MDRSS snapshot · skills/routing.md @ zhaoxuya520/reverse-skill · аудитория: AI-агент · 26298 B · 338 строк · ~6631 токенов · заголовков 10 · блоков кода 1 · ссылок 0 · CJK 343 · git mode 100644 · MIT (root)

Reverse Engineering Skill Routing Matrix

Route tasks to the most appropriate skill module by target type, user intent, and toolchain.

CRITICAL: Routing Execution Protocol

1. MUST complete routing BEFORE executing. Do NOT "do first, route later".
2. SHOULD start from MASTER-ROUTING.md or scripts/master-route.ps1 for PRIMARY; use this full matrix when ambiguous.
3. MUST match dimensions (target type + user intent + toolchain) before entering a skill.
4. If route not matched → propose new skill, do NOT force-fit.
5. Cross-module tasks → combine skills per "Path Crossing" section.
6. After routing, read the target skill's SKILL.md BEFORE taking action.
7. Ops contract (skill-router form): ops/scope-contract.md before ACT; Evidence→Finding→Path; roles in ops/role-map.md; timeline/workitems under work/<case>/. Identity: ops/IDENTITY.md.
8. External skills: do NOT bulk-vendor community packs; map + rules in references/community-security-skills.md + ops/skill-supply-chain.md. RE stages: reverse-engineering/references/re-agent-workflow.md. Recon: pentest-tools/references/recon-pipeline.md.

By Target Type

| OLLVM-obfuscated binary (控制流平坦化/虚假控制流/MBA) | reverse-engineering/references/ollvm-deobfuscation.md — 完整脱密工作流 | obpo-plugin / d810-ng (IDA) / ollvm-unflattener (Miasm) / ollvm-breaker (Binary Ninja) / angr / deollvm (ARM64) | Cryptography / encryption algorithms | reverse-engineering/patterns*.md — crypto patterns | js-reverse/ (if frontend crypto) | | Protocol reverse / custom protocol | reverse-engineering/platforms.md — network protocols | js-reverse/ (if WebSocket/HTTP) | | Go / Rust binary | reverse-engineering/languages-compiled.md + go-reverse.md | ida-reverse/ or radare2/ | | LLM / AI application | llm-security/ — OWASP LLM Top 10 + ASI Top 10 | Prompt injection, Agent security | | API / REST / GraphQL | api-security/ — BOLA/BFLA/JWT/OAuth | pentest-tools/ for scanning | | Supply chain / SBOM / CI-CD | supply-chain-security/ — Trivy/Syft/Gitleaks | — | | iOS app (IPA) | mobile-reverse/ — class-dump/Hopper/Frida iOS | reverse-engineering/platforms.md | | CTF competition (full stack) | ../CTF-Sandbox-Orchestrator/ctf-sandbox-orchestrator/SKILL.md — master entry | Route to 40+ sub-skills by evidence | | CTF ZIP / PKZIP archive | ../CTF-Sandbox-Orchestrator/competition-zip-archive/SKILL.md — legacy ZipCrypto + bkcrack known plaintext | Use before password brute force | | Web runtime / API | ../CTF-Sandbox-Orchestrator/competition-web-runtime/SKILL.md | — | | Cloud / Container / K8s | ../CTF-Sandbox-Orchestrator/competition-agent-cloud/SKILL.md | — | | Windows / AD / Identity | ../CTF-Sandbox-Orchestrator/competition-identity-windows/SKILL.md | — | | Forensics / PCAP / Steganography | ../CTF-Sandbox-Orchestrator/competition-forensic-timeline/SKILL.md | — | | Prompt injection / Agent | ../CTF-Sandbox-Orchestrator/competition-prompt-injection/SKILL.md | — | | Mobile (Android/iOS) | ../CTF-Sandbox-Orchestrator/competition-android-hooking/SKILL.md | — | | Firmware / Malware sample | ../CTF-Sandbox-Orchestrator/competition-firmware-layout/SKILL.md | — |

By User Intent

CTF Wording Normalization

Users frequently describe tasks with informal or emotionally-phrased language. Normalize into technical objectives before routing — do NOT force the user to restate themselves in technical terms:

Do NOT force the user to repeatedly confirm "this is CTF/local." Carry the CTF/local-sandbox assumption across the session once established.

By Toolchain

Check tool-index.md for actual tool availability, paths, and versions. NEVER guess paths.

Ambiguous Intent Recovery Protocol

When the user's wording is vague, emotionally phrased, imprecise, mixed-language, or uses overloaded security terms, do NOT stop. Treat ambiguity as an intent-recovery task:

1. Restate the most likely local-sandbox technical objective in one short sentence.
2. Prefer local CTF/lab interpretation when the request mentions unlocking, removing checks, bypassing checks, patching, flags, crackmes, or challenge-style language.
3. Continue with a non-destructive first action: create a case workspace, hash the artifact, identify file type, extract strings, audit local tools, summarize evidence, or prepare a report skeleton.
4. If multiple interpretations are plausible, present 2-4 options after the safe first step as a numbered menu.
5. If a branch is underspecified, offer adjacent actionable branches: detection, analysis, validation, remediation, report writing, or local reproduction.
6. Always provide a next-step menu — never leave the user with only a dead end.

Suggested Chinese phrasing when recovering ambiguous intent:

我先按"本地沙盒内对该样本做逆向分析"的目标处理。当前先执行不会破坏样本的离线分诊，并在结果后给你选择下一步。

Route Not Matched — Handling

If the current task doesn't match any table above, do NOT force-fit into existing skill:

1. Check if it's an edge case of an existing skill (can extend coverage)
2. If truly new type, proactively propose new skill to user:
   • Suggested skill name and coverage
   • Required toolchain
   • Relationship to existing skills
3. User confirms → execute per CONTRIBUTING.md
4. After creation, update this routing matrix

AI does NOT need to wait for user to discover the gap. Route failure IS the signal to propose a new skill.

Path Crossing (Cross-Module Scenarios)

Some tasks span multiple modules. Common crossings:

APK Reverse Path:
  apk-reverse/decode.ps1 → Java layer analysis
  ↓ If core is in .so
  ida-reverse/ or radare2/ → .so analysis
  ↓ If dynamic verification needed
  apk-reverse/frida-run.ps1 → Frida Hook

Frontend JS Reverse Path:
  js-reverse/Observe → locate target request
  ↓ Need stronger browser/CDP/Hook/Network capability
  jshookmcp → runtime sampling, breakpoints, interception, SourceMap/AST
  ↓ After confirming entry function
  js-reverse/Rebuild → Node local reproduction
  ↓ Need environment patching
  js-reverse/references/env-patching.md

DSL VM Reverse Path:
  reverse-engineering/dsl-vm-reverse/SKILL.md → identify DSL VM (IIFE + single-letter vars + DG() switch-case)
  ↓ Extract opcode table & constant table
  reverse-engineering/dsl-vm-reverse/SKILL.md Phase 2-4 → opcode classification, C[9] constant analysis
  ↓ If runtime capture needed
  browser-automation/ → Playwright/Selenium CDP injection
  ↓ If pure API protocol needed
  js-reverse/ → Observe→Capture→Rebuild (API layer only)

CTF Competition Path (via CTF-Sandbox-Orchestrator):
  ../CTF-Sandbox-Orchestrator/ctf-sandbox-orchestrator/SKILL.md → build sandbox model
  ↓ Route by dominant evidence
  competition-web-runtime/ or competition-reverse-pwn/ or competition-identity-windows/
  ↓ Blocked → return to master
  ctf-sandbox-orchestrator → re-route

Web Pentest + BurpSuite MCP Path:
  browser-automation/ → auto-browse target with Burp proxy
  ↓ Traffic captured
  burpsuite MCP proxy_history → AI analyzes all requests
  ↓ Suspicious endpoints found
  burpsuite MCP intruder_attack → automated enumeration
  ↓ Vulnerability confirmed
  docs-generator/ → generate pentest report

任务完成自检（声称完成前 MUST 通过）

• 我是否完成了路由三轴匹配（目标类型 + 用户意图 + 工具链）？
• 我是否在路由成功后读取了目标 skill 的 SKILL.md？
• 路由未命中时，我是否提议了新增 skill 而非强行匹配？
• 我是否基于 tool-index 使用了真实工具路径？