gateway-guard（网关守卫）

Gateway Guard

Description

Ensures OpenClaw gateway auth consistency and can auto-prompt "continue" when a run error (Unhandled stop reason: error) appears in gateway logs. Use when checking or fixing gateway token/password mismatch, device_token_mismatch errors, or before delegating to sub-agents.

Ensures OpenClaw gateway auth consistency and can auto-prompt "continue" when a run error (Unhandled stop reason: error) appears in gateway logs. Use when checking or fixing gateway token/password mismatch, device_token_mismatch errors, or before delegating to sub-agents.

Gateway Guard

Keeps OpenClaw gateway authentication in sync with openclaw.json. Use when the user or agent sees gateway auth issues, device_token_mismatch, or needs to ensure the gateway is running with the correct token/password before spawning sub-agents.

Metadata: This skill uses always: false in _meta.json. It is not forced into every agent run; the orchestrator invokes it when needed (e.g. before delegating to sub-agents). Optional persistence (LaunchAgent) is installed only when you run the install scripts; see "Before installing" below.

Before installing

• Backup openclaw.json — The script may add or correct gateway.auth (token/password) when missing or wrong. Make a copy before running ensure --apply.
• Test read-only first — Run python3 scripts/gateway_guard.py status --json and python3 scripts/gateway_guard.py ensure --json (without --apply) to see what it would do before allowing restarts or config writes.
• Understand continue delivery — The watcher can run openclaw agent --message continue --deliver when a run error appears in gateway.log. Confirm that automatically sending that message is acceptable in your environment.
• LaunchAgent is optional — Persistence (watcher every 30s) is installed only if you run install_watcher.sh. The installer copies the plist from the skill directory into ~/Library/LaunchAgents and runs launchctl load; only run it if you accept that. The plist is included in this package: scripts/com.openclaw.gateway-guard.watcher.plist (and scripts/com.openclaw.gateway-guard.continue-on-error.plist). Ensure OPENCLAW_HOME and OPENCLAW_BIN resolve to your intended paths before installing the watcher.
• Try in a non-production environment first if you are unsure.

Package contents (file manifest)

Included in this skill so installers do not error:

• scripts/gateway_guard.py — Main script (status, ensure, continue-on-error, watch).
• scripts/install_watcher.sh — Installs the single combined LaunchAgent (token sync + continue-on-error).
• scripts/install_continue_on_error.sh — Redirects to install_watcher.sh.
• scripts/com.openclaw.gateway-guard.watcher.plist — LaunchAgent plist template (install_watcher.sh copies and substitutes paths).
• scripts/com.openclaw.gateway-guard.continue-on-error.plist — Legacy plist (optional; install_watcher.sh replaces with the combined watcher).

Usage

• User or logs report "Gateway auth issue", "device_token_mismatch", or "unauthorized"
• Before running the router and sessions_spawn (orchestrator flow): check gateway status first
• After installing or updating OpenClaw: verify gateway and config match
• When the TUI disconnects or won't connect: fix auth and restart gateway
• Run error (Unhandled stop reason: error): run continue-on-error --loop (e.g. via LaunchAgent or cron) so the guard auto-sends "continue" to the agent when this appears in gateway.log

python3 <skill-dir>/scripts/gateway_guard.py status [--json]
python3 <skill-dir>/scripts/gateway_guard.py ensure [--apply] [--wait] [--json]
python3 <skill-dir>/scripts/ensure_gateway_then.sh [command ...]
python3 <skill-dir>/scripts/gateway_guard.py continue-on-error [--once] [--loop] [--interval 30] [--json]

• status — Report whether the running gateway's auth matches openclaw.json. Exit 0 if ok, 1 if mismatch.
• ensure — Same check; if mismatch and --apply, restart the gateway with credentials from config. Writes gateway.auth to openclaw.json only when it is missing or wrong (never overwrites correct config). Use --wait after --apply to block until the gateway port is open (up to 30s), so clients can connect immediately after.
• ensure_gateway_then.sh — Detect and connect automatically: ensures the gateway is running (starts it if needed, waits for port), then runs your command. Example: ensure_gateway_then.sh openclaw tui or ensure_gateway_then.sh (just ensure and wait).
• continue-on-error — When gateway.log contains Unhandled stop reason: error (run error), send continue to the agent via openclaw agent --message continue --deliver. Use --once to check once and exit, or --loop to run every --interval seconds. Cooldown 90s between triggers. State: logs/gateway-guard.continue-state.json.
• watch — Single combined daemon (one LaunchAgent). Each run: (0) token sync — ensure --apply so gateway auth matches config (prevents device_token_mismatch); (1) gateway back → what-just-happened summary; (2) continue-on-error check. Install one daemon: bash <skill-dir>/scripts/install_watcher.sh (or install_continue_on_error.sh). This unloads the old separate what-just-happened and continue-on-error LaunchAgents and loads com.openclaw.gateway-guard.watcher so users only need one. For periodic gateway recovery (check every 10s, restart if not ok), use the separate gateway-watchdog skill.

Behavior

• Reads openclaw.json → gateway.auth (token or password) and gateway.port.
• Compares with the process listening on that port (and optional guard state file).
• If ensure --apply: restarts gateway via openclaw gateway stop then openclaw gateway --port N --auth token|password --token|--password SECRET.
• If token is missing in config (token mode only): generates a token, writes it to config once, then proceeds. Does not overwrite config when it is already correct.
• continue-on-error: Tails OPENCLAW_HOME/logs/gateway.log for the string Unhandled stop reason: error. When found (and not in cooldown), runs openclaw agent --message continue --deliver so the agent receives "continue" and can resume. Run install_continue_on_error.sh to install a LaunchAgent that checks every 30s. If the error appears in the TUI but the watcher never triggers, the gateway may not be writing run errors to gateway.log — ensure run/stream errors are logged there.

JSON output (for orchestration)

• status --json / ensure --json: ok, secretMatchesConfig, running, pid, reason, recommendedAction, configPath, authMode, gatewayPort. When not ok, recommendedAction is "run gateway_guard.py ensure --apply and restart client session".

Requirements

• OpenClaw openclaw.json with gateway.auth (mode token or password) and gateway.port.
• CLI / system: openclaw CLI on PATH (for ensure --apply and continue-on-error); lsof and ps (macOS/Unix); launchctl on macOS when using the LaunchAgent install scripts.
• Environment (optional): OPENCLAW_HOME — OpenClaw home directory (default: ~/.openclaw). OPENCLAW_BIN — Path or name of openclaw binary (default: openclaw).

Privileged actions (what you accept)

This skill may: read and modify openclaw.json (including writing gateway.auth when missing or wrong); write state and log files under OPENCLAW_HOME/logs/; restart the gateway via the OpenClaw CLI; and, if the watcher is installed, invoke openclaw agent --message continue --deliver automatically when a run error is detected. These are privileged local actions; run only if you accept them.