What the plugin adds
The paperclip-plugin-jev project page links to a pinned README that describes two surfaces. Issue triage returns a recommended owner, priority, issue type, missing-context probability and blocker probability. A separate Paperclip agent tool chooses one browser operation and one observed target from a minimized page snapshot.
The browser tool returns an operation such as CLICK, TYPE_TEXT, SELECT, SCROLL, WAIT, DONE or BLOCKED. Jev does not generate the text for TYPE_TEXT, and the executor is expected to resolve the target against the same live snapshot and verify the result.
Understand the automation modes
| Mode | Documented behavior | Safe starting question |
|---|---|---|
| Advisory | Runs when a user selects “Analyze with Jev” and never changes issue fields | Do the recommendations help a human triage the queue? |
| Auto when confident | Processes new issues and updates owner or priority only when thresholds pass | Are the thresholds and review samples appropriate for this issue mix? |
| Always auto | Applies every valid owner and priority recommendation for new issues | What is the rollback path when the taxonomy or provider changes? |
Plugin updates listen only for issue.created, and the documented idempotency marker prevents a duplicate create event from consuming another call. Automatic processing does not change status, description, comments, attachments or relationships. Those boundaries should remain part of your review policy.
Keep browser execution behind a separate gate
Browser decisions have their own modes: Observe only, Confirm mutations and Auto safe actions. The README recommends starting with observation and comparing suggestions with human decisions. Click, text and select operations require confirmation in the middle mode, while sensitive controls remain gated in the automatic mode.
Do not treat DONE as proof that the goal succeeded. The executor must reject stale or obscured targets, log the attempted operation before execution and verify the resulting state afterward. This is a Paperclip integration contract, not a claim that Jev can safely operate arbitrary sites.
Check the data boundary
For triage, the documented payload includes issue title, description, status and priority plus eligible agent names, roles, titles, statuses and capabilities. The plugin says it does not send comments, attachments, run logs, workspace files or credentials. Browser decisions send the goal, URL, element roles and labels, select options and a short action history; visible page text is opt-in and screenshots, DOM selectors and input values are excluded.
The provider base URL is fixed to TypeSafe's endpoint, and the Paperclip secret reference resolves the API key in the plugin worker. Review the current pinned source and your own Paperclip configuration before sending real issue text or enabling an automatic mode.
Start with a reversible setup
- Install Paperclip with plugin support and confirm the current plugin publication status.
- Store the TypeSafe key in Paperclip Secrets rather than source code, comments or shell history.
- Enable
Observe onlyand the advisory triage mode. - Review a labeled sample of recommendations, missing-context cases and blocked browser actions.
- Only then choose thresholds and a narrowly scoped automatic mode.
For other coding-agent integrations, compare the Jev agent guide. For credentials and endpoint differences, use the API and SDK guide; the plugin is community-maintained and is not endorsed by Paperclip or TypeSafe.