Canvast 0.1.0 — First Public Release / 首个公开发布版本

English

First public release of Canvast — a self-built, open-source (Apache-2.0) loop harness for AI coding agents, built for large projects.

Traditional harnesses lose the project: every turn starts nearly blind, context compaction silently deletes decisions, and long tasks drift into forgetting, goal loss, and unverifiable "done". Canvast hooks the agent loop itself and re-grounds every turn in persistent whole-project state.

Signature capability — Graph Canvas. The project's living map: File, Plan, Decision, and AgentRun nodes with typed links, persistent across sessions and compactions, queryable by the agent, traceable end-to-end, and exportable to JSON, Markdown, Mermaid, SVG, and HTML.

A complete advanced harness toolkit. Automatic DAG workflow decomposition, bounded sub-agent dispatch, sidecar continuation, plan mode, sandboxed execution with typed permission profiles, runtime observability (receipts, queues, lifecycle, tool runs, approvals), resource-safety fencing, and a sealed runtime — across two surfaces sharing one persisted project state.

Measured, not marketed. Evaluated in paired runs against Claude Code as the reference toolchain — same tasks, same protocol, same DeepSeek model backend on both sides. The README evidence block is machine-regenerated from the checked-in record; see docs/EFFECTIVENESS_EVIDENCE.md for the full methodology and current measurements.

Two ways to run:

• Terminal (TUI) — canvast-tui-0.1.0.tgz: extract, npm install, cp .env.example .env (add DEEPSEEK_API_KEY), then ./canvast.sh. Requires Node.js 22.19.0+.
• macOS App — Canvast-macos-0.1.0.zip: extract and move Canvast.app to Applications. macOS 13+. The App embeds the sealed runtime (no separate Node.js install). Ad-hoc signed — first launch via right-click → Open. Not notarized; see docs/guides/macos-public-notarization.md for the optional path.

Known issues: in the published paired evaluation Canvast's median run duration is higher than the reference toolchain's; exact values are disclosed in the README evidence block and docs/EFFECTIVENESS_EVIDENCE.md, and the overhead is recorded as an optimization target.

中文

Canvast 首个公开发布版本——自研、开源（Apache-2.0）的 loop 架构 AI 编程 agent harness，为大型项目而生。

传统 harness 会丢失项目：每一轮几乎都是半盲启动，上下文压缩静默删除决策，长任务在遗忘、目标漂移和无法验证的"完成"中失控。Canvast 钩住 agent loop 本身，让每一轮都重新锚定到持久化的全项目状态。

标志性能力——Graph Canvas 治理画布。 项目的活地图：File、Plan、Decision、AgentRun 四类节点与 typed link，跨 session、跨压缩持久存在，agent 可查询，全程可溯源，可导出 JSON、Markdown、Mermaid、SVG、HTML。

全量高级 harness 能力。 自动 DAG workflow 拆分、有界子 agent 配发、sidecar 接续、plan mode、带 typed 权限 profile 的沙箱执行、运行时观测（回执、队列、生命周期、工具运行、审批）、资源安全围栏与密封运行时——两个界面共享同一份持久化项目状态。

只用实测说话。 以 Claude Code 为参考工具链做配对运行——同任务、同协议、两侧相同的 DeepSeek 模型后端。README 证据区块由机器从已检入记录重新生成；完整方法论与当前测量见 docs/EFFECTIVENESS_EVIDENCE.md。

两种使用方式：

• 终端（TUI）——canvast-tui-0.1.0.tgz：解压后 npm install，cp .env.example .env（写入 DEEPSEEK_API_KEY），然后 ./canvast.sh。需要 Node.js 22.19.0+。
• macOS App——Canvast-macos-0.1.0.zip：解压后将 Canvast.app 移入"应用程序"。需要 macOS 13+，App 内嵌密封运行时（无需单独安装 Node.js）。ad-hoc 签名，首次启动请右键 → 打开。未公证；可选公证路径见 docs/guides/macos-public-notarization.md。

已知问题： 在已发布的配对评估中，Canvast 的中位运行时长高于参考工具链；具体数值已在 README 证据区块与 docs/EFFECTIVENESS_EVIDENCE.md 公开，该开销已记录为优化目标。