openclaw - 💡(How to fix) Fix [Bug]: 2026.5.22 Discord takes a long time to respond

The new slowdown is this: this Discord session is now ~86k tokens, and the Codex app-server path has a hard native-thread reuse cap at 70k. So every turn logs:

codex app-server native transcript exceeded active token limit; starting a fresh thread

That means the runtime keeps throwing away the warmed Codex native thread and starting fresh, which kills the fast path and causes Gateway CPU/event-loop stalls. The new delay is session bloat + native thread rotation.

---

The Codex-side session wrapper is losing its warm thread.

The 70k tokens are probably the accumulated local transcript that OpenClaw is passing into the Codex app-server for this Discord channel: system/context + memory excerpts + chat history + tool logs. After the update/debugging, that transcript got huge. The app-server then says “too large for active reuse” and starts a fresh native Codex thread each turn.

---

My agent provided me the following information:
Environment:
- Ubuntu 24.04.4 LTS (noble)
- OpenClaw CLI: 2026.5.22
- Model path: openai-codex/gpt-5.5, thinking=xhigh

Initial post-update Discord issue:
- Discord bot did not appear online after update.
- Doctor reported stale managed npm Discord plugin shadowing bundled Discord plugin:
  ~/.openclaw/npm/node_modules/@openclaw/discord
  package was @openclaw/discord@2026.5.4
- Removing the stale managed plugin/install record let bundled discord@2026.5.22 load.
- Also needed:
  openclaw config set plugins.entries.discord.enabled true --strict-json
- After restart, Discord provider started and bot probe resolved.

Second issue: severe response latency / event loop stalls:
- After Discord recovered, Gateway had major event-loop starvation.
- Logs include:
  [fetch-timeout] fetch timeout after 10000ms (elapsed 35372ms) timer delayed 25372ms, likely event-loop starvation operation=fetchWithTimeout url=https://discord.com/api/v10/users/@me
  [discord] gateway: Gateway websocket closed: 1000
  [diagnostic] liveness warning: reasons=event_loop_delay,cpu ... eventLoopDelayP99Ms=9026.1 ... eventLoopUtilization=0.933 cpuCoreRatio=0.977
- Gateway process was around 40-45% CPU while mostly idle/responding.
- Active Memory was disabled and Gateway restarted, but new turns still log:
  [agent/embedded] codex app-server native transcript exceeded active token limit; starting a fresh thread

Current session state:
- Session key: agent:main:discord:channel:<channel_id>
- sessions.json reports totalTokens=70911, totalTokensFresh=true
- Source constant in dist/run-attempt-Dgb9l7L8.js:
  CODEX_APP_SERVER_NATIVE_THREAD_MAX_TOKENS = 7e4
- Relevant source behavior:
  when tokenCount >= CODEX_APP_SERVER_NATIVE_THREAD_MAX_TOKENS, OpenClaw clears the Codex app-server binding and starts a fresh native thread.
- This seems to break warm-thread reuse every turn once the Discord transcript crosses ~70k tokens, causing cold-path latency and Gateway event-loop pressure.

Important caveat:
- The 70k limit is not the Codex model context window. It appears to be an OpenClaw native-thread reuse/cache threshold.