Adom Hydrogen (macOS)

@kyle — collected on anastasia's Mac on 2026-10-26 (Apple Silicon, 16 GB RAM, Darwin 25.6.0).

What the user saw

• The Claude Code panel in Hydrogen's VS Code would not start a session.
• She restarted the workspace container a few times. After that, VS Code also lost the Adom theme.
• The Mac was under heavy memory pressure: 7.0 of 8 GB swap used, Premiere Pro at about 4.1 GB, and the Lima VM at about 4.1 GB.

Findings

1. Two Hydrogen instances were running at the same time

The second instance launched without any single-instance check. Both write to the same ~/Library/Application Support/adom-hydrogen/adom-hydrogen.log. The second one failed to bind every port:

14:32:03.563 INF [direct-api] :8770 unavailable (Address already in use) — binding ephemeral fallback
14:32:03.563 ERR [discover] Failed to bind 0.0.0.0:47080 — Address already in use
14:32:04.424 ERR [bridge-supervisor] liveness beacon bind :9001 failed: Address already in use (Adom Bridge embedded monitor may flap)
14:32:04.765 ERR [shell] FAILED to bind http://127.0.0.1:1431 (Address already in use) — window will not load
14:32:04.887 WRN Proxy server error: Address already in use
14:32:04.887 ERR [control] bind 0.0.0.0:47084 attempt 1/8 … 8/8 failed

Even with no ports, the dev.4 instance still:

• restarted code-server: CS_UNIT_RESTARTED_NEW_FLAGS
• re-merged VS Code settings
• overwrote ~/.adom/hydrogen-direct-api-port with 0.0.0.0:56338, its ephemeral fallback port
• took Bridge ownership: embedded.json shows owner_pid: 0 and entered_via: runtime-adopt

The result was two workbench clients against one code-server, with two Claude extension hosts (exthost1 and exthost2) each spawning claude.

Ask: enforce a single instance, such as a lock file or a probe of :1431/:47084 before starting. If a second launch happens anyway, it should focus the running instance or offer to replace it, not run headless and mutate shared state.

2. The workspace WebSocket flapped for about 2 hours

[workspace] WS error … Keepalive timeout — no pong received
[workspace] (still retrying — 81 consecutive failures, same error)

This repeated about every 60 s against ws://127.0.0.1:8765 from 14:33 to 16:30. That was probably the two instances, plus the VM under memory pressure.

3. code-server restarts, user-triggered

• At 16:23 and 16:28: cs-watch: code-server restarted … unannounced and cs-pane lifecycle: machine stopped externally.
• Every Hydrogen launch also logs CS_UNIT_RESTARTED_NEW_FLAGS, so code-server restarts on each launch.

4. The Adom themes are gone

• settings.json has "workbench.colorTheme": "Adom Studio", and Hydrogen re-merges it on every start.
• No installed extension contributes that theme. ~/.local/share/code-server/extensions/ contains only adom.adom-agent-bar-0.2.6, anthropic.claude-code-2.1.310, google.google-antigravity, moonshot-ai.kimi-code and openai.chatgpt.
• The Hydrogen log prints [macos] adom-vscode macOS line: MISSING on every start, so the extension that provides the theme looks like it isn't installed on the mac line, or got dropped on a container restart.
• The Hydrogen page lists the color schemes as Adom Studio (default), Adom Studio Dark, Kickstand and Slate, which Windows Hydrogen has in VS Code. None of them are present on the Mac workspace, so VS Code falls back to its stock theme. Ask: ship the adom-vscode theme extension on the macOS line, the same as Windows.

5. Claude itself is healthy

• Inside Adom-Workspace, claude -p "reply with the single word OK" returns OK, and OAuth tokens are found.
• The extension logs show launch_claude going to Launching Claude on channel … with no spawn or auth errors.

That points the failure at the webview, the extension host, or the workbench connection, not the CLI. The extension logs do contain these noise items:

• claude-vscode notification channel error: Received a response for an unknown message ID {"id":0}. This could fit two clients sharing one backend.
• [uds-messaging] Failed to set up sockets directory /tmp/cc-socks-1001 … a sockets-directory component is not a pr[ivate dir], because /tmp in the nspawn container is a tmpfs owned by root.
• YAML frontmatter fails to parse in the synced skills adom-bridge-notify, adom-bridge-embedded and adom-bridge-web-integration.
• claude-artifacts MCP returns 404.

What I did

1. Sent TERM to both instances (16214 and 8758), which freed all the ports.
2. Relaunched only 0.1.421-dev.4 from the DMG, which is what the user wanted. Every bind succeeded (serving … on http://127.0.0.1:1431).
3. There is now a single exthost, and the extension launched Claude at 16:42:12 with no errors.

Status after the relaunch: the user confirms the Claude panel works again with only dev.4 running. The Adom themes are still missing.

Logs

• Mac: ~/Library/Application Support/adom-hydrogen/adom-hydrogen.log and .log.1
• Container: /home/adom/.local/share/code-server/logs/20261026T143215, …T162307, …T162807 and …T164200 (exthost*/Anthropic.claude-code/Claude VSCode.log)

Filed via Claude Code on anastasia's behalf.