plucky grove
I’m on OpenClaw 2026.3.13. A few days ago gpt-5.5 was fine working for me, but now it shows up as configured but missing in the model list and calls fail with "unknown model", then fall back to gpt-5.4. I have tried switching from openai-codex/gpt-5.5 to openai/gpt-5.5 but still shows blank fields / missing auth in the table. Is this fixable on 2026.3.13 or has something changed and is a version issue now?
finite owlBOT
proven cryptBOT
I’m on OpenClaw 2026.3.13. A few days ago gpt-5.5 was fine working for me, but n...
I checked the current docs/source and the old 2026.3.13 docs. This looks like a version/support drift issue, not something I’d try to fix by swapping prefixes.
On 2026.3.13, the OpenAI/Codex docs were still centered around the older Codex model route. Current docs say GPT-5.5 is available via subscription/OAuth routes: openai-codex/gpt-5.5 for the default PI runner, or openai/gpt-5.5 only if you are using the native Codex app-server harness. Plain openai/gpt-5.5 with normal OpenAI API-key auth is still described as future/direct-API-ready, not the normal subscription path.
I’d do this:
slender finchopenclaw update
openclaw doctor
openclaw gateway restart
openclaw models auth login --provider openai-codex
openclaw config set agents.defaults.model.primary openai-codex/gpt-5.5
openclaw models status
openclaw models list --provider openai-codex
The blank/missing auth fields when you switch to openai/gpt-5.5 are expected if your working auth is Codex OAuth: openai and openai-codex are different provider/auth routes. Keep the openai-codex provider for subscription auth unless you intentionally configure the Codex harness.
proven cryptBOT
I’m on OpenClaw 2026.3.13. A few days ago gpt-5.5 was fine working for me, but n...
plucky grove
So it sounds like I need to keep the openai-codex/gpt-5.5 path then for the version I am on. The version I am on is stable so I am reluctant to jump straight to an update in case it breaks something else. Is it possible to fix this on 2026.3.13? I am on an OpenAI subscription so using OAuth. Can be gpt-5.5 still be accessed via the codex model route using OAuth?
proven cryptBOT
So it sounds like I need to keep the openai-codex/gpt-5.5 path then for the vers...
For 2026.3.13 specifically: I would not expect openai-codex/gpt-5.5 to be reliably fixable in-place.
I checked the v2026.3.13-1 tag: its docs/code only forward-compat the Codex OAuth route to openai-codex/gpt-5.4; gpt-5.5 is not in that release’s Codex model compatibility path. Current OpenClaw docs now say GPT-5.5 is available through Codex OAuth as openai-codex/gpt-5.5, but that support landed after 2026.3.13.
So the practical answer is:
openai-codex/gpt-5.5, not plain openai/gpt-5.5.openai-codex/gpt-5.4 if you want to stay pinned.openai-codex/gpt-5.5, update OpenClaw. That is the supported fix.If you are nervous about jumping versions, do a dry run first:
openclaw update --dry-run
Then if it looks sane:
proven cryptBOT
So it sounds like I need to keep the openai-codex/gpt-5.5 path then for the vers...
openclaw update
openclaw doctor
openclaw gateway restart
openclaw models auth login --provider openai-codex
openclaw config set agents.defaults.model.primary openai-codex/gpt-5.5
openclaw models status
Docs: https://docs.openclaw.ai/install/updating and https://docs.openclaw.ai/providers/openai
plucky grove
So I upgraded from 2026.3.13 → 2026.5.12 (many suggested this was a stable version).
Before upgrade: openai-codex/gpt-5.4 worked, but gpt-5.5 did not. After upgrade: neither works. All OpenAI calls fall back to my fallback models.
openclaw models status --probe output for OpenAI targets:
openai-codex/gpt-5.5 openai-codex:default (oauth) unknown
↳ Requested agent harness "codex" is not registered.
openai/gpt-4 env (api_key) unknown
↳ Requested agent harness "codex" is not registered.
Same error on both OAuth and API-key paths. OAuth profile valid (ok expires in 8d, 97% of 5h window left).
Other providers (zai, openrouter) probe fine.
Current config:
agents.defaults.model.primary = openai-codex/gpt-5.5
plugins.entries: telegram, openai, zai, openrouter all enabled
No agentRuntime set anywhere
No codex plugin entry
auth.profiles has both openai:default (api_key) and openai-codex:default (oauth)
grep -rn "agentRuntime|codex" on the live config returns nothing relevant — only historical references in memory/session logs.
Most users in this Discord seem to run 5.5 successfully on 2026.5.x. What am I missing?
proven cryptBOT
So I upgraded from 2026.3.13 → 2026.5.12 (many suggested this was a stable versi...
This doesn’t look like an OpenAI auth problem. The key clue is that both openai-codex/* and openai/* are failing with:
Requested agent harness "codex" is not registered
That means something is forcing the Codex app-server harness, but the codex harness plugin is not loaded. openai-codex/gpt-5.5 normally means “Codex OAuth through the PI runner”; it does not require the Codex app-server harness.
I’d check these first:
openclaw config get agents.defaults.embeddedHarness
openclaw config get agents.list
env | grep OPENCLAW_AGENT
openclaw plugins list --enabled | grep codex
openclaw status --all
Most likely one of these is set outside the config file you grepped:
proven cryptBOT
So I upgraded from 2026.3.13 → 2026.5.12 (many suggested this was a stable versi...
OPENCLAW_AGENT_RUNTIME=codex
OPENCLAW_AGENT_HARNESS_FALLBACK=none
Env wins over config, so a daemon/service env var can cause this even when your JSON5 config has no embeddedHarness.
If you want the normal OAuth route, keep this shape:
openclaw config set agents.defaults.model.primary openai-codex/gpt-5.5
openclaw config set agents.defaults.embeddedHarness.runtime auto
openclaw config set agents.defaults.embeddedHarness.fallback pi
Then remove any OPENCLAW_AGENT_RUNTIME=codex / OPENCLAW_AGENT_HARNESS_FALLBACK=none from the gateway service env, restart the gateway, and start a fresh chat session with /new or /reset because harness choice can be pinned per session.
proven cryptBOT
So I upgraded from 2026.3.13 → 2026.5.12 (many suggested this was a stable versi...
If you actually meant to use the native Codex app-server harness, then you need the other route: enable/load the codex plugin and use openai/gpt-5.5 with embeddedHarness.runtime: "codex".
Docs that line up with this:
https://docs.openclaw.ai/providers/openai
https://docs.openclaw.ai/plugins/codex-harness
https://docs.openclaw.ai/gateway/configuration-reference#agentsdefaultsembeddedharness
plucky grove
I tried this and it didn't fix it so I ran OpenClaw doctor --fix. It made some changes to config and installed the codex plugin.
Restarted gateway, and re-ran model probes, and it's giving me a different error now.
Auth probes
┌──────────────────┬────────────────────────┬───────────────────────────────────────────────────────────────────────┐
│ Model │ Profile │ Status │
├──────────────────┼────────────────────────┼───────────────────────────────────────────────────────────────────────┤
│ github-copilot/ │ env (api_key) │ auth · 11.9s │
│ claude-haiku-4.5 │ │ ↳ Copilot token exchange failed: HTTP 404 │
│ openai/gpt-5.5 │ env (api_key) │ unknown · 4.9s │
│ │ │ ↳ (0 , _agentHarnessRuntime.hasBeforeToolCallPolicy) is not a │
│ │ │ function │
│ openai-codex/ │ openai-codex:default │ unknown · 2.5s │
│ gpt-5.4 │ (oauth) │ ↳ (0 , _agentHarnessRuntime.hasBeforeToolCallPolicy) is not a │
│ │ │ function
proven cryptBOT
I tried this and it didn't fix it so I ran OpenClaw doctor --fix. It made some c...
Yep, that new error is a different layer. I checked the source tags: the hasBeforeToolCallPolicy export is present in v2026.5.18+, but not in the v2026.5.12-beta.* tags I can see.
So you likely now have a Codex app-server plugin/runtime newer than your OpenClaw core. Doctor fixed the “harness not registered” problem by installing/loading Codex, but that exposed a plugin/core ABI mismatch.
Two practical paths:
If you only want openai-codex/gpt-5.5 OAuth through the normal PI runner:
disable/remove the Codex harness plugin and stop forcing the Codex harness. You do not need the Codex app-server plugin for openai-codex/*.
openclaw plugins disable codex
openclaw plugins disable openclaw-codex-app-server
openclaw config set agents.defaults.embeddedHarness.runtime auto
openclaw config set agents.defaults.embeddedHarness.fallback pi
openclaw gateway restart
openclaw models status --probe
If one of those plugin ids doesn’t exist, that’s fine; confirm the actual id with:
openclaw plugins list --enabled
If you want the native Codex app-server harness:
update OpenClaw core to at least 2026.5.18 / current latest so the plugin API has hasBeforeToolCallPolicy, then restart:
openclaw --version
openclaw update
openclaw gateway restart
openclaw models status --probe
Also: openai/gpt-5.5 env (api_key) probably won’t work unless OpenAI has enabled GPT-5.5 on your API account. For subscription/OAuth, the normal route is openai-codex/gpt-5.5; for native Codex harness, it’s openai/gpt-5.5 plus embeddedHarness.runtime: "codex".
proven cryptBOT
I tried this and it didn't fix it so I ran OpenClaw doctor --fix. It made some c...
plucky grove
Managed to get this working, posting the fix in case it helps anyone else with similar issues:
The fix that worked was to stay on 2026.5.19 but revert back to my original config file and remove everything the doctor did:
Now gpt-5.5 is working on 2026.5.19 and using the Codex OAuth on the PI runner. No more errors. Got there in the end.
proven cryptBOT
Managed to get this working, posting the fix in case it helps anyone else with s...
Thanks for writing this up, Parkzo — that’s a useful recovery path for anyone who got caught between the openai-codex/* config style and the newer codex plugin/runtime migration.
The key bit I’d underline for future readers: on 2026.5.19, your working setup was still the legacy openai-codex/gpt-5.5 / openai-codex/gpt-5.4 refs with no agentRuntime: { id: "codex" } blocks and no enabled codex plugin entry.
If this is the fix, please right-click your solution message → Apps → ✅ Mark Solution so Answer Overflow indexes it cleanly for the next person.