[Feature]: Jev as the System-One lane for computer use

Summary

Small follow-up split from the JEV comment on RFC NousResearch/hermes-agent#112639 (comment).

OMP tracking issue — mirrors NousResearch/hermes-agent#113850.

Use Jev (TypeSafe System One) as an optional cheap backend for bounded computer-use decisions. Do not replace the planner, and do not block computer-use work on Jev.

Related: browser / computer-use tooling in OMP (browser relay, eval browser); #12372 / #12373 (official Jev skill suggestion + first-class /login jev); #12374 (other Jev follow-ups). Upstream: NousResearch/hermes-agent#112639 (CU RFC), #113008 (DecisionProvider RFC), #113837 (Jev provider listing).

Do now

Keep the first typed questions small and factorized:

• ACTION: click / type / key / scroll / wait / done / escalate
• TARGET: one currently visible semantic element
• NEEDS_VISION: yes / no
• NEEDS_GENERATION: yes / no
• DONE: yes / no

Do not flatten those into one giant choice space (click_element_1, type_element_2, …). Overlapping choices dilute confidence.

Backend order (Jev is last, not required):

deterministic rules
→ local semantic reranker
→ existing fast auxiliary model
→ Jev when TYPESAFE_API_KEY is set

Fail open to the current computer-use planner when the decision is uncertain, Jev is missing, or verification disagrees.

Out of scope

• Persistent GUI state, runahead, dependency scheduler (NousResearch/hermes-agent#112639)
• Host DecisionProvider runtime / replay service (NousResearch/hermes-agent#113008)
• Jev as the session chat model (#12372 / #12373 — never /model jev)
• Changing computer-use approval / safety paths

Acceptance

• Computer use can ask those typed questions without a frontier-model round trip when confidence is high
• Jev is optional: no key → same path as today
• Replayable decision packets: semantic state, candidates, scores, chosen action, latency, verifier outcome — no typed secrets or raw screenshots by default
• Same captured states can be A/B'd across backends (rules / reranker / aux / Jev)

Reference experiment (not an OMP benchmark): awlevin/typesafe-computer-use