Skip to content

Reverse Engineering / Authorized Penetration Testing / Security Research Skill Router Pack AI-powered routing + On-demand toolchain bootstrapping + Self-evolving knowledge base Supports Claude Code, Kiro, Cursor, Cline, and other AI coding clients 逆向/渗透/安全技能路由包 - AI 自动路由 + 按需自举工具链 + 自动进化经验库 | 支持 Claude Code / Kiro / Cursor / Cline 等代码 AI 客户端

Shell
14.3 KB
Raw Blame

Changelog

All notable changes to reverse-skill are documented in this file.

Format follows Keep a Changelog. Versioning follows Semantic Versioning.

Unreleased

Added

  • CI runs remaining unwired suites — test-p0-friction.ps1 on the Windows leg of routing-tests (Windows PowerShell 5.1); case-review/tests/test_review_case.py in the Linux case-contract job. test-workflow-title-safety.ps1 was already wired.
  • Binary Ninja route and skill — added binary-ninja-reverse for HLIL/MLIL/LLIL, Python API, and an explicitly enabled loopback community MCP bridge; Binary Ninja remains a manual commercial dependency.
  • Optional Codex adapter plugin — added plugins/reverse-skill/ without changing the client-neutral core or auto-registering MCP servers.
  • Benchmark coverage for R40 (case-evidence-review) — the route had a single English case; added three cases covering its previously-untested branches (Chinese 证据链/证据图/可追溯性/案件审查/案例审计, and English evidence.?graph/fixity.?check/case.?audit), each verified through both the PowerShell and Bash routers.
  • Route↔benchmark coverage gate — verify-routing-coherence.ps1 now fails when any routing.json route has no routing-benchmark.json case, enforcing the benchmark meta's "add a case when you add a route" rule (the reverse of the existing ghost-expect check).

Fixed

  • macOS BSD sed in bash workflow tests — test-bash-workflow.sh used GNU-only sed -i 's/…/…/'. BSD sed (macOS) treats the expression as a backup suffix and aborts with invalid command code, so Test 4/5 (case-guard reject paths) never ran on macOS. Replaced with a portable mktemp + sed + mv helper and wired test-bash-workflow.sh into macos-bash-compat.yml so the reject-path contract is covered on macOS CI (closes #135).
  • Kali tool discovery: yara — yara is in the Kali bootstrap manifest and the Windows tool catalog and is referenced by four skills, but was missing from kali/scripts/lib/tool-discovery.sh, so refresh-tool-index.sh never reported its path on Kali. Added the catalog entry (malware-analysis), matching its sibling analysis CLIs (binwalk, bkcrack).
  • binwalk could never auto-install on Windows — skills/scripts/bootstrap-manifest.json declared "bootstrapKind": "winget-package" with "wingetId": "ReFirmLabs.binwalk", but that id does not exist in the winget community repository (winget show --id ReFirmLabs.binwalk --exact → 0x8A150014 NO_APPLICATIONS_FOUND), and binwalk ships no prebuilt binaries on any GitHub release (assets: 0 for v3.1.0 and the seven preceding releases), so bootstrap-reverse.ps1 -Capability binwalk could only ever fail with a package-not-found error. The Windows entry is now "bootstrapKind": "manual" + "canAutoInstall": false with a manualInstallHint pointing at the Rust toolchain (winget install Rustlang.Rustup → cargo install binwalk), matching what firmware-pentest/SKILL.md already documents. Linux / Kali / macOS are unaffected: bootstrap-reverse.sh installs binwalk through apt/brew and does not read this entry. tool-index.md.template updated to match.
  • IDA MCP HTTP stall — run-supervisor.py patches stock idalib_supervisor onto ThreadingHTTPServer and accepts Streamable HTTP GET /mcp (patch failure is skipped, supervisor still starts). Keep-alive deadlock is time since last healthy tools/list, not process CreationDate; open.ps1 holds opening.lock so in-flight opens are never -Forced. New recover.ps1 is an immediate -Force path. Never taskkills ida.exe.
  • Broken internal doc links + guard — removed a dangling phishing-case-study.md reference and redirected the missing payloader 反弹shell.md index entry to its tracked raw data. Added skills/scripts/verify-doc-links.py, wired into CI, so internal Markdown links are checked from Git index blobs even when Defender quarantines a working-tree payload file.
  • Windows PowerShell 5.1 encoding — added a UTF-8 BOM to five non-ASCII .ps1 scripts (skills/scripts/verify-doc-facts.ps1, apk-reverse/scripts/frida-run.ps1, apk-reverse/scripts/rebuild-sign-install.ps1, ida-reverse/scripts/start.ps1, radare2/scripts/recon.ps1). Without a BOM, PS 5.1 parses these files as the system ANSI codepage and garbles their Chinese / em-dash string literals; verify-doc-facts.ps1 was failing four checks under 5.1 (CI only ran it under pwsh, which defaults to UTF-8). CI now guards every non-ASCII .ps1 for a BOM.
  • Evidence-consolidation test fixed + two suites wired into CI — test-consolidate-evidence.ps1 printed its success marker but leaked exit 1: it ran review_case.py --verify-hashes on a case whose consolidation had (by design) rewritten E-001.md, so hash fixity could never pass. Dropped --verify-hashes, added an explicit exit-code assertion, and wired both it and test-bootstrap-codex-encoding.ps1 into the Windows leg of routing-tests (both shipped in the repo but were never run by CI).

Changed

  • Coherence clamp (identity-preserving) — RULES.md hot path is master-route → case-init → PRIMARY. routing.json remains the only route table; MASTER-ROUTING.md priority order is verified against JSON. routing.md is advisory. precedent-auth.md no longer grants auth.
  • IDA open — lock files may force a temp copy; .i64 / .idb are never deleted.
  • IDA MCP keep-alive — start.ps1 reuses a healthy HTTP server, launches idalib_supervisor via windowless Python, never taskkills ida.exe (no /T). A listening 13337 with tools/list timeout is treated as busy, not dead, so the 1-minute watchdog cannot kill a supervisor mid-idb_open. open.ps1 talks ida-pro-mcp 2.x idb_open/idb_list.
  • IDA discovery — ToolDiscovery.ps1 now catalogs idalib-mcp, ida-pro-mcp, and ida with Program Files + per-user Python fallbacks.

Added

  • ida-reverse watchdog / scheduled-task installer / GUI launcher / supervisor wrapper (watchdog.ps1, install-autostart.ps1, start-gui.ps1, run-supervisor.py) plus portable LOCAL-SETUP.md.

1.0.1 — 2026-08-08

Added

  • Routing single source of truth — skills/config/routing.json (R0–R39 keyword rules with must / mustAll / exclude semantics). master-route.ps1 now reads this file; hardcoded routing tables removed from scripts. Routing knowledge lives in one place.
  • Routing regression benchmark — skills/tests/routing-benchmark.json (163 bilingual cases, 40 quick) + skills/scripts/test-routing.ps1 runner. Any routing change must keep the benchmark green.
  • Routing keyword coverage expansion (benchmark-driven): burp suite family, pcap/wireshark, root-detection/certificate-pinning, buffer overflow, .so/native/JNI, go binaries (中文), js-encrypt, webshell, privilege escalation, S3/object storage, memory dump, incident response, Bluetooth/BLE, USB, Unity/game reverse, security assessment, and more.
  • Supply-chain pin gate — verify-routing-coherence.ps1 now fails on any auto-install capability lacking pinnedVersion / pinnedCommit / pinPolicy / asset hash. Pinned: frida-tools 14.10.4, pwntools 4.15.0, agent-browser 0.31.1, ida-pro-mcp @commit, SecLists/ProxyCat @commit, nuclei v3.9.0; winget sources annotated with winget-latest policy.
  • Client-neutral integration contract — routing, tests, manifests, and case workflows remain independent of Claude Code, Codex, Cursor, OpenCode, or any other client; client adapters are optional and must not define repository identity.
  • Skill navigation index — skills/INDEX.md auto-generated from SKILL.md frontmatter by extract-summaries.ps1 (-Check mode for CI drift detection).
  • CI pipeline — .github/workflows/ci.yml: Windows + Ubuntu matrix (PowerShell shim for Linux) running test-routing / verify / smoke / INDEX check / JSON validation, plus bash -n syntax checks.
  • Example case — examples/ctf-demo/ full workflow walkthrough (route → scope gate → timeline → evidence → report).
  • frontmatter completion — dsl-vm-reverse/SKILL.md gained name/description frontmatter (was the only module missing it).
  • README refresh — updated the multilingual project overview, release badge, current capabilities, and sponsor showcase layout.
  • case-review/: read-only Evidence Graph Review with scope, timeline, work item, Finding, Path, and optional SHA-256 fixity checks
  • Domain skills R21–R27, R29–R30: protocol-reverse, ghidra-reverse, cloud-k8s, windows-ad, digital-forensics, code-audit, threat-hunting, wifi-wireless, browser-extension-reverse
  • High-quality skills R28, R31–R38: ot-ics, macos-reverse, thick-client, go-rust-reverse, hardware-security, database-security, email-security, identity-federation, radio-sdr
  • Wired into MASTER-ROUTING.md, master-route.ps1, routing tables, domain map, role-map, coherence tests

Removed

  • game-reverse/ (not a product focus; Unity/IL2CPP remains via reverse-engineering + seed-014)

Fixed

  • Upstream mixed-EOL files — 3 markdown files committed with CRLF while .gitattributes declares *.md eol=lf; normalized to LF so git status stays clean on fresh clones.
  • Routing: sigma vs malware, LLM 越狱 vs iOS 越狱, 完整渗透/打到域控 vs AD 域控, forensics vs OT ics; master-route.ps1 rewritten UTF-8 BOM for PS 5.1 CJK
  • Linux/macOS bootstrap: register PentestSwarm MCP with a verified executable path after Go install or when already installed

Security

  • Core scripts do not write client-global instruction files; client-specific integration remains outside the routing core.
  • Added docs/PACKAGE-SECURITY-AUDIT.md: static audit of package executables (no backdoor / no auto DB wipe found)
  • Pin supply-chain floating tags: jshook @0.3.4, pentestswarm v0.1.0
  • Bootstrap integrity: GitHub zip/jar downloads verify assetSha256 (manifest) or GitHub API digest; mismatch deletes file and fails
  • Pin jadx v1.5.6 and apktool v3.0.2 with published SHA256
  • Remove shell evaluation from Kali user-home resolution and pass Frida hosts through argument arrays
  • Write the Burp MCP token atomically with owner-only permissions on POSIX filesystems
  • Reconnect the Burp MCP bridge when Burp starts after the bridge and parse one stdio message per line
  • Enable authentication for bootstrapped Anything Analyzer MCP servers and register the bearer token with supported clients
  • Stop each stale IDA MCP process individually before starting a replacement
  • Add Bash case initialization, authorization guard, and a structured router that reads the same routing.json as PowerShell
  • Verify Bash routing parity in CI without introducing a client-specific plugin manifest
  • Enforce immutable Kali/Windows bootstrap sources for Frida, IDA MCP, Agent Browser, ProxyCat, Nuclei, and pwntools
  • Reject path-like Bash case names and scope authorization checks to the contract's auth/network/signoff sections
  • Align Bash network defaults with PowerShell: authorized URLs use authorized_target_only, while offline readiness requires an explicit local sample
  • Pin GitHub Actions checkout to the immutable v4.2.2 commit and keep the CI case count synchronized
  • Create a functional Kali proxycat wrapper after installing the pinned source checkout
  • Scope PowerShell authorization fields to their contract sections and reject unsupported network modes in both guards
  • Reject unsupported network profiles during case initialization so invalid scopes are never emitted as ready
  • Reject unknown case presets in Bash and PowerShell before creating case artifacts, preventing mistyped presets from silently falling back to pending/offline defaults
  • Generate skills/INDEX.md from tracked skills only, excluding ignored local modules so clean-clone CI stays reproducible
  • Fail routing coherence when a configured skill is missing or only exists as an untracked local file

1.0.0 — 2026-07-18

First formal public release of the reverse-skill skill-router pack.

Added

Ops / combat contract layer (skills/ops/)

  • IDENTITY.md — product identity: lightweight skill router + bootstrap + field-journal (not a Z3r0-style platform)
  • scope-contract.md — case scope + network_profile; auth not granted → no ACT on target
  • evidence-finding-path.md — Evidence → Finding → Path chain
  • role-map.md — lead / specialist role mapping and handoff
  • timeline-workitem.md — timeline + workitem coverage
  • sandbox-profile.md — tool profile mapping
  • skill-supply-chain.md — Agent Skill / MCP install gate (AST10-lite)

PRIMARY routing & case tooling (skills/scripts/)

  • master-route.ps1 — PRIMARY route from task hint
  • case-init.ps1 / case-guard.ps1 — case bootstrap + scope guard
  • append-evidence.ps1 — structured evidence append
  • smoke.ps1 — package smoke checks
  • verify-routing-coherence.ps1 — routing / ops coherence verification
  • test-p0-friction.ps1 — P0 client-side lab friction regression tests

Core skills & docs

  • Full skill matrix: APK / IDA / radare2 / JS / .NET / mobile / malware / pwn / firmware / EDR / pentest / API / LLM / supply-chain / crypto / binary-diff / patch-diff / attack-chain / docs & diagrams
  • MASTER-ROUTING.md + routing.md / routing_zh.md three-axis matrix
  • Bootstrap + tool-index pipeline (bootstrap-reverse.ps1 / .sh, refresh-tool-index)
  • field-journal precedent library + completion checklist
  • Multi-platform paths: Windows primary, Linux / macOS / Kali docs and scripts
  • CTF-Sandbox-Orchestrator competition sub-skills
  • Burp MCP extension package (burp-mcp-full/)

Quality / localization

  • UTF-8 integrity for Chinese docs (RULES_zh, routing_zh, related guides)
  • Client-side lab playbook and recon pipeline references for authorized testing friction reduction

Notes

  • skills/tool-index.md / tool-index.json are machine-local and intentionally gitignored; generate via refresh-tool-index after clone.
  • This tag freezes the skill-router product surface at commit 9fc280b plus this release metadata.

Links