 — synchronized ask check and Read:
da5c2df2-8039-49a4-a399-988ace3fc55f
13:07:24 — Read after sign-out/sign-in:
49b99858-7b10-481d-acb1-a6f121c18561
13:18:03 — Read after proxy trial:
73ce8d6d-33dc-4217-af1b-763296839351
Additional Information
Permission synchronization
Before reinstalling, the cloud configuration was reported as per-machine ask, while local Read still returned Never.
After reinstalling, the Bot reported the following updated cloud configuration at 13:29:50, with modification time 13:28:15.874651011 +0800. Device IDs below are placeholders; the actual values are available privately:
{
"localToolPermission": null,
"localToolPermissionCeiling": null,
"localToolPermissionByMachineId": {
"<new-machine-id>": "always",
"<old-machine-id>": "ask"
}
}
The refreshed connected-device list contained only the new machine ID. Independently inspecting the Mac’s ~/.grokbot/settings.json confirmed local permission always with no ceiling.
At 13:35:18, after confirming desktop version 0.71.2, valid application signature, and macOS Full Disk Access granted (auth_value=2 in TCC), a further Read using the new machine ID still returned the same Never error.
Checks and fixes already attempted
- Confirmed the target exists, is owned by the logged-in user, and has mode
0600; parent directory mode is0700. The file is 1,649 bytes and remained unchanged. - Checked macOS permissions. Full Disk Access was granted for the latest 0.71.2 test. Accessibility had been granted in earlier failing tests; it was not re-granted for the final file-only check.
- Restarted the desktop client, re-saved local execution settings, and signed out/back in to the same account.
- Backed up and regenerated the local daemon’s sealed credential and connection caches. The replacement daemon connected, but reads still failed.
- Performed a clean local reinstall: removed active configuration/user data, deleted the app-specific Safe Storage keychain item, and reset macOS app permissions before reinstalling. Old configuration was not restored. A new machine ID appeared.
- Completed the official update from 0.71.1 to 0.71.2 and verified its signature, then retested.
- Desktop/cloud-computer update indicators reported current; the actual cloud tool-runtime build was not independently obtained.
- The same-account dashboard privacy endpoint returned
PRIVACY_MODE_NO_TRAINING. Legacy Privacy Mode has not been established as the cause. - A startup log showed the privacy lookup timeout below. A trial of the client’s built-in system-proxy support did not resolve the Read failure and was rolled back. Explicit direct and proxied API-root requests both returned HTTP 200, which does not prove authenticated registration RPCs succeeded.
[sand:privacy] privacy-mode lookup failed, using privacy-safe fallback backend=https://api2.cursor.sh/ error=DeadlineExceeded
No target-file/personality edits, application-code patches, approval bypasses, or cloud-computer reset were performed.
What needs investigation
Could engineering trace the latest request and identify the effective permission and its exact source at the rejecting layer?
- Did the account’s
GetUserPrivacyMode,RegisterSandMachine, andListSandMachinescalls succeed? What canonical registration and permission values were returned for the new device? - Which component generated Never: server authorization, the cloud tool wrapper, or the local daemon? Is missing/unavailable permission being reported as an explicit user-selected Never?
- Is this connection server-authoritative, and is the synchronized per-machine map actually used by the rejecting runtime?
- Could a null global field, stale denial/cache, wrong data root/account/team scope, privacy/storage gate, or policy override take precedence over the per-machine value?
- Could there be a client/cloud permission-protocol mismatch? The desktop shows the generic local-execution dropdown rather than a concrete registered-computer row.
These are investigation questions, not confirmed root causes.
Evidence boundary
Local configuration, TCC grants, version/signature, daemon connectivity, and file metadata/hash were independently checked. Device-list/cloud-config/tool-error details above are relayed through Bot replies; independent raw tool-call objects were not available. The Bot reports no public tool for querying effective-permission provenance.
The issue is still unresolved. The acceptance criterion is a fresh Grok Bot local-tool read returning the actual contents of the target file. A changed UI value, synchronized JSON, or connected-device status alone does not meet that criterion.
Related reports
Related macOS Never-permission report
These reports have similar symptoms; I am not assuming they have the same root cause. This topic adds my clean-reinstall/new-device test and traceable request IDs.
AI tools helped organize my troubleshooting notes. This is a real issue on my own Mac; verified observations and unconfirmed hypotheses are separated above.
Does this stop you from using Cursor
No - Cursor works, but with this issue