The aim of this put up is to document how coding agents have and haven’t helped with the decompilation process. While a lot has been written about LLMs, far much less has been written about decompilation,1 so I’m adding one other data point. A couple of classes may apply beyond the N64. If you’ve received strategies to improve the workflow, please let me know! What's matching decompilation? Snowboard Kids 2 was written in C and compiled to MIPS machine code. The matching decompilation process includes analysing the MIPS assembly, inferring its behaviour, and writing C that, when compiled with the identical toolchain and settings, reproduces the exact code: same registers, delay slots, and instruction order.
Of course, this nonetheless doesn’t tell us what the function is for within the broader codebase. 58164 to a minimum of four? Matching the bytes, nonetheless, is just the first step. A true match requires understanding the function’s intent by way of naming conventions, cross-references, and evaluation of how the call sites behave. Matches also range in quality. An excellent match is extra than just C code that compiles to the appropriate bytes.Four It should appear like one thing an N64-era developer would plausibly have written: simple, idiomatic C management circulate and wise information structures.5 Prioritising plausibility pays off: later features naturally share helpers and struct layouts, and you avoid artefacts that only exist to trigger 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, supplies a solid base for tooling and, most significantly, it’s enjoyable! 2. Attempt to produce matching code. 3. If I get stuck, import the perform domestically and bring in agents. 4. If LIVEKING88 matching percentage reaches 100%, the perform is considered decompiled and work can cease. For comfort, it additionally starts a new Claude instance with directions to read the aforementioned file and ‘use thinking’. You will also get a rating, with a score of 100% indicating an ideal match.
2. Search for an area the place the control move and instructions don't match. Consider what the original builders in all probability meant to write down given the function’s broader function. Print out an explanation for why they don’t match. 5. In case your potential answer didn't enhance the match percentage, use instruments to analyse what went wrong and summarise your theory. Then apply this principle to bettering the match in your subsequent try. The aim of this workflow is to encourage the agent to determine and check small, incremental improvements. Claude would in any other case try to repair every thing at once, solely to be confounded by the drop in match charge, which had too many attainable causes to diagnose effectively.
Equally, every try is made in a new file so that we can easily recuperate good matches. We want to avoid situations the place an agent achieves a great match however subsequently edits the file and degrades its high quality. The trail of attempts is also a useful reference for the agent. The requests in Step 2 and Step 5 for a proof are intended to encourage the agent to (respectively) anticipate the effect of its adjustments and reflect on their outcomes.
Disassembly: dump a binary to MIPS meeting. Object file diff: disassemble and diff two object recordsdata with register-name normalisation. Line mapping: greatest-effort mapping from disassembly strains to C lines (via the candidate’s debug symbols). Most of those utilities had been initially scaffolded by Claude and then customised for this venture. They’re all out there on Github. Brokers excel at bootstrapping one-off CLI tools: setup is quick, and ongoing upkeep is minimal as a result of the agent can regenerate or patch them as wanted.
My workflow is still human-driven, with brokers acting as research assistants. I pick functions and try to decompile them first. A part of that is emotional. It’s more satisfying to figure issues out yourself than to stare at a coding agent. Even in case you wished to, relying solely on an agent won’t get you far. We’re a long way from ‘vibe coding’ being viable for matching decompilation. Agents are great at spotting patterns and generally make stunning leaps. Where I see 2,000 strains of noise, Claude would possibly spot an audio processing routine. They’re also patient sufficient to try dozens of variations to see what nudges the match. Brokers strategy issues in distinct ways and infrequently produce totally different answers.
I steadily run Claude and Codex in parallel to widen the search area and diversify ideas. Agents usually stumble on primary arithmetic and bookkeeping that a human would do quickly. For example, they wrestle to compute the right byte offset of a struct field. They also repeat small syntactic mistakes despite prompting. Claude is instructed to make use of C89 (i.e., the 1989 standard of C), however it typically forgets this, which ends up in trivial syntax errors.