pi-molagent bundles 12 skills covering the full OPLS-AA parameterization, polymer building, LAMMPS data prep, DFT charge / dihedral refinement, and plotting pipeline — plus steering content and HPC guards that keep the pi coding agent disciplined on login nodes.
Out of the box they invent force-field parameters, propose unconverged k-points, and forget your group's conventions. A plausible wrong answer burns weeks of compute.
Each skill is a self-contained SKILL.md plus helper scripts. The pi agent loads them on demand when the task matches.
Mol2 / PDB structure analysis, ring detection, atom typing.
Build oligomers and polymer chains from monomer mol2.
OPLS-AA parameterization via LigParGen + BOSS.
End-to-end OPLS-AA pipeline: param → data → validate.
Pack solvent boxes with Packmol + LAMMPS equilibration.
Merge multiple LAMMPS data files into one consistent box.
Wrap / unwrap LAMMPS trajectories across PBC.
Extract and renumber molecules from bulk LAMMPS data.
DFT charge derivation (Gaussian g16, Hirshfeld / RESP).
OPLS-AA dihedral fitting: DFT scan → LAMMPS coefficients.
Publication-quality matplotlib figure standards.
List and describe available molagent skills.
The skills compose into validated, group-tested pipelines. Each diagram below is exactly the sequence the agent runs — SLURM submits, gates, human checkpoints and all. Click a tab to inspect.
// every “compute” step runs through SLURM; only stdlib-Python helpers ever execute on the login node (RULE #0).
Skills are only half the story. pi-molagent ships steering content that encodes how a careful computational chemist actually works.
A tool_call hook blocks Python-on-login-node before SLURM admins do. Configurable: warn / strict / off.
RULE #1 forbids the agent from running pip install or conda install without your explicit "yes". The cluster has it already.
RULE #2: validated user scripts are off-limits to "improvements". The agent adds drivers on top instead of rewriting internals.
Subagents for devil's-advocate review and blind re-derivation of numerical claims. Invoke with /critic and /verify.
/hpc, /lammps, /gaussian load curated references for partitions, OPLS defaults, g16 conventions.
All paths, binaries, SLURM settings, and policy live in one molagent.json. /molagent:doctor validates it.
Compute skills emit SLURM job scripts and watch squeue/sacct. Auto-restart on failure where it makes sense.
Past-incident notes (LAMMPS gotchas, Colvars/PLUMED traps, plot discipline) so the agent stops repeating mistakes.
pi-molagent is a pi coding agent package. Pi discovers skills, prompts, and extensions automatically. The only manual step is the steering symlink.
$ npm install -g @earendil-works/pi-coding-agent
# Global install $ pi install git:git@github.com:your-org/pi-molagent # Project-local (recommended for shared HPC projects) $ pi install -l git:git@github.com:your-org/pi-molagent
$ pi list # find install path $ ln -s /abs/path/pi-molagent/steering/AGENTS.md \ ~/.pi/agent/AGENTS.md
› /molagent:init # inside pi › /molagent:doctor # env + skill readiness report
These are not suggestions in a prompt. RULE #0 is enforced by a pi extension
hook; RULE #1 and #2 are baked into steering/AGENTS.md and loaded
before every session.
No Python with numpy / scipy / torch / RDKit on the login node. No
conda activate followed by anything that spawns Python. No xargs -P.
When in doubt, SLURM. Enforced by a tool_call hook with
warn / strict / off policy.
No pip install, no conda install, no source builds,
no ~/.bashrc edits. The cluster has every major scientific package
preinstalled — find the module, ask if it fails.
Validated user scripts are off-limits to "improvements". When automation is requested, the agent adds a driver on top — it does not rewrite internals.
Your scientists shouldn't be spending their days writing input files. Hand the boilerplate to pi-molagent and keep them on the science.