On this page
How it works
Lookout runs on Jev, Venice's decision model. Jev writes nothing. It takes the session as data plus typed yes-or-no questions and returns a probability for each one. When an answer clears its bar, Lookout puts a card or a line in front of you. That is the whole job.
Every bar was set from 160 labelled examples: sessions with a real problem, and sessions built to look like one without being one. On the four tuned questions, every problem clears its bar and no look-alike does.
Before a turn
The moment you send a message, Lookout reads it with four questions: how big is the job, is it too vague to start, does it touch data, deployment, secrets or money, and what one detail would help most. The agent starts regardless.
- A clear request gets silence.
- A vague one gets a pointed question.
- A big one gets Plan first, which switches the run to plan mode so the agent proposes before it edits.
After a turn
The recent turns go out with five questions:
| Question | What it catches | The card offers |
|---|---|---|
| Looping | The same command, edit or error, with no progress. | Back to the request, Stop |
| Drift | Work toward a goal you never asked for. | Your request, quoted back |
| Stuck | Retrying an approach its own results already ruled out. | Re-plan |
| Unbacked | A claim no tool output shows: tests pass, file written, deployed. | Back to the request |
| Looks finished | Everything you asked for is done. | Review changes |
When a turn edits files, the change is also held to your project's own rules. Lookout splits the rule lines in CLAUDE.md and AGENTS.md into single questions and asks each one against the diff. A rule the change likely breaks gets a card that names it, with Ask to fix.
At an approval
When the agent asks permission, Lookout reads the pending call against your request. Does it reach outside the task? Can it be undone? Does it send something out? Each yes adds one line to the approval card. Allow, Deny and Cancel do not move.
What it cannot do
Lookout cannot approve a call, message the agent, change your model or hide an approval. When Venice does not answer, the check is skipped and the run carries on as if Lookout were off.
Two switches let it go further. Both are off until you turn them on:
- Set effort from the request sets how long the Claude engine thinks, never which model runs. It applies only on the Claude engine, and a size read under 50% sure leaves effort where it was.
- Pause after three looping turns stops a run that keeps looping, the way Stop does.
What it sends
Only what the questions read: your request, the recent turns and tool output up to a fixed size, the paths a turn edited, and the pending call at an approval. The rules check adds the rule lines and the diff, with credentials masked.
- Files named like
.envor a private key are never sent. - Nothing is sent for an incognito chat or an end-to-end encrypted session.
- Nothing is sent while Lookout is off. A question you turn off is never sent.
| Route | Model | Tier |
|---|---|---|
| Your session | A private model on Venice | Private: prompt and response are not retained after the request, by contract. |
| Lookout's checks | Jev, Venice's decision model, on your key | Anonymized: returns probabilities and keeps neither the state nor the answers. One step below private. |
Turning Lookout on sends excerpts of a private session to the second route. That is the trade, and it is why every question has its own switch. Privacy & data has the full row for every route.
Speed and cost
Your agent never waits for it.
Turn it on
Open Settings, find Lookout, and turn on Watch each session. Every question has its own switch.
If it flags something it should not, or misses something it should catch, tell us. That is how the next bar gets set.