Game Localization MVR (Minimum Viable Rules) v2.1

A robust, automated workflow system for game localization with strict validation, AI translation/repair, glossary management, and multi-format export.

Core Principle: Input rows == Output rows ALWAYS. No silent data loss.

---

🤖 For AI Coding Agents

Quick Commands for Agents:

# 1. Verify LLM connectivity (MUST run first)
python scripts/llm_ping.py

# 2. Validate workflow configuration (dry-run)
python scripts/translate_llm.py --input input.csv --output output.csv --style workflow/style_guide.md --glossary glossary/compiled.yaml --style-profile workflow/style_profile.generated.yaml --dry-run

# 3. Run E2E test
python scripts/test_e2e_workflow.py

Environment Variables (REQUIRED):

LLM_BASE_URL=https://api.example.com/v1
LLM_API_KEY=sk-your-key
LLM_MODEL=gpt-4.1-mini
LLM_TRACE_PATH=data/llm_trace.jsonl

Key Rules for Agents:

1. Never hardcode API keys - Use environment variables only
2. Run llm_ping.py first - Fail-fast if LLM unavailable
3. Check WORKSPACE_RULES.md - See docs/WORKSPACE_RULES.md for hard constraints
4. Row preservation is P0 - Empty source rows must be preserved with status=skipped_empty
5. Glossary is mandatory - glossary/compiled.yaml must exist before translation

---

🚀 Pipeline Overview

Input CSV → Normalize → Translate → QA_Hard → Repair → Export
                ↓
            Glossary (required)

Step	Script	Purpose	Blocking?
0	llm_ping.py	🔌 LLM connectivity check	YES
1	normalize_guard.py	🧊 Freeze placeholders → tokens	YES
2-4	extract_terms.py → glossary_compile.py	📖 Build glossary	YES
5	translate_llm.py	🤖 AI Translation	YES
6	qa_hard.py	🛡️ Validate tokens/patterns	YES
7	repair_loop.py	🔧 Auto-repair hard errors	-
8	soft_qa_llm.py	🧠 Quality review	-
10	rehydrate_export.py	💧 Restore tokens → placeholders	YES

---

📁 Project Structure

loc-mvr/
├── config/
│   ├── llm_routing.yaml    # Model routing per step
│   └── pricing.yaml        # Cost calculation
├── glossary/
│   ├── compiled.yaml       # Active glossary (generated)
│   └── generic_terms_zh.txt # Blacklist for extraction
├── scripts/
│   ├── llm_ping.py         # ★ Run first - connectivity check
│   ├── normalize_guard.py  # Step 1: Placeholder freezing
│   ├── translate_llm.py    # Step 5: Translation
│   ├── qa_hard.py          # Step 6: Hard validation
│   ├── repair_loop.py      # Step 7: Auto-repair
│   └── runtime_adapter.py  # LLM client with routing
├── workflow/
│   ├── style_guide.md      # Translation style rules
│   ├── forbidden_patterns.txt
│   └── placeholder_schema.yaml
└── docs/
    └── WORKSPACE_RULES.md  # ★ Hard constraints for agents

---

🔧 Quick Start (Human)

1. Setup

git clone https://github.com/Charpup/game-localization-mvr.git
cd game-localization-mvr
pip install pyyaml requests numpy pandas jieba

2. Configure LLM (推荐持久化)

# Windows PowerShell
$env:LLM_BASE_URL="https://api.apiyi.com/v1"
$env:LLM_API_KEY="sk-your-key"
$env:LLM_MODEL="gpt-4.1-mini"

也可在本地持久化文件中配置（优先于环境变量自动读取）：

# 在 main_worktree/.llm_credentials 创建
LLM_BASE_URL=https://api.apiyi.com/v1
LLM_API_KEY=sk-your-key

当前加载顺序：LLM_API_KEY_FILE -> ./.llm_credentials/./.llm_env/./config/llm_credentials.env/~/.game-localization-mvr/.llm_credentials -> LLM_API_KEY

4. Dependency + Environment Quick Check (before every smoke run)

python - <<'PY'
import os
import importlib

for pkg in ["requests", "numpy", "yaml", "pandas"]:
    try:
        importlib.import_module(pkg)
        print(f"[OK] {pkg}")
    except Exception:
        print(f"[MISSING] {pkg}")

for key in ["LLM_BASE_URL", "LLM_API_KEY", "LLM_MODEL"]:
    print(f"{key}={'SET' if os.getenv(key) else 'MISSING'}")
PY

If any dependency shows MISSING or env variable shows MISSING, do not start smoke run yet.

PowerShell 快速检查：

$missing = @()
foreach ($m in @("requests","numpy","yaml","pandas","jieba")) {
  try {
    python -c "import importlib.util; print(bool(importlib.util.find_spec('$m')))"
    Write-Host "[OK] $m"
  } catch {
    $missing += $m
    Write-Host "[MISSING] $m"
  }
}
Write-Host "LLM_BASE_URL=$([bool]$env:LLM_BASE_URL)"
Write-Host "LLM_API_KEY=$([bool]$env:LLM_API_KEY)"
Write-Host "LLM_MODEL=$([bool]$env:LLM_MODEL)"

3. Run Pipeline

# Bootstrap tracked style assets once per clean worktree
python scripts/style_guide_bootstrap.py --dry-run

# Verify LLM
python scripts/llm_ping.py

# Normalize → Translate → QA → Export
python scripts/normalize_guard.py input.csv normalized.csv map.json workflow/placeholder_schema.yaml
python scripts/translate_llm.py --input normalized.csv --output translated.csv --style workflow/style_guide.md --glossary glossary/compiled.yaml --style-profile workflow/style_profile.generated.yaml
python scripts/qa_hard.py translated.csv qa_report.json map.json
python scripts/rehydrate_export.py translated.csv map.json final.csv

3.1 Smoke Pipeline (Manifest + Issue Record)

# Full smoke pass with manifest output + issue recording
python scripts/run_smoke_pipeline.py --input "D:\\Dev_Env\\loc-mvr 测试文档\\test_input_200-row.csv" --target-lang en-US
# 可选：仅做预检
python scripts/run_smoke_pipeline.py --input "D:\\Dev_Env\\loc-mvr 测试文档\\test_input_200-row.csv" --target-lang en-US --verify-mode preflight

This command:

• auto-bootstraps workflow/style_profile.generated.yaml if the clean worktree does not have one yet
• runs llm_ping -> normalize_guard -> translate_llm -> qa_hard -> rehydrate_export
• generates a run manifest: data/smoke_run_<timestamp>/run_manifest.json
• runs smoke_verify --manifest ...
• records issues to reports/smoke_issues_<run-id>.json and .jsonl
• emits manifest.stage_artifacts with:
  • connectivity_log
  • normalize_log
  • translate_log
  • qa_hard_report
  • final_csv
  • smoke_verify_log
• verify_mode supports preflight|full，默认 full（含行数/QA 统计）

建议每次冒烟固定检查以下产物：

• D:\Dev_Env\GPT_Codex_Workspace\game-localization-mvr\main_worktree\data\smoke_runs\<run>\run_manifest.json
• D:\Dev_Env\GPT_Codex_Workspace\game-localization-mvr\main_worktree\reports\smoke_issues_<run_id>.json
• D:\Dev_Env\GPT_Codex_Workspace\game-localization-mvr\main_worktree\reports\smoke_verify_<run_id>.json

---

⚡ Key Features

• Row Preservation: Empty rows kept with status=skipped_empty
• Drift Guard: Refresh stage blocks non-placeholder text changes
• Progress Reporting: --progress_every N for translation progress
• Router-based Models: Configure per-step models in llm_routing.yaml
• LLM Tracing: All calls logged to LLM_TRACE_PATH for billing

---

📋 Testing

# Unit tests
python scripts/test_normalize.py
python scripts/test_qa_hard.py
python scripts/test_rehydrate.py

# E2E test (small dataset)
python scripts/test_e2e_workflow.py

# Dry-run validation
python scripts/translate_llm.py --input input.csv --output out.csv --style workflow/style_guide.md --glossary glossary/compiled.yaml --style-profile workflow/style_profile.generated.yaml --dry-run

---

📄 License

MIT License. Built for game localization automation.

---

🔗 Links

• Workspace Rules: docs/WORKSPACE_RULES.md
• Demo Walkthrough: docs/demo.md