Problem

OpenClaw uses a file lock ${sessionFile}.lock for session writes. In current build (npm openclaw@2026.2.9), the lock defaults are:

• timeoutMs = 10000 (10s)
• staleMs = 30min

However, call sites invoke it without overrides:

• acquireSessionWriteLock({ sessionFile })
  so timeoutMs/staleMs are effectively hardcoded defaults and not configurable via openclaw.json.

Additionally, the lock appears to be held across a broad portion of the embedded run (including LLM/tool execution), not just during the actual transcript append/flush. Under concurrent inbound messages to the same session, this produces:

• Error: session file locked (timeout 10000ms)
• which then propagates as All models failed / failover, even though this is not a model/provider error.

Why it matters

If two channels (e.g., Telegram + Webchat) end up hitting the same session concurrently, the second request times out after 10s and fails the run, causing user-visible outages. This is concurrency/locking contention, not provider failure.

Requested changes

1. Configurable lock timeouts

   • Allow configuring sessionWriteLock.timeoutMs and sessionWriteLock.staleMs via openclaw.json (global defaults), and/or via env.

2. Narrow lock scope

   • Hold the lock only during critical sections that read/repair/prewarm/append to the session file.
   • Avoid holding the lock across the entire LLM run.
   • Alternative: per-session async queue / single-writer actor to serialize writes without blocking the whole run.

3. Better error handling

   • Distinguish lock-timeout from model/provider failures.
   • Prefer a retry/backoff or a graceful 409/429-style response with “please retry” instead of surfacing All models failed.

Evidence (from dist bundle)

In dist/reply-*.js:

• Defaults inside acquireSessionWriteLock() include timeoutMs ?? 1e4 and staleMs ?? 1800*1e3.
• Multiple call sites use acquireSessionWriteLock({ sessionFile }) (no overrides).
• The lock is acquired early and released after the run finishes.

Workarounds we are using

• External cleanup of PID-dead .jsonl.lock files.
• Avoiding cross-channel contention by ensuring different channels don’t share the same sessionId.

Happy to provide exact line snippets from the bundle if you want, or test a PR.