Roadmap: implemented vs planned

June 15, 2026 · View on GitHub

A snapshot of what the v0.2.0 codebase delivers and what remains, against the development plan.

Implemented (verified in CI on toy/synthetic systems)

AreaModule(s)Verification
Shared core drivercore/driver.py, engine.py, progress.py, selection.pyfull-driver smoke test on the toy engine
Multi-GPU device poolcore/parallel.py, backend engine.pyper-backend device-dispatch tests; G=1 ≡ serial
Toy Wolfe–Quapp enginecore/toy.pyused throughout the suite
Strategy profilescore/strategy.pyprofile-resolution + guard tests
Data-driven CV (SPIB)cv/features.py, cv/spib.py2-state recovery + driver integration
Non-linear search (RRT/Connect)search/rrt.py, core/policy.pyRRT reaches the opposite WQ basin; Connect links both
Roadmap graphsearch/roadmap.pyDijkstra + Yen k-shortest correctness
Agentic controlleragent/controller.pyescalate/relax/stop/frontier rules
Weighted Ensemblesampling/weighted_ensemble.pyweight conservation + WQ FES recovery
OPES (core + PLUMED interface)sampling/opes.pytoy WQ marginal recovery; PLUMED input generation
Path sampling (TPS/TIS) bridgesampling/path_sampling.pyseed prep / reactive-path extraction / interfaces (OPS run needs the pathsampling extra)
Downstream wiringsampling/runner.py, backendstoy run_downstream test

Run pytest -q to reproduce (69 tests; SPIB tests need the ml extra).

Environment caveats (not runnable in a CPU-only sandbox)

These code paths are implemented and exercised through the shared, tested helpers, but a live run needs hardware/binaries unavailable here:

  • Real-MD multi-GPU runs of the AMBER (pmemd.cuda) and GROMACS (gmx mdrun) backends.
  • Downstream stages launched from a real backend (WE/OPES on solvated systems via pathgennie.downstream).
  • OPES on real MD, which needs a PLUMED-patched engine exposing run_plumed (see opes.md).
  • TPS/TIS runs, which need the pathsampling extra (openpathsampling) and an OPS engine; only the dependency-free seed preparation is CI-tested (see path-sampling.md).

Planned / future work

  • A run_plumed capability on the MD engines (PLUMED patch / CustomCVForce) so OPESStage(mode="plumed") runs end-to-end on real systems.
  • RRT* rewiring of the roadmap (edge cost = -log(transition fraction)) to bias toward minimum-free-energy tubes.
  • A learning controller (contextual bandit / RL) and an optional LLM meta-controller behind the existing Controller surface.
  • Process-pool OpenMM execution (one CUDA context per GPU) for in-process multi-GPU.