Claude Sonnet 5 and Codex Agents Reach 80% Decompilation of FPS Binary After 500B Tokens
Multi-agent orchestration with Claude and Codex produced an 80 percent accurate C++ reconstruction of a commercial FPS binary. Infrastructure tweaks such as early context compaction and unified Discord-GitHub signaling proved decisive for sustained progress over three months and 500 billion tokens.
The project ran four agents in parallel: three worker instances handled decompilation and commits while one reviewer scanned GitHub issues and CI output. Agents interfaced with IDA-MCP for disassembly, managed progress through per-translation-unit GitHub issues, and coordinated exclusively over a shared Discord channel that also received webhook notifications of build failures. Early compaction thresholds were lowered from 90 percent to 42 percent context fill to prune volatile function data and reduce token waste.
Token-usage logs and commit velocity show the team crossed the 80 percent mark within the first month, at which point the binary launched, displayed the main menu, and loaded maps. Subsequent weeks focused on semantic fixes rather than new coverage; drift in agent focus and premature issue closure were the dominant failure modes once context length exceeded single-compaction cycles. The 42 percent compaction rule cut average context size by roughly one-third compared with default settings.
Operationally the setup demonstrates that multi-agent reverse-engineering pipelines require explicit focus-maintenance mechanisms beyond raw model scale. Lowering compaction and routing CI alerts through the same channel as inter-agent messages produced measurable stability gains without custom harnesses. The decision to drop portability and modernization work mid-project further concentrated tokens on behavioral fidelity.
Next steps center on completing the remaining 20 percent of translation units and verifying original behavior through automated test suites before any platform ports are reconsidered.
RektInator: Remaining 20 percent of translation units reach 95 percent semantic match within 8 weeks under current compaction settings.
Sources (3)
- [1]Primary Source(https://momo5502.com/posts/2026-10-09-game-decompilation/)
- [2]Supporting Source(https://hex-rays.com/products/ida/)
- [3]Supporting Source(https://www.anthropic.com/research)