# Slide Run Planner for PathML

> Before a computational-pathology run, paste the whole-slide image manifest and the slide, tiling
> and model numbers. A free in-browser check runs the pathml agent skill's own manifest validator,
> PathML 3.0.5 tile planner and inference batch planner; two metered lanes do the judgement from
> counts and plans only.

URL: https://pathml-desk.skillsafe.ai/
API: https://pathml-desk.skillsafe.ai/api.html
Tokens: https://pathml-desk.skillsafe.ai/tokens.html
Model: gpt-terra (the balanced GPT tier alias on SkillSafe)
Affiliation: Independent tool, not affiliated with or endorsed by the PathML project, Dana-Farber Cancer Institute or Weill Cornell Medicine. "PathML" names the open-source library this page plans runs for.
Source skill: @k-dense-ai/pathml (k-dense-ai/scientific-agent-skills, MIT) - local, research-only
computational pathology with PathML.

## What runs free, in the browser

JavaScript ports of three standard-library CLIs bundled with the skill, with the same JSON output and
error messages as the Python (checked on fuzzed inputs against CPython; see NOTICE.txt):

- slide_manifest.py validate - required columns slide_id, patient_id, path; pseudonymous identifier
  syntax; unique slide ids and resolved paths; no URLs; paths inside --root with a slide suffix;
  split values from --splits; no patient in more than one split. Stops at the first problem.
- plan_pipeline.py - PathML 3.0.5 tile-count arithmetic for a width, height, level downsample, tile
  size, stride and padding; overlap and gaps; mask/label/count outputs of the pipeline stages;
  uncompressed payload in GiB; physical tile size from microns per pixel.
- plan_inference.py - batch memory from numbers or a strict JSON model card (never a checkpoint):
  bytes per tile, peak memory against a budget, the largest batch under the estimate, total output.

The one difference, stated on the page: a browser cannot open the slide files, so each path is
checked as text (URL, NUL byte, ~ expansion, normalisation, escape from --root, duplicates, suffix)
and assumed to exist as a regular file. The page adds its own labelled work: a pass over every
manifest row (the tool stops at the first), identifier and path heuristics, per-slide tile plans when
the manifest has width/height columns, cross-checks between the tiling plan and the model card, and a
status: blocked, fix_first or pilot_ready.

## The two metered lanes (field `task`)

- review - what blocks a bounded pilot. Status blocked, fix_before_pilot or ready_for_pilot; a
  response to every flag; findings with the numbers they rest on and a fix; a pilot plan (scope,
  steps, checks on training slides only); PathML 3.0.5 pilot code bounded with islice.
- methods - a methods paragraph and a 16-item provenance record (version, formats and backend,
  slides/patients/splits, split unit, de-identification, tiling, stages, physical tile size, tile
  count and payload, model and checksum, input/dtype/batch, memory, QC, seeds, network), with
  "[to fill: ...]" wherever the facts cannot supply a value. It can take the plan review as input.

Both read only counts, tool reports, flags and what the user types - never a slide, a file name, a
slide id or a patient id. Every reply is reconciled in the browser: all flags answered once, tool
refusals and patient-split leaks never dismissed, a status no looser than the stances, every cited
number present in what was sent, the pilot code using the plan's own tile size, stride and stages.

## Limits, stated plainly

- Research use only. PathML is not a medical device; nothing here supports a diagnosis.
- The plans model PathML 3.0.5, the version the skill pins. PyPI's latest release is 3.0.8
  (2026-08-14, checked 2026-09-27); re-check arithmetic against the version you install.
- Tile counts, payloads and memory are dry-run estimates, not measurements.

## Citation

Kassis, T., Agarwal, V., He, Y., Patel, D., & Brueckner, A. M. (2026). Scientific Agent Skills: A
Library of Procedural Knowledge for Research Agents. arXiv:2609.00065.
https://doi.org/10.48550/arXiv.2609.00065