io.github.xixi048/create-all-cg-save

create-all-cg-save

Create an all-CG save — unlock every CG, gallery, and recollection room in your own offline single-player saves.

An agent skill (plus a small zero-dependency Python toolkit) that turns "I want all CGs unlocked" into a precise, reversible byte patch instead of a downloaded stranger's save file.

中文说明见 README.zh-CN.md.

What it does

• Scans the usual save locations on this machine and finds candidate save files for a game
• Identifies the engine (RPG Maker, Ren'Py, Kirikiri, Unity, Siglus, BGI, …)
• For parseable formats (e.g. RPG Maker's Ruby Marshal) it computes the exact byte offset of every unlock flag
• For opaque private binaries it reconstructs the unlock flags by diffing a "before" and an "after" save, fits the start + stride array, and extrapolates to the whole gallery
• Emits a human-readable unlock plan with dry-run, verification, automatic backup, and one-command rollback

It never downloads the game, never touches CG images or game assets, and never replaces your save with someone else's — it only makes a minimal, surgical edit to a save you already own.

Two routes

Decision tree lives in skills/create-all-cg-save/SKILL.md.

Quick start

# 1. Find the save file
python skills/create-all-cg-save/scripts/save_scout.py --game <game name>

# 2. RPG Maker VX / VX Ace / XP: inspect structure and current switch state
python skills/create-all-cg-save/scripts/engines/rgss_save.py inspect "Save01.rvdata2"

# 3. Build an unlock plan (flip the given switches ON)
python skills/create-all-cg-save/scripts/engines/rgss_save.py plan-switches \
    "Save01.rvdata2" --ids 1,3,5-40 --out plan.json

# 4. Dry-run → apply → verify (apply backs up automatically)
python skills/create-all-cg-save/scripts/apply_plan.py plan.json --dry-run
python skills/create-all-cg-save/scripts/apply_plan.py plan.json
python skills/create-all-cg-save/scripts/apply_plan.py plan.json verify

# Roll back if anything looks wrong
python skills/create-all-cg-save/scripts/apply_plan.py plan.json rollback

Unknown engine? Go through the diff path:

python skills/create-all-cg-save/scripts/save_diff.py A.save B.save --json ab.diff.json
python skills/create-all-cg-save/scripts/infer_plan.py ab.diff.json \
    --target A.save --stride 4 --count 128 --out unlock-plan.json

--stride / --start matter. With only one or two observed flag positions the fitted stride is ambiguous (a stride-4 array where slots 0 and 5 were unlocked makes GCD report 20). The script warns loudly; pin the layout explicitly when you can.

Verified target

RPG Maker VX Ace (RGSS3) — e.g. Save0X.rvdata2:

• The save is Ruby Marshal with two concatenated Marshal streams (stream 1 is make_save_header for the load menu, stream 2 is the 11-key make_save_contents hash)
• :switches → @data is a boolean array; 'T' / 'F' / '0' (nil) are one byte each, so the patch is an equal-length replacement — file size not changed by a single byte
• Result: 59 switch slots flipped ON, size 617,987 → 617,987 bytes, re-parsed as genuinely true, and rollback restored the file byte-for-byte

Repository layout

create-all-cg-save/                  # repository
├── .claude-plugin/
│   └── marketplace.json            plugin-marketplace manifest (`/plugin marketplace add`)
└── skills/
    └── create-all-cg-save/         the skill itself (auto-indexed by skills.sh & friends)
        ├── SKILL.md                skill entry point (the workflow an agent reads)
        ├── scripts/
        │   ├── save_scout.py       locate candidate saves + engine fingerprinting
        │   ├── save_diff.py        byte-level diff of two saves
        │   ├── infer_plan.py       infer the flag array → unlock plan
        │   ├── apply_plan.py       apply / verify / roll back
        │   ├── common.py           backup, atomic write, hashing
        │   └── engines/
        │       ├── ruby_marshal.py dependency-free Ruby Marshal parser + encoders
        │       └── rgss_save.py    RPG Maker RGSS save adapter
        └── references/
            ├── engines.md          per-engine save locations and formats
            ├── diff-method.md      the full diff-method manual
            ├── unlock-plan-schema.md  diff / plan JSON schemas
            └── safety.md           backup, integrity, and compliance red lines

Install as a skill

Straight from the public repo (no clone needed):

npx skills add xixi048/create-all-cg-save           # skills.sh / `npx skills` CLI

/plugin marketplace add xixi048/create-all-cg-save  # Claude Code plugin marketplace

Or copy the folder into your agent's skills directory:

Drop the whole skills/create-all-cg-save/ directory into your agent's skills directory:

• User level: ~/.workbuddy/skills/create-all-cg-save/
• Project level: <repo>/.workbuddy/skills/create-all-cg-save/

The directory name must match the name field in SKILL.md frontmatter (create-all-cg-save). The frontmatter follows the standard agent-skill shape: name, description (what + when, with Chinese and English trigger keywords), license, and metadata.

Why the extra skills/ level? Directory sites that crawl GitHub (skills.sh, SkillsMP, npx skills, and Claude Code's plugin marketplaces) only look for skills/<name>/SKILL.md. A skill folder sitting at the repo root is never discovered, no matter how good it is.

Note: all relative paths inside SKILL.md are relative to the skill directory, not your working directory. cd into it first, or prefix the commands with the absolute skill path.

Use as an MCP server

The whole toolkit is also exposed over MCP (stdio, zero dependencies), so any MCP-capable agent can drive it without knowing the repo layout:

{
  "mcpServers": {
    "create-all-cg-save": {
      "command": "python",
      "args": ["C:/path/to/create-all-cg-save/mcp/server.py"]
    }
  }
}

Tools: save_scout · engine_inspect · plan_unlock_switches · diff_saves · infer_unlock_plan · apply_plan (apply_plan takes action: dry_run | apply | verify | rollback; apply always backs up first).

Machine-discoverable index:

https://xixi048.github.io/create-all-cg-save/.well-known/agent-skills/index.json

Requirements

Python 3.8+, zero third-party dependencies (standard library only).

On a Windows console with a legacy code page you may need chcp 65001 (or PYTHONIOENCODING=utf-8) for the Chinese progress output to render correctly.

Safety model

• Every write is backed up first as <save>.allcg-bak-<timestamp> — never an in-place overwrite
• sha256_before is verified before patching; a plan built for a different save is refused
• Patches that would change the file length are refused unless --allow-resize is passed
• Always dry-run before applying, and always load the game afterwards to confirm

See references/safety.md for the full list.

Known limits

• RPG Maker XP / VX / VX Ace has full automation; MV / MZ is easy when the save is plain JSON
• Ren'Py, Kirikiri, Unity, Siglus, BGI etc. usually need the diff method, and some saves are compressed/encrypted end-to-end, where plain diffing simply fails
• Cloud-only saves must be synced locally first
• Unlocking every switch can also trip story branches or game-over conditions — start with the IDs you have verified inside the game

Disclaimer

This tool is for personal research and modification of offline single-player games you legally own. Do not use it on online / multiplayer games, leaderboards, or anything where it could affect other people. You are responsible for whatever happens to your save — every write is backed up, though.

License

MIT — see LICENSE.