OpenClaw Issue Draft — CLI backend fails silently on root systemd gateway

Repo: openclaw/openclaw Labels (suggested): bug, cli-backend, auth, linux-systemd Title: claude-cli backend silently skipped on systemd-managed root gateway, never spawns subprocess

Summary

On a Linux VPS where the OpenClaw gateway runs as root under systemd, the bundled claude-cli CLI backend is listed as primary model but is silently skipped by the fallback planner with no [agent/cli-backend] cli exec event, no error log, and no actionable diagnostic. Traffic goes straight to agents.defaults.model.fallbacks[0].

Following Ziwen's setup guide (Twitter 2046679352977580437) and the docs at /gateway/cli-backends and /providers/anthropic does not produce a working state on this topology.

Environment

• OpenClaw: 2026.4.15 (041266a) (global npm install at /usr/lib/node_modules/openclaw/)
• Node: v22.22.2 (wrapper with --no-warnings)
• OS: Ubuntu 24.04 (KVM VPS, Hostinger)
• Gateway: systemd service running as User=root
• Claude CLI: 2.1.88 at /usr/bin/claude, claude auth status → {"loggedIn":true,"authMethod":"oauth_token"} with CLAUDE_CODE_OAUTH_TOKEN env exported
• /root/.claude/.credentials.json present (555 bytes, claudeAiOauth key)

Expected

Following the Ziwen video and /gateway/cli-backends beginner quickstart, a message in Discord should trigger [agent/cli-backend] cli exec: provider=claude-cli and return a response from the local Claude CLI subscription.

Actual

1. openclaw models status --json correctly shows:

   • providersWithOAuth: ["anthropic (1)", "claude-cli (1)", "openai-codex (2)"]
   • allowed includes all claude-cli/* models
   • default: "claude-cli/claude-sonnet-4-6"

2. Any agent turn (Discord or openclaw agent -m "..." --agent main) produces this fallback decision chain without any [agent/cli-backend] event:

   [model-fallback/decision] decision=candidate_failed requested=claude-cli/claude-sonnet-4-6
                             candidate=openai-codex/gpt-5.4 reason=rate_limit

   Notice: requested=claude-cli/... but candidate=openai-codex/... — the primary is never attempted.

3. When I briefly got the CLI to actually spawn (before adding IS_SANDBOX=1), the failure surfaced in logs as:

   decision=candidate_failed requested=claude-cli/claude-sonnet-4-6
     detail=--dangerously-skip-permissions cannot be used with root/sudo privileges for security reasons

4. sessions.json under agents/main/sessions/ had cached per-channel model values from previous fallbacks (e.g. gemini-2.5-pro for a Discord channel) that override agents.defaults.model.primary silently on subsequent turns. Resetting the file does not survive — the gateway re-writes each session's model to whatever candidate candidate_succeeded for that turn (so Anthropic fallback re-cements gemini).

Steps to reproduce

1. Install OpenClaw 2026.4.15 on a Linux host; run as root via systemd.
2. Install Claude Code CLI; claude auth login with a Pro/Max subscription.
3. Export CLAUDE_CODE_OAUTH_TOKEN in the gateway's environment file.
4. In openclaw.json, add agents.defaults.cliBackends.claude-cli per the quickstart and set agents.defaults.model.primary = "claude-cli/claude-sonnet-4-6".
5. systemctl restart openclaw-gateway.
6. Send a message on any channel → agent responds using a fallback model, never the CLI.

What I tried (none worked)

Hypothesis

The silent skip is happening in resolveCooldownDecision at dist/model-selection-*.js (the subagent I delegated to found :72722-72740 on a nearby hash). When the claude-cli candidate is evaluated, something — possibly the root-detection propagated from the Claude CLI failure mode, or a session-level model cache invalidation — causes type:"skip" without emitting a user-visible log. The lack of [agent/cli-backend] in this path makes operator-side diagnosis impossible.

Request

1. Add an info-level log when a CLI backend is skipped pre-dispatch: [cli-backend/skipped] provider=<> reason=<auth|cooldown|root_detected|...>.
2. Document the root/systemd constraint in /gateway/cli-backends — readers running openclaw gateway under systemd as root will hit this and have no hint. Suggest either "run as non-root user" or a supported bypass env var.
3. Stop re-writing sessions.json with the fallback model when the primary isn't even attempted — at minimum gate that write behind candidate_succeeded on the primary ref.
4. (Optional) Expose cliBackends.allowRoot: true (or similar) that adds the correct flags to the child process so operators can opt-in.

Logs snippet (sanitized)

2026-04-22T12:00:00 [cron] payload.model 'gemini/gemini-2.5-flash-lite' not allowed, falling back to agent defaults
2026-04-22T12:00:00 [agent/cli-backend] cli exec: provider=claude-cli model=sonnet promptChars=381
(no further claude-cli log; ~68s later graph-memory recall runs as if turn finished via another provider)

2026-04-22T12:00:22 [model-fallback/decision] decision=candidate_failed requested=claude-cli/claude-sonnet-4-6
                     candidate=openai-codex/gpt-5.4 reason=rate_limit detail=...

Happy to share more logs or run diagnostic commands — please file a triage path.

cc: @steipete (OpenClaw maintainer)