openclaw - 💡(How to fix) Fix Codex runtime persists openai-codex as session modelProvider, causing recurring legacy route doctor warning

Bug type

Doctor / session routing / Codex runtime

Summary

When using the canonical OpenAI model route with Codex runtime authentication, OpenClaw persists openai-codex into the session's top-level modelProvider after a normal agent run.

This causes openclaw doctor to repeatedly report:

Legacy openai-codex/* session route state detected

Running openclaw doctor --fix clears the warning temporarily, but the warning returns after the next successful agent run.

Steps to reproduce

1. Configure an agent to use a canonical OpenAI model route:

   • primary model: openai/gpt-5.5
   • fallback model: openai/gpt-5.4
   • no explicit agents.defaults.modelProvider

2. Configure OpenAI auth to use a Codex OAuth profile:

   • routing chain observed as: openai -> codex runtime -> openai-codex auth provider

3. Run: openclaw config validate

4. Confirm config is valid and runtime auth is usable.

5. Run: openclaw doctor --fix --non-interactive

6. Run a normal agent turn through Telegram/TUI.

7. Run: openclaw doctor --non-interactive

8. Inspect the affected session entry in sessions.json.

Expected behavior

The session's top-level modelProvider should remain the user-facing/canonical provider, e.g. openai.

The Codex OAuth auth provider may remain represented separately as an auth profile, e.g. openai-codex:*, but it should not be persisted as the top-level session modelProvider.

After doctor --fix, the legacy route warning should not reappear after a normal successful run.

Actual behavior

After a normal successful run, affected session entries are persisted with:

modelProvider: "openai-codex" model: "gpt-5.5"

The auth profile is also Codex OAuth based, which is expected, but the top-level modelProvider being openai-codex causes doctor to report:

Legacy openai-codex/* session route state detected

The warning returns after each normal run, even after /new, restart, and doctor --fix.

OpenClaw version

OpenClaw 2026.5.27 (27ae826)

OS

Ubuntu 26.04 LTS

Install method

Global npm install

[email protected]

Model

Primary: openai/gpt-5.5 Fallback: openai/gpt-5.4

Provider / routing chain

Configured/user-facing model route:

openai/gpt-5.5

Runtime/auth route observed:

openai -> codex runtime -> openai-codex auth provider

The Codex OAuth auth provider appears usable and valid. This does not look like an auth failure.

Logs / evidence

openclaw config validate passes.

openclaw doctor --non-interactive reports:

Legacy openai-codex/* session route state detected Affected sessions: 2

Affected session entries include top-level:

modelProvider: "openai-codex" model: "gpt-5.5"

The warning reappears after normal agent execution even after running:

openclaw doctor --fix --non-interactive

Local code-path inspection suggests the run result provider is being persisted into the session's top-level runtime model:

• runner derives providerUsed from runResult.meta.agentMeta.provider
• session usage persistence writes that provider through persistRunSessionUsage(...)
• session runtime model storage writes it into top-level entry.modelProvider
• doctor flags entry.modelProvider === "openai-codex" as legacy session route state

So this appears to be a mismatch between execution/auth provider and user-facing session route provider.

Impact / severity

Low to medium.

This does not appear to break agent execution, but it makes doctor report a recurring warning that cannot be durably fixed with doctor --fix. It also makes Codex recovery/session route state look stale even when the configured model route is already canonical.

Related issues

Related but not identical:

• #78407
• #79461
• #79440

Those appear to cover doctor/migration rewriting or handling openai-codex/* routes. This report is about a normal successful Codex-runtime run persisting openai-codex back into the session's top-level modelProvider, causing the doctor warning to recur.