Cristallin2006/ghidra-skill-for-dsh

This repository is a reverse-engineering agent-skill family built specifically for DeepSeek Harness (dsh). It bundles a persistent Ghidra 12.x headless RPC daemon — based on the vendored Cellebrite Labs ghidra-rpc engine plus dsh-specific patches, with no Jython, GUI, or MCP dependency — alongside eight scenario skills: re-triage, re-unpack, ghidra-static, vuln-audit, re-dynamic, pwn-exploit, traffic-analysis, and android-re. Installation copies the nine directories into ~/.dsh/skills/, creates a Python ≥3.11 virtual environment, and editable-installs the engine submodule; the optional marketplace path uses the git reference v0.9-market (a curated build that intentionally omits pwn-exploit). The unified engine entry is ghidra-core/scripts/rpc_driver.py, exposing ensure, triage, decompile, rename-function, and version-track subcommands, and using an absolute path prefixed with @ as the first argument writes the full JSON response to disk while write operations commit immediately and auto-save the project.

The typical workflow begins with rpc_driver.py triage, which auto-classifies the binary through a three-signal cross-check (section names, magic bytes, structural heuristics) covering UPX, ASPack, Themida, VMProtect, multi-layer packers, and PyInstaller. From there, routes diverge by goal: CTF reversing flows through Go/Rust stripped fingerprinting and pseudocode linting in the static skill; malware triage routes through qiling function-level oracles plus Frida hooks on time and randomness sources in the dynamic skill; vulnerability auditing follows a reachability-first checklist across eight memory-corruption and injection categories; pwn challenges pass through pwn_triage's protection-matrix gate plus a mandatory remote-echo ledger entry that enforces "local pass ≠ remote pass"; pcap forensics uses zero-dependency tshark scripts for DNS/ICMP/USB HID/WPA; APK analysis uses a four-tier jadx decompile plus Toast-anchor localization. The system targets CTF players, malware analysts, vulnerability researchers, and Android security engineers. Rather than relying on agent self-discipline, fourteen iron rules are enforced by mechanical gates: ledger.py blocks same-region revisits without --delta, locks conclusions on write, requires evidence for resolve (otherwise exit 2), and mandates L2 left-inverse anchor binding for --kind model claims via confirmed/killed/waived hypothesis state machines; read_views.py diffs rendered text against real bytes; and crypto_sanity.py validates legality of inverse operations, with any violation hard-failing the run.

Dependencies and limits are explicit. The engine requires Python ≥3.11, Ghidra 12.x via GHIDRA_INSTALL_DIR, and JDK 21+ via JAVA_HOME; DSH_GHIDRA_WS is optional, since ProjectLocator rejects path elements starting with . and the base automatically creates a junction at ~/dsh-ghidra-workspace. The platform badge names Windows and WSL only. Dynamic debugging is delegated outward to Frida, GDB, qemu-user, Qiling, and angr — and angr is gated by gate_explore: no frame slot or emulation evidence means no symbolic execution. The pwn-exploit scenario requires the WSL toolchain to actually run exploits. The optional discipline-enforcement layer lives in dsh-hooks/, mounted through cordis.patch.yml so that dsh's built-in hooks-claude-code bridge turns SessionStart compressed rule cards, PreToolUse gates (sample/explore/longrun/stuck/churn), and Stop stop_check reconciliation into mechanical hooks; changes only take effect after restarting the dsh service and opening a new session. Source code is MIT-licensed, derived from ghidra-rpc, zhaoxuya520/reverse-skill, wgpsec/AboutSecurity, ljagiello/ctf-skills, yaklang/hack-skills, and Und3rf10w/ai-ghidra-tools (Apache-2.0 fragment for Go/Rust/crypto identification). First-run users should run python ~/.dsh/skills/ghidra-core/scripts/doctor.py for environment self-check, then validate conclusion independence through same-problem replay plus the --program-accept ledger entry.