v2.7.0 — Preset & Search Improvements

Added\n\n- Explicit custom-vs-built-in style preset precedence via \prefer_custom.\n\n## Fixed\n\n- Vector search now clamps invalid \ op_k\ values to a safe range.\n\n## Tests\n\n- 372 tests passed.

v2.6.0 — Restore & UI Hardening

Fixed

• Validated project names stored in trash metadata before restore or compensation, preventing restore-path traversal.
• Escaped dashboard labels and values before rendering them into HTML.

Tests

• 370 tests passed, including regression coverage for malicious trash metadata.

v2.5.0 — Isolation & Recovery Hardening

Fixed

• Hardened Web Studio project resolution so unregistered or escaping novel names cannot access storage outside the novels root.
• Validated character names on every read and mutation path to prevent filename traversal.
• Prevented a timed-out shutdown from re-queuing a handler that later completed, avoiding duplicate side effects.
• Preserved proxy environment configuration for API users.
• Excluded symbolic links from complete-project ZIP exports and rejected malformed state.json archives.
• Kept vector-search results from different novels from colliding across the shared index.
• Returned protocol-level MCP tool errors with isError=true.

Tests

• 369 tests passed, including regression coverage for traversal, malformed imports, and shutdown recovery.

Novel Workspace MCP v2.4.0

Highlights

• Added optional Web Studio access-token protection with Basic, Bearer, and dedicated-header authentication.
• Replaced wildcard CORS with exact origin allowlists and made Docker Compose loopback-only by default.
• Added sanitized configuration diagnostics and verified manual project backups.
• Packaged the complete Codex Skill in wheel and source distributions, discoverable with novel-workspace skill-path.
• Added PEP 517 metadata so uv installs console commands from a clean checkout.
• Restored reliable Python 3.10 dependency resolution after newer ONNX Runtime releases dropped CPython 3.10 wheels.
• Fixed backup retention collisions, glob-name handling, relative exclusions, in-project targets, and symbolic-link inclusion.

Verification

• 366 automated tests passed on Python 3.10 and 3.12 in GitHub Actions.
• Wheel and source distribution contents were audited with zero runtime-data entries.
• Installed-wheel doctor, CLI, and Codex Skill resource checks passed.

See CHANGELOG.md for the complete release history.

v2.3.0 — Diagnostics, portable installs, and container publishing

Highlights

• English is now the default language for GitHub, PyPI metadata, and the Web Studio; Chinese and Japanese remain available.
• Added the novel-workspace CLI with serve, mcp, doctor, and version commands.
• Added health and readiness probes for Docker and process supervisors.
• Published a verified GHCR image and attached wheel, source archive, and SHA-256 checksums to this release.
• Installed packages now keep mutable data in the operating system user-data directory instead of site-packages.

Fixes

• Fixed reversible English, Chinese, and Japanese UI switching.
• Fixed MCP stdio corruption by routing diagnostics to stderr.
• Fixed Windows-only GPU controls on Linux and macOS.
• Fixed Docker Compose startup when no local .env file exists.
• Unified package, MCP, and Web API version reporting.

Verification

• 352 tests pass on Python 3.10 and Python 3.12.
• CLI diagnostics, frontend JavaScript, wheel installation, Docker image publishing, and release artifacts are verified by GitHub Actions.

See CHANGELOG.md for the complete history.

v2.2.1 — 稳定性修复版

修复

• 自动加载项目根目录 .env，且不覆盖宿主环境变量
• API 模式严格校验配置的模型 ID，避免错模型被误判为可用
• 模型服务探测失败时正确释放 HTTP client，避免网页/MCP 重试造成连接累积

验证

• 全量测试：346 passed
• AST 语法检查通过
• Codex Skill 结构校验通过

配置

继续参考 .env.example，支持本地 LM Studio 和远程 OpenAI-compatible API。

v2.2.0 — 三入口公开版

亮点

• 提供本地网页工作台、MCP Server、Codex Skill 三种入口
• 模型后端支持本地 LM Studio 和远程 OpenAI-compatible API
• API 密钥、模型名、嵌入模型和服务地址全部支持环境变量配置
• 增强公开仓库边界：忽略本地小说数据、数据库、日志、缓存、虚拟环境和模型资产
• 补充 MIT License、.env.example 和公开使用文档

验证

• 完整测试：344 passed
• Codex Skill 结构校验通过

安装

详见 README：https://github.com/LiZh132707/novel-workspace-mcp#readme