v1.2.0-beta.1 — 蜂群与修复测试版

本测试版基于已合并的 main 39d1d19，发布源提交为 22cabc791305c9ccb3a92dcfbe7d0b70b5dbb651。Python 包版本 1.2.0b1，GitHub 标记为 Pre-release。

推荐下载 reverse-skill-evolver-v1.2.0-beta.1.zip（约 5 MB）。完整包包含 skills/、.omp/skills/、Python wheel、源码包、Burp JAR、安装指南和验证记录。单独 wheel 仅安装 Python CLI。

这次可测试：

• Agent 蜂群：协调 → 并行分析 → 汇总；显式逻辑调用预算、证据快照、取消与失败记录。
• IDA 服务超时与离线诊断、Windows 启动器的 PID/就绪管理。
• 路由演化的预算、进度、部分 checkpoint 和 not_promoted。
• Burp bridge 恢复与错误反馈、PowerShell 别名与能力发现、OMP 原生 workflows。

Windows PowerShell 安装：解压 ZIP 并进入其中目录，然后执行：

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install .\packages\reverse_skill_evolver_cli-1.2.0b1-py3-none-any.whl
.\.venv\Scripts\reverse-skill.exe --version
.\.venv\Scripts\reverse-skill-swarm.exe --help

Python 3.10+；真实蜂群需要已配置模型的 OMP CLI。先按 BETA-TEST.md 测试三次调用、一个任务，再测试并发及取消。IDA 需要已许可的 IDA/idalib；Burp 使用 Java 21+，加载 packages/burp-mcp-full.jar（既有组件版本 2.0.0）。

发布前验证：30 项版本/蜂群回归通过；在独立 venv 安装本 wheel 后，两个 CLI 帮助、版本、三次受控 OMP 子进程调用、证据字节/哈希、目录防覆盖及预算必填均通过；Gradle Java 21 的 check jar 成功。完整 ZIP 已检查 CRC、skills 内容、wheel 源代码与 tag 一致、内嵌安装包与独立资产一致。基础 main 的 Linux CI 为 364 passed、44 skipped。

真实模型协作质量、OMP skill 自动发现、首次用户上手、Burp GUI、Linux/macOS 安装入口及设备/远端同步仍待本次测试；模型输出始终标记 advisory。完整边界见 BETA-VALIDATION.md。

下载后可用 SHA256SUMS.txt 校验。反馈请到仓库 Issues，附版本、实际命令与脱敏日志。

v2.0.0-beta.4 — deterministic index and MCP 2 workflow

Beta test release

This is the 2.0.0-beta.4 test release of the corrected Python-main-chain workflow.

Verified

• GitHub Actions: Ubuntu 24.04.4 LTS (ubuntu-latest) and Windows Server 2025 (windows-latest)
• Python: 3.10 and 3.13 on both runners
• AIGX 1.2.0: clean
• Repository gates: clean
• Test suite: 136 passed, 1 skipped
• Wheel and source distribution built successfully

Scope

• Deterministic SQLite index with BM25/tree/hybrid retrieval
• MCP 2.0 thin read-only adapter
• Explicit tree-sitter provider/cache flow
• IDA JSON/xref ingestion
• No vector database, RAG/model call, implicit grammar download, or duplicate MCP surface

Compatibility note

Ubuntu 20.04 was not part of this CI run. The tested Linux target is Ubuntu 24.04.4 LTS; this beta should not be read as an Ubuntu 20.04 certification.

v2.0.0-beta.1 - AIGX-first workflow beta

Beta highlights

• AIGX is now the mandatory project context and edit-boundary plane.
• Deterministic route_task preflight fails closed on missing artifact input, capability, service, authorization, project context, or structural gate.
• Adds portable IDA 9.4/idalib entrypoints, plugin capability validation, read-only Teams preflight, and isolated worktree-lab contracts.
• Resolves project-owned Code Intel/Sentrux scope through AIGX; no repository-root fallback and no automatic baseline save.
• Keeps target paths, credentials, IDB files, runtime traces, and analysis evidence outside the generic skill repository.

Breaking change

RULES files are compatibility bootstraps only. Start with .aigx/protocol.aigx, then use skills/scripts/route_task.py before opening a child skill.

Known beta boundary

IDA Teams management and binary analysis are currently two explicit stages. A combined Teams + artifact-analysis request returns composite_workflow_not_supported instead of claiming success; run Teams preflight/lab first, then a separate artifact route with input_path.

v1.0.0 - First packaged release

First tagged release of the Cybersecurity Skills Router / Reverse-Engineering Skill Routing Pack.

What's included

• Evolution control plane (skills/evolution/): Goal Contract -> capability graph -> macro/micro route -> TraceCard -> oracle -> promotion gate. Wraps, does not replace, domain skills.
• 22 domain skills: apk-reverse, ida-reverse, ghidra-reverse, js-reverse, radare2, malware-analysis, mobile-reverse, firmware-pentest, pwn-chain, patch-diff-exploit, binary-diff, edr-bypass-re, supply-chain-security, llm-security, api-security, pentest-tools, attack-chain, browser-automation, reverse-engineering, docs-generator, diagram-generator, and the field-journal memory system itself.
• Field-journal memory tiers (skills/field-journal/): validated / candidate / forensic, each entry requires provenance (source trace, tool versions, verification time, environment).
• MCP/wheel integration examples: kali/mcp-kali-example.json (kali-server, metasploit-mcp, hexstrike, pentestswarm, jshook, anything-analyzer, idapro, ghidra), burp-mcp-full/ (Burp Suite MCP bridge), .kiro/settings/mcp.json (live client config example).
• Routing tables: skills/routing.md / skills/routing.json (+ zh variant), skills/CONTRIBUTING.md skill-addition criteria.

New in this release

• ghidra-reverse skill: new skill for GhydraMCP (starsong-consulting/GhydraMCP), sourced from upstream README/code. Flags a real discrepancy found in the pre-existing but previously undocumented .kiro/settings/mcp.json ghidra entry (configured port/transport does not match GhydraMCP's documented defaults) so it doesn't get silently trusted.
• js-reverse CDP tool mapping: documents the mcp__plugin_ecc_chrome-devtools__* tool surface alongside the existing js-reverse_*/jshookmcp mappings.
• Field-journal methodology entry (candidate tier): generalized lessons from a private-API reverse-engineering session -- treat a target's own front-end column/grid-definition JS as a higher-signal field-semantics source than guessing from network traffic, live-probe the real endpoint and diff against the inferred spec rather than trusting it blindly, parse by column name not position, fail loudly on missing/malformed fields.
• Fleet sync script (skills/scripts/fleet-sync.ps1): collision-safe additive sync of field-journal candidate/validated entries, capability-graph.json, and tool-index.md between two machines over SSH/SCP. Never overwrites, never deletes; reports collisions for manual resolution.