AGENTS.md git:20261028.f695e6aA

Immutable. This exact content never changes and is served at /api/v1/blob/36deccc4f059958e.

# Mathodology Agent Guide

Mathodology is a skills pack for mathematical modeling contests: prompts that
support mathematical judgment, reproducible computation and clear scientific
communication. Adapt the amount of work to the actual problem and deadline;
quality follows from the argument and evidence, not from phase gates, panel
scores or a predicted prize.

## Skills and roles

- `mathodology-whole-project`: entry point, installation, backup and repository use.
- `mathodology-agent-pipeline`: modeling prompts and focused collaboration.
- `mathodology-evidence-search`: literature, data, citations and licensed references.
- `mathodology-figure-presets`: scientific figure selection, design, image2 and examples.
- `mathodology-award-gates`: substantive mathematical and editorial review questions.
- `mathodology-workbench`: executable workspace checks, recorded runs, results contract and hooks.
- `mathodology-project-orientation`: skills-only repository boundaries.
- `mathodology-skill-authoring`: maintain skills and their metadata.
- `mathodology-dev-test-release`: optional repository maintenance checks.

Roles in `.claude/agents/` (Codex: `.codex/agents/`) are for parallel or
isolated work: brief one with a concrete question and evidence, and ask for
findings, artifacts and limitations in ordinary prose. A small task is done by
a single agent.

## Figures

Load `mathodology-figure-presets` before figure work. It owns the image2
availability question (ask once per task; reuse any answer already in the
conversation) and the default data-driven rendering route.

## Workbench and hooks

`mathodology-workbench` is the executable part of the pack: `scripts/mtd.py`
(standard library only) records runs and checks that results, figures and the
paper agree. `.claude/settings.json` registers its Claude Code hooks; they act
only inside a contest workspace (`.mathodology/`, or `problem/` with `paper/`)
and stay silent elsewhere, including in this repository. Results come from code:
do not edit `results/*.json` by hand to match a paper.

## Repository boundary

Maintain skills in `.claude/skills/`, roles in `.claude/agents/` and
`.codex/agents/`, short workflow prompts in `.claude/workflows/` and the shared
hooks in `.claude/settings.json` (generated by `mtd.py install-hooks`). This is
the sole source of truth; `.agents/skills/` is a gitignored local mirror for
Codex. Back up this project's mirror before refreshing it; do not modify other
skills or global settings.

Keep skill references, licensed exemplars and small optional utilities within
the owning skill. Contest outputs belong in a separate working project or the
ignored `work/` directory. `.mcp.json` configures evidence search; keep
credentials and machine-specific paths out of it.

The one application is Mathodology Studio in `apps/studio/`: an Electron
desktop workbench that drives locally installed agent CLIs (Claude Code, Codex,
Antigravity CLI, Qoder CLI, pi, oh-my-pi and the ACP agents in its catalog,
each through its own programmatic interface) and installs the contest skills
into each project. Its manifest, lockfile and sources stay inside that folder;
see its [README](apps/studio/README.md). The one workflow,
`.github/workflows/studio-windows.yml`, builds and signs its Windows installer
(SignPath signs only builds from GitHub-hosted runners). Do not add other
application source, datasets, deployment or build outputs, and never commit
`node_modules/`, `out/` or `release/`.

See [the workflow](docs/WORKFLOWS.md) and [installation](docs/INSTALL.md).