---

name: easy12306 description: "Routes easy12306 12306 captcha recognition tasks across pretrained inference, data preparation, text-model training, and image-model training workflows." disable-model-invocation: true metadata: disco-role: operating license: Artistic 2.0

easy12306 repo skill

Use this repo skill when a task involves easy12306, the legacy Python/OpenCV/Keras project for recognizing 12306 captcha prompts and image tiles. It is a router: load the focused sub-skill that matches the user's goal, then use that sub-skill's bundled references and scripts instead of reopening source files from an original checkout.

Quick fit

Read this skill when the user asks about:

• Running or adapting 12306 captcha recognition from model.h5, 12306.image.model.h5, texts.txt, and a captcha image.
• Validating easy12306 image/model/label/data artifacts before running legacy Keras code.
• Recreating prompt-text crops, eight image tiles, perceptual hashes, .npz datasets, or OCR-assisted labels.
• Training or diagnosing the text prompt classifier or image-tile classifier.
• Troubleshooting Keras/TensorFlow, OpenCV, missing model/data files, Baidu OCR credentials, or geometry/label mismatches for this repo.

Avoid this skill when the request is only generic CAPTCHA solving, modern object detection/OCR unrelated to easy12306 artifacts, or editing a different repository.

Compatibility baseline

The repository is a flat script collection rather than an installable Python package. The verified inspection baseline was:

Python 3.11 + TensorFlow/Keras 2.15-compatible APIs + OpenCV + NumPy + SciPy + scikit-learn + matplotlib + requests

Setup

Create a Python 3.11 environment, then install the legacy-compatible runtime dependencies explicitly:

python3 -m pip install \
  "tensorflow-cpu==2.15.1" "keras==2.15.0" \
  opencv-python-headless numpy scipy scikit-learn matplotlib requests

Use opencv-python instead of opencv-python-headless only when the task needs GUI/image-window features. The source dependency list used the deprecated sklearn package name; use scikit-learn in new environments.

Minimal verification command:

python scripts/check_environment.py

Keras 3 can break legacy imports because the image-model script imports keras.preprocessing.image.ImageDataGenerator. Use references/troubleshooting.md and scripts/check_environment.py before running model code in a new environment.

Route map

User task	Load
Use pretrained artifacts for a full captcha or single image-tile prediction; validate image/model/label prerequisites; interpret row/column output.	sub-skills/inference/SKILL.md
Prepare captcha data, prompt-text crops, eight object tiles, perceptual hashes, label vocabulary files, Baidu OCR labeling notes, or .npz data schemas.	sub-skills/data-preparation/SKILL.md
Train, fine-tune, inspect, or troubleshoot the cropped prompt-text classifier and model.h5/model.v*.h5 artifacts.	sub-skills/text-modeling/SKILL.md
Train, inspect, or troubleshoot the VGG16-based image-tile classifier and 12306.image.model.h5 artifact.	sub-skills/image-modeling/SKILL.md

Shared references

• references/model-artifacts.md explains required and generated model/data files across all sub-skills.
• references/label-vocabulary.md preserves the exact 80-class Chinese label order from texts.txt.
• references/troubleshooting.md covers cross-cutting install/import, artifact, credential, and safety failures.
• references/repo-provenance.md records the source snapshot and evidence paths for refresh decisions.
• references/repo-routing-metadata.json is structured metadata for managed repo-skill routing.

Shared script

• python scripts/check_environment.py checks that the active Python environment has the dependency imports and legacy Keras surface expected by easy12306. It does not read user model/data files.

Operating rules

1. Start with validation scripts before model loading or training. Model files and datasets are not bundled with this skill.
2. Treat network, credential, and long-training workflows as explicit user actions, not smoke tests. The 12306 download loop, Baidu OCR calls, VGG16 weight downloads, and 100/400-epoch training runs should not run automatically.
3. Preserve OpenCV BGR assumptions, the 80-row label order, the 19x57 text crop, and eight 67x67 tile geometry unless the user explicitly asks to adapt the repo to a new captcha layout.
4. Do not import credentialed OCR code during diagnostics; use the data-preparation reference for a safe credential pattern.
5. Keep generated skill use self-contained. When source filenames are mentioned, treat them as provenance and artifact-name evidence; use the bundled scripts/references for actionable operations.