terminated | Friends of the Crustacean 🦞🤝 | Page 1

sturdy sparrow Apr 23, 2026, 1:32 PM

#

I am getting Terminated as a response of very simple prompts like Hello using GPT5.4 model

round forgeBOT Apr 23, 2026, 1:32 PM

#

slim magnetBOT Apr 23, 2026, 1:32 PM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

round forgeBOT Apr 23, 2026, 1:32 PM

#

exotic hemlockBOT Apr 23, 2026, 1:36 PM

#

That usually means the run is dying before the model answers, not that GPT is literally replying with “Terminated”.

First thing, please paste:

openclaw --version

Then these:

openclaw status --all
openclaw gateway status
openclaw models status
openclaw logs --follow

#

After logs --follow is running, send one tiny prompt like hello, then paste the log lines around that failure.

Also, please confirm the exact model id you selected, because this matters a lot:

openai/gpt-5.4 = direct OpenAI API
openai-codex/gpt-5.4 = ChatGPT/Codex OAuth

Those are different paths, and the Codex/OAuth path has had a few regressions recently.

One extra quick repro that helps a lot:

openclaw infer model run --model <exact-model-id> --prompt "hello" --json

If you paste those outputs, I can narrow it down pretty fast. Docs: https://docs.openclaw.ai/help/faq and https://docs.openclaw.ai/gateway/troubleshooting

sturdy sparrow Apr 23, 2026, 1:51 PM

#

openai-codex/gpt-5.4

exotic hemlockBOT Apr 23, 2026, 1:52 PM

#

Ah, okay. That narrows it down a lot.

openai-codex/gpt-5.4 is the Codex OAuth path, and there have been a few recent regressions specifically on that route. So this is probably not your prompt, and possibly not your config either.

Can you paste these next:

openclaw --version
openclaw models status --probe
openclaw infer model run --local --model "openai-codex/gpt-5.4" --prompt "pong" --json

Then tail logs and run one more tiny prompt:

openclaw logs --follow

#

What I want to see is whether it shows one of these patterns:

404 terminated
502 Bad Gateway
HTML coming back instead of model text
auth/quota issues on the codex route

If you need a quick workaround right now, best options are:

Best workaround: use openai/gpt-5.4 with a real OpenAI API key, if you have one
Otherwise, switch off the Codex route temporarily to another working model/provider
If this reproduces on the latest version, it is probably GitHub-issue territory

Docs: https://docs.openclaw.ai/help/faq
Troubleshooting: https://docs.openclaw.ai/gateway/troubleshooting

Paste those outputs and I’ll help pin it down.

sturdy sparrow Apr 23, 2026, 1:53 PM

#

OpenClaw 2026.4.21 (f788c88)

exotic hemlockBOT Apr 23, 2026, 1:53 PM

#

Thanks, that version is recent enough that I would not assume “just update” here.

At this point I’d treat it as likely a Codex route regression until proven otherwise.

Please run these two and paste the output:

openclaw models status --probe
openclaw infer model run --local --model "openai-codex/gpt-5.4" --prompt "pong" --json

Then:

openclaw logs --follow

#

and trigger one tiny prompt so we can catch the matching log lines.

Why I’m asking for exactly those:

if models status --probe looks bad, it’s auth/quota
if infer model run returns terminated, 404, 502, or weird HTML, it’s very likely an OpenClaw bug on the openai-codex/gpt-5.4 path
if another provider works but this one fails, that narrows it even more

If that repro is cleanly failing on 2026.4.21, I’d file a GitHub issue with:

openclaw --version
the exact infer model run output
the matching logs --follow snippet

Issue tracker: https://github.com/openclaw/openclaw/issues

sturdy sparrow Apr 23, 2026, 1:56 PM

#

OpenClaw 2026.4.21 (f788c88)

15:54:07 [agent/embedded] workspace bootstrap file TOOLS.md is 14540 chars (limit 12000); truncating in injected context (sessionKey=probe-openai-codex-6ee31710-605d-4ac0-abab-1b8015edaf9f)

Config : ~/.openclaw/openclaw.json
Agent dir : ~/.openclaw/agents/main/agent
Default : openai-codex/gpt-5.4
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (0) : -
Configured models (2): openai-codex/gpt-5.4, openai-codex/gpt-5.2
Auth overview
Auth store : ~/.openclaw/agents/main/agent/auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (1): openai-codex (1)

openai-codex effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=1, token=0, api_key=0) | openai-codex:luis_tabla@hotmail.com=OAuth (luis_tabla@hotmail.com)

OAuth/token status

openai-codex usage: 5h 92% left ⏱3h 58m · Week 98% left ⏱5d 6h
- openai-codex:luis_tabla@hotmail.com (luis_tabla@hotmail.com) ok expires in 3d

Auth probes
┌──────────────────────┬──────────────────────────────────────────────────────────────────────┬────────────┐
│ Model │ Profile │ Status │
├──────────────────────┼──────────────────────────────────────────────────────────────────────┼────────────┤
│ openai-codex/gpt-5.4 │ openai-codex:luis_tabla@hotmail.com (luis_tabla@hotmail.com) (oauth) │ ok · 10.9s │
└──────────────────────┴──────────────────────────────────────────────────────────────────────┴────────────┘
Probed 1 target in 21.3s
15:54:35 [agent/embedded] workspace bootstrap file TOOLS.md is 14540 chars (limit 12000); truncating in injected context (sessionKey=agent:main:main)
{
"ok": true,
"capability": "model.run",
"transport": "local",
"provider": "openai-codex",
"model": "gpt-5.4",
"attempts": [],
"outputs": [
{
"text": "ping 🦊",
"mediaUrl": null
}
]
}

exotic hemlockBOT Apr 23, 2026, 1:57 PM

#

This is actually good news.

Your output says:

auth is ok
probe is ok
direct model.run with openai-codex/gpt-5.4 is ok
TOOLS.md truncation is just informational, not the problem

So the Codex route itself is working. That points much more at the session/thread state, not the model.

Best next step, in that same Discord thread:

/new

#

Then send hello again.

Why: if infer model run works but the normal chat thread fails, the usual culprit is a stale/corrupted session history in that thread. The docs explicitly recommend starting a fresh session for weird per-thread failures like this: https://docs.openclaw.ai/help/faq

If /new fixes it, that was the issue.

If /new does not fix it, then the next thing I want is:

run openclaw logs --follow
send one normal Discord message in the thread
paste the matching log lines

You can also test in a brand new thread as a quick control.

sturdy sparrow Apr 23, 2026, 1:59 PM

#

13:58:25+00:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ logs.tail 149ms conn=16f9e131…be07 id=681e410e…ea4b
13:58:58+00:00 warn agent/embedded {"subsystem":"agent/embedded"} {"event":"embedded_run_agent_end","tags":["error_handling","lifecycle","agent_end","assistant_error"],"runId":"9077acfc-99ef-4605-bb73-0062a7dd1f51","isError":true,"error":"terminated","failoverReason":null,"model":"gpt-5.4","provider":"openai-codex","rawErrorPreview":"terminated","rawErrorHash":"sha256:e8c95a2a7271","providerRuntimeFailureKind":"unknown"} embedded run agent end

exotic hemlockBOT Apr 23, 2026, 2:00 PM

#

Yep, that log is the smoking gun.

models status --probe and direct infer model run are healthy, but the embedded agent run is dying with:

error: "terminated"
provider: "openai-codex"
providerRuntimeFailureKind: "unknown"

So this does not look like bad auth, bad quota, or “GPT-5.4 just doesn’t work”. It looks like an OpenClaw embedded-run bug on the openai-codex/gpt-5.4 path.

I’d do exactly this next:

Try /new once in that thread and retest
If it still happens, file a GitHub issue with:
- openclaw --version
- your openclaw models status --probe
- your successful openclaw infer model run --local --model "openai-codex/gpt-5.4" --prompt "pong" --json

#

this failing log line from agent/embedded

Issue tracker: https://github.com/openclaw/openclaw/issues

A good summary would be:

openai-codex/gpt-5.4 works via openclaw infer model run, but normal embedded chat runs fail with error:"terminated" and providerRuntimeFailureKind:"unknown" on OpenClaw 2026.4.21.

If you want, I can help you turn what you already pasted into a clean issue report here in-thread.

sturdy sparrow Apr 23, 2026, 2:01 PM

#

please

exotic hemlockBOT Apr 23, 2026, 2:02 PM

#

Paste this into GitHub and fill the OS/install placeholders if you want:

Title:
[Bug]: openai-codex/gpt-5.4 works via infer model run, but embedded chat runs fail with "terminated" on 2026.4.21

Body:

### Bug type

Behavior bug (incorrect output/state without crash)

### Summary

On OpenClaw 2026.4.21, `openai-codex/gpt-5.4` works when called directly with `openclaw infer model run`, and `openclaw models status --probe` reports the model/profile as healthy, but normal chat/embedded runs fail with a generic `terminated` error.

#

This makes simple prompts like `hello` fail in the normal chat path even though the direct model route succeeds.

### Steps to reproduce

1. Configure OpenClaw to use `openai-codex/gpt-5.4` as the default model.
2. Verify model health:
   ```bash
   openclaw models status --probe

Verify direct inference works:

openclaw infer model run --local --model "openai-codex/gpt-5.4" --prompt "pong" --json

Send a simple prompt like hello through a normal chat/thread conversation.
Observe that the embedded agent run fails with terminated.

#


### Expected behavior

If `openai-codex/gpt-5.4` is healthy and `openclaw infer model run` succeeds, normal embedded chat runs should also complete and return a normal assistant reply.

### Actual behavior

- `openclaw models status --probe` reports the auth probe as `ok`
- direct `infer model run` succeeds
- normal embedded chat runs fail with:
  - `error: "terminated"`
  - `providerRuntimeFailureKind: "unknown"`

### OpenClaw version

#

`2026.4.21 (f788c88)`

### Operating system

`<fill this in>`

### Install method

`<fill this in, e.g. npm global / source / installer>`

### Model

`openai-codex/gpt-5.4`

### Provider / routing chain

#


normal embedded chat run -> `openai-codex/gpt-5.4`

### Logs, screenshots, and evidence

`openclaw models status --probe` showed the `openai-codex/gpt-5.4` auth probe as healthy (`ok`).

Direct model test succeeded:

```json
{
  "ok": true,
  "capability": "model.run",
  "transport": "local",
  "provider": "openai-codex",
  "model": "gpt-5.4",

#

"attempts": [],
"outputs": [
{
"text": "ping 🦊",
"mediaUrl": null
}
]
}


But the embedded chat path logged:

```text
warn agent/embedded {"subsystem":"agent/embedded"} {"event":"embedded_run_agent_end","tags":["error_handling","lifecycle","agent_end","assistant_error"],"runId":"9077acfc-99ef-4605-bb73-0062a7dd1f51","isError":true,"error":"terminated","failoverReason":null,"model":"gpt-5.4","provider":"openai-codex","rawErrorPreview":"terminated","rawErrorHash":"sha256:e8c95a2a7271","providerRuntimeFailureKind":"unknown"} embedded run agent end

#

### Impact and severity

This breaks normal chat usage for `openai-codex/gpt-5.4` even though the direct model route appears healthy.

### Additional information

I also saw bootstrap truncation warnings for `TOOLS.md`, but direct `infer model run` still succeeded, so that does not appear to be the primary failure.

Small privacy note: I’d leave your OAuth email/profile name redacted, like above.

If you want, I can also make this into a shorter Discord-sized summary before you file it.

sturdy sparrow Apr 23, 2026, 2:08 PM

#

created

#

https://github.com/openclaw/openclaw/issues/70627

exotic hemlockBOT Apr 23, 2026, 2:09 PM

#

Nice, thanks for filing it. 🦐

If you want, drop the GitHub issue link here so other people hitting the same terminated bug can track it too.

Glad we got it narrowed down to the embedded run path instead of you chasing auth/config ghosts.

#terminated