Restart Gemini (LazyGravity → Antigravity control chain)

Use when Antigravity / LazyGravity needs restarting — typically the Discord error Failed to connect to workspace: Workbench page for workspace "X" not found within 30 seconds, or the user says「重啟 gemini / antigravity / lazygravity」.

Just run it

bash ~/.hermes/skills/restart-gemini/scripts/restart.sh

Idempotent: safe to run anytime. Flags: --no-bot leaves the Discord bot untouched (only fixes Xvfb + Antigravity); --no-clean keeps orphan blank windows instead of closing them. Then tell the user to retry the workspace command in Discord.

Not every failure is a restart problem

Read step [5/5] of the output before telling the user to retry. A restart fixes a dead/frozen IDE. It cannot fix a stale binding, and the two look nothing alike:

Discord error	Meaning	Fix
Workbench page for workspace "X" not found within 30 seconds	Antigravity/Xvfb down	this skill
Timeout calling CDP method Runtime.enable	renderers frozen (GPU process died)	this skill (step 2 auto-retries)
Failed to activate session "X" after N attempt(s) (direct: Chat title not found in side panel; past: Conversation not found in Past Conversations)	the conversation is gone from Antigravity	/new in that channel — restarting is useless

The third case is real: on 2026-07-21 all 7 named sessions in antigravity.db were stale — Past Conversations held ~22 chats, none newer than ~1 month and none labeled src — while the stack itself was perfectly healthy. Bot-created conversations do not survive in the history the way workspace-opened ones do, so bindings rot silently. Step [5/5] (check_sessions.cjs) now detects this by driving the real Past Conversations widget per bound session, so the skill stops recommending a pointless restart.

Why this is needed (the architecture)

LazyGravity (Discord bot Ado#4549, ~/src/LazyGravity) drives the Antigravity IDE remotely over Chrome DevTools Protocol. As of 2026-07-05 the whole stack is systemd --user, tied together by lazygravity.target:

Unit	What
openclaw-xvfb.service	headless X display :99
lazygravity-antigravity.service	Antigravity IDE on :99, CDP 9223 (Requires xvfb; forces X11 ozone)
lazygravity-bot.service	the Discord bot (DISPLAY=:99); ordered After= antigravity
lazygravity-autoapprove.service	clicks "Always Allow" browser prompts
lazygravity.target	pulls in all four; WantedBy=default.target (starts at boot)

All enabled → survive reboot. The antigravity unit holds in activating (via ExecStartPost polling) until CDP 9223 actually answers, so the bot won't start talking to a not-yet-ready IDE.

Historical gotcha (now solved by the target): in the task path the bot does NOT auto-launch Antigravity — it only launches it via the open command. Before it was a service, if Antigravity wasn't up on CDP 9223 every workspace command timed out with "workbench page not found". Two causes were: Xvfb :99 down, or Antigravity not running. Now the target keeps both up (and Restart=always respawns Antigravity if it crashes).

Key facts:

• Port 9223 is fixed by ANTIGRAVITY_ACCOUNTS=default:9223 in ~/src/LazyGravity/.env (9222 is Hermes's own Chrome — don't touch it).
• The antigravity unit runs the Electron binary directly (/usr/share/antigravity/antigravity), not the bin/antigravity CLI wrapper (which detaches). It must force X11 (--ozone-platform=x11 + XDG_SESSION_TYPE=x11); the login session is Wayland and the binary otherwise picks Wayland and segfaults.

What the script does

1. Restarts lazygravity.target (or, with --no-bot, just Xvfb + Antigravity) and waits for CDP 9223 to answer HTTP.
2. Renderer liveness probe (scripts/probe_cdp.cjs): attaches a WebSocket to EVERY workbench page and runs a real Runtime.evaluate. This matters because the HTTP /json endpoints are served by the browser process and stay green even when every renderer is frozen — seen 2026-07-18: Antigravity's GPU process died on startup, all renderers blocked waiting for a GPU channel, 12 page targets listed but every Runtime.enable timed out (Discord error: Failed to connect to workspace: Timeout calling CDP method Runtime.enable). If any page hangs, the script automatically restarts Antigravity once more and re-probes (2 attempts max, then tells you to check the GPU process: ps -eo pid,args | grep -F -- --type=gpu-process).
3. Closes orphan blank windows (scripts/clean_windows.cjs, --close; skip with --no-clean). Restarts leave behind folder-less windows with an untouched new-chat panel — useless to the bot (it routes by workspace) but each keeps a renderer alive. On 2026-07-21 there were 32 of them holding ~20.4 GB RSS; closing them dropped Antigravity to 3.0 GB, which matters because system-wide earlyoom SIGKILLs at 1.5 GB free during GPU jobs. A window counts as blank only if document.title is exactly Antigravity, the Explorer has 0 roots, and the agent panel still shows the empty composer; unresponsive pages are left alone. They come back: Antigravity session-restores its window list, so closing them via CDP does not stick across a restart — that is why this runs on every invocation rather than as a one-off fix.
4. Prints each unit's active state, CDP status, and the open workbench pages (workspaces the bot can reach, e.g. ComfyUI - Antigravity - …).
5. Verifies bound chat sessions still exist (scripts/check_sessions.cjs) — see "Not every failure is a restart problem" above. Reads antigravity.db, and for each named session drives the real Past Conversations widget in that workspace's window. Prints ok / STALE / ? per session.

Manual verification (if needed)

curl -s http://localhost:9223/json/version                  # CDP alive?
curl -s http://localhost:9223/json/list | python3 -c "import sys,json;[print(p['type'],p.get('title')) for p in json.load(sys.stdin)]"
cd ~/src/LazyGravity && node dist/bin/cli.js doctor          # full check (CDP ports)

To prove the chain end-to-end you can connect to a workbench page's webSocketDebuggerUrl and Runtime.evaluate (e.g. check document.querySelector('.monaco-workbench')).

Antigravity UI selectors (verified 2026-07-21)

Needed by check_sessions.cjs; re-verify these if Antigravity updates:

• Agent panel: .antigravity-agent-side-panel
• Past Conversations toggle: [data-past-conversations-toggle] (the clock icon)
• The conversation picker is a custom React widget, .jetski-fast-pick — not a VS Code quick-pick. Its rows are plain divs, so .monaco-list-row, [role="option"] and li all miss them (this is why LazyGravity's own buildActivateViaPastConversationsScript fails to select a conversation).
• Search box inside it: input[placeholder="Select a conversation"]; it is React- controlled, so type with CDP Input.dispatchKeyEvent char events — setting .value does not filter the list.
• Empty result renders the literal text No items found.
• Beware: typing without focusing that input first lands in VS Code's quick-open overlay, which also shows "No items found" — an easy way to fake a false STALE.

Notes / caveats

• (Outdated note removed: Antigravity is a systemd --user unit now — lazygravity-antigravity.service, Restart=always — see the table above.)
• Workspaces live under WORKSPACE_BASE_DIR=~/src (e.g. ~/src/ComfyUI).
• Don't kill CDP 9222 — that's Hermes's Chrome, not Antigravity.