The purpose of this publish is to doc how coding brokers have and haven’t helped with the decompilation course of. While a lot has been written about LLMs, far less has been written about decompilation,1 so I’m including another information point. A number of lessons may apply past the N64. If you’ve got suggestions to enhance the workflow, please let me know! What's matching decompilation? Snowboard Youngsters 2 was written in C and compiled to MIPS machine code. The matching decompilation process involves analysing the MIPS assembly, inferring its behaviour, and writing C that, when compiled with the identical toolchain and settings, reproduces the precise code: same registers, delay slots, and instruction order.
In fact, this still doesn’t tell us what the operate is for within the broader codebase. 58164 to a minimal of 4? Matching the bytes, nevertheless, is just the first step. A true match requires understanding the function’s intent through naming conventions, cross-references, and analysis of how the decision websites behave. Matches also vary in quality. A very good match is extra than just C code that compiles to the right bytes.4 It should appear to be one thing an N64-period developer would plausibly have written: simple, idiomatic C control stream and wise data buildings.5 Prioritising plausibility pays off: later functions naturally share helpers and struct layouts, and also you avoid artefacts that only exist to set off some quirk of the compiler.
Why do a matching decompilation in any respect? It deepens understanding of the engine, opens the door to new content, gives a solid base for tooling and, most significantly, it’s fun! 2. Strive to provide matching code. 3. If I get caught, import the perform regionally and herald agents. 4. If the matching percentage reaches 100%, the function is taken into account decompiled and work can stop. For convenience, it additionally begins a brand new Claude occasion with instructions to learn the aforementioned file and ‘use thinking’. You will also get a rating, with a score of 100% indicating a perfect match.
2. Search for an space the place the management circulation and directions do not match. Consider what the original developers most likely meant to write given the function’s broader purpose. Print out a proof for why they don’t match. 5. If your potential resolution did not enhance the match share, use tools to analyse what went incorrect and summarise your concept. Then apply this idea to bettering the match in your subsequent attempt. The aim of this workflow is to encourage the agent to determine and check small, incremental enhancements. play Pragmatic slots at BOLAKING88 would in any other case try to fix the whole lot at once, solely to be confounded by the drop in match price, which had too many attainable causes to diagnose successfully.
Equally, every try is made in a brand new file in order that we will simply get well good matches. We wish to avoid conditions the place an agent achieves a superb match however subsequently edits the file and degrades its high quality. The path of attempts can also be a useful reference for the agent. The requests in Step 2 and Step 5 for an explanation are meant to encourage the agent to (respectively) anticipate the effect of its changes and replicate on their outcomes.
Disassembly: dump a binary to MIPS assembly. Object file diff: disassemble and diff two object information with register-title normalisation. Line mapping: greatest-effort mapping from disassembly strains to C traces (by way of the candidate’s debug symbols). Most of these utilities have been initially scaffolded by Claude and then customised for this challenge. They’re all obtainable on Github. Agents excel at bootstrapping one-off CLI instruments: setup is quick, and ongoing upkeep is minimal as a result of the agent can regenerate or patch them as needed.
My workflow remains to be human-driven, with agents acting as analysis assistants. I pick features and attempt to decompile them first. A part of that's emotional. It’s extra satisfying to figure things out yourself than to stare at a coding agent. Even in the event you wanted to, relying solely on an agent won’t get you far. We’re a good distance from ‘vibe coding’ being viable for matching decompilation. Brokers are nice at spotting patterns and typically make surprising leaps. Where I see 2,000 lines of noise, Claude would possibly spot an audio processing routine. They’re additionally patient enough to attempt dozens of variations to see what nudges the match. Agents strategy issues in distinct methods and often produce totally different answers.

I incessantly run Claude and Codex in parallel to widen the search space and diversify ideas. Brokers typically stumble on fundamental arithmetic and bookkeeping that a human would do shortly. For example, they struggle to compute the correct byte offset of a struct discipline. In addition they repeat small syntactic errors despite prompting. Claude is instructed to make use of C89 (i.e., the 1989 commonplace of C), but it usually forgets this, which results in trivial syntax errors.