https://github.com/ThisIsJeron/openclaw-codex-watchdog
OpenClaw plugin that blocks Codex narrative-loop replies when no tool calls were made.
https://github.com/ThisIsJeron/openclaw-codex-watchdog
Last synced: 20 days ago
JSON representation
OpenClaw plugin that blocks Codex narrative-loop replies when no tool calls were made.
- Host: GitHub
- URL: https://github.com/ThisIsJeron/openclaw-codex-watchdog
- Owner: ThisIsJeron
- Created: 2026-04-11T01:44:28.000Z (2 months ago)
- Default Branch: main
- Last Pushed: 2026-04-25T20:26:39.000Z (about 2 months ago)
- Last Synced: 2026-05-01T20:37:09.073Z (about 1 month ago)
- Size: 5.86 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- awesome-openclaw-plugins - codex-watchdog - Guardrail plugin that blocks Codex narrative-loop replies when action-oriented runs make zero tool calls, replacing fake progress with an honest watchdog message. By [@ThisIsJeron](https://github.com/ThisIsJeron). `openclaw plugins install https://github.com/ThisIsJeron/openclaw-codex-watchdog.git` (Skills vs Plugins: What's the Difference? / Bundled Plugins)
README
# openclaw-codex-watchdog
> Status: monitoring mode. This watchdog was built for an OpenClaw instance where `openai-codex/gpt-5.4` sometimes produced text-only "work happened" replies. With GPT-5.5 behavior improving, this repo is likely to become deprecated if the watchdog is no longer needed.
A small OpenClaw plugin that blocks Codex narrative-loop replies when a run looks action-oriented but the model makes zero tool calls.
## What it does
If a Codex run:
- uses `openai-codex/gpt-5.3-codex` or `openai-codex/gpt-5.4`
- looks like the user asked for an action
- makes zero tool calls
- returns text that sounds like fake progress or completion
then the plugin blocks that reply and replaces it with an honest watchdog message.
## Why
Sometimes Codex returns text like "I'll do that" or "done" without actually using tools or performing work. This plugin is a first-pass guardrail against that failure mode.
GPT-5.5 may have improved this behavior enough that the guardrail can be removed. For now, keep this plugin enabled only where monitoring shows the older failure mode still matters.
## Install
Add the plugin path to your OpenClaw config, for example:
```json
{
"plugins": {
"load": {
"paths": [
"/path/to/openclaw-codex-watchdog"
]
},
"entries": {
"openclaw-codex-watchdog": {
"enabled": true
}
}
}
}
```
If you use a restrictive `plugins.allow` list, include `discord` and any other bundled plugins you still want, plus this plugin if needed by your setup.
## Files
- `openclaw.plugin.json` - plugin manifest
- `dist/index.js` - plugin entry
## Status
Monitoring mode; likely to be deprecated if GPT-5.5 continues to reliably execute action-oriented requests with tool calls.
This is an early, practical first pass. It is intentionally simple and focused on blocking obviously fake "work happened" replies for affected older Codex model routes.