mx3-log

# The notebook Work on one sample runs over weeks. Without a record, every session starts by re-deriving what was already settled — and the person least able to reload that context is the one who did not write the scripts. ```bash N=${CLAUDE_PLUGIN_ROOT}/skills/mx3-log/scripts/notebook.py python3 $N --root . scan # find every run and record it python3 $N --root . note "Hc converged at 4 nm" python3 $N --root . note "does the frame width matter?" --open python3 $N --root . open # what is still unresolved python3 $N --root . resolve "frame width" --answer "no, <1% over 5-20 nm" python3 $N --root . show ``` Everything lives in one `MX3-NOTEBOOK.md` beside the simulations. Plain markdown — readable, editable and committable without this tool. ## What scan reconstructs Each mumax3 output directory carries `log.txt`, which holds the script exactly as executed and the build that ran it. So the run history comes from **evidence, not memory**: mesh, material parameters, how many rows and snapshots, when it ran, and a `DIVERGED (NaN)` flag where the table went bad. `scan` is idempotent — re-running it adds only what is new. ## When to write - **After a conclusion, not after a run.** "Ran sweep_3.mx3" is already in the directory. "Coercivity is mesh-converged below 4 nm" is not. - **Open questions as they appear.** The assumption you had to make and did not test is exactly what you will forget. - **At the end of a session**, record where things stand, so the next one does not start cold. ## At the start of a session Read the notebook before proposing anything: ```bash python3 $N --root . show python3 $N --root . open ``` If a user asks something the notebook already settled, say so and cite the note rather than re-running it. If an open question is relevant to what they just asked, raise it. ## Related skills - **mx3-run** — produces the directories this reads - **mx3-check** — its convergence verdicts are the notes worth keeping - **mx3-paper** — a reproduction spec is a more formal version of the same idea