openclaw - 💡(How to fix) Fix [Bug]: Subagent killed by LLM provider streaming idle timeout (~60s) — no retry or graceful handling [1 comments, 2 participants]

Problem

When a subagent is processing a task that requires extended thinking (e.g., reading large codebases, complex analysis), the LLM provider's streaming connection may have an idle timeout (e.g., ~60s for GitHub Copilot API). If the model is thinking and produces no tokens within this window, the connection is silently severed.

The subagent dies without any error message, retry, or notification to the parent session. From the parent's perspective, the subagent simply "timed out" even though runTimeoutSeconds was set much higher (e.g., 900s).

Observed Behavior

• Subagent starts processing a complex task
• Model enters extended thinking/reasoning phase (>60s without token output)
• LLM provider drops the streaming connection due to idle timeout
• Subagent session ends with timed out status
• No retry attempt is made
• Parent session receives a completion event with truncated output

Evidence

Measured across multiple runs:

• Failed runs: max gap between streaming events = 60.9s (connection dropped)
• Successful runs: events are continuous with no gaps >30s
• runTimeoutSeconds: 900 has no effect — the kill comes from the provider, not OpenClaw

Expected Behavior

1. Detection: OpenClaw should distinguish between runTimeoutSeconds expiry and provider-level connection drops
2. Retry: On provider connection drop, retry the request (at least once) before declaring failure
3. Transparency: The completion event should indicate the failure reason ("provider connection dropped" vs "run timeout exceeded")
4. Workaround config: Allow configuring provider-level keepalive or heartbeat tokens to prevent idle disconnection

Environment

• OpenClaw: v2026.4.2
• Provider: GitHub Copilot API (but likely affects any provider with streaming idle timeouts)
• Model: claude-opus-4.6 via custom-localhost proxy

Related Issues

• #44925 — Subagent completion silently lost
• #58786 — Subagent announce timeout destabilize gateway
• #53202 — Subagent announce hits repeated gateway timeouts in cron

Workaround

Currently splitting tasks into smaller chunks to ensure continuous token output within the idle window. This works but limits the complexity of tasks that can be delegated to subagents.