openclaw - 💡(How to fix) Fix OpenClaw 2026.5.27: onboard and doctor --fix did not recover Codex provider/runtime mismatch after update

Summary

After updating one OpenClaw server from 2026.5.22 to 2026.5.27, embedded Codex agent runs failed before replying because the Gateway/Codex harness and configured model provider were out of sync.

Running openclaw onboard did not repair the model provider/runtime mismatch. The operator also ran openclaw doctor --fix, but it did not recover the installation. The server only recovered after the managed @openclaw/[email protected] plugin was reinstalled/refreshed, the Gateway restarted, and the primary model was aligned to codex/gpt-5.5.

A second server updated through the same general process did not hit the problem, so this looks like a stale managed plugin / registry / route-state convergence issue rather than a universal update failure.

Environment

• OpenClaw: 2026.5.27 (27ae826) after update
• Previous OpenClaw version: 2026.5.22
• OS: Linux 6.17.0-29-generic x64
• Node: v24.15.0
• npm: 11.15.0
• Gateway service: systemd user service
• Codex managed plugin after recovery: @openclaw/[email protected]

Symptoms

Gateway repeatedly started with model refs such as:

agent model: openai-codex/gpt-5.5
agent model: openai/gpt-5.5

Agent runs then failed before producing a reply:

Requested agent harness "codex" does not support openai-codex/gpt-5.5 (provider is not one of: codex).

and later:

Requested agent harness "codex" does not support openai/gpt-5.5 (provider is not one of: codex).

What was tried

1. Updated OpenClaw from 2026.5.22 to 2026.5.27.
2. Ran openclaw onboard manually.
3. Ran openclaw doctor --fix manually. The operator reports this also did not resolve the issue.
4. Reset/restarted sessions/Gateway during troubleshooting.
5. Final recovery happened only after the managed Codex plugin was refreshed/reinstalled and the Gateway restarted.

Evidence from preserved state/logs

openclaw.json backups show openclaw onboard ran but did not repair the active primary model route:

21:50 KST backup:
  wizard.lastRunCommand = onboard
  agents.defaults.model.primary = openai-codex/gpt-5.5

22:12 KST backup:
  wizard.lastRunCommand = onboard
  agents.defaults.model.primary = openai/gpt-5.5

22:15 KST current/recovered:
  agents.defaults.model.primary = codex/gpt-5.5
  agents.defaults.models["codex/gpt-5.5"].agentRuntime.id = codex

The successful recovery coincided with a managed Codex plugin reinstall/refresh:

~/.openclaw/npm/node_modules/@openclaw/codex/package.json modified at 22:14:55 KST
~/.openclaw/npm/node_modules/@openclaw/codex/dist/harness.js modified at 22:14:55 KST
~/.openclaw/plugins/installs.json modified at 22:15:11 KST
Gateway restarted and then reported: agent model: codex/gpt-5.5

npm logs from that window show @openclaw/[email protected] being installed/replaced:

npm install ... @openclaw/[email protected]
silly reify mark retired ~/.openclaw/npm/node_modules/@openclaw/codex
silly CHANGE node_modules/@openclaw/codex

Expected behavior

At least one of these should happen:

• openclaw onboard should converge the configured primary/fallback model refs, Codex runtime policy, and managed @openclaw/codex plugin state.
• openclaw doctor --fix should detect and repair this stale Codex provider/runtime/plugin mismatch.
• If auto-repair is unsafe, the error should clearly tell the operator to refresh/reinstall the managed Codex plugin / registry and restart the Gateway, instead of only surfacing provider is not one of: codex.

Actual behavior

onboard and doctor --fix did not recover the server. The Gateway continued to start with openai-codex/gpt-5.5 or openai/gpt-5.5, while the selected Codex harness rejected those providers and agent runs failed before reply.

Recovery that worked

The server recovered after:

• managed @openclaw/[email protected] was reinstalled/refreshed,
• plugin install/registry state was updated,
• primary model was aligned to codex/gpt-5.5, and
• Gateway was restarted.

Suggested fix area

This may need a stronger convergence check across:

• OpenClaw core version,
• managed @openclaw/codex plugin package version and loaded runtime,
• plugin registry/install state,
• agents.defaults.model.primary / fallbacks,
• agents.defaults.models[*].agentRuntime, and
• stale session route state.

A doctor check specifically for “Codex harness provider set does not accept configured primary/fallback model provider” would have made the problem much easier to diagnose and repair.