OpenCode tools →Jev (optional)→ LiteLLM policy → Headroom → Model
File reads return to OpenCode. Its next standard LLM request passes through our proxy.
Meet Jev: smarter history decisions
EXPERIMENTAL
Headroom compresses tool context. Jev is a small decision model, accessed through OpenRouter, that checks whether completed older work is still relevant before LiteLLM receives a request.
Agent pruning stays Off by default. Short or single-task sessions usually keep their history. Try a prepared example below to see a real Jev decision without running a long agent session.
Jev considers older read-only tool context only after at least five minutes of session inactivity. It preserves completed-task summaries and checks recent references, cache risk, and savings beyond Headroom. A pause does not prove the cache has expired.
Creating two fresh sessions. This can take a few minutes.
The cards below compare this turn. Each branch also shows totals across all follow-ups.
Compression within policy-on run—
Total input before → after
Estimated task cost difference—
Baseline − compressed, input + output
Task latency difference—
Compressed − baseline
Exact evidence checks—
Baseline / Headroom · default task only
ⓘ
Agent paths may differ. Compare complete tasks as well as compression per request. Automatic OpenCode compaction and pruning are disabled. Repository sessions remain available briefly for follow-ups; reports and sessions have bounded retention and disappear on restart. These task checks are not a general correctness benchmark.
Keep the prefix cache warm
EXPERIMENTAL
While you are away, Jev judges once whether this conversation’s final state warrants keeping the prefix cache warm — a keep-alive verdict needs no further model calls per cycle.