Release v2.46.0 · aebrer/dreb

Evidence-Based Subagent Routing

dreb can now research scoped models and optionally arbitrate every subagent route before spawning.

Model routing guide. /skill:model-routing-guide generates or incrementally updates a researched guide for explicitly configured models, combining external evidence with sanitized local subagent history.
Thinking overrides. Subagent requests support off, minimal, low, medium, high, and xhigh, including per-task overrides for parallel and chain modes.
Dispatch Arbiter. An opt-in, global-only, headless model call can revise the agent, scoped provider/model, and thinking level before each child spawn.
Fail-closed routing. Invalid configuration, stale guides, unavailable models, malformed decisions, provider failures, and unsupported thinking levels prevent spawning rather than bypassing arbitration.

Configure arbitration through interactive /settings, dashboard Settings, settings JSON, or RPC.
Changed, unchanged, and failed decisions appear in TUI, JSON, RPC, session records, and dashboard views.
Final routed identity, canonical model, effective thinking, and chain-step metadata remain consistent across lifecycle events.
The arbiter sees a size-limited excerpt of the parent session, including recent tool output. dreb removes recognized secrets before sending it. Only the final agent/model/thinking choice is recorded—not the arbiter's prompt, response, or reasoning.

Preserves loud model-fallback diagnostics alongside final arbitrated routes.
Adds extensive offline coverage for single, parallel, and chain dispatch; cancellation; concurrency; scope and guide validation; settings recovery; persistence; hydration; and UI rendering.
Updates root and package documentation for routing guides, thinking overrides, settings, sessions, JSON/RPC events, and dashboard behavior.

Implemented in PRs 405 and 409.