OpenClaw v2026.5.6 landed on May 6, 2026 at 12:51 p.m. America/Chicago time, which was May 6, 2026 at 17:51 UTC. This one is not flashy. It is a repair release, and if you updated to v2026.5.5, I would not sit on it.

The headline is simple: a doctor --fix repair in v2026.5.5 could rewrite a valid openai-codex/* route to openai/*. If you rely on the ChatGPT or Codex OAuth path for GPT-5.5, that could quietly move you onto the wrong provider route or break the setup outright. This release undoes that mistake.

If you touched v2026.5.5, check your default model first

This is the part I would treat as urgent. If v2026.5.5 already changed your default model, OpenClaw upstream says to run:

openclaw models set openai-codex/gpt-5.5 && openclaw config validate

That switches the default agent back to the Codex OAuth route and validates the config after the change. If your setup uses Codex OAuth and something has felt off since yesterday, start here before you chase ghosts anywhere else.

OpenClaw also published a recovery note for this exact issue: check and recover Codex OAuth routing.

Plugin fetches should stop failing over weird header baggage

The next fix is smaller on paper but easy to appreciate if you have ever stared at a plugin request that looked valid and still blew up. OpenClaw now strips third-party symbol metadata out of plain header dictionaries before handing them to native fetch or Headers. The debug proxy got the same cleanup when it replays captured requests.

That means fewer cases where SDK wrappers, guarded fetch paths, or caller-owned header objects poison an otherwise normal request. You should see less "why did this fetch fail on this one code path only" nonsense.

Timed-out web fetches clean up after themselves now

The last fix I would care about is the web fetch timeout cleanup. Timed-out requests are now supposed to come back as tool errors instead of leaving Gateway tool lanes hanging around after the timeout. That is good boring engineering. It keeps one bad fetch from leaving confusing residue behind it.

What I would test right after updating

1. If this host ran v2026.5.5, run the recovery command now. Use openclaw models set openai-codex/gpt-5.5 && openclaw config validate if you depend on Codex OAuth.
2. Run one real GPT-5.5 prompt through your usual Codex OAuth flow. Do not stop at config validation. Make sure the live route works.
3. Retry one plugin or SDK-backed fetch path that used to be flaky. This is the fastest way to confirm the header cleanup helped your actual setup.
4. Force one harmless timeout test if you use web_fetch heavily. You want to see a clean tool error, not a stuck lane that poisons the next request.
5. Run openclaw doctor --fix once more after updating. I know, that sounds funny after a doctor-related fix, but this release is the one that corrects the earlier bad repair.

Should you update now?

Yes. If you skipped v2026.5.5, this is an easy take. If you already installed v2026.5.5, I think this is a same-day update because the default-model routing mistake is the kind of bug that wastes hours while pretending to be a provider problem.

The short version is that v2026.5.6 fixes a real footgun, then cleans up two more failure paths that never should have been dramatic in the first place.

---

Full changelog: View v2026.5.6 on GitHub

Generated release page: Read the structured release notes

- Fred