ReAgent (auto-re-agent) is an autonomous reverse-engineering agent that reconstructs C and C++ functions from compiled binaries. It pulls evidence from Ghidra through a companion Ghidra Bridge, has one model write a candidate implementation and an independent model check it, then builds, tests, and analyses the candidate before accepting it.
Each run selects target functions (in dependency order, easiest first, or by impact) and gathers context: decompiled code, cross-references, structs, enums, vtables, globals, strings, normalized high P-code, control-flow graphs, assembly, and nearby project source. A reverser, checker, and fix loop runs for a bounded number of rounds, and the reverser can request further read-only evidence during the loop.
Acceptance is deliberately conservative. A reversal can require four independent conditions: the checker model returns a pass, a structural verifier finds no strong mismatch against the decompile, assembly, and P-code, the candidate passes the project's configured build, test, and runtime gates in an isolated copy of the project, and an 11-signal parity engine does not flag it red. The project describes this as conservative verification, not proof of semantic equivalence, and it never patches the original source tree.
Reverser and checker roles can use different providers: the Claude API, an authenticated Claude Code CLI, the Codex CLI, or any OpenAI-compatible API. Results are written as reports, per-call logs, round checkpoints, session history, and a knowledge graph linking functions, calls, globals, and strings.
Features
- Reverser and checker loop: independent models generate and review candidate C/C++ bodies over bounded rounds with targeted evidence requests
- Ghidra evidence: decompile, xrefs, structs, enums, vtables, globals, strings, P-code, and CFG data through the Ghidra Bridge
- Structural verifier: compares generated code against decompile, assembly, CFG, and P-code and fails only on strong mismatches
- Candidate validation: builds and tests each candidate in an isolated project copy against configured acceptance gates
- Parity engine: 11 heuristic signals graded red, yellow, or info, such as stub markers, call-count mismatches, and missing floating-point logic
- Manifests: plan bounded batches of functions across classes without model calls, then reverse them with cumulative validation
- Evidence export: stored evidence exported as linked JSON packets and TSV indexes, plus a knowledge graph of functions and calls
- Project profiles: presets for generic C++, Windows x64, and existing reversing projects such as gta-reversed and OpenRCT2
- Provider mix: separate models per role across the Claude API, Claude Code CLI, Codex CLI, and OpenAI-compatible APIs
