ArcVellum v0.9.4

ArcVellum v0.9.4

本次更新

• 叙事星仪支持受限的中键拖动伪 3D 视角旋转。
• 视角旋转以当前聚焦章节或视野中心叙事节点为枢轴，减少长篇作品在旋转时漂移出视野的问题。
• 新增一键复位伪 3D 视角按钮，并让 Canvas 与静态降级渲染保持一致。
• 章节聚焦改为依据章节全部场景节点的重心定位。
• 完成 Windows NSIS 安装包与 Tauri updater 签名产物。

验证

• 前端测试：13 个测试文件、49 个测试通过。
• 生产构建：Vue 类型检查、Vite 构建、桌面同步和 v0.9 构建校验通过。

ArcVellum v0.9.3

ArcVellum v0.9.3

本次更新

• 修复自动推进任务暂停后无法成功续期的问题。新的授权范围会为当前运行建立新的时长计量窗口，避免历史暂停时间被重复计入上限。
• 在自动推进面板前置“按当前范围续期并继续”操作，并在恢复后继续连接正式任务流。
• 新增 Agent 执行中心：以可读状态展示当前任务、受控执行阶段、会话与近期事件，不暴露密钥、提示词正文或项目路径。
• 章节与场景观察视图会展开同章节内每个场景的分支、审查和读者承诺证据；缺失的正式证据会明确标为待完成，不再静默消失。
• 改进空间叙事仪的章节聚焦、曲线布局和紧凑浮窗布局。

验证

• Vue 生产构建及桌面资源同步通过。
• 叙事投影、自动推进、授权续期与 Agent 观测的 28 项回归测试通过。
• 正式发行包将由 packaging/build_desktop.ps1 重新冻结 Python sidecar、生成溯源记录与签名 NSIS 安装程序。

ArcVellum v0.9.2

ArcVellum v0.9.2

Desktop startup reliability hotfix

This release fixes an installation-blocking Windows desktop defect. The frozen Python sidecar previously gave Uvicorn a string import target. That import is resolved by a secondary dynamic import, which is not reliable inside a PyInstaller-frozen executable; the desktop window could therefore remain on its loading scene while the local API was unavailable.

The sidecar now creates the FastAPI application in-process and passes the ASGI application object directly to Uvicorn. The Tauri shell also retains sidecar stderr and termination information long enough to show a clear startup error on the loading scene instead of leaving the user with an indefinite wait.

Verification

• Source serve smoke test reaches the authenticated /health endpoint.
• Frozen sidecar smoke test reaches the authenticated /health endpoint.
• CLI regression test verifies that serve gives Uvicorn an in-process ASGI object.
• Frontend tests and Rust compilation remain required before packaging.

Existing temporary v0.8 smoke-test copies are not update targets. Install the v0.9.2 setup program after it is built, then start ArcVellum from the installed application entry.

ArcVellum v0.7.0

ArcVellum v0.7.0

ArcVellum v0.7.0 turns the autonomous literary-engineering beta into a place where a creator can also read, understand and accompany a growing long-form work.

Highlights

• Complete formal-manuscript reader with chapter and continuous modes, lazy loading, search, bookmarks, typography controls, themes, fullscreen and persisted reading position.
• Reader Manifest v1 with deterministic order, source precedence, coverage deduplication and clean-body projection.
• Five built-in advisor personas, custom persona editing and a layered prompt contract that cannot widen advisor permissions.
• Persistent proactive advisor inbox with trigger policy, deduplication, severity modes, quiet hours and safe suggested actions.
• Narrative Observatory v1 for real scene, branch, character, Canon, reader-question and promise/payoff evidence.
• Book, chapter and scene-level aggregation with accessible list fallback and reduced-motion support.
• Windows hidden-process execution, cached Runner probes, deferred provider loading and a single-document Tauri startup bridge.
• Default local work library at Documents/ArcVellum/Works, with user-selectable future project location.
• Repository and update identity migrated to o-1717986918/arcvellum while preserving the installed application identifier and data location.

Reader Contract

Only formal promoted or exported prose enters the Reader. Candidate drafts, Agent tasks, review reports, state patches and Canon annotations remain outside the reading surface. Exported chapter coverage replaces its underlying promoted scenes; chapters without trustworthy coverage do not hide scene prose.

The manifest contains metadata and per-unit endpoints rather than the full book body. A 100-chapter, 500,000-Chinese-character fixture is covered by automated regression tests.

Advisor Contract

Persona changes affect language, attention and conversational stance only. The advisor remains a read-only consumer of project projections and can suggest only whitelisted actions. Proactive notices are deterministic projections of workflow events; they are persisted, deduplicated and subject to the user's notification mode and quiet hours.

Observatory Contract

Every displayed node and edge is derived from formal project facts or workflow state. Animation communicates real progression and never becomes a second source of truth. Large works default to aggregation, while keyboard-accessible list views preserve the same information without motion or SVG interaction.

Upgrade Notes

• Existing projects are not moved. The default works location applies to new projects unless changed by the user.
• The configuration schema is migrated in place and backed up by the existing application migration path.
• The Tauri identifier remains cn.literary-engineering.studio, so application data and update continuity are preserved.
• Model provider catalogs now load when the connection page is opened or explicitly refreshed.

Known Boundaries

• This release is Windows x64 Beta software.
• Model inference requires network access and a valid provider configured through the bundled Agent Runner.
• Updater artifacts use the existing Tauri minisign chain; Windows Authenticode commercial signing is not included.
• Clean Windows 10/11 virtual-machine install, uninstall and long-running real-model projects remain v1.0 evidence goals.

Literary Engineering Studio v0.3.0

Literary Engineering Studio v0.3.0

Release Summary

v0.3.0 turns the standalone literary workflow repository into an installable Windows application. The release bundles the literary engineering engine, a frozen local application service, the OpenCode Agent Runner, the Chinese project client, and the formal sandbox/writeback pipeline.

Delivered

• Bundled OpenCode 1.18.3 with pinned archive checksum and MIT license notice.
• Separate Agent Runner and Model Connection contracts and settings.
• Frontend provider connect, select, probe and disconnect flows; disconnecting the active custom provider safely restores the built-in starter model.
• Durable SQLite job/event store, route locks, idempotency, leases and restart recovery.
• Normalized streaming events with SSE cursor replay and polling fallback.
• Controlled Worker execution with expected-output-only sandboxes.
• Typed dual-track dispatch: deterministic-cli tasks run only through the embedded engine inside the sandbox; creative and review tasks run through the selected Agent Runner.
• Human-readable writeback diffs, approval, rejection, stale-target detection and rollback.
• Exact Prompt assets for seventeen high-risk literary tasks and a deterministic regression harness.
• Read-only project advisor with immutable snapshots, citations and persistent sessions.
• Tauri 2 desktop shell, native folder selection, single-instance protection, window-state persistence, random localhost port, per-launch authentication and sidecar parent monitoring.
• NSIS Windows installer containing the frozen Python service, embedded engine, frontend and OpenCode runtime.

Verification

• 53 Python unit and API tests passed.
• Prompt Registry: 25 assets, 61 task prompt ids, no errors or warnings.
• Prompt evaluation: 17/17 high-risk exact assets passed.
• JavaScript syntax, Python compilation, git diff --check and Rust cargo check passed.
• Frozen sidecar reported engine ready and rejected unauthenticated API requests.
• Frozen OpenCode probe completed with opencode/big-pickle, matched the requested model and emitted normalized events.
• A real longform fixture completed deterministic budgeting, OpenCode-authored expansion, writeback preview, approval, task submission, task completion and route audit.
• Tauri desktop window, WebView, random-port sidecar startup and forced-parent-exit cleanup were exercised successfully.
• The final NSIS artifact installed silently as version 0.3.0, enforced single-instance startup, served the authenticated local client, uninstalled cleanly and preserved the application data directory.
• Frontend was visually checked at desktop and 375px mobile widths without horizontal overflow.

Windows Artifact

GitHub Release download: literary-engineering-agent-studio_0.3.0_x64-setup.exe

desktop/src-tauri/target/release/bundle/nsis/文学工程 Agent Studio_0.3.0_x64-setup.exe

• Size: 69,593,961 bytes (66.37 MiB)
• SHA256: A94D8D82060D6F63DB972AD364A0D4B0BD091804C69466F655B7CBD13030A559

The generated installer is intentionally excluded from Git. Rebuild it with:

powershell -NoProfile -ExecutionPolicy Bypass -File packaging/build_desktop.ps1

Remaining Release Operations

• Sign the installer and executable with a trusted Windows code-signing certificate.
• Repeat install, upgrade, repair and uninstall checks across the supported clean Windows virtual-machine matrix.
• Add an offline WebView2 installer variant for environments without a current WebView2 runtime.
• Continue large-project performance and long unattended generation tests before recommending fully autonomous book-length production.