Summary
The Codex app lets me insert/tag [@chrome](plugin://chrome@openai-bundled), but fresh chats do not receive chrome@openai-bundled as an active callable tool.
The UI and runtime disagree:
- UI:
chrome@openai-bundled can be selected/mentioned.
- Runtime: the model reports that the Chrome tool is not available in active tools.
- Result: requests tagged with
@chrome either fail or fall back to Browser / in-app browser.
Environment
- Platform: Windows x64
- App: Codex desktop app
- Plugins enabled in
~/.codex/config.toml:
[plugins."chrome@openai-bundled"]
enabled = true
[plugins."browser-use@openai-bundled"]
enabled = true
Local plugin files exist at:
~/.codex/plugins/cache/openai-bundled/chrome/0.1.7
Reproduction
- Start a fresh Codex chat.
- Insert/select the Chrome plugin mention from the UI:
[@chrome](plugin://chrome@openai-bundled).
- Ask the agent not to use Browser/in-app browser and to report which browser tools are actually active.
Prompt used:
Use only [@chrome](plugin://chrome@openai-bundled). If chrome is not actually available in active tools, do not use Browser/in-app browser. Tell me which browser tools you see.
Observed behavior
The model replies that chrome@openai-bundled is not present in active tools and that only Browser / browser-use / in-app browser is available.
In a separate attempt, asking [@chrome](plugin://chrome@openai-bundled) ajde idi na google.com caused the agent to use the in-app Browser plugin instead. When asked afterward, it confirmed it used Browser / in-app browser, not Chrome.
Expected behavior
If [@chrome](plugin://chrome@openai-bundled) can be selected in the UI and the Chrome extension is connected, the Chrome tool should be exposed to the model runtime.
If the Chrome tool cannot be exposed, the UI should not allow selecting the mention or should show a clear unavailable state before sending.
Additional diagnostics
The Chrome/extension backend can be reached manually from an existing diagnostic thread via the bundled chrome browser-client. In that diagnostic thread, the backend could list user tabs and open https://example.com successfully.
This suggests the browser backend itself can work, but new chats are not receiving chrome@openai-bundled in their active tool set.
Related issues that look similar in class:
Impact
This makes @chrome unreliable: the UI allows selecting it, but the runtime may silently lack the actual Chrome tool, causing accidental fallback to the in-app browser or failed Chrome-specific workflows.
Summary
The Codex app lets me insert/tag
[@chrome](plugin://chrome@openai-bundled), but fresh chats do not receivechrome@openai-bundledas an active callable tool.The UI and runtime disagree:
chrome@openai-bundledcan be selected/mentioned.@chromeeither fail or fall back to Browser / in-app browser.Environment
~/.codex/config.toml:Local plugin files exist at:
Reproduction
[@chrome](plugin://chrome@openai-bundled).Prompt used:
Observed behavior
The model replies that
chrome@openai-bundledis not present in active tools and that only Browser / browser-use / in-app browser is available.In a separate attempt, asking
[@chrome](plugin://chrome@openai-bundled) ajde idi na google.comcaused the agent to use the in-app Browser plugin instead. When asked afterward, it confirmed it used Browser / in-app browser, not Chrome.Expected behavior
If
[@chrome](plugin://chrome@openai-bundled)can be selected in the UI and the Chrome extension is connected, the Chrome tool should be exposed to the model runtime.If the Chrome tool cannot be exposed, the UI should not allow selecting the mention or should show a clear unavailable state before sending.
Additional diagnostics
The Chrome/extension backend can be reached manually from an existing diagnostic thread via the bundled chrome browser-client. In that diagnostic thread, the backend could list user tabs and open
https://example.comsuccessfully.This suggests the browser backend itself can work, but new chats are not receiving
chrome@openai-bundledin their active tool set.Related issues that look similar in class:
Impact
This makes
@chromeunreliable: the UI allows selecting it, but the runtime may silently lack the actual Chrome tool, causing accidental fallback to the in-app browser or failed Chrome-specific workflows.