Token-efficiency release. The deterministic orchestrator is now on by default — validated on a real multi-host pivot range with −87% provider overloads and no loss of reasoning depth.
New
- Orchestrator by default — the coordinator dispatches subagent waves through a bounded rolling pipeline (a slot refills the instant one finishes), keeping depth while capping concurrency. Disable with
OPENCODE_DISABLE_ORCHESTRATOR=true. - Cheap-model offload — large raw tool output (nmap/nuclei/gobuster) is digested by a configurable
small_modelbefore entering the transcript; raw stays retrievable by ref. Reasoning always runs on the main model. New TUI picker forsmall_model. - Resolved-vectors ledger — dead-end attack vectors are remembered so the agent stops re-testing them.
Improved
- Provider overload cap — per-provider concurrency limiter stops fan-outs from hitting 429/529.
- Leaner prompts & context — objective-based task prompts, cached system prefix, smaller per-turn state.
abandon(let it finish) split fromkill(hard-stop) for subagents.
Fixed
- Config
invalidatenow refreshes the live session (small-model changes apply immediately). - Offload failures surface inline instead of silently disappearing.
- TUI subagent bar tracks live child sessions.