- name
SKILL.md
El archivo de la skill permanece en el repositorio del publicador.
Pega el comando copiado en una terminal. Instala este SKILL.md en ese agente.
npx skills add fcakyon/labrat --skill debug -a cursor -yPega el comando copiado en una terminal. Instala este SKILL.md en ese agente.
npx skills add fcakyon/labrat --skill debug -a claude-code -yPega el comando copiado en una terminal. Instala este SKILL.md en ese agente.
npx skills add fcakyon/labrat --skill debug -a codex -yPega el comando copiado en una terminal. Instala este SKILL.md en ese agente.
npx skills add fcakyon/labrat --skill debug -a grok -yEnvía el mensaje copiado en un chat de Grok Bot. El Bot obtiene este SKILL.md de GitHub y lo añade a las skills que comparten todos tus Bots.
Install this skill: https://github.com/fcakyon/labrat/tree/HEAD/plugin/skills/debugDel publicador
The most expensive class of mistake in ML debugging is asserting a cause based on plausibility, then attempting a "fix" that masks the real problem. This skill enforces the discipline of probe → hypothesis → smoke → controls → claim, in that order.
The agentic Stop hook routes here from reason when an assistant claims a cause without backing tool output.
The user just said any of:
Before forming any hypothesis, gather the cheap evidence. None of these cost more than a few seconds:
Process state:
ps aux | grep -E '(python|train|torchrun|accelerate)' | grep -v grep
Is the process still running? Zombie? Defunct? Multiple instances?
Kernel / system events:
dmesg | tail -100 # OOM kills, hardware errors, NFS errors
journalctl -xe --since "1 hour ago" | tail -50
GPU state:
nvidia-smi
nvidia-smi --query-gpu=utilization.gpu,memory.used,temperature.gpu --format=csv
Is the GPU even being used? Idle GPU during "training" means the process is blocked on data loading or has died.
Disk / filesystem:
df -h /path/to/run-dir
du -sh /path/to/run-dir/*
Out of disk? Checkpoints not being written?
Log scrollback: Read the last few hundred lines of the training log. Don't trust the user's summary, they may have skimmed. Look for:
Checkpoint state:
ls -la /path/to/run-dir/checkpoints/
When was the last checkpoint written? What does its size suggest? An empty .pt is different from a 2GB one cut short.
After the probe, state what might be happening, explicitly framed as a hypothesis:
"Hypothesis: the run is OOMing because dmesg shows
oom-kill3 minutes ago and the process is gone. Alternative hypotheses I haven't ruled out: (a) NFS write timeout, (b) explicit kill from a sibling process."
Never skip to "the cause is X." The hypothesis labels what you don't yet know.
The cheapest way to confirm or refute a hypothesis is to reproduce the failure shape under a controlled condition:
batch_size=1 for 1 step. If it survives, OOM is confirmed; if it fails the same way, OOM is wrong.eval() mode. Loss finite? Outputs sane?lr=0. If the loss still explodes, the loss itself is broken (not the optimizer).A 30-second smoke beats a 30-minute restart-and-pray.
If the smoke is ambiguous, run a control: change exactly one variable from the failing config and rerun the smoke. The differences narrow what mechanism is responsible.
Common control axes (change one at a time):
Only after evidence stacks up, probe, smoke, control, do you assert a cause. The claim should cite the specific tool output that proves it:
"Root cause: NFS write timeout. Evidence: dmesg shows
nfs server X not respondingat 14:23 (the same minute the last checkpoint was written), and the smoke withbatch=1reproduces the timeout. Recommended fix: bind-mount a local scratch dir for checkpoints and rsync to NFS at end of epoch."
If the evidence isn't stacking up, do not promote a hypothesis to a cause. Say "I don't yet know" and propose the next probe.
A concise diagnostic report: (1) what the probes showed, (2) the hypothesis, (3) the smoke outcome, (4) the cause-or-uncertain verdict, (5) the recommended next action. Each claim cites the tool output that backs it.
Evidence-before-action diagnosis of failing ML experiments. Probes the system before guessing causes, process list, dmesg, GPU stats, log scrollback, checkpoint state, then states a hypothesis as a hy
---name: debugdescription: Evidence-before-action diagnosis of failing ML experiments. Probes the system before guessing causes, process…---# plugin/skills/debug/SKILL.mdResumen general
Evidence-before-action diagnosis of failing ML experiments. Probes the system before guessing causes, process list, dmesg, GPU stats, log scrollback, checkpoint state, then states a hypothesis as a hypothesis and runs a smoke before claiming a root cause. Use when the user asks why a run is failing, diverging, OOMing, hanging, slow, producing weird metrics, has crashed, or asks to debug, diagnose, troubleshoot, or investigate a training issue.
Preguntas frecuentes
Descarga plugin/skills/debug/SKILL.md desde el publicador, o copia la ruta y abre el repositorio. PlugMCP lista la skill y no aloja el archivo.
El publicador lo mantiene en https://github.com/fcakyon/labrat.
Claude, Cursor, Codex y Grok Build pueden leer un SKILL.md que añadas desde el repositorio del publicador. Grok Bot lo instala cuando envías en un chat el enlace a la carpeta de la skill en GitHub.