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)
| Area | Module(s) | Verification |
|---|---|---|
| Shared core driver | core/driver.py, engine.py, progress.py, selection.py | full-driver smoke test on the toy engine |
| Multi-GPU device pool | core/parallel.py, backend engine.py | per-backend device-dispatch tests; G=1 ≡ serial |
| Toy Wolfe–Quapp engine | core/toy.py | used throughout the suite |
| Strategy profiles | core/strategy.py | profile-resolution + guard tests |
| Data-driven CV (SPIB) | cv/features.py, cv/spib.py | 2-state recovery + driver integration |
| Non-linear search (RRT/Connect) | search/rrt.py, core/policy.py | RRT reaches the opposite WQ basin; Connect links both |
| Roadmap graph | search/roadmap.py | Dijkstra + Yen k-shortest correctness |
| Agentic controller | agent/controller.py | escalate/relax/stop/frontier rules |
| Weighted Ensemble | sampling/weighted_ensemble.py | weight conservation + WQ FES recovery |
| OPES (core + PLUMED interface) | sampling/opes.py | toy WQ marginal recovery; PLUMED input generation |
| Path sampling (TPS/TIS) bridge | sampling/path_sampling.py | seed prep / reactive-path extraction / interfaces (OPS run needs the pathsampling extra) |
| Downstream wiring | sampling/runner.py, backends | toy 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
pathsamplingextra (openpathsampling) and an OPS engine; only the dependency-free seed preparation is CI-tested (see path-sampling.md).
Planned / future work
- A
run_plumedcapability on the MD engines (PLUMED patch /CustomCVForce) soOPESStage(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
Controllersurface. - Process-pool OpenMM execution (one CUDA context per GPU) for in-process multi-GPU.