Get fully up to date on a project, or build its weekly / PI-meeting update bundle, from INSIDE the project directory. Use when an agent lands cold in a project and needs the current state (open workstreams, what moved, open PI decisions, blockers), or when asked to "build the weekly update", "prep the PI meeting", or "aggregate the relevant files into the update dir". Mines only THIS repo, driven by its own .sync/manifest.yaml. No dependency on LabSync.
Run Quantum ESPRESSO DFT calculations. Use when asked to perform first-principles calculations, SCF, structural relaxation, band structure, DOS, phonons, or any ab initio quantum mechanical calculation.
Decide where and how to run compute jobs across available backends (HPC, cloud GPUs, local). Use when starting any compute-intensive task, choosing a backend or partition, deciding job parameters (walltime, GPU type, parallelism), or debugging compute infrastructure issues. Embeds the smoke-first iteration pattern — validate cheaply before scaling.
Orchestrate one or more long-running compute campaigns (HPC simulations, multi-stage MD pipelines, ensemble runs) where stages take hours to days and failures must not halt the pipeline. Use when running multiple parallel campaigns that share a stage structure (smoke → equilibration → production → analysis), when waiting on long jobs makes interactive driving impractical, or when designing a self-healing tick-based supervisor for compute work. Reads/writes per-campaign WORKFLOW.md files as durable state.
Analyze simulation data and compute properties. Use when asked to parse LAMMPS/QE output, calculate diffusion coefficients, RDF, MSD, energies, or generate plots and visualizations.
Generate candidate crystal structures from a chemical formula and assess their stability using ggen. Use when you need trial structures for a composition with no known experimental structure, or want stability-ranked polymorph candidates before running MD/DFT.
Run compute jobs (QE DFT, LAMMPS/MD, MLIP) that may outlive a single agent turn, without parking-and-dying. Use whenever a task launches work that could take more than a few minutes — decide block-vs-detach, detach long jobs durably, keep state in files, and let the asw-loop supervisor re-invoke you across ticks. Read BEFORE launching any simulation whose runtime you are unsure of.
Run molecular dynamics and materials calculations with ML interatomic potentials (MACE, CHGNet, M3GNet) via ASE — near-DFT accuracy at classical-MD cost. Use when asked to run simulations with ML/universal potentials, or when DFT would be too expensive (or is unavailable) but accuracy matters. For high-throughput screening of 10+ structures, see torch-sim.