Opencode Jev Orchestrator

It keeps an OpenCode session on one cheap sticky model so the prompt cache stays warm, and only burns scarce models such as Luna inside temporary child subagents when a turn is flagged hard, instead of switching the parent model on every turn.

Jev flags each turn as hard, easy or unsure and never silently swaps the parent onto another model. The parent stays on Muse Spark; when escalation happens it calls the jev_escalate tool, and a strong child keeps working until Jev confidently says easy or is down or unsure.

Run `opencode plugin opencode-jev-orchestrator -g`, or add the plugin to ~/.config/opencode/opencode.json; the README documents pinning version 0.2.1. After restarting OpenCode, /jev-status should show key present, orchestration on and sticky parent Muse.

Because OpenCode often misses shell exports, the README recommends writing the key to ~/.config/opencode/opencode-jev-orchestrator.key with chmod 600 rather than relying on JEV_KEY being visible in the environment.

/jev-status shows key, orchestration and sticky parent; /jev-explain reports the last action; /jev-on and /jev-off toggle the session; /jev-quota lists exhausted models and /jev-exhausted marks one; /jev-reset clears marks.

The jev_parallel tool runs up to three concurrent cheap children for mechanical subtasks, and resumes send delta context only. Escalations give the strong child near-full context, and the result merges back through the tool return.

Yes — `use luna` and unmanaged picker pins still win, and models not listed in the tier configuration pin by default. The README recommends listing Claude/Codex models in the tier config or starting from one of the example configuration files.