What problem does it solve? Running an Arbor-style autonomous research workflow requires coordinating many phases—intake, ideation, executor dispatch, merge evaluation, related-work search, resume, and reporting—without losing run state or violating experiment discipline. This Skill acts as the top-level controller that loads the right phase skill at the right time and enforces the non-negotiable invariants of the workflow. ## Core Features & Use Cases - Phase Loading Order: Defines exactly when to load setup-intake, coordinator, ideate, executor, merge-eval, search, plugins/HITL/budget, resume-report, and tools skills. - Invariant Enforcement: Protects B_test from contamination, keeps the main branch protected, requires Idea Tree metadata before executor dispatch, and mandates durable artifacts. - Smoke and Forward-Test Mode: Propagates smoke-only behavior through contracts, executor prompts, and reports so the suite can be validated without running expensive training jobs. - Use Case: A user asks to recreate the open-source AutoResearch workflow on a new benchmark; this orchestrator sequences the phase skills, maintains the Idea Tree as durable memory, and produces a final REPORT.md with preserved evidence. ## Quick Start Ask the agent to run an Arbor-style research cycle using the arbor-agent-orchestrator, starting with intake and a smoke-only validation run.
