Summary

When agents.defaults.sandbox.mode is set to "non-main", sub-agents spawned via sessions_spawn fail to initialize. Sessions are accepted (return valid childSessionKey and runId) but never start executing — zero tokens, zero messages, no transcript files created. No errors appear in gateway logs.

Environment

• OpenClaw Version: 2026.3.2 (85377a2)
• OS: macOS 15.3.0 (Apple Silicon, arm64)
• Node: v25.8.0
• Docker Desktop: running (coding agent sandbox works fine)

Steps to Reproduce

1. Set agents.defaults.sandbox.mode: "non-main" in config
2. Restart gateway
3. From main agent, spawn a sub-agent:

   sessions_spawn(task="Reply with: test", model="sonnet")
   

4. Check sessions_history or subagents list after 30+ seconds

Expected Behavior

Sub-agent starts within seconds, systemSent: true, tokens accumulate, task completes.

Actual Behavior

• sessions_spawn returns status: "accepted" with valid session key
• Sub-agent appears in subagents list as running
• sessions_history returns messages: [] indefinitely
• No transcript file created on disk
• systemSent never set to true
• totalTokens never set
• Zero error output in gateway logs
• Eventually times out via runTimeoutSeconds

Bisect Results

Systematically tested with a simple one-line task (Reply with exactly: ...):

The issue is not caused by specific Docker settings (network, readOnlyRoot, tmpfs). Any sandbox.mode that enables Docker for subagents causes silent failure.

Notes

• Cron isolated sessions work fine with sandbox.mode: "non-main" on the same config — they appear to use a different code path that bypasses container initialization.
• Per-agent sandbox overrides are unaffected: a coding agent with sandbox.mode: "all" and its own Docker settings works correctly.
• Possibly related to Bug: agentToAgent.enabled: true breaks sessions_spawn - sub-agents never start #5813 (similar symptoms, different attributed cause).
• The silent failure with zero log output makes this very difficult to diagnose.

Workaround

Set agents.defaults.sandbox.mode: "off" and use per-agent sandbox.mode: "all" overrides for agents that specifically need sandboxing (e.g., coding agents with tools.exec.host: "sandbox").