| name | solve |
| description | Use when the user brings a concrete quantum many-body research problem — phrases like "ground state of Heisenberg N = 20", "is kagome a spin liquid", "Hubbard at U/t = 8", "compute the gap for TFIM at h = 1", "what does VMC give for J1-J2 at 0.5". Also fires when `/onboard` routes here or when the user pivots to a new problem mid-session. |
Solve
Interactive problem-solving loop: infer problem, run the right harness path, report, offer next steps.
User is technical, expects concise results and a plot, and should not answer setup questions when defaults are clear.
When
- User states a concrete QMB problem.
/onboard routes here.- User pivots to a new problem.
Loop
intake -> match -> run -> report -> options -> repeat
Intake
Infer defaults from the user prompt and model/physics cards. Ask one narrow question only when the prompt maps to multiple model cards or materially different method branches.
Consult:
.knowledge/models/<name>/MODEL.md- relevant
.knowledge/physics/<topic>/PHYSICS.md
- the selected method card and stack contract
If a card redirects to dynamics, finite-T, or another stub, follow the redirect immediately.
Run
Use the selected method card's code shape. Install missing canonical stack profiles before compute; route remote/non-trivial compute through /using-slurm. Save scripts under scripts/, results under results/, and generate a convergence/stability plot.
Frontier card flag: run literature search first, present the literature-backed interpretation as the solve result, and offer compute as a next step.
Off-scope: use the closest card only if honest; label output "Off-skill — not harness-verified" or ask the user to choose among real redirects.
Report
Keep it to <=3 lines plus the plot: primary quantity, method/reason, verification state. Include the rerun command, e.g. julia --project=julia-env scripts/<name>.jl.
Options
Offer 2-3 real next steps, recommended first, with Done always available. Common options: richer visualization (a publication figure, or — for site-resolved data on a lattice — an interactive 3D view: emit a lattix scene JSON per viz/README.md and render it through /report's lattice3d block), parameter scan, deeper observable, cross-method check, literature context, writeup, different problem.
Done
When the user explicitly stops or picks Done, save results/session_summary.md, list produced scripts/results, and offer writeup handoff if appropriate.
