Problem

When using openai-codex/gpt-5.2-codex as a subagent model, ~10% of runs get stuck: the tool call completes and returns a result, but the subsequent API call to the provider never returns. No error, no timeout — the connection just hangs indefinitely.

Current Behavior

• timeoutSeconds (agent-level) is the only timeout that eventually kills these stuck runs
• But it's a blunt instrument: it also limits successful long-running tasks
• noOutputTimeoutMs exists for CLI tools (exec watchdog) but not for API provider streaming responses

Proposed Fix

Add noOutputTimeoutMs config for API provider connections:

• If no data (tokens/chunks) arrives from the provider within X ms, abort and retry
• Could be configured per-provider or globally
• Suggested default: 120000ms (2 min)

Data

• 21 Codex subagent runs today, 2 stuck (10% failure rate)
• Both stuck runs: last tool result returned successfully, then next API call hung
• No error in logs, no stream error, just silence
• Only killed by service restart

Environment

• OpenClaw v2026.2.26
• Provider: openai-codex/gpt-5.2-codex
• thinking=xhigh