re-agent/docs/reverse-engineering-tool-coverage.md

Reverse-engineering tool coverage

This page separates RE-Agent operations that are part of the source tree from optional host tools, local development installs, and work that is still only a proposal. Installing a tool does not automatically register it as an RE-Agent operation. A provider counts as integrated here only when its adapter is wired to the shared CLI/MCP workflow and produces retained Evidence.

Current capability map

Analysis area	Integrated in RE-Agent	Depends on the host	Limits
Static native analysis	Ghidra provider; read-only Hopper and IDA provider adapters; native artifact inspection; PE/ELF/Mach-O and other format-specific workflows	Ghidra is bring-your-own; Hopper and IDA are proprietary and optional	Provider support, language coverage, symbols, and decompiler output vary by target. A discovered function is an analysis result, not proof of complete function recovery.
Managed analysis	Execution-free PE/CLI metadata inspection and member analysis	ILSpy package/profile and runtime dependencies	Managed analysis does not execute the target and is not a native Windows runtime test.
Dynamic analysis	Owned-process workflows, native call observation, Frida-backed trace correlation, process capture, browser and native UI scenarios	OS permissions, compatible Frida device/runtime, LLDB, or browser installation	Instrumentation is platform- and target-dependent. An unavailable hook is reported as unavailable; memory writes are not part of the ordinary observation workflow.
Memory and state evidence	Process snapshots, native call argument observations, selected state and file observations	Compatible operating-system process APIs and permissions	These capture selected observable state. They do not establish whole-process equivalence or recover all memory semantics.
Symbolic and differential verification	Bounded Z3 checks for supported scalar expressions; independent differential tests, regression evidence, and bounded repair workflows	Z3 and a runnable original/replica pair	Formal results apply only to the extracted expression and explicit assumptions. Test-suite agreement is empirical, not universal equivalence.
Game-specific verification	General binary, process, UI, image, and differential workflows can support selected game investigations	Game assets, graphics/audio capture tools, input devices, and often a Windows worker	No universal gameplay, physics, rendering, or audio equivalence provider is present. Cross-platform screenshots alone are not an engine proof.

The existing interfaces and evidence model are documented in the Ghidra provider guide, autonomous tool bootstrap guide, and verification reports.

Adapters not currently exposed as public RE-Agent operations

The working tree may contain experimental or incomplete adapter code. It is not a supported operation until its input/output contracts, provider composition, CLI/MCP exposure, lifecycle tests, and a real-tool verification lane are complete. In particular, CAPA and radare2 source drafts must not be described as registered engine capabilities until those gates pass.

This distinction is important because upstream tools do different jobs:

• radare2 provides static disassembly, analysis, debugging, and scripting. Its official documentation describes JSON output and r2pipe; an RE-Agent adapter must still constrain accepted commands and preserve target/evidence identity.
• Mandiant CAPA matches static or supported sandbox-report features against rules. Its JSON output is a capability hypothesis set; unmatched behavior is unknown, not absent.
• Frida supports injected, embedded, and preloaded instrumentation. Because injection can affect a target process, each workflow needs explicit target selection, bounded lifetime, and cleanup evidence; its existence on a host is not proof of an RE-Agent integration.
• Ghidra headless analysis can batch-analyze binaries. RE-Agent's current Ghidra contract remains the supported path; a duplicate Ghidra MCP server is not needed merely to add another frontend.

Development-host snapshot

Observed on macOS Apple Silicon on 2026-10-15 using command -v and version probes. This is a local host snapshot, not a release manifest or a statement that these tools ship with a clone.

Tool	Host observation
Ghidra	ghidraRun resolves to Homebrew Ghidra 12.1.2
radare2	r2 reports 6.2.0
LLDB	/usr/bin/lldb, version lldb-1600.0.39.109
ripgrep	rg reports 15.2.0
Capstone	Python package 5.0.9 in the system Python probe
Pillow	Python package 12.3.0 in the system Python probe
Frida CLI, GDB, CAPA CLI, YARA, RenderDoc CLI	Not found on the probed PATH
Frida, LIEF, Z3, angr	Not found in the probed system Python; a separate bootstrap environment is documented in docs/autonomy-tool-bootstrap.md

The tool bootstrap report in that guide records verified wheel provenance and smoke-test outcomes for its own environment. Do not combine that report with this system-Python snapshot as if they came from the same installation.

Evidence required before marking an integration complete

For each adapter, retain:

1. The exact input artifact SHA-256 and the external tool path, version, and executable SHA-256.
2. A real provider request through the supported RE-Agent CLI or MCP surface.
3. The complete request outcome, process exit status, bounded stdout/stderr, cleanup result, and resulting Evidence identity.
4. Tests for malformed output, cancellation, artifact mutation, output limits, and cleanup failure at the production boundary.
5. A statement of what the tool did not establish.

Passing a schema test or showing that a tool is installed does not satisfy these integration checks.