Bug type

Regression (worked before, now fails)

Summary

CLI gateway probe reports "missing scope: operator.read" even though paired.json and device-auth.json contain operator.read and operator.write in all device and token scopes (2026.3.13).

Steps to reproduce

1. Run OpenClaw 2026.3.13 on Ubuntu 24.04 with systemd gateway (local, loopback).
2. Run over SSH: ssh user@vps "npx openclaw status --all" or npx openclaw security audit --deep.
3. Observe "missing scope: operator.read" in Gateway health / probe output.

Expected behavior

CLI should have full operator access (including RPC methods used by security audit) when the device token and paired.json have operator.read, operator.write, operator.admin, operator.approvals, operator.pairing.

Actual behavior

Gateway connects (Reachable: yes, Connect: ok) but RPC is limited: "RPC: limited - missing scope: operator.read". Security audit --deep and status --all report "Gateway: unreachable (missing scope: operator.read)".

OpenClaw version

2026.3.13

Operating system

Ubuntu 24.04 (linux 6.8.0-101-generic x64)

Install method

npm global / npx

Model

N/A - gateway auth bug

Provider / routing chain

N/A - gateway auth

Config file / key location

~/.openclaw/openclaw.json, ~/.openclaw/devices/paired.json, ~/.openclaw/identity/device-auth.json

Additional provider/model setup details

N/A

Logs, screenshots, and evidence

openclaw gateway probe output:
- Reachable: yes
- Connect: ok (50ms)
- RPC: limited - missing scope: operator.read

paired.json (CLI device, redacted) has:
- device.scopes: ["operator.read","operator.write","operator.admin","operator.approvals","operator.pairing"]
- tokens.operator.scopes: same

device-auth.json has same scopes. Tried devices rotate, manual patch, clientId gateway-client→cli, gateway restart — no fix.

Journal evidence (gateway logs):
[ws] ⇄ res ✗ status 0ms errorCode=INVALID_REQUEST errorMessage=missing scope: operator.read [ws] ⇄ res ✗ system-presence 0ms errorCode=INVALID_REQUEST errorMessage=missing scope: operator.read [ws] ⇄ res ✗ config.get 0ms errorCode=INVALID_REQUEST errorMessage=missing scope: operator.read

Meanwhile channels.status and doctor.memory.status succeed (different conn IDs), so scope check is per-connection and some RPCs require operator.read.

Impact and severity

Affected: Operators running CLI over SSH (e.g. make check, security audit --deep)
Severity: Medium — blocks full gateway probe and deep security audit
Frequency: 100% reproducible
Consequence: Cannot run openclaw security audit --deep; status --all shows "unreachable (missing scope: operator.read)"

Related: #23006, #21688, #22193. All documented workarounds (rotate, patch paired.json, clientId) applied without success.

Additional information

Docs indicate clients must declare scopes in connect handshake. Hypothesis: CLI may send a different/cached scope list in connect.params.scopes than device-auth.json, or gateway validates scopes differently than stored paired.json.

Gateway Protocol: https://docs.openclaw.ai/gateway/protocol