---
trigger: always_on
---

# EPUB Bilingual Translator - File Architecture & Workspace Rules

## 1. Directory Structure
The project follows a strict modular structure. Code is in `src/`, execution scripts in `pipeline/`, and intermediate data in `.work/`.

.
├── main.py                     # Entry point (orchestrator)
├── pipeline/                   # Executable scripts for each stage
│   ├── 01_preprocess.py        # Step 1: Clean & Extract
│   ├── 02_translate.py         # Step 2: LLM Translation
│   └── 03_assemble.py          # Step 3: Backfill & Build
├── src/                        # Source modules
│   ├── common/                 # shared utils, config, paths.py
│   ├── preprocessing/          # epub_cleaner, text_extractor, profiler
│   ├── translation/            # llm_client, translator_engine, manifest_manager
│   └── assembly/               # backfiller, builder, format_restorer
├── config/
│   ├── config.yaml             # System configuration (LLM, Translation)
│   └── prompts.json            # LLM Prompts
└── .work/                      # Working Directory (Gitignored)
    └── {book_name}/            # One folder per book
        ├── book_structure.json # Structural skeleton (created by Step 1)
        ├── manifest.json       # Translation source of truth
        ├── assets/             # Extracted images/css
        └── chunks/             # Debug chunks from translation (before/after)

## 2. Path Resolution Rule
ALWAYS use `src.common.paths.get_work_dirs(input_path)` to resolve paths.
DO NOT hardcode `work/`, `.work/`, or `tmp/` paths in scripts.

## 3. Data Flow
1. Preprocess: EPUB -> .work/{book}/book_structure.json + .work/{book}/manifest.json
2. Translate: .work/{book}/manifest.json (read/write) -> .work/{book}/chunks/ (logs)
3. Assemble: .work/{book}/manifest.json + .work/{book}/book_structure.json -> output/{book}_bilingual.epub