openclaw - 💡(How to fix) Fix [Bug]: Session compaction deadlock blocks child subagent SessionManager.open() [1 comments, 2 participants]

Bug Description

When a session reaches the compaction threshold (~200k tokens), the parent agent's compaction process acquires a WRITE lock on the session file. If a child subagent spawn is attempted simultaneously, SessionManager.open() tries to acquire a READ lock on the same file. The child blocks indefinitely because the parent's WRITE lock is held for the duration of the compaction/summarization process.

This creates a structural deadlock where:

1. Parent session is in compaction (WRITE lock held)
2. Child subagent spawn requests READ lock → blocks
3. Subagent times out with zero output (0 tokens, empty result)
4. Parent session logs "stuck session" warning after 120s

Repro Steps

1. Use a model with small context window or heavy tool usage to reach compaction threshold
2. While session is near compaction, spawn a subagent via sessions_spawn with default context
3. Observe subagent times out with no output
4. Check logs for stuck session and InteractionEventListener slow-listener warnings

Expected Behavior

Subagent spawn should succeed even when parent session is compacting, OR subagent spawn should be rejected immediately with a clear error rather than silently timing out.

Actual Behavior

Subagent spawn blocks indefinitely until timeout. The subagent receives no tokens and produces no output. The parent session appears "stuck" in processing state for 120s+.

Environment

• OpenClaw: 2026.4.23
• Runtime: ollama/deepseek-v4-pro:cloud
• Context: ~200k tokens at deadlock
• Config: agents.defaults.subagents.runTimeoutSeconds: 900

Workaround

Use lightContext: true when spawning subagents to avoid triggering compaction during the spawn.

Related Issues

• #9855 — Session lock deadlock on timeout during compaction (related but focuses on timeout cleanup)
• #7688 — Auto-compaction blocks other sessions
• #7630 — Session lane deadlock after LLM timeout

Severity

High — breaks all subagent delegation when session approaches compaction threshold. Affects any workload using sessions_spawn with normal context.

Suggested Fix

1. Short-term: Allow READ lock acquisition during compaction (upgradeable locks, or use a separate compaction lock)
2. Medium-term: Reject subagent spawn with clear error if parent is compacting, instead of blocking
3. Long-term: Move compaction to a background process that doesn't hold the session lock