| name | hal0-service-management |
| category | homelab-ops |
| description | Operating and troubleshooting hal0-managed user-level systemd services. Use when a hal0 platform service (hermes-gateway, hal0-agent@*, etc.) is misbehaving — missing env vars, not starting, not connecting to Telegram/Discord, etc. Encodes the structural fact that hal0 provisioners write secrets files but do NOT wire service units to load them. |
hal0 service management
Hal0 runs platform services as user-level systemd units under the agent's user account (often root's user systemd at /root/.config/systemd/user/). The orchestrator's provisioners (e.g. hermes_provision.py) write secrets to a canonical env file, but they do not add the EnvironmentFile= directive to the service unit. The unit must be wired by hand. This is the single most common reason a hal0 service comes up silent.
When to use this skill
- A hal0 platform service is logged as "started" but its platform adapters (Telegram, Discord, etc.) report "no platforms enabled" / "not configured" at boot.
/proc/<pid>/environ for the service process is missing variables that exist in /var/lib/hal0/secrets/agents/<service>.env.- A token / secret was added or rotated, but the service doesn't pick it up after restart.
- A user-level systemd unit is failing silently (no journal output, no error).
- A
delegate_task subagent reports "no shell execution tool available" — see the allow_code_exec and max_spawn_depth pitfalls below.
- A dashboard plugin tab shows "Could not load this plugin's script" — see
references/dashboard-plugin-js-loading.md.
For the Claude Code web-development handoff pipeline (writing handoff JSON, launching Claude Code with correct flags, turn-limit calibration), see the hermes-claude-workflow skill.
Structural fact (the core lesson)
hal0 provisioner ──writes──▶ /var/lib/hal0/secrets/agents/<service>.env
│
│ EnvironmentFile=... ◀── YOU must add this
▼
service unit ──sources──▶ process env ──read by──▶ service code (os.environ)
The service code reads tokens from os.environ only. Nothing on disk is sourced at startup unless the unit says so. The provisioner writes the file but does not edit the unit — that gap is yours to close.
Diagnostic procedure
Run these in order. Stop at the first one that explains the problem.
Critical: gateway logs are FILE-BASED, not journald
The hermes gateway writes its application output to /var/lib/hal0/agents/hermes/logs/gateway.log (fd 13 of the process). journalctl --user -u hermes-gateway shows ONLY systemd lifecycle messages (Started, Stopped, Main process exited) — NOT the platform connection status, NOT the "Channel directory built" count, NOT the Telegram/Discord connect/disconnect lines. Do NOT use journalctl to diagnose gateway platform issues; read the file.
tail -50 /var/lib/hal0/agents/hermes/logs/gateway.log
The companion log /var/lib/hal0/agents/hermes/logs/gateway-exit-diag.log records JSON-line lifecycle events (gateway.start, gateway.exit_nonzero, atexit.hook) for every PID that has ever run. Useful for spotting crash loops across restarts.
-
Confirm the service is running:
systemctl --user status <service>.service
If it isn't, fix the unit / deps first. If it is, continue.
-
Read the gateway application log (NOT journalctl):
tail -50 /var/lib/hal0/agents/hermes/logs/gateway.log
"No messaging platforms enabled" / "X not configured" / "missing required env" all point at wiring.
"Telegram network error, scheduling reconnect: Bad Gateway" means DNS/proxy routing failed — the gateway's Telegram fallback IP mechanism usually resolves this on the next retry.
"No user allowlists configured" means TELEGRAM_ALLOWED_USERS / DISCORD_ALLOWED_USERS are empty or unset.
-
Compare disk secrets vs process env — this is the smoking-gun check:
ls -la /var/lib/hal0/secrets/agents/<service>.env
sudo cat /var/lib/hal0/secrets/agents/<service>.env
PID=$(systemctl --user show -p MainPID --value <service>.service)
sudo tr '\0' '\n' </proc/$PID/environ | grep -E '^[A-Z_]+=' | sort
sudo python3 -c "
import os
d = open('/proc/$PID/environ','rb').read().split(b'\x00')
for e in d:
if b'=' in e and any(k in e for k in (b'TELEGRAM_', b'DISCORD_', b'GATEWAY_', b'HERMES_', b'HAL0_')):
k,v = e.split(b'=',1); m = v[:4]+b'***'+v[-4:] if len(v)>12 else b'***'
print(f'{k.decode()} = {m.decode()} (len={len(v)})')
"
Variables present in the file but missing from /proc/$PID/environ are not wired.
-
Inspect the unit file for an EnvironmentFile= directive:
systemctl --user cat <service>.service
If EnvironmentFile= is missing or points to the wrong path, that's the bug.
-
Verify the secret is still valid (for token-rotation cases):
curl -s "https://api.telegram.org/bot<TOKEN>/getMe"
curl -s -H "Authorization: Bot <TOKEN>" https://discord.com/api/v10/users/@me
A reusable wrapper for steps 3–4 lives at scripts/check-service-env.sh. Run it before editing the unit to confirm the hypothesis, after editing to confirm the fix.
The worked example for the most common case (hermes-gateway platform tokens) is at references/hermes-gateway-platform-tokens.md.
Fix procedure
Preferred: use a drop-in (NOT the main unit)
hermes_cli regenerates the main .service file on every gateway startup (the refresh_systemd_unit_if_needed() call). Any EnvironmentFile= added directly to the main unit will be clobbered. The correct pattern is a drop-in under .service.d/:
-
Create the drop-in directory and conf file:
mkdir -p /root/.config/systemd/user/<service>.service.d
cat > /root/.config/systemd/user/<service>.service.d/10-hal0-secrets.conf << 'EOF'
[Service]
EnvironmentFile=/var/lib/hal0/secrets/agents/<service>.env
EOF
-
Reload and restart:
systemctl --user daemon-reload
systemctl --user restart <service>.service
-
Verify the drop-in took effect:
systemctl --user show <service> -p EnvironmentFiles
Fallback: edit the main unit directly
Only use this if the drop-in pattern isn't possible (legacy unit, read-only drop-in dir, etc.). The unit WILL be overwritten on next gateway startup.
Access control: platform allowlists
Tokens alone don't decide who can talk to the bot. Three vars per platform, plus one global bypass, gate incoming traffic:
| Var | Effect |
|---|
<PLATFORM>_ALLOWED_USERS | Comma-separated allowlist of user IDs. Empty = nobody gets through. |
<PLATFORM>_HOME_CHANNEL | Where cron output, notifications, and the "home" delivery target go. Separate from access control. |
GATEWAY_ALLOW_ALL_USERS | true bypasses the per-platform allowlists entirely. Default-off. Flipping it to true is a free-for-all. |
Default-on (provisioner-shipped) state is usually GATEWAY_ALLOW_ALL_USERS=true, which is the opposite of safe — it silently allows anyone to talk to the bot. The right end state for any internet-reachable bot is:
GATEWAY_ALLOW_ALL_USERS=false
DISCORD_ALLOWED_USERS=<your_id>,<friend_id_1>,<friend_id_2>
TELEGRAM_ALLOWED_USERS=<your_telegram_id>
DISCORD_HOME_CHANNEL=<channel_id_where_bot_lives>
Editing the allowlist is the same edit-restart cycle as rotating a token — see "Fix procedure" above. The provisioner owns the file, so prefer the provisioner's write path if one exists; otherwise a careful patch / write_file on the secrets file is fine (don't change ownership/perms — they need to stay 0600 root).
Caveat: DISCORD_ALLOWED_USERS and TELEGRAM_ALLOWED_USERS use different ID formats. Discord wants snowflake IDs (17–19 digit strings). Telegram wants numeric user IDs (8–10 digit strings). They're not interchangeable — putting a Discord snowflake in TELEGRAM_ALLOWED_USERS will silently match nobody. The Channel directory built: N target(s) line in gateway.log reflects the resolved allowlist; if N=0 after a fix, the values are probably wrong-format rather than missing.
Pitfalls
- Don't trust secrets-file mtime as a signal. A file that hasn't changed in months can still be the source of the bug — what changed is the unit, not the file.
- Don't write to the secrets file by hand for token rotation if the provisioner has a credential-write path. Use
hal0-admin:provider_credential_write (or whatever the orchestrator's blessed path is) so the file's ownership/permissions stay correct. Hand-editing can silently break secrets-file consumers.
- Don't add
EnvironmentFile= to /etc/systemd/system/ — these are user units, not system units. The path is /root/.config/systemd/user/.
systemctl --user only works in the right user context. If you're root, root's user systemd is what hal0 uses, and the unit is at /root/.config/systemd/user/. Don't try to manage it from a different user's session.- Don't
hermes setup to fix gateway wiring. It will overwrite the whole model wiring. Use hal0 agent bootstrap hermes --repair (or the equivalent) for surgical fixes.
- A service can be "active" and "running" while every adapter silently no-ops. Always check
/proc/$PID/environ, not just the unit state.
- User-mode journald persists across reboots, but only for the user that owns it. If you
journalctl --user and see nothing, check loginctl session state.
- The
patch tool can no-op on systemd unit files. On this host, patch on /root/.config/systemd/user/<svc>.service has been observed to report success: true with a clean diff while the file on disk stays byte-identical to the pre-edit state — the mtime bumps but the content does not. Symptom: gateway log says "No user allowlists configured" / "No messaging platforms enabled" on restart even though the secrets file has the keys and the unit file looks right in your cat. Workaround: use write_file to overwrite the whole unit file atomically, then daemon-reload + restart. Confirm with systemctl --user show <unit> -p EnvironmentFiles that the new directive is registered before trusting the restart.
- If
tr '\0' '\n' </proc/$PID/environ | grep ... returns empty for vars you expect to see, fall back to Python. Shell pipelines against /proc/<pid>/environ occasionally lose NUL bytes (buffering, sigpipe, or env-block layout). A two-line Python check is the ground truth: read bytes, split on b'\x00', iterate. Don't trust an empty shell grep as proof the var is missing.
delegate_task subagents cannot execute shell commands by default. Even with toolsets=["terminal","file"], subagents will report "no shell execution tool available" or produce TimeoutError on MCP calls. The fix: in /var/lib/hal0/agents/hermes/config.yaml, set delegation.allow_code_exec: true (add the delegation: block if absent). After editing, restart: systemctl --user restart hermes-gateway. Without this, every subagent is MCP-only — it can call hal0-admin/hal0-memory tools but cannot run which, systemctl, cat, or any shell command.max_spawn_depth gates orchestrator subagents. If delegation.max_spawn_depth is 1 (or absent; default), all subagents are forced to leaf — even ones requested with role="orchestrator". Leaf subagents cannot call delegate_task, execute_code, memory, or clarify. To enable orchestrator subagents that spawn their own workers, raise delegation.max_spawn_depth to 2 or higher in /var/lib/hal0/agents/hermes/config.yaml. This is critical for Claude Code swarm workflows that rely on opus-orchestrator subagents delegating to implementation workers.- Telegram "Bad Gateway" errors can self-resolve via fallback IPs. The gateway auto-discovers Telegram fallback IPs (e.g., 149.154.166.110) at startup and bypasses DNS/proxy routing when the primary path fails. Gateway log shows:
Auto-discovered Telegram fallback IPs: 149.154.166.110 → Connected to Telegram (polling mode). If you see "Bad Gateway" followed by a successful connection within the same startup, the fallback IPs kicked in. No manual intervention needed. If the error persists across multiple restarts, check that outbound HTTPS to 149.154.167.* is not firewalled.
- Dashboard plugin
entry defaults to "dist/index.js" when absent from manifest.json. If a plugin's React components are built into the main dashboard JS bundle (the common case for bundled plugins like kanban), but the manifest doesn't declare entry at all, the dashboard tries to dynamically load a non-existent external JS file → 404 → "Could not load this plugin's script." Do NOT set "entry": "" — the frontend loader has no empty-entry guard and will still 404 on /dashboard-plugins/<name>/. Instead, either (a) remove the tab declaration from the manifest to make the plugin API-only, or (b) upgrade hermes-agent to a version that ships the plugin's JS bundle. See references/dashboard-plugin-js-loading.md for the full diagnostic procedure and xI()/sv() internals. The gateway creates runtime files (gateway.lock, gateway.pid, gateway_state.json, config.yaml, channel_directory.json, session files) owned by root. When the dashboard (running as hal0) tries to open them, /api/status crashes with a 500 PermissionError on gateway.lock. The fix: find /var/lib/hal0/agents/hermes -user root -exec chown hal0:hal0 {} \;. This is a recurring issue — the gateway will keep creating root-owned files on every restart. See references/dashboard-permission-mismatch.md for the full traceback and long-term fix options.
Token rotation — separate workflow
Rotating a Telegram or Discord token is a different class of action (and a security-relevant one). Don't fold it into "wiring." See references/hermes-gateway-platform-tokens.md for the procedure.
Files this skill touches
/root/.config/systemd/user/*.service — service units (edit, never write from scratch)/root/.config/systemd/user/*.service.d/*.conf — drop-in configs (preferred place for EnvironmentFile=)/var/lib/hal0/secrets/agents/*.env — secrets env files (provisioner-owned; avoid hand-edits)/var/lib/hal0/state/<service>/ — service state, logs, channel directory/var/lib/hal0/agents/hermes/config.yaml — hermes-agent config; the delegation.allow_code_exec key controls whether subagents get shell access/var/lib/hal0/agents/hermes/logs/gateway.log — primary diagnostic target; application output for platform connections, cron, kanban/var/lib/hal0/agents/hermes/logs/gateway-exit-diag.log — JSON-line lifecycle events per PID; useful for spotting crash loops/var/lib/hal0/agents/hermes/logs/gateway-shutdown-diag.log — verbose teardown diagnostics from last shutdown
For diagnostic traceback of the dashboard 500 error caused by root-owned gateway files, see references/dashboard-permission-mismatch.md.
For understanding the dashboard plugin JS loading system (how plugins load scripts, why entry: "" doesn't fix things, the xI() / sv() internals), see references/dashboard-plugin-js-loading.md.
For detailed quirks of gateway connectivity, see references/gateway-connectivity-quirks.md.
For navigating the hal0 source tree (/opt/hal0/), API routes, dispatcher, agent provisioner templates, and the Hermes↔hal0 integration flow, see references/hal0-codebase-map.md.
Verification
After any fix, the service should:
- Show
active (running) in systemctl --user status
- Log the expected platform adapters in gateway.log (not journalctl):
✓ telegram connected / ✓ discord connected / Gateway running with N platform(s) / Channel directory built: N target(s)
- Have all expected env vars visible in
/proc/<main_pid>/environ
- Round-trip a test message via the platform adapter
Don't declare victory on 1+2 alone — 3 is the real proof the wiring is right. Also, the gateway writes NO application output to journald; if journalctl --user -u hermes-gateway is silent beyond systemd lifecycle messages, that's normal — read gateway.log instead.
