Headless VPS setup (Ubuntu 24, DigitalOcean) — multiple issues after migration from miniPCRunning O

[thorny schooner]

Headless VPS setup (Ubuntu 24, DigitalOcean) — multiple issues after migration from miniPC
Running OpenClaw 2026.4.21 gateway / 2026.4.14 CLI, Discord + Telegram, OpenRouter/auto as primary model, Open Brain via Supabase MCP (streamable-http).
Resolved today:

Gateway token mismatch (fixed via gateway.remote.token sync)
Grok-3-mini stuck as fallback model causing all embedded runs to fail
Memory embeddings broken (google:default missing from auth-profiles.json)
Systemd user service not working headless (fixed via loginctl enable-linger)
Orphan sessions and split state from botched miniPC migration

Still broken:

Edit tool fails silently — oldText not found errors even when text is visually identical. Claw cannot reliably edit files.
Subagent loops — Claw spawns subagents that run prohibited commands (sudo), hang, time out, and Claw spawns replacements without realizing. No native kill mechanism.
Read tool drops between sessions — works after /reset but fails mid-session with no error.
Open Brain not appearing in connected tools — MCP streamable-http server configured correctly but tools missing from /tools list intermittently.

[green elm]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[tame aspen]

Let me verify against the actual docs/source so I don’t make up keys or behaviors.

I found docs confirming two useful points already: /tools is runtime-only, and there is a native subagent kill path (/subagents kill ...). I’m digging into the edit/read/MCP bits now.

A few concrete things from the docs/source:

1. First, align versions
   You’re on gateway 2026.4.21 and CLI 2026.4.14. That may not be the whole problem, but I would not trust flaky session/tool behavior until those match.

2. Subagents do have a native kill path
   This part is the agent being wrong, not you. Current docs show:

   • /subagents list
   • /subagents kill <id>
   • /subagents kill all
   • /subagents log <id>
   • /subagents steer <id> <message>

   Also worth setting agents.defaults.subagents.runTimeoutSeconds, and if you want less chaos, keep maxSpawnDepth: 1.

3. Edit tool
   edit is still exact-match, so CRLF/LF or hidden whitespace will absolutely break it.

I also found recent hardening in source around:

• better edits[] handling
• better recovery after upstream edit throws
• adding current file contents to mismatch errors

For now, I’d strongly prefer apply_patch over edit for real code changes. If edit is still silently bad on 2026.4.21 after version alignment, that smells like a real bug.

4. Open Brain / MCP tools missing from /tools
   This one looks especially real.
   Docs say /tools is runtime-scoped, not config-scoped, and mcp.servers only saves config, it does not prove reachability.
   Also:

   • coding / messaging profiles expose configured MCP tools
   • minimal hides them
   • tools.deny: ["bundle-mcp"] disables them

   And I found an upstream fix from today:
   fix(agents): expose configured MCP tools in Pi profiles

That feels very relevant to your intermittent Open Brain symptom.

5. Read tool “drops”
   I’d want to split this into:

   • read disappears from /tools
   • read is listed, but the model stops calling it
   • read runs, but the result is swallowed/truncated

   Since /tools is session-scoped, compare it in a good state vs broken state. Also check /status to see if the active runtime model/fallback changed.

Best next move

I’d do this in order:

• update so CLI + gateway match
• retest in a fresh session
• when it breaks, capture:
  • /tools