Inspect the original
Pin the build. Read the formats. Run the program. When only a binary exists, use REA to recover evidence.
survey → reference.lock AN AGENT SKILL · RUST / WGPU
Same game.
New engine.
Show the proof.
Port games and apps to Rust with an agent that starts from the original. Real content. Small playable slices. Measured compatibility.
OPEN SOURCE / APACHE-2.0 / SIX AGENT HOSTS / EVIDENCE WORKBENCH
COMPATIBILITY WORKBENCH● LOCAL
01 / ORIGINALREFERENCE BUILD
STATE + INPUT + CONTENT↓COMPARE WHAT MATTERS
Rust runtimeSame scenario. Inspect the differences.
⌖ILLUSTRATION, NOT A TEST RESULTrp / 001
ONE SKILL.
SIX HOSTS.
SIX HOSTS.
Codex
Claude Code
OpenCode
Pi
DSH
Hermes
01 / THE PROCESSA port is more than
A port is more than
a translation.
Keep the original as the reference. Rebuild one observable piece at a time.
Make a real slice
Load an original map, execute a vehicle script, or round-trip a save. Connect the actual runtime early.
content → core → frameMeasure the match
Feed the same inputs to both sides. Compare state, files and frames. Keep unknowns in the ledger.
original ↔ Rust → diffPort available source
Capture original outputs, port a headless core, then bring up the platform and renderer.
Rebuild around existing content
Original maps, formats, scripts and saves. A workflow for projects in the style of Horizon and openOMSI.
Continue a port
Resume the compatibility ledger. Reproduce the next gap and fix the path the player actually runs.
02 / PUT IT TO WORKYour agent.
Your agent.
One shared skill.
Choose a host for its install command and invocation. The installer keeps scripts, templates and instructions together.
Install from main
One shared skill, native invocation for your host. Use --dry-run to inspect the installation before writing files.
TERMINAL / PYTHON 3.9+
THEN, IN YOUR AGENT
02 / THE WORKBENCHProve it with
Prove it with
repeatable evidence.
An opt-in toolkit for version-locked original builds, function mapping and first-divergence trace comparison.
Exact original build
SHA-256 identity without launching the original.
SQLite ledger
Research status, Rust location, runtime wiring and stale evidence checks.
Independent scenarios
Explicit opt-in dual execution and strict JSONL trace differences.
Measured trace parity is not complete-game parity. The dual-runner is not a security sandbox.
NATIVE LAB / DEEP RESEARCHDecode the original.
Decode the original.
Bound the evidence.
Headless Ghidra recovery, optional CPU emulation and seeded original-vs-Rust differential probes.
Ghidra exports
Function addresses, call graph and reproducible executable hashes.
Unicorn probes
Bounded x86-64 slices with no host runtime or OS loader.
Find the first mismatch
Seeded JSON harnesses and smaller replayable failures.
Opt-in tools. Optional dependencies. Static analysis is not gameplay observation.
03 / FIELD NOTESTested. With the
Tested. With the
limits left in.
These are actual integration experiments. Passing one small case never proves a whole engine.
EXPERIMENTWHAT HAPPENEDVERDICT
Six-host installerTemporary homes and projects15 local tests: layouts, complete bundles, arguments, existing-file preservation.LOCAL PASS ↗C → Rust CLIActual Codex implementation13 original goldens, bidirectional save/load, and seeded differential traces.SCOPED PASS ↗Binary → Rust helperopenMMD · arm64 releaseVPD name rules recovered from disassembly. 3 static tests; original-side comparison unavailable.UNVERIFIED ↗Interactive end-to-end behavior has been tested in Codex. Installation layouts for all six hosts are covered; all six interactive clients and whole-engine reconstruction are not verified.
“Never claim parity
you haven't measured.”
Read the instructions. Inspect the examples.
Bring a program worth rebuilding.