Wolf AgentLab

Wolf AgentLab is a local Windows workflow for translating Wolf RPG Editor games with an Agent-driven pipeline.

The project is built around att-wolf, a CLI adapted from the A.T.T tooling line. It does not depend on existing .trans files. The normal flow is:

1. Register a Wolf game.
2. Prepare and unpack Data with UberWolf/WolfDec fallback.
3. Dump text through WolfTL.
4. Build a stable text index.
5. Translate with LLM batches.
6. Audit and restore runtime-sensitive strings.
7. Patch back through WolfTL.
8. Generate and finalize a clean patch package.

For old Wolf games, insert the Pro conversion gate before write-back: unpack the original Data.wolf, remove every .wolf from the game directory, run EditorPro.exe until Backup_Before_Ver3\ConvertLog.txt records successful conversion, then re-dump the converted Data and patch that converted dump. Do not patch a pre-conversion dump into a converted Data directory.

Repository Layout

.
├── wolf-source/              # Python/Rust source for att-wolf
├── scripts/                  # repository-level packaging helpers
├── game/                     # local input games, ignored by git
├── workspace/                # generated workspace, ignored by git
├── patches/                  # generated patch output, ignored by git
├── tools/                    # local WolfTL/UberWolf/WolfDec binaries, ignored by git
├── att-wolf-auto.bat         # Windows entry point
├── att-wolf-wizard.ps1       # safe workflow wizard
├── WOLF_FIXED_WORKFLOW.md    # fixed production workflow
└── PATCH_FINALIZATION.md     # release package finalization rules

Requirements

• Windows
• Python 3.14
• uv
• Rust stable and MSVC build tools, for the native extension
• WolfTL.exe
• UberWolf.exe, optional but preferred
• WolfDec.exe, fallback unpacker
• An OpenAI-compatible model endpoint

Tool paths can be configured in wolf-source\setting.toml, through environment variables, or by placing binaries in G:\wolf\tools.

Environment variables:

• ATT_WOLF_WOLFTL_PATH
• ATT_WOLF_UBERWOLF_PATH
• ATT_WOLF_WOLFDEC_PATH

Quick Start

cd G:\wolf\wolf-source
uv sync --locked --dev
uv run maturin develop --release
uv run att-wolf --agent-mode doctor --no-check-llm --json

Prepare a game without translating or writing back:

G:\wolf\att-wolf-auto.bat -GamePath "G:\wolf\game\<game-dir>" -PrepareOnly -SkipLlmCheck

Run a one-batch smoke translation:

G:\wolf\att-wolf-auto.bat -GameTitle "<game-title>" -SmallBatchMaxBatches 1 -SkipLlmCheck

Run full translation without write-back:

G:\wolf\att-wolf-auto.bat -GameTitle "<game-title>" -RunFullTranslation -SkipLlmCheck

Write back only after all quality gates pass:

G:\wolf\att-wolf-auto.bat -GameTitle "<game-title>" -RunFullTranslation -WriteBack -SkipLlmCheck

Safety Gates

The wizard intentionally does not write back by default.

Before write-back, the workflow runs or expects:

• translation-status with no pending text.
• restore-runtime.
• quality-report.
• audit-runtime.
• WolfTL re-dump after patching.
• Real game-window QA before release.

Runtime-sensitive strings are protected by default, including labels, common-event names, resource paths, map references, database internal IDs, common-event parameter keywords, and control tokens such as \s[9], \cself[...], \self[...], \v[...], \i[...], and <管理番号...>.

Wolf database structure names are not translated by default. Keep types/*/fields/*, types/*/name, and types/*/description intact unless a game-specific whitelist has been tested in the real game window.

Development Checks

cd G:\wolf\wolf-source
uv run python -m pytest tests/test_wolf_pipeline.py tests/test_cli_json_output.py::test_parser_commands_have_dispatch_handlers -q
uv run python -m compileall app

Release Package

After make-patch, finalize the patch directory:

G:\wolf\finalize-patch-package.bat "G:\wolf\patches\<patch-dir>"

Default password: sstm

See WOLF_FIXED_WORKFLOW.md and PATCH_FINALIZATION.md for the fixed production process.