What problem does it solve? Interactive coding agents are turn-based and cannot block for the days an HPC ensemble takes, so a multi-phase model calibration loop stalls between turns. This Skill acts as the run-to-convergence driver: it loads the persisted WorkflowStateOffline, resolves the next action, dispatches to the matching phase skill, saves state, and repeats until Phase-7 CONVERGED or a loop limit. ## Core Features & Use Cases - State-driven dispatch: Loads WorkflowStateOffline, calls resolve_next_action(), and routes to phase0 through phase6 skills, round-close steps, or housekeeping without re-implementing them. - Three valid stops only: Classifies every turn end as done (converged), wait (HPC ensemble in flight with monitors armed), or gate (one of four human decision forks), preventing silent halts with work outstanding. - Harness-neutral operation: Runs on any coding agent using only on-disk state, AGENTS.md rule 10, and whatever re-invocation trigger the harness provides, with no dependency on platform-specific features. - Use Case: A researcher calibrating an EcoSIM or PFLOTRAN site case says "keep calibrating until the targets are met"; the driver advances phases, pauses at the Phase-6 converge/redesign fork and curated-KB write gates, waits through HPC ensembles, and produces the final converged configuration. ## Quick Start Tell the agent to run the calibration to convergence for the current site case and let it drive the phase loop until the validation targets are met.
