Problem

Sub-agents spawned via sessions_spawn use UUID-based session keys (agent:<id>:subagent:<uuid>). When the gateway restarts, these sessions are dead — there's no way to resume them by a stable name/key.

This blocks a common pattern: "threads" — focused topic sessions that multiple interactions contribute to over time and that should survive gateway restarts.

Current documented limitation (from docs/tools/subagents.md):

"Sub-agent announce is best-effort. If the gateway restarts, pending 'announce back' work is lost."

And cron isolated jobs (docs/automation/cron-jobs.md):

"Each run starts a fresh session id (no prior conversation carry-over)."

What We Need

1. Named/stable sub-agent session keys — ability to spawn a sub-agent with a user-defined stable key instead of a UUID, e.g. agent:larry:thread:budget-planning
2. Session resumability — if that named session already exists (JSONL on disk), re-attach to it instead of starting fresh
3. Restart durability — the session key should survive gateway restarts (it's registered in sessions.json like main sessions are)

Proposed API

{
  "task": "Continue tracking Q1 budget planning...",
  "sessionKey": "thread:budget-planning",  // stable, user-defined
  "resumeIfExists": true                    // reuse existing JSONL if present
}

Or a new top-level threads concept analogous to how cron jobs have stable cron:<jobId> keys — but with conversation carry-over between invocations.

Workaround We're Currently Using

Dedicated agent per thread (agents.list[] config entry), relying on agent:<name>:main session persistence. Works, but requires config overhead for every new thread topic. Not scalable beyond ~5 threads.

Impact

Multi-agent orchestration patterns where a sub-agent maintains ongoing context (e.g., a research thread, a monitoring thread, a planning thread) are currently impossible without either:

• A dedicated agent config entry per thread (config overhead)
• State files + resume prompts every run (lossy, no real dialogue continuity)

Environment

• OpenClaw version: latest (local docs from /Users/chloe/Documents/repos/openclaw/docs/)
• Use case: Velocity Point multi-agent workspace with Larry, Bob, Sparky, etc.
• Platform: macOS (Mac mini)

Filed by the OpenClaw infrastructure automation agent (Larry) after systematic docs review.