openclaw - 💡(How to fix) Fix [Bug]: 2026.5.12 Codex runtime path causes severe latency, high CPU, and stuck sessions

Bug type

Regression / performance degradation after update

Beta release blocker

No

Summary

After updating to OpenClaw 2026.5.12, the local gateway became effectively unusable when running agents through the Codex runtime path.

The main symptoms were:

• very large response-time regression: agent replies either took extremely long or never completed;
• high CPU usage while the gateway was stuck;
• repeated stalled/stuck session diagnostics in logs;
• openclaw status started failing with gateway unreachable / timeout;
• attempts to switch back to PI did not appear to take effect: logs/status still indicated the Codex runtime/harness path was being used.

Rolling back from 2026.5.12 restored usable behavior.

Steps to reproduce

1. Start from a working OpenClaw install before 2026.5.12.
2. Update to 2026.5.12.
3. Restart gateway.
4. Send normal agent messages through the main session.
5. Observe that Gateway and agents on Codex are mostly unusable.

Expected behavior

After updating, normal agent turns should remain responsive.

If the operator switches back to PI, future sessions should clearly use the PI runtime, or OpenClaw should explain why the current session is still pinned to Codex and what action is required (/new, /reset, config/env override cleanup, etc.).

openclaw status should remain reachable unless the gateway is genuinely down, and diagnostics should make the active runtime/pin state obvious.

Actual behavior

After the update, agent turns became effectively unusable:

• responses stalled or took too long to be practical;
• CPU usage increased while sessions were stuck;
• logs showed stalled/stuck session state;
• openclaw status failed with gateway unreachable / timeout;
• switching back to PI did not appear to work, because logs still showed Codex being used.

The only reliable recovery was rolling back.

OpenClaw version

2026.5.12

Operating system

Linux

Server specs

• CPU: 4 cores
• RAM: 8 GB

Install method

npm global

Model

openai/gpt-5.5 / openai-codex path

Provider / routing chain

Telegram / main agent session -> OpenClaw gateway -> OpenAI/Codex runtime path

Additional provider/model setup details

The install had previously been usable before the update.

The suspected regression is related to Codex runtime becoming active after the update. Attempts were made to switch back to PI, but runtime selection still appeared to use Codex according to logs/status.

Logs, screenshots, and evidence

Representative symptoms from local logs/status:

openclaw status -> gateway unreachable / timeout

diagnostic: stalled/stuck session state
diagnostic: session stuck / stale session / active embedded run
gateway high CPU while agent response is pending

runtime/harness still indicates Codex after attempting to switch back to PI

More exact sanitized log lines can be added if needed.

Impact and severity

High for affected local installs.

The gateway remains technically running, but normal agent use becomes impractical: messages do not complete reliably, CPU usage rises, sessions get stuck, and even openclaw status can time out.

Rollback was required to restore a usable system.

Additional information

Possibly related issues:

• #78947 — Codex native runtime has very high latency for trivial turns.
• #79495 — Codex app-server shared client eviction across agents can cause CPU/reconnect churn.
• #81132 — Telegram polling / gateway behavior under event-loop starvation.
• #81326 — agentRuntime.id=codex can be configured without clear plugin/runtime diagnostics.
• #81888 — Codex harness auth/runtime mismatch fixed on main but not necessarily in a release.
• #81941 — 2026.5.12 Codex OAuth/runtime failures despite healthy profile status.

Report was prepared by an agent and carefully reviewed by a human.