GitHub Issue Form Fields — OpenClaw Codex timeout

Issue URL: https://github.com/openclaw/openclaw/issues/new?template=bug_report.yml

Title

[Bug]: Codex OAuth /responses requests still hang until 300s timeout on 2026.5.25-beta.1

Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

Codex OAuth /responses requests intermittently hang until OpenClaw aborts at 300s on 2026.5.25-beta.1, causing Telegram agents to return the generic “Something went wrong” error while the gateway remains healthy.

Steps to reproduce

1. Run OpenClaw 2026.5.25-beta.1 (f5b52bd) on Windows with a Telegram direct bot agent using openai-codex/gpt-5.5 through the ChatGPT/Codex OAuth route.
2. Send normal Telegram requests to the affected agents after the beta upgrade.
3. Observe intermittent generic Telegram error: Something went wrong while processing your request. Please try again, or use /new to start a fresh session.
4. Confirm gateway logs show https://chatgpt.com/backend-api/codex/responses hanging until the 300s timeout.

Expected behavior

Codex responses should either stream/return normally or fail quickly with an actionable upstream/runtime error. Telegram should not receive a generic failure after a 300s model-call hang when the gateway itself is healthy.

Actual behavior

Requests intermittently hang at the ChatGPT/Codex OAuth /responses endpoint until OpenClaw aborts at ~300s. Telegram then surfaces the generic “Something went wrong” error. The gateway remains healthy and a local context/runtime policy audit passes.

OpenClaw version

2026.5.25-beta.1 (f5b52bd)

Operating system

Windows_NT 10.0.26200 x64

Install method

npm global (openclaw@beta), gateway on local Windows host

Model

openai-codex/gpt-5.5

Provider / routing chain

OpenClaw Telegram agent -> openai-codex provider -> ChatGPT/Codex OAuth transport -> https://chatgpt.com/backend-api/codex/responses

Additional provider/model setup details

• Runtime pin: PI via agents.defaults.models[openai-codex/gpt-5.5].agentRuntime.id = "pi" and per-agent model runtime pins where configured.
• Fallbacks are configured, but observed failures are on primary openai-codex/gpt-5.5 calls.
• node scripts/audit_agent_context_policy.js returns { ok: true, failures: [], warnings: [] } after the beta upgrade.
• No API keys/tokens included.

Logs, screenshots, and evidence

# Version / audit
OpenClaw 2026.5.25-beta.1 (f5b52bd)
node scripts/audit_agent_context_policy.js -> { ok: true, failures: [], warnings: [] }

# Failure 1 — Home agent, post-beta
2026-05-26T22:20:10+02:00 stalled session: sessionKey=agent:home:direct:<redacted-peer-id> state=processing age=287s reason=active_work_without_progress activeWorkKind=model_call
2026-05-26T22:20:24+02:00 fetch timeout reached; aborting operation url=https://chatgpt.com/backend-api/codex/responses elapsedMs=300008 timeoutMs=300000
2026-05-26T22:20:24+02:00 [model-fetch] error provider=openai-codex api=openai-codex-responses model=gpt-5.5 elapsedMs=300043 name=TimeoutError message=request timed out
2026-05-26T22:20:24+02:00 [responses] error provider=openai-codex api=openai-codex-responses model=gpt-5.5 message=Request was aborted.
2026-05-26T22:20:24+02:00 EmbeddedAttemptSessionTakeoverError: LLM idle timeout (300s): no response from model

# Failure 2 — control agent, post-beta
2026-05-26T22:30:01+02:00 fetch timeout reached; aborting operation url=https://chatgpt.com/backend-api/codex/responses elapsedMs=300001 timeoutMs=300000
2026-05-26T22:30:02+02:00 [model-fetch] error provider=openai-codex api=openai-codex-responses model=gpt-5.5 elapsedMs=300409 name=TimeoutError message=request timed out
2026-05-26T22:30:02+02:00 [responses] error provider=openai-codex api=openai-codex-responses model=gpt-5.5 message=Request was aborted.

# User-visible Telegram error
⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session.

Impact and severity

Affected: Telegram direct bot agents using openai-codex/gpt-5.5 through the ChatGPT/Codex OAuth transport.
Severity: High for operations because it affects the command/control agent and other agents.
Frequency: Intermittent; observed both before beta and twice after beta upgrade.
Consequence: Agent responses fail after ~300s, user sees generic error, and model calls can tie up sessions until timeout.

Additional information

• Last known stable state for this transport is NOT_ENOUGH_INFO.
• Observed on stable 2026.5.22 and still reproduced after upgrading to beta 2026.5.25-beta.1.
• Controlled gateway restart at 18:00 SAST succeeded but did not prevent recurrence.
• Related public reports noted during investigation:
  • Codex /responses timeouts on the ChatGPT subscription path (May 23-24) while status shows green - anyone else? openai/codex#24374
  • https://community.openai.com/t/bug-report-app-hang-tokio-runtime-parks-indefinitely/1381294
• Temporary public paste removed after redaction review.