Bug

When openclaw agent times out at the gateway and falls back to the embedded runner, the embedded runner reuses the same sessionId. The gateway process still holds the session write lock, so the embedded runner fails immediately with:

session file locked (timeout 10000ms): pid=XXXX *.jsonl.lock

This cascades to all fallback models, resulting in FallbackSummaryError: All models failed.

Root Cause

src/commands/agent-via-gateway.ts — the fallback path passes the same sessionId to agentCommand() that the gateway already has locked:

} catch (err) {
  // gateway still holds the lock on localOpts.sessionId
  return await agentCommand(localOpts, runtime, deps)  // ← same sessionId
}

Reproduction

openclaw config set agents.defaults.timeoutSeconds 60
openclaw agent --agent main --message "<long task requiring >60s to respond>"

Full error output:

Gateway agent failed; falling back to embedded: Error: gateway timeout after 90000ms
session file locked (timeout 10000ms): pid=47321 ...9e051a52.jsonl.lock
FallbackSummaryError: All models failed (3): qiniu/deepseek-v3: session file locked ... (timeout) | qiniu/kimi-k2: Request was aborted (timeout) | qiniu/glm-4.5: Request was aborted (timeout)

Confirmed on 2026.4.7-1 and 2026.4.8.

Expected Behavior

Embedded fallback should use a new sessionId (fresh session), or wait for the gateway to release the lock before attempting to acquire it.