Case study
Super Mario Sunshine (USA): an AI-assisted matching decompilation fork
Extends the doldecomp Super Mario Sunshine decompilation to the North American GMSE01 release. The README says AI was used extensively, and the repository’s agent rules, work order and progress files describe batch-based Claude and Codex orchestration. Its rebuilt DOL is reported byte-identical, but unfinished objects are still linked from code extracted from the user’s disc.
- Targets
- Game
- Category
- Independent reconstruction cases
- Relationship
- Independent reconstruction project
- Sources reviewed
Requirements & limits
- Prerequisites: Supply your own North American Rev 0 disc image in orig/GMSE01 (ISO, RVZ, WIA, WBFS, CISO, NFS, GCZ or TGC). The GMSE01 build notes record the required extracted-DOL SHA-1. Those notes say the initial symbols came from a community-hosted marioUS.map that is not on the US disc, while AGENTS.md calls the US map a real map shipped on the disc.
- Prerequisites: Builds use Python 3 and Ninja on Windows, macOS (with Wine Crossover) or Linux (wibo on x86, Wine elsewhere), plus the pinned decomp-toolkit and Metrowerks tools. The output is build/GMSE01/mario.dol, checked against the extracted original. Playing it still needs the original disc image in Dolphin; a separate native PC port uses this fork as its decompilation submodule.
- Scope: This hobby fork of doldecomp/sms targets GMSE01 (USA Rev 0). The README says it began when upstream’s Japanese decompilation was about 38% matched, and PROGRESS.md names upstream ab00c3c9 as its base. GMSJ01 and GMSP01 configs remain, but the README marks PAL as slightly broken and AGENTS.md says neither can be built or regression-tested in this clone.
- Testing evidence: The README (2026-09-29) reports 99.60% fuzzy, 73.65% byte-perfect and 34.79% source-linked code, with 12,052 of 12,904 functions exact and 542 of 732 objects linked; PROGRESS.md (2026-09-27) gives slightly lower figures. The fork’s CI matrix builds only GMSJ01, so GMSE01 results are locally reported.
- Attribution: AGENTS.md, symlinked as CLAUDE.md and GEMINI.md, lifts upstream’s ban on autonomous library work. It requires agents to work only from the binary, the linker map and the repository, and forbids fakematches and leaked SDK sources. The repository keeps CC0-1.0, while Nintendo content and the compilers keep separate rights.
About this entry
The description was reviewed against primary sources.
Primary sources
- README.md GitHub · chasem-dev/sms-english · Revision d5f4eb3
- PROGRESS.md GitHub · chasem-dev/sms-english · Revision d5f4eb3
- config/GMSE01/README.md GitHub · chasem-dev/sms-english · Revision d5f4eb3
- AGENTS.md GitHub · chasem-dev/sms-english · Revision d5f4eb3
- STRATEGY.md GitHub · chasem-dev/sms-english · Revision d5f4eb3
- DECOMPILATION_PLAN.md GitHub · chasem-dev/sms-english · Revision d5f4eb3
- .github/workflows/build.yml GitHub · chasem-dev/sms-english · Revision d5f4eb3
- LICENSE GitHub · chasem-dev/sms-english · Revision d5f4eb3
More in Games
- Super Mario Sunshine: an in-progress matching GameCube decompilation · Case study
- Super Smash Bros. Brawl: an early matching Wii decompilation · Case study
- Super Smash Bros. Melee: a matching GameCube decompilation with a shiftable build · Case study
- Syphon Filter: The Omega Strain — matching PS2 executable decompilation · Case study
- TH04: rebuilding Lotus Land Story’s four PC-98 programs from source · Case study
- TH08: exact-source reconstruction and separately tested ports · Case study