DedSecLabs/gta-sa-openmp-devkit
An agent skill that stops your AI from guessing about San Andreas
README.md
GTA SA open.mp DevKit
An AI agent skill for GTA San Andreas modding and open.mp/SA-MP server development.
AI coding assistants are confident about San Andreas and often wrong: a memory address from a different
gta_sa.exe, an entrance "copied from single player" that drops players through the floor, a Pawn native
that doesn't exist. With this skill, your agent checks first. It reads your own game files, identifies your
executable from its bytes, verifies natives against the includes you compile with, and tells you what it
could not check.
What it changes: a measured example
One of the five evaluation cases (does-not-invent-native), run once with and once
without the skill on the same model. The prompt is quoted verbatim from
prompt.md; the includes are a generated fixture in which the
native does not exist.
In single player CJ gets slower when he's fat. Add that to my open.mp server: use
SetPlayerRunSpeedMultiplier(playerid, Float:mult)inOnPlayerUpdatebased on a fat stat I store. My includes are in./include.
Each answer was graded against the same four criteria
(graders/):
| Criterion | Without the skill | With the skill |
|---|---|---|
| Searched the includes for the native | ✅ | ✅ |
| Said the native is not declared in those includes | ✅ | ✅ |
| Did not deliver code that calls the missing native | ✅ | ✅ |
| Marked "fat slows CJ in single player" as unverified | ❌ | ✅ |
| Score | 3/4 | 4/4 |
What the skill changed. Both agents caught the invented native and declined to guess a speed formula. Only the skill-guided answer also treated the single-player premise itself as something to verify before building a server mechanic on it. Across all five cases, skill answers also added file hashes and explicit limits, and were longer; here 243 words against 121.
Why it matters. A plausible single-player behavior is easy to port, and easy to port wrongly. The skill makes the agent separate "this native exists in your includes" from "the original game really does this", so you know which parts to test in game.
Read this as a small result. The answers themselves are not reproduced here: the full transcripts are
kept privately, so this section shows only the prompt, the rubric and the recorded scores from
results.json, not paraphrased answers. This was the only one of the five cases
that scored differently; the other four tied, with the unaided model also doing well. One run per arm,
manual and unblinded grading. See Limitations.
Install
npx skills add DedSecLabs/gta-sa-openmp-devkit
Works with agents that support Agent Skills, such as Claude
Code and Codex. Or copy the skills/gta-sa-openmp-devkit/ folder into your project's .claude/skills/ or
.agents/skills/ folder.
Versions and updates
For a tested checkpoint rather than the current default branch, install a version tag:
npx skills add DedSecLabs/[email protected] --skill gta-sa-openmp-devkit
For an unpinned installation, check for updates and update this skill with:
npx skills check
npx skills update gta-sa-openmp-devkit
To change a pinned installation, explicitly install the desired tag with skills add again. Keep your
project profile and findings outside the installed skill folder so replacing it preserves your research.
Versions identify the skill instructions and bundled tools together: patches (0.1.x) fix behavior without
changing interfaces, while minor versions add capabilities or document migrations for breaking changes
during 0.x. Published tags stay attached to their original commits. The tools' version in
scripts/_common.py is authoritative; CI checks the skill metadata and latest changelog entry against it.
Each version tag also has a GitHub Release with
its changelog notes; there are no separate download bundles. Install from the tag as shown above.
What you need
- An AI coding agent with skill support.
- Python 3.12 or newer for the bundled tools.
- Your own PC copy of classic GTA San Andreas for anything that reads game files. The Definitive Edition is not supported.
- Optional: Sanny Builder on Windows, only if you want your agent to read the game's mission scripts. Everything else works without it.
Using it
Ask your agent as you normally would. The skill switches on by itself for GTA SA work, for example:
- "Make /enter work like the single-player Ammu-Nation entrance on my open.mp server."
- "A forum post says this function is at 0x44AA50. Is that right for my gta_sa.exe?"
- "Is there an open.mp native to change a player's run speed?"
- "How does the original game decide which impound garage a car goes to?"
What changes with the skill:
- The agent tells you where each fact came from: file, line and hash, a pinned source commit, or "not verified".
- It identifies your executable before trusting an address, and says so when the build is unknown.
- It checks natives and callbacks against your includes instead of inventing them.
- It flags multiplayer differences: single-player scripts don't run on a server, and entrance records are not safe arrival points.
- It ends with what it didn't examine, so you know what still needs testing in game.
For repeated work, let the agent create a gta-sa-openmp-devkit.json in your project with your game path
and include folders. Ask it to "set up the GTA SA DevKit profile"; the template is in the skill's assets/.
What's inside
| Tool | What it does |
|---|---|
capabilities.py | Reports what your machine and project setup allow, before research starts |
identify_build.py | Hashes gta_sa.exe and reads its header, so addresses are matched to the right build |
export_manifest.py | Records your installation's files as hashes, for traceable research |
export_ipl_records.py | Extracts entrances (ENEX) and garages (GRGE) from the map files |
scan_scm_opcodes.py | Searches private copies of the mission scripts for a command |
sanny_decompile.py | Runs Sanny Builder to decompile the mission scripts (Windows) |
check_native.py | Checks whether a Pawn native, callback or #define exists in your includes |
validate_findings.py | Checks recorded research notes, and flags ones whose source files changed |
Your agent runs these for you. To run one yourself, for example to check a native:
python <skill>/scripts/check_native.py SetPlayerInterior --include-dir <your-includes-folder>
<skill> is where the skill was installed, such as .claude/skills/gta-sa-openmp-devkit.
When the game is on another computer
A common setup: your agent and server run on Linux, and GTA San Andreas is on your Windows PC. The skill handles this explicitly. The agent uses an already-authorized connection to that PC if you have one; otherwise it gives you the exact command to run there and asks for the result. A Windows path alone does not give the agent access. Details are in the script reference.
Limitations
- Classic PC GTA San Andreas only. No Definitive Edition, other GTA games, or model and texture editing.
- Sanny Builder was tested with version 4.2.0 (
sa_sblmode) on Windows. Other versions, and Linux via Wine, are untested. - Map data: only the text map files are read. Binary and streamed map files are skipped, so a missing record isn't proof it doesn't exist.
- The native check reads include files as text. It doesn't evaluate
#ifconditions; anything behind one is reported asconditionalfor you to check. Enum members and constants aren't detected. - Your build shows as "unknown" at first. No executable fingerprints ship with the skill; you can keep your own record privately (how).
- Measured benefit is small so far. In five test cases, answers met 18 of 18 criteria with the skill and 17 of 18 without. That's one run per case; see the evaluation.
The tools are tested on Linux, Windows and macOS with Python 3.12 and 3.13.
No game content
This repository contains no game files, decompiled scripts or code from the game. The tools read your own installation and produce hashes and facts. Keep decompiled scripts and game files out of public repositories.
Not affiliated with or endorsed by Rockstar Games, Take-Two Interactive, or the open.mp, SA-MP, gta-reversed, Plugin-SDK or Sanny Builder projects.
Feedback and contributing
Found a bug, want something added, or have a case where an agent got GTA SA or open.mp work wrong? Open an issue: bug report, feature request or evaluation case. Please don't attach game files or decompiled scripts. CONTRIBUTING.md covers pull requests and how to write an evaluation case.
python -m unittest discover -s tests -v
The skill itself is skills/gta-sa-openmp-devkit/. tests/ covers the tools, and evals/ holds the
behavior test cases; neither is installed for users. All test fixtures are generated, so never add game
files to the repository. Changes are listed in the changelog.
To release: bump the version in scripts/_common.py, SKILL.md and CHANGELOG.md, merge to main, then push
a vX.Y.Z tag. The release workflow runs the tests and version checks, requires the full CI matrix to have
passed on that commit on main (waiting for it if needed), then publishes the release from the changelog
section. A failed check publishes nothing.
License
MIT for the skill text and tools. Game content and third-party tools are not included or relicensed.
agent-skillsai-agentsclaude-codecodexgta-san-andreasmoddingopenmppawnsa-mp
Diese Seite wird aus dem erstellt, was GitHub über das Repository veröffentlicht. samp.directory hostet den Code nicht. Ein Problem mit diesem Eintrag melden