0.13.0

Stela v0.13.0

Compared with v0.12.0, this release makes long Agent workflows easier to control and inspect, replaces partial Canvas data refresh with atomic Agent-led re-analysis, and adds native multi-block editing to Markdown notes.

English

Highlights

• Multi-block Markdown editing: Drag through the editor gutter to select complete paragraphs, headings, lists, RunSQL blocks, tables, and other blocks as one range. The selection works with copy, cut, delete, replacement, undo, and drag-and-drop instead of falling back to an incomplete browser text selection.
• Atomic Agent-led Canvas refresh: Whole-Canvas and single-source refresh actions now open a dedicated Agent run. The Agent reruns the relevant queries, re-evaluates dependent KPI, chart, table, narrative, and flow content, and commits the Canvas only after every targeted source has a successful audited result from that run.
• Controllable model reasoning: Every AI profile has a reasoning-effort setting from off through max, constrained to the selected built-in model's supported levels. Agent Dashboard records requested and effective effort so runs and DataAgentBench results remain comparable.
• Visible Agent progress: Ordinary model-step narration appears while a long run is in progress, without exposing hidden reasoning or tool-call payloads. The last process message becomes the final answer in place, while earlier process messages and strategy reviews collapse after completion.
• Action-oriented execution traces: Agent Dashboard focuses its trajectory on model calls, tool executions, approvals, strategy reviews, and context compaction. Model details now own readable input and output, token usage, cache tokens, context-window occupancy, reasoning effort, status, and timing.

Improvements

• Faster live schema lookup: Schema retrieval prefers a connector's direct multi-table describe capability and falls back safely when a connector does not implement it, reducing repeated remote round trips.
• More focused Agent behavior: The system prompt now follows a compact evidence hierarchy and conditional analysis workflow. Routine lookup, schema, and query tasks avoid unnecessary planning, retrieval, and adjacent investigation.
• Safer Skill discovery and maintenance: Skills are classified as fresh, stale, or untracked. Routine discovery hides stale entries, while explicit knowledge maintenance can inspect old drafts, verify them against live evidence, and save source-bound replacements.
• Useful Agent empty state: The empty Agent Panel offers context-aware actions for creating or refreshing a Canvas, summarizing or auditing the current document, and maintaining the knowledge library.
• Cleaner maintenance UX: Skill listing and search share one paginated interface, knowledge maintenance no longer injects every Skill summary into the initial prompt, and stale Skills fail early without biasing routine Agent work.

Upgrade Notes

• No manual data migration is required when upgrading from v0.12.0. Existing notes, Canvas files, SQL templates, connections, execution history, Agent history, and Skills remain readable.
• Existing AI profiles without a saved reasoning-effort value default to medium. Custom OpenAI-compatible endpoints must support the standard reasoning_effort field when a non-off level is selected; otherwise select off for that profile.
• The Canvas file schema remains version 1. Refresh now requires an available model and connection and may take longer than direct SQL-only refresh, but a failed target query leaves the Canvas unchanged.
• Completed process narration is stored in bounded device-local Agent history. Streaming token fragments and hidden reasoning are not persisted as conversation events, and the existing history-retention policy is unchanged.

中文

重点更新

• Markdown 多块编辑： 在编辑器边缘拖动，可将段落、标题、列表、RunSQL Block、表格等完整内容块选为一个范围。选区可直接复制、剪切、删除、替换、撤销和拖放，不再退化为不完整的浏览器文本选区。
• Agent 原子刷新 Canvas： 整个 Canvas 和单一数据源的刷新都会启动专用 Agent 任务。Agent 重新执行相关查询，并重新评估 KPI、图表、表格、说明和流程内容；只有当所有目标数据源都绑定到本轮审计通过的成功结果后，才会一次性更新 Canvas。
• 可控的模型 Reasoning： 每个 AI Profile 都可选择从 off 到 max 的 Reasoning Effort，内置模型会限制为其实际支持的强度。Agent Dashboard 会记录请求强度和实际强度，便于对比不同运行与 DataAgentBench 结果。
• 可见的 Agent 执行进度： 长任务执行时会展示模型各步产生的普通过程说明，但不暴露隐藏 Reasoning 或工具调用参数。最后一条过程消息会原位升级为最终回答，早期过程消息和策略复盘在完成后默认折叠。
• 面向动作的执行轨迹： Agent Dashboard 的主轨迹专注于模型调用、工具执行、用户审批、策略复盘和上下文压缩。模型步骤详情会统一展示可读的输入输出、Token 用量、Cache Token、Context Window 占用、Reasoning 强度、状态和耗时。

改进

• 更快的实时 Schema 查询： Schema 获取优先使用 Connector 直接提供的多表 Describe 能力；Connector 未实现时会安全回退，从而减少重复的远程往返。
• 更聚焦的 Agent 行为： System Prompt 改为精简的证据层级和按需执行的分析流程。常规的定位、Schema 检查和查询任务会减少不必要的计划、检索和旁支调查。
• 更安全的 Skill 发现与维护： Skill 会被分为 fresh、stale 和 untracked 三种状态。日常发现会隐藏 stale 条目；显式的知识维护任务则可以查看旧草稿，用实时证据验证后保存绑定来源的替代内容。
• 更实用的 Agent 空白状态： Agent Panel 会根据当前文档提供创建或刷新 Canvas、总结或审查当前文档、整理知识库等可直接执行的入口。
• 更干净的知识维护体验： Skill 列表与搜索合并为一个可分页接口，知识维护不再把所有 Skill 简介注入初始 Prompt，stale Skill 也不会再干扰日常 Agent 工作。

升级说明

• 从 v0.12.0 升级无需手动迁移数据。现有笔记、Canvas、SQL 模板、数据库连接、执行历史、Agent 历史和 Skill 均可继续读取。
• 没有保存 Reasoning Effort 的现有 AI Profile 会默认使用 medium。当 Custom OpenAI-compatible Endpoint 选择非 off 强度时，需要支持标准 reasoning_effort 字段；如果不支持，请将该 Profile 设为 off。
• Canvas 文件 Schema 仍为版本 1。刷新现在依赖可用的模型和数据连接，可能比仅执行 SQL 更慢，但任一目标查询失败都不会修改 Canvas。
• 完成的过程说明会保存到有界的本机 Agent 历史中。流式 Token 片段和隐藏 Reasoning 不会作为对话事件持久化，现有历史保留策略不变。

0.12.0

Stela v0.12.0

Compared with v0.11.0, this release makes the Data Agent substantially more capable and inspectable. It can work across SQL and MongoDB sources, analyze larger query artifacts in a sandboxed Python runtime, keep ordered resource references in the conversation, and expose its session-level execution path in Agent Dashboard.

English

Highlights

• Sandboxed Python data analysis: The Agent can use Pyodide with DuckDB, pandas, and NumPy to calculate statistics, join query artifacts, and inspect data that would be too large or awkward to pass through chat text. The runtime is bundled with Stela, requires no system Python installation, and runs without host filesystem or network access.
• Session-scoped query artifacts: Read-only SQL and MongoDB queries can retain bounded local artifacts for later Agent steps. Python analysis reads those artifacts by session alias, so results from different connections can be combined without embedding full datasets in model messages or durable Canvas files.
• MongoDB and structured read-only queries: MongoDB joins MySQL and PostgreSQL as a bundled connector. The Agent uses one structured read-only query interface for SQL statements and MongoDB aggregation pipelines, with connector-aware validation and existing mutation approval boundaries kept intact.
• Session-oriented Agent Dashboard: Inspect a session by turn, scan the full-width execution waterfall, and select model or tool steps to view formatted payloads, results, status, and timing. Cache-hit telemetry is included so prompt-cache improvements can be measured rather than assumed.
• Ordered Agent composer: The Agent input now uses ProseMirror. Tables, notes, Canvas files, and RunSQL blocks stay as inline resource pills at the exact position where they were mentioned, while the currently open tab is supplied as implicit context without adding repetitive visible pills.

Improvements

• Prompt-cache-friendly requests: Stable Agent instructions are kept ahead of turn-specific context, supported providers receive cache-retention hints, and sessions keep a stable cache identity. This improves latency and token cost when the provider supports prompt caching.
• Agent quick actions: RunSQL rewrite, question, and error-repair actions open the Agent Panel and reliably insert their prepared message. SQL repair includes the execution error instead of sending the query alone.
• Adaptive strategy review: Repeated or low-yield query patterns can trigger one bounded, tool-free strategy checkpoint. The Agent either continues with explicit justification or changes direction, reducing time spent circling around similar attempts.
• Analysis efficiency: The Agent reuses validated query evidence, reviews repeated query families, and records efficiency diagnostics used by DataAgentBench. Failed runs now retain usage data where the provider reports it.
• Dashboard readability: Session lists use a narrower navigation column and a wider details pane. String payloads preserve line breaks, JSON payloads and results wrap long values, and low-value truncated summary panels have been removed.
• Model and connection controls: The Agent model selector is wider and switches profiles through one validated settings path. Settings can open the new-connection form reliably.
• Cross-platform packaging: Pyodide assets are prepared before builds and copied with Windows-safe normalized glob paths.

Upgrade Notes

• No manual migration is required when upgrading from v0.11.0. Existing notes, Canvas files, SQL templates, connections, AI profiles, execution history, and Agent history remain compatible.
• Stela now bundles approximately 31 MB of Pyodide runtime and Python package assets before installer compression. This enables execute_python without a local Python environment and increases the packaged application size.
• Python inputs are temporary, session-scoped query artifacts rather than Vault documents. They are not intended for Git synchronization; if an artifact is no longer available, rerun its source query before using it in Python.
• Prompt-cache savings depend on the selected provider and model. Agent Dashboard reports cache reads when the provider exposes them; a zero rate does not imply a local Stela error.

中文

重点更新

• 沙箱 Python 数据分析： Agent 可以通过 Pyodide 使用 DuckDB、pandas 与 NumPy 进行统计计算、关联查询产物，并检查不适合完整塞进对话文本的大规模数据。运行时随 Stela 一起提供，无需安装系统 Python，也无法访问宿主文件系统和网络。
• 会话级查询产物： 只读 SQL 与 MongoDB 查询可以为后续 Agent 步骤保留有界的本地临时产物。Python 分析通过会话别名读取这些产物，因此可以组合不同连接的结果，而无需把完整数据集放进模型消息或持久化 Canvas 文件。
• MongoDB 与结构化只读查询： MongoDB 现在与 MySQL、PostgreSQL 一样作为内置连接器提供。Agent 使用统一的结构化只读查询接口执行 SQL 语句和 MongoDB 聚合管道，同时保留针对不同连接器的校验以及现有的数据修改审批边界。
• 按会话组织的 Agent Dashboard： 可以按 Turn 检查会话，通过横跨顶部的执行时间线浏览全局过程，再选择模型或工具步骤查看格式化参数、结果、状态和耗时。Dashboard 还会记录缓存命中指标，让 Prompt Cache 优化可以被实际衡量。
• 有序的 Agent 编辑器： Agent 输入框改用 ProseMirror。表、笔记、Canvas 与 RunSQL Block 会作为内联资源 Pill 保留在用户提及它们的准确位置；当前打开的标签页则作为隐式上下文提供，不再重复增加显式 Pill。

改进

• 适配 Prompt Cache 的请求结构： 稳定的 Agent 指令位于每轮动态上下文之前；支持的 Provider 会收到缓存保留提示；同一会话保持稳定的缓存身份。Provider 支持 Prompt Cache 时，可以降低延迟与 Token 成本。
• Agent 快捷操作： RunSQL 改写、问询和错误修复操作会打开 Agent Panel，并可靠地插入准备好的消息。修复 SQL 时会同时提供执行错误，而不再只发送 SQL。
• 自适应策略复核： 重复或低收益的查询模式可以触发一次有界且不调用工具的策略检查。Agent 需要明确说明继续理由或改变方向，减少围绕相似尝试反复打转。
• 分析效率： Agent 会复用已经验证的查询证据，检查重复查询族，并记录供 DataAgentBench 使用的效率诊断；如果 Provider 提供 Usage，失败运行也会保留相关统计。
• Dashboard 可读性： 会话事件列表改为更窄的导航栏，详情面板得到更大空间。字符串参数保留换行，JSON 参数和结果会对长内容自动折行，并移除了价值较低且经常被截断的摘要面板。
• 模型与连接控制： Agent 模型选择器更宽，并通过单一、经过校验的设置链路切换配置；设置面板也可以可靠地打开新建连接表单。
• 跨平台打包： 构建前会准备 Pyodide 资源，并使用兼容 Windows 的规范化 Glob 路径完成复制。

升级说明

• 从 v0.11.0 升级无需手动迁移。已有笔记、Canvas、SQL 模板、数据库连接、AI 配置、执行历史和 Agent 历史保持兼容。
• Stela 现在会在安装包压缩前携带大约 31 MB 的 Pyodide 运行时和 Python 包资源。这使 execute_python 不依赖本地 Python 环境，但也会增加应用包体积。
• Python 输入是按会话管理的临时查询产物，不是 Vault 文档，也不用于 Git 同步；如果产物已经不可用，需要先重新执行来源查询，再在 Python 中使用。
• Prompt Cache 的收益取决于所选 Provider 和模型。只有 Provider 返回相关统计时，Agent Dashboard 才能显示缓存读取；命中率为零不一定表示 Stela 本地出现错误。

0.11.0

Stela v0.11.0

English

Highlights

• Analysis Canvas: Turn audited query results into a Git-trackable *.stela.canvas with KPI, chart, table, narrative, and flow cards. Canvas sources keep their SQL and latest successful run reference without copying result rows into the artifact. Refresh one source or the whole Canvas explicitly; a failed refresh keeps the last successful snapshot.
• Charts tied to query results: The Agent can visualize an exact RunSQL result in the conversation or use it in a durable Canvas. Stela's controlled chart grammar covers trends, rankings, composition, distributions, correlations, funnels, retention, comparisons, and bounded layered views, with consistent number, percent, currency, date, and duration formatting.
• Offline Canvas export: Export a Canvas as one standalone HTML file. ECharts views retain tooltips, hover emphasis, legend interaction, and responsive resizing without a network connection. Tables and flow diagrams are stable snapshots, and the supporting SQL remains inspectable.
• Agent Dashboard: Review 7-, 30-, or 90-day activity, completion rate, latency, token usage, tool reliability, Skill selection and usage, knowledge-maintenance outcomes, and paginated redacted traces. The dashboard is Vault-local, Git-ignored, clearable, and never sends telemetry to Stela.
• SQL template editing: Create a template directly from the library and edit its name, description, connection, and RunSQL content in the normal editor. Insert it with Mod+Alt+T; repeated {{variables}} stay linked while editing, and Tab / Shift+Tab moves between variable groups.
• Bilingual Demo Vault: Try Demo Vault now creates one English-and-Chinese Northstar Outfitters commerce review instead of a single welcome note. It includes business context, audited SQL evidence, saved results, Canvas, Agent history, Skills, SQL templates, and an optional Docker-based MySQL fixture for rerunning the analysis.

Improvements

• Agent knowledge maintenance: The main answer now finishes and unlocks before background Skill maintenance begins. Maintenance is Vault-scoped and bounded to 60 seconds and five model turns, while source hashes and retrieval anchors allow stale knowledge to be refreshed when supporting notes change.
• Canvas flow diagrams: Flow cards use a readable natural-size, scrollable preview instead of shrinking the whole diagram. Expand a flow for zooming and panning, then opt into layout adjustment for node dragging, TB/LR direction, and deterministic auto-layout. Open Canvas tabs also refresh when the Agent updates them.
• Export path feedback: Canvas HTML export now reports the saved local path, matching the existing Markdown export experience.
• SQL template drafts: The template library no longer keeps name and description fields in a separate creation form. New drafts open in the editor, use sortable timestamp-based filenames, and recover with a stable fallback name if closed before naming.
• Shortcut documentation: Settings and docs/keybindings.md now cover the product-defined workspace, editor, search, RunSQL, template-variable, and image-preview shortcuts.
• macOS and Windows fixes: macOS now waits for the native Vault watcher to unsubscribe before process teardown, preventing the FSEvents crash seen on quit. Demo validation also handles Windows CRLF checkouts.

Upgrade Notes

• No manual migration is required when upgrading from v0.10.3. Existing Markdown notes, connections, execution history, Git settings, and SQL templates remain compatible.
• Agent Dashboard data starts accumulating after the upgrade; previous Agent history is not backfilled. Metrics and redacted traces live in .stela/agent-metrics.local.sqlite, are ignored by Git, expire after 90 days, and can be cleared from the dashboard.
• Canvas files (*.stela.canvas) and SQL template Markdown files under .stela/sql-templates/ are normal Vault artifacts and may be committed to Git. Query result rows continue to live in Stela's existing JSONL history and local SQLite cache rather than inside Canvas files.
• A Canvas HTML export containing charts embeds the pinned ECharts runtime and is therefore roughly 1.1 MB larger. It runs embedded JavaScript for chart interaction, but makes no network requests and has no access to Stela IPC, database connectors, or the source Vault.
• The prepared Demo Vault can be browsed immediately from its saved results. Docker is required only to rerun the bundled MySQL queries, and an AI provider is required only for the live Agent and new Dashboard activity steps. Recreating the Demo in the same destination adds missing files but does not overwrite files you have edited.

中文

重点更新

• 分析画布： 将经过审计的查询结果组织成可由 Git 跟踪的 *.stela.canvas，在同一画布中展示 KPI、图表、表格、分析文字和流程图。Canvas 数据源保存 SQL 与最近一次成功执行的引用，不把结果行复制进画布文件；你可以显式刷新单个或全部数据源，刷新失败时仍保留上一次成功快照。
• 与查询结果绑定的图表： Agent 可以直接把某次 RunSQL 的准确结果可视化在对话中，也可以将其用于持久化 Canvas。Stela 的受控图表语法支持趋势、排名、构成、分布、相关性、漏斗、留存、对比和有限的分层图表，并统一处理数字、百分比、货币、日期和时长格式。
• 离线 Canvas 导出： Canvas 可以导出为单个独立 HTML 文件。ECharts 图表在无网络环境下仍保留 Tooltip、Hover 高亮、图例交互和响应式缩放。表格与流程图导出为稳定快照，支撑分析的 SQL 也可以展开查看。
• Agent Dashboard： 查看最近 7、30 或 90 天的 Agent 活跃情况、完成率、耗时、Token 用量、工具可靠性、Skill 匹配与使用、知识维护结果，以及可分页查看的脱敏 Trace。Dashboard 数据仅保存在当前 Vault 的本机 Git 忽略文件中，可以随时清空，不会向 Stela 发送遥测数据。
• SQL 模板编辑： 从模板库创建模板后，直接在普通编辑器里修改名称、描述、连接和 RunSQL 内容。按 Mod+Alt+T 插入模板；重复的 {{变量}} 会保持联动编辑，并可用 Tab / Shift+Tab 在变量组之间跳转。
• 双语 Demo Vault： Try Demo Vault 不再只创建一篇欢迎笔记，而是在同一个 Vault 中提供 Northstar Outfitters 电商经营复盘的中英文完整流程，包括业务背景、可审计 SQL 证据、已保存结果、Canvas、Agent 历史、Skills、SQL 模板，以及用于重新执行分析的可选 Docker MySQL 数据集。

改进

• Agent 知识维护： Agent 主回答完成后会立即解除会话占用，再独立执行后台 Skill 维护。维护任务按 Vault 隔离，最多运行 60 秒和 5 个模型轮次；通过来源哈希与检索锚点，支撑笔记发生变化时可以识别并刷新过期知识。
• Canvas 流程图： 流程图卡片改为自然尺寸、可滚动的清晰预览，不再为了显示完整流程而整体缩小。展开后可以缩放和平移，并按需进入布局调整模式，拖动节点、切换 TB/LR 方向或执行确定性的自动布局。Agent 更新画布时，已经打开的 Canvas 标签页也会同步刷新。
• 导出路径提示： Canvas HTML 导出成功后会显示本地保存路径，与现有 Markdown 导出的体验保持一致。
• SQL 模板草稿： 模板库不再用单独的创建表单重复填写名称和描述。新草稿会直接进入编辑器，使用可排序的时间戳文件名；如果尚未命名就关闭，也会生成稳定的兜底名称，避免丢失草稿。
• 快捷键说明： Settings 与 docs/keybindings.md 现在完整列出工作区、编辑器、搜索、RunSQL、模板变量和图片预览相关的产品快捷键。
• macOS 与 Windows 修复： macOS 退出时会等待原生 Vault watcher 完成退订，避免进程析构阶段发生 FSEvents 崩溃；Demo 校验也已兼容 Windows 的 CRLF checkout。

升级说明

• 从 v0.10.3 升级无需手动迁移。已有 Markdown 笔记、数据库连接、执行历史、Git 设置和 SQL 模板保持兼容。
• Agent Dashboard 从升级后开始积累数据，不会回填旧的 Agent 历史。指标与脱敏 Trace 保存在 .stela/agent-metrics.local.sqlite，默认不进入 Git，保留 90 天，并可在 Dashboard 中清空。
• Canvas 文件（*.stela.canvas）以及 .stela/sql-templates/ 下的 SQL 模板 Markdown 都是普通 Vault 文件，可以提交到 Git。查询结果行仍保存在现有 JSONL 历史与本地 SQLite 缓存中，不会写入 Canvas 文件。
• 包含图表的 Canvas HTML 会内嵌固定版本的 ECharts runtime，因此文件大约会增加 1.1 MB。它会运行内嵌 JavaScript 来保留图表交互，但不会发起网络请求，也无法访问 Stela IPC、数据库连接或源 Vault。
• Demo Vault 的已保存结果可以直接浏览。只有重新执行内置 MySQL 查询时才需要 Docker；只有体验真实 Agent 和新增 Dashboard 活跃记录时才需要配置 AI Provider。在同一个目标位置重新创建 Demo 只会补齐缺失文件，不会覆盖你已经编辑过的内容。

0.10.3

Stela v0.10.3

English

Highlights

• Reusable SQL templates — Create Vault-scoped SQL templates as Markdown and insert them directly into RunSQL blocks. Templates support
  {{variable}} placeholders, linked editing, keyboard navigation, and optional connection metadata.
• Persistent Agent session history — Agent conversations now survive app restarts through device-sharded JSONL history stored in the Vault. Sessions
  from another device remain read-only until continued as a new local fork.
• Agent execution plans — For multi-step analysis, the Agent can maintain and display a structured execution plan, making progress and remaining
  work easier to follow.
• Live database schema authority — Agent table search and schema inspection now use the live connector instead of potentially stale schema
  snapshots. Connectors can also provide column descriptions and comments through the new optional describeTables capability.
• More efficient Vault watching — Stela now uses a native recursive watcher, reducing filesystem resource usage and improving shutdown behavior for
  larger Vaults.

Improvements

• Agent answers are more concise by default while retaining the key evidence needed to verify a result.
• Agent tool activity and execution-plan progress are presented more clearly in the Agent panel.
• Automatic Skill maintenance is now evidence-gated, create-only, and limited to concise reusable knowledge.
• Skill distillation can use related SQL usage and favors recently updated Vault documentation when historical guidance conflicts.
• SQL templates are stored under .stela/sql-templates/, remain Git-friendly, and do not appear as ordinary notes or pollute SQL usage search.
• Added keyboard shortcuts for opening the template picker and navigating template variables.
• Improved Windows CI compatibility and expanded automated coverage for Agent history, execution plans, schema resolution, Skills, Vault watching, and
  SQL templates.

Upgrade Notes

No manual migration is required.

Agent history created after upgrading is stored under .stela/agent-history/<device>/ and follows the Vault through Git. Stela retains the 20 most
recently updated Agent sessions for each device.

SQL templates are stored under .stela/sql-templates/. This directory is created when templates are added.

中文

重点更新

• 可复用 SQL 模板 —— 可以在 Vault 中以 Markdown 创建 SQL 模板，并直接插入 RunSQL 代码块。模板支持 {{variable}} 变量、同名变量联动编辑、键盘切换以
  及可选的连接信息。
• Agent 会话历史持久化 —— Agent 对话现在可在应用重启后恢复。历史以设备分片的 JSONL 文件保存在 Vault 中；来自其他设备的会话保持只读，继续对话时会创建
  新的本机会话分支。
• Agent 执行计划 —— 进行多步骤分析时，Agent 可以维护并展示结构化执行计划，让当前进度和剩余工作更加清晰。
• 实时数据库 Schema 权威 —— Agent 的表搜索和结构读取现在以实时连接为准，不再依赖可能过期的 schema 快照。连接器还可通过新的可选 describeTables 能力
  提供字段描述和注释。
• 更高效的 Vault 文件监听 —— Stela 改用原生递归文件监听，降低大型 Vault 的文件系统资源占用，并改善应用退出速度。

改进

• Agent 默认输出更加简洁，同时保留验证结果所需的关键依据。
• Agent 面板可以更清楚地展示工具调用和执行计划进度。
• 自动 Skill 维护现在必须基于实际工具证据，仅创建新 Skill，并限制为简洁、可复用的知识。
• Skill 提炼可以参考相关 SQL 使用记录；历史说明发生冲突时，优先采用最近更新的 Vault 文档。
• SQL 模板保存在 .stela/sql-templates/，便于 Git 同步，同时不会显示为普通笔记或干扰 SQL 用法检索。
• 新增打开模板选择器及切换模板变量的键盘快捷键。
• 改善 Windows CI 兼容性，并扩充 Agent 历史、执行计划、Schema 解析、Skills、Vault 监听及 SQL 模板的自动化测试。

升级说明

本次升级无需手动迁移。

升级后产生的 Agent 历史会保存在 .stela/agent-history/<设备>/，并随 Vault 通过 Git 同步。每台设备默认保留最近更新的 20 个 Agent 会话。

SQL 模板保存在 .stela/sql-templates/，首次添加模板时会创建该目录。

Review 结论：发布范围里没有需要声明的破坏性数据迁移。需要特别告知用户的行为变化主要是 Agent 历史会进入 Vault/Git，以及每台设备保留最近 20 个会话。

ADRs：none。
Docs：新增双语 release note 模板。
验证：git diff --check 通过；当前干净分支没有 npm run checks 脚本。

0.10.2

Stela 0.10 — Release Notes

Versions: v0.10.0 → v0.10.2
Theme: A sharper Data Agent — multi-provider AI, SQL inline completion, better retrieval, and vault-scoped Experience Knowledge.

---

Highlights

Multi-provider AI profiles

Configure multiple AI providers in one vault (OpenAI, DeepSeek, MiniMax, custom OpenAI-compatible endpoints, and more). Switch the active profile from Settings or the Agent panel. API keys stay encrypted per profile via safeStorage.

SQL inline completion in runsql

Get streamed ghost-text suggestions while writing SQL in notes. Completion is conservative (layout-stable, less noisy) and uses real schema/column context from the editor cache plus nearby note context — not just local schema docs.

Faster, clearer Agent turns

• Read-oriented Agent tools can run in parallel; only note-edit proposals stay sequential.
• Ranked lexical retrieval improves vault/schema/history search without embeddings or RAG.
• When the Agent is unsure, it can ask a clarification question as a first-class proposal (approve / answer / reject), instead of silently guessing.
• Assistant replies show clearer thinking / proposal diffs in the UI.

Experience Knowledge (Skills)

The Agent can now retain reusable data knowledge in your vault as Markdown Skills under .stela/skills/:

• Metric definitions, triage runbooks, domain gotchas
• Explicit save during chat, plus automatic post-run maintenance when appropriate
• Browse, review, and delete entries from the Experience Knowledge dialog

Skills are vault-scoped Markdown instructions (Git-friendly), not executable plugins.

Polish & packaging

• Quit checkpoint feedback while the final Git commit runs on exit
• Safer “reveal in Finder/Explorer” for exported files outside the vault
• Product Hunt assets, refreshed README / landing visuals, and a richer demo vault with sample Skills and weekly-release notes

---

What's new by version

v0.10.2

• Product Hunt launch assets and gallery
• README / docs site refresh
• Demo vault updates: sample Skills, richer MySQL/Postgres seed data, and walkthrough notes

v0.10.1

• Vault-scoped Agent Skills / Experience Knowledge
• load_skill / search_skills / save_skill Agent tools
• Experience Knowledge dialog (list, inspect, delete)
• Explicit + automatic Skill maintenance

v0.10.0

• Multi-provider AI profiles
• Streamed SQL inline completion
• Parallel Agent tools (except propose_edit)
• Ranked lexical retrieval for Agent tools
• Agent clarification questions
• Quit-checkpoint feedback and export reveal token
• Agent panel / proposal UX improvements

---

Upgrade notes

1. Update to the latest 0.10.x build from GitHub Releases.
2. Open Settings → AI and migrate to profiles if you still have a single endpoint configuration.
3. Optional: open the demo vault or create Skills under .stela/skills/<name>/SKILL.md to seed Experience Knowledge.
4. Inline completion and the Agent use your active AI profile — ensure a model/API key is configured.

---

English summary (short)

Stela 0.10 turns the Data Agent into a daily driver: switch AI providers freely, get schema-aware SQL autocomplete in notes, retrieve the right context faster, ask when ambiguous, and keep reusable analysis knowledge in the vault as Experience Knowledge — still local-first, Markdown-native, and Git-friendly.

---

Stela 0.10 — 更新说明

版本： v0.10.0 → v0.10.2
主题： 更强的 Data Agent — 多模型配置、SQL 行内补全、更好的检索，以及 Vault 内的「经验知识」。

---

重点更新

多 Provider AI 配置

在同一个 Vault 里配置多个 AI Provider（OpenAI、DeepSeek、MiniMax、自定义 OpenAI 兼容接口等），可在设置或 Agent 面板切换。各 Profile 的 API Key 仍经 safeStorage 加密保存。

runsql 中的 SQL 行内补全

在笔记里写 SQL 时可获得流式 ghost text 建议。补全策略偏保守（布局更稳、干扰更少），并会结合编辑器侧真实列缓存与笔记上下文，而不仅依赖本地 schema 文档。

更快、更清楚的 Agent 回合

• 只读类工具可并行执行；仅笔记编辑提案保持串行
• 排序后的词法检索增强 Vault / Schema / 历史搜索（无 embedding、无 RAG）
• 不确定时，Agent 可发起澄清类提案，而不是默默猜一个答案
• 助手回复、思考过程与提案 diff 的展示更清晰

经验知识（Experience Knowledge / Skills）

Agent 可以把可复用的数据分析知识沉淀到 Vault 的 .stela/skills/：

• 指标定义、排查手册、领域坑点等
• 对话中可显式保存；合适时也会在回合结束后自动维护
• 通过 经验知识 对话框浏览、查看、删除

Skills 是 Vault 内的 Markdown 说明（便于 Git），不是可执行插件。

体验与发布物料

• 退出时最终 Git checkpoint 的进度反馈
• Vault 外导出文件更安全的「在访达 / 资源管理器中显示」
• Product Hunt 物料、README / 落地页更新，以及更完整的 demo vault（示例 Skills 与周报类笔记）

---

分版本说明

v0.10.2

• Product Hunt 发布物料与 gallery
• README / 文档站刷新
• Demo vault：示例 Skills、更丰富的 MySQL/Postgres 种子数据与导读笔记

v0.10.1

• Vault 级 Agent Skills / 经验知识
• Agent 工具：load_skill / search_skills / save_skill
• 经验知识对话框（列表、查看、删除）
• 显式 + 自动的 Skill 维护

v0.10.0

• 多 Provider AI profiles
• 流式 SQL 行内补全
• Agent 工具并行（propose_edit 除外）
• Agent 词法检索排序
• Agent 澄清提问
• 退出 checkpoint 反馈与导出 reveal token
• Agent 面板 / 提案体验优化

---

升级提示

1. 从 GitHub Releases 安装最新 0.10.x。
2. 打开 设置 → AI；若仍是单 endpoint 配置，迁移到 profiles。
3. 可选：打开 demo vault，或在 .stela/skills/<name>/SKILL.md 下创建 Skills。
4. 行内补全与 Agent 都使用当前激活的 AI profile，请确认已配置模型与 API Key。

---

中文摘要（短）

Stela 0.10 让 Data Agent 更适合日常分析：自由切换模型、在笔记里做 schema 感知的 SQL 补全、更快找到相关上下文、不确定时会先问你，并把可复用的分析经验沉淀为 Vault 内的经验知识 — 依然本地优先、Markdown 原生、对 Git 友好。

0.10.0

Stela v0.10.0

English

Highlights

• Multi-provider AI profiles — Configure and switch between multiple AI providers and models, including built-in providers and custom OpenAI-compatible endpoints. API keys remain device-local and protected by the system credential store.
• SQL inline completion — Get lightweight, streaming ghost-text suggestions directly in RunSQL blocks. Suggestions use the current SQL scope, available table schemas, and nearby note context.
• Smarter Agent retrieval — The Agent now ranks results across the whole vault, supports CJK/Chinese keyword matching, reads column comments, and can search existing SQL usage.
• Agent clarification questions — When the data is ambiguous, the Agent can pause and ask a structured question instead of silently guessing.
• Parallel Agent reads — Read-only Agent tools can run in parallel for faster research, while note edits and mutation SQL remain sequential and require approval.
• Improved exports and quit feedback — Exported files can be revealed in the system file manager, and Stela now shows progress while completing its final Git checkpoint on exit.

Improvements

• Agent responses now render thinking and proposal diffs more clearly.
• Agent-executed SQL is included in the local run history, marked with an agent: block identifier.
• SQL completion can use cached live columns even when no local schema document exists.
• Improved Agent sidebar collapse behavior and cross-platform scrollbar layout stability.
• Added retrieval, inline-completion, and Agent clarification evaluation scripts.

Upgrade Notes

Existing single-provider AI settings are migrated automatically into a default profile. Your stored API key remains local to the device.

---

中文

重点更新

• 多 AI Provider 配置 —— 支持配置并切换多个 AI Provider 与模型，包括内置 Provider 和自定义 OpenAI 兼容端点。API Key 仍仅保存在当前设备，并受系统凭据存储保护。
• SQL 行内补全 —— 在 RunSQL 代码块中提供流式幽灵文本补全建议，并结合当前 SQL 范围、可用表结构和附近笔记上下文生成结果。
• 更智能的 Agent 检索 —— Agent 现在会在整个 Vault 中进行排序检索，支持中文/CJK 关键词匹配、读取字段注释，并可搜索已有 SQL 用法。
• Agent 澄清提问 —— 当数据含义存在歧义时，Agent 可以暂停并以结构化问题向你确认，而不是自行猜测。
• Agent 并行读取 —— 只读工具可并行执行以加快检索；笔记修改和写入型 SQL 仍会顺序执行，并必须经过用户确认。
• 导出与退出体验优化 —— 导出后可直接在系统文件管理器中定位文件；退出时会显示最终 Git checkpoint 的进行状态。

改进

• 优化 Agent 的思考过程与提案 Diff 展示。
• Agent 执行的 SQL 会写入本地运行历史，并以 agent: block 标识区分。
• 即使没有本地 schema 文档，SQL 补全也可使用已缓存的真实字段信息。
• 改善 Agent 侧边栏折叠行为及跨平台滚动条布局稳定性。
• 新增检索、行内补全和 Agent 澄清提问的评估脚本。

升级说明

旧版的单 Provider AI 配置会自动迁移为默认 Profile；已保存的 API Key 会继续保持仅本机可用。

0.9.5

Stela v0.9.5

中文

Highlights

• pi Agent Harness：Agent 接入 @earendil-works/pi-ai / AgentHarness 新运行时（ADR-0018），工具审批与编辑流程更稳
• 导航前进/后退：标题栏增加后退/前进，工作区历史可回溯
• Windows 原生感：title bar overlay，窗口按钮色随主题同步
• 自动更新：macOS / Windows 更新链路与状态指示更清晰；CI / release workflow 统一

Improvements

• 国际化：更多对话框、表格、面板文案接入 i18n（中/英）
• AutoGit：自动 commit 防抖由 10s 调整为 2 分钟；退出前再做一次 commit-only flush，减少长防抖窗口丢 checkpoint
• Agent 输入体验：@ / [[ 候选弹层贴窗口边界；[[ 文档 pill 淡黄色并显示闭合 ]]
• Agent 写笔记约束：明确要求可执行 SQL 用 runsql，勿擅自改写 <detail>

Fixes

• 更新状态指示相关问题
• Windows CI / release 构建若干修复

Notes

• 退出时的 AutoGit flush 只 commit、不 push，避免卡在凭据提示
• 公开仓库 release gate 相关约定见 ADR-0019

---

English

Highlights

• pi Agent Harness: Agent runtime now uses @earendil-works/pi-ai / AgentHarness (ADR-0018), with more reliable tool approval and edit proposals
• Back / Forward navigation: Title-bar history navigation for workspace browsing
• Windows polish: Native title bar overlay; window control colors follow the active theme
• Auto-update: Clearer update status for macOS / Windows; unified CI and release workflows

Improvements

• i18n: More dialogs, tables, and panels localized (zh / en)
• AutoGit: Auto-commit debounce increased from 10s to 2 minutes; on quit, a commit-only flush runs so long debounce windows don’t drop the last checkpoint
• Agent input UX: @ / [[ suggestion popovers stay within the window; [[ note pills use a light-yellow style and show the closing ]]
• Agent note-writing rules: Executable SQL must use runsql fences; do not invent or rewrite <detail> blocks

Fixes

• Update status indicator issues
• Several Windows CI / release build fixes

Notes

• Quit-time AutoGit flush is commit-only (no push), to avoid hanging on credential prompts
• Public-release gate conventions: see ADR-0019

0.9.4

Highlights

• App-wide DockBar — Git, Vault, Settings, and sidebar toggle live in a persistent bottom bar (also on Welcome; survives sidebar collapse).
• Vault name is clickable to switch vaults; Git shows the branch name.
• With a file open, the dock shows a static Backlinks (N) count (expandable panel removed).
• Sidebar chrome cleaned up; Connections / Theme shortcuts live in Settings.

Improvements

• Copy button on AI/Agent markdown code blocks.
• Agent attachment chips moved above the prompt input.
• Editor path bar height aligned with sidebar / Agent chrome.
• Removed Mod+B sidebar toggle (use the DockBar control).

Fixes

• Vault watcher coalesces duplicate fsevents in the same batch.
• Welcome page no longer puts -webkit-app-region: drag on a scrollable container (fixes unclickable UI after scrolling to the bottom).

v0.9.1

0.9.1