Question 1

Why can I not just ask the teacher a follow-up question?

Accepted Answer

Because every consult is a fresh codex exec with no resume and no thread id, so the teacher keeps nothing. The measured effect is that the file count and total bytes under the session directory stay unchanged. You have to paste the previous reply into the new request explicitly. On top of that, the budget allows only one follow-up or one escalation bump per task, never both, because the third slot is shared.

Question 2

Can the teacher touch my code?

Accepted Answer

No, and the restriction is enforced rather than requested. Consults run with the read-only sandbox flag applied by the codex sandbox itself, not by prompt wording, and the README records a measured attempt to make the teacher write a file, which came back BLOCKED with no file created. The teachers also cannot see the agent's conversation or talk to each other; each reply is self-contained.

Question 3

What exactly does Jev decide in this setup?

Accepted Answer

Only whether asking is worth it. The advisory tool returns plan, expert or none with three probabilities, and Jev can never initiate a consult itself. Two things gate it: a free pre-filter that must pass first, and an advisory budget of at most two calls per task, the second only if failures accumulate or a new architectural fork appears. The state sent is a whitelist with no transcript, diff or tool output, trimmed at about 2,000 tokens and skipped rather than sent truncated.

Question 4

What happens when a teacher call fails?

Accepted Answer

You get an explicit error, the budget slot is refunded, and nothing is retried automatically. If Jev itself fails, times out or has no key, the advisory is simply unavailable and the agent decides on its own, since that layer is meant to be non-blocking. If a model configuration is unavailable, that role refuses outright rather than quietly substituting a different model, and model and effort combinations are validated against the local model cache before each consult.

Question 5

Which teachers and models does it use?

Accepted Answer

A plan teacher on gpt-6-astra at low effort and an expert teacher on gpt-6.1-sol, at low for primary and high for escalation; the plan teacher answers with PLAN:, RISKS:, and VERIFY FIRST:.

Question 6

How does the free prefilter decide whether to call Jev?

Accepted Answer

It only considers advisory when the user explicitly asks for a plan, architecture, design, migration, or refactor, or the task spans modules, has two or more forks, has two consecutive failures, or has an unexplained blocker.

Question 7

How is the per-task budget structured?

Accepted Answer

Each human user task gets one plan, one expert primary, and one shared slot for a follow-up or an escalation, for at most three replies; the fourth request is refused before any process starts.

Dsh Teacher Consult

Highlights

Questions about this entry

More Jev resources like this

From the Jev community