Summary

After upgrade from 2026.4.23 to 2026.5.18 on WSL2 Ubuntu-24.04, the openclaw-gateway enters a perpetual restart loop, killing itself every ~30-90 seconds via SIGTERM. When a Telegram inbound message arrives, the gateway:

1. Receives the message correctly ([telegram] Inbound message ... (direct, 15 chars))
2. Triggers cli exec to claude-cli
3. Gets SIGTERMed ~1.5 min later, before claude-cli emits its first token
4. User never receives a response

Result: bot has been functionally unresponsive for 16+ days even though it accepts and acknowledges incoming messages.

Environment

• OS: Ubuntu 24.04 on WSL2 (Windows 11 Pro)
• Node: v24.15.0
• OpenClaw: 2026.5.18 (50a2481), upgraded from 2026.4.23
• Channel: Telegram (single bot, no other consumers)
• CLI backend: claude-cli (Claude Code 2.1.119) via custom wrapper
• Agent profile: `messaging`

Reproduction

1. Install openclaw 2026.4.23, configure Telegram channel + claude-cli backend.
2. Use it normally for some weeks (bot responds OK).
3. Upgrade to 2026.5.18 (`npm i -g openclaw@latest`).
4. Restart `openclaw-gateway` via systemd.
5. Observe gateway log: ready → telegram provider starts → SIGTERM at ~30-90s. Loops.
6. Send DM to bot → `[telegram] Inbound message` appears, `[agent/cli-backend] cli exec` fires, but SIGTERM kills gateway before claude-cli output. No outbound `sendMessage` logged. User receives nothing.

Evidence

Diagnostics bundle: attached (`openclaw-diagnostics-2026-05-20T14-48-45-951Z-1942.zip`, 10.7 KiB, 8 files, payload-free).

Representative log excerpt (one full cycle where a real user message was received and processed):

```
2026-05-20T13:10:34.347Z [gateway] loading configuration…
2026-05-20T13:10:38.937Z [gateway] ready
2026-05-20T13:10:39.052Z [telegram] [default] starting provider (@Festinnbot)
2026-05-20T13:10:40.258Z [telegram] [diag] isolated polling ingress started spool=...
2026-05-20T13:10:40.287Z [telegram] Inbound message telegram: -> @Festinnbot (direct, 15 chars)
2026-05-20T13:10:40.368Z tools policy: profile "messaging" (agent "main") has configured tool sections (tools.exec / tools.fs) that no longer implicitly widen the profile. Add alsoAllow: ["exec", "process", "read", "write", "edit"] explicitly if these tools should be available. See #47487.
2026-05-20T13:10:52.166Z [agent/cli-backend] cli session reset: provider=claude-cli reason=system-prompt
2026-05-20T13:10:52.360Z [agent/cli-backend] cli exec: provider=claude-cli model=sonnet promptChars=4320 trigger=user useResume=false session=none resumeSession=none reuse=invalidated:system-prompt historyPrompt=present
2026-05-20T13:11:39.900Z [health:debug] channel { ... }
2026-05-20T13:11:40.580Z [health:debug] probe.bot { channel: 'telegram', accountId: 'default', username: 'Festinnbot' }
2026-05-20T13:12:18.002Z [gateway] signal SIGTERM received
2026-05-20T13:12:18.016Z [gateway] received SIGTERM; shutting down
2026-05-20T13:12:18.037Z [shutdown] started: gateway stopping
2026-05-20T13:12:18.186Z [shutdown] completed cleanly in 147ms
```

Pattern across 20+ cycles in last 2h: SIGTERM lands ~85-90s after `[telegram] [default] starting provider`, regardless of whether a message was received. Always preceded by `[health:debug] probe.bot` returning OK with the correct `@Festinnbot` username.

journalctl --user output is empty in this WSL2 install (no user systemd journal). System journalctl was also queried, no matching entries for the gateway service shutdown initiator.

What I tried (did not resolve)

The `tools policy` warning correlation is suspicious. The new v2026.5.x breaking change requires explicit `alsoAllow`, but the proper structural form (e.g., `tools.profile` as object with `{name, alsoAllow}`) is rejected by the config schema:

```
tools.profile: Invalid input (allowed: "minimal", "coding", "messaging", "full")
```

The workaround was to add `alsoAllow` at `tools` root level. That schema-accepted, but unclear if it is the intended location.

Hypothesis (best guess)

CLI watchdog (`noOutputTimeoutMs`) fires when claude-cli takes >~85s to emit first token. claude-cli in stream-json mode with thinking=off should normally emit within seconds for short prompts, but with the wrapper script and the `--add-dir` traversal of large repos, first-token latency can spike. Suggesting:

1. Either the `noOutputTimeoutMs` default is too aggressive for stream-json mode with large workspaces, or
2. Something else (gateway supervisor, parent process) is sending SIGTERM independently of cli watchdog.

The exact code path emitting SIGTERM is opaque in the minified `dist/`. From source map names: `onSigterm` handler at `run-loop.ts:461` only logs and processes the signal — it does not originate it. The originator remains unidentified.

Requested

• Documented env var or openclaw.json key to extend `noOutputTimeoutMs` (tried `agents.defaults.cliBackends.claude-cli.noOutputTimeoutMs`, schema rejects).
• Confirmation of intended location for `tools.alsoAllow`.
• Investigation of supervisor that sends SIGTERM independent of the user's `Restart=on-failure` systemd unit (parent process supervises the child gateway).

This is reproducible 100% on this install. Happy to attach more bundles or run specific diagnostics.