---

name: detikzify description: "Load, generate, compile, rasterize, evaluate, train, and serve DeTikZify and TikZero workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

DeTikZify Repo Skill

Use this skill when the task names DeTikZify, detikzify, TikZ synthesis, sketch-to-TikZ, figure reconstruction, TikZero, MCTS-based generation, or this repository's web UI, training, evaluation, or dataset workflows.

This skill is for the public detikzify Python package and its repo-owned workflows. It is self-contained: do not depend on the original checkout remaining available once the skill is generated.

Quick Start

Install the package with the extras that match the task:

pip install "detikzify[examples]"

Add legacy when you need the v1 / timm-backed model paths:

pip install "detikzify[examples,legacy]"

For compile/rasterize or web UI workflows, make sure TeX Live, latexmk, Ghostscript, and Poppler are available on the host.

Minimal import check:

python scripts/api_smoke.py

Minimal compile smoke:

python scripts/tikz_smoke.py

Route By Task

• Programmatic inference, model loading, adapters, pipelines, compile/rasterize/save, and MCTS-backed sampling: use sub-skills/inference-and-rendering/SKILL.md.
• Gradio UI, python -m detikzify.webui, model selection, sharing, and runtime UI options: use sub-skills/webui/SKILL.md.
• Training, pretraining, GRPO refinement, sketchification, TikZero adapter workflows, checkpoints, and distributed launch patterns: use sub-skills/training-and-adapters/SKILL.md.
• Metric wrappers, evaluation scoring, redacted outputs, and the examples/eval.py workflow: use sub-skills/evaluation-and-metrics/SKILL.md.
• Dataset builders/loaders and the generic Node / MonteCarlo tree-search engine: use sub-skills/datasets-and-mcts/SKILL.md.

Common Decisions

• Use load_adapter(...) or an already adapter-augmented processor whenever the task needs text-conditioned generation. Plain image-only loading is not enough for text prompts.
• Treat a successful CPU import as not sufficient proof of GPU readiness. Check torch.cuda.is_available() and a tiny CUDA allocation when the workflow depends on CUDA.
• For TikzDocument, compile success, rasterizability, and non-empty output are distinct checks. A document can compile and still rasterize to an empty page.
• If a task mentions examples/refine.py, remember that it needs the TRL vision-support path plus TeX-backed compilation during reward computation.
• If a task mentions v1 models or other legacy paths, check the legacy extra and timm availability before assuming the model loader can resolve them.

Bundled References And Helpers

Read the smallest bundled reference or helper that matches the task:

• references/installation.md: package extras, system dependencies, and install variants.
• references/troubleshooting.md: cross-cutting install/import, CUDA, TeX, adapter, web UI, training, evaluation, and dataset issues.
• references/repo-provenance.md: source revision, branch, package version, and evidence paths.
• references/repo-routing-metadata.json: structured router metadata consumed during import.
• scripts/api_smoke.py: safe import/signature snapshot for the public API surface.
• scripts/tikz_smoke.py: safe compile/rasterize smoke for a tiny TikZ document.
• scripts/mcts_smoke.py: safe dummy-state sanity check for the MCTS engine.
• scripts/webui_help.sh: safe python -m detikzify.webui --help wrapper.

Safety And Scope

• Do not tell future agents to run the original repository's examples or tests as the runtime skill. Use the bundled scripts and references instead.
• Do not assume system TeX tools or optional model extras are available unless the task already confirmed them.
• Do not claim CUDA, TeX, or adapter support from a CPU-only import check.
• Do not leak local paths, environment names, or the original checkout into runtime guidance.