Main no longer opens desktop business SQLite. A single utility-process
storage owner holds the Assistant, Knowledge, attachment, document-result,
remote-binding and native model-ledger databases and runs whole domain
transactions; Main and Runtime callers use explicit async domain operations.
- Remove the Main readWithFallback path and the supervision worker's
separate business write connection; manifests are scanned read-only and
committed by the sole writer.
- Remote events are ACKed only after commit; process loss reports
unconfirmed writes instead of replaying them.
- Use native utilityProcess messaging with one in-flight bound (32) and
ordered waiting instead of capacity rejection and retries.
- Start the storage process before the window so startup overlaps.
- Keep short tool output in memory and store full output once in owned
storage; paging, branching, restart and deletion reuse that copy.
- Put required runtime launch files under goodbuddy-prefixed owned temp
directories and reclaim confirmed stale ones; stop writing duplicate
parsed.md.
- Add lint guards for SQLite ownership and architecture, concurrency
acceptance fixtures, and maintenance/architecture documentation.
Validation: typecheck and lint pass. Alternating baseline (edf019aa) vs
candidate runs show no regression on single-session, C01, C02, C04 and a
5-minute C06; Main p99 drops from ~21 ms to ~16.7 ms under concurrency.
Fresh portable copy (1.75M events) upgrade content unchanged; real-model
output paging passes with 4 external requests. Full npm test runs had
1-9 different timing-sensitive failures per run that pass when rerun.
Not covered: C03 review publication, C07 source body, app-level C05.
Main module evaluation drops from ~420 ms to ~330 ms. Types stay as import type; the DeepSeek Harness chunks stay static because the protocol chunk shares code with Main startup paths.
Add a System tasks tab to run history that splits background usage (supervisor, auto-supervision, magic notes, knowledge base, delegation) from conversation usage, with a manual refresh button and automatic refresh when usage is recorded. Record previously dropped usage from supervisor reviews, image regeneration, knowledge graph extraction, and remote embeddings.
Add composer shortcuts for local Continue and OpenCode terminals and the official DeepSeek Harness Web client, using the current project and model bridge. Manage browser service reuse, stop actions and terminal cleanup. Upgrade DeepSeek Harness to 0.1.7-rc.2 and Cordis to 4.0.4.
Include the coordinated IPC handler extraction and related tests and documentation. Remote native-client coordinator wiring, managed Node consolidation, Ask MCP mappings and full desktop/package acceptance remain unfinished; this commit records the implemented local workflow and supporting SSH groundwork.
Validation: typecheck and lint passed; 267 focused tests passed and 5 skipped. No new real-provider calls were made for this commit. Prior real-model evidence and remaining acceptance work are documented in the feature progress files.
Release note: 本地项目选择 Continue 或 OpenCode 后，可从输入框下方打开原生终端；选择 DS 后，可打开使用当前项目和模型的官方 Web 客户端。客户端使用独立会话，DS 网页关闭后服务继续运行，可在 GoodBuddy 中停止；远程入口尚未完成。
Keep interactive questions and answered cards at their original position in
conversation history, including remote reception, recovery, and reopening.
Preserve legacy history placement without a database migration.
Move supplementary guidance into shared accessible contextual-help popovers
across pages and settings. Add keyboard, modal, accessibility, and Electron
regressions, plus bounded opt-in live canvas-analysis validation.
Remove unused Agent, Runtime, and in-memory graph-search helpers, obsolete
mocks and browser copy; reuse the existing package-size formatter. Keep
production execution paths, protocol schemas, and released data migrations.
Validation: targeted cleanup tests, typecheck, lint, shared Linux Host
attach/read/re-attach, and added help/Electron checks passed. Final full-suite
verification passed 4,858 tests with 67 skipped and zero failures (412 files
passed, 9 skipped). Focused help checks passed 54 tests, including Electron
style comparisons and Chromium accessibility descriptions. Details and
validation boundaries are recorded in feature docs.
Release note: 修复交互式问答在回答后排到后续输出末尾的问题，重新打开会话仍保留提问位置。页面和设置中的补充说明改为标题或字段旁的按需帮助，支持鼠标与键盘操作。
Graph extraction could exhaust its fixed 8192-token budget before returning
structured content. OpenAI requests now omit output-token limits, while
Anthropic uses the model connection limit required by its protocol.
Provider limits, timeout, cancellation and bounded response handling remain.
Use the shared accessible icon button for image downloads in chat and the
image viewer. Add graph request regression coverage and document the limits.
Release note: 知识图谱抽取不再额外限制为 8192 tokens；OpenAI 由模型服务决定输出上限，Anthropic 沿用模型连接配置。聊天和图片预览中的下载按钮统一为带提示的图标按钮。
Add HTTP PaddleOCR-VL settings and shared document previews for chat and knowledge imports. Preserve attachment originals and parsed assets across drafts, messages and queues; support explicit image OCR, selecting document images into a conversation draft, and restoring queued input for editing.
Protect unsaved note bodies and titles during navigation and asynchronous updates. Match canvas analysis to source revisions, invalidate affected comments, preserve annotations during pagination, and provide a resizable record index with saved width preferences.
Existing OCR settings retain local behavior. Historical attachments without saved originals are not reconstructed. Advanced OCR options depend on the configured service; image sending depends on the effective model and Runtime.
Validation: full typecheck and lint passed; 477 focused tests passed across 12 files; staged whitespace checks passed. The full repository suite was not rerun for this commit.
Release note: Keep note drafts safe while switching workspaces, resize the record index, and revisit saved document attachments with OCR previews and selected images in chat.
Image-only PPTX attachments failed with no indexable text even when a local
OCR model was configured. PPTX parsing now recognizes embedded PNG, JPEG
and WebP images through the existing local OCR path, preserves native text
and slide locators, and honors the selected parsing workflow.
OCR workers previously retained model sessions and WASM memory indefinitely
after successful parsing. Release the worker after 60 seconds with no active
or queued requests, reuse it for continuous work, and reload on demand.
Fast text/index modes still skip OCR. This does not render Office layouts,
recognize inherited master images, or change Agent execution permissions.
Validated four real PPTX attachments (40 pages), idle memory release and a
nine-page reimport after model reload on Windows x64 with PP-OCRv6 Tiny.
49 local OCR requests and zero text-model calls. All 4293 tests passed
with 67 skipped; typecheck, lint and the isolated Electron build passed.
Release note: 修复已配置本地 OCR 时图片型 PPTX 仍无法导入的问题，识别结果保留幻灯片定位。OCR 空闲 60 秒后自动释放模型运行内存，连续解析复用模型，下次使用时自动重新加载。
Binding checks loaded and parsed every external knowledge binding to answer a
single-row question. requireLibrary did this on more than thirty call sites
including one inside a task reconciliation loop, and the knowledge snapshot
scanned the table twice per library across its seven callers.
Add keyed store lookups (getBinding, hasBinding, getBindingsForInstance) that
use the existing primary key, build the snapshot's bindings once into a Map,
and count bindings per instance in a single pass.
Memoize the conversation activity index in the renderer. It was rebuilt on
every render, including once per streaming token, which discarded half the
benefit of the memoized activity summary directly above it.
Release note: 知识库页面、对话预检索和跨项目活动列表在绑定较多时响应更快，流式回答期间的
界面开销降低。
Conversation retrieval called the binding verification helper, so every
question sent the user query as a test query and RAGFlow bindings with graph
or knowledge compilation enabled spent an extra detail request before
retrieving. A failed capability check could also surface
EXTERNAL_KB_CAPABILITY_UNAVAILABLE during a normal conversation.
Retrieval and verification are now separate service entry points sharing one
snippet-bounding helper. The capability pre-check stays in the verification
path and keeps using the binding's configured request timeout.
Replace the five configuration snapshot comparisons and the
EXTERNAL_KB_CONFIG_CHANGED code with the existing per-instance cancellation.
A connection test cancelled by a configuration change now rejects without
writing back a probe result, so disabling an instance no longer records it as
a failed probe. Duplicate binding safety relies on a re-check plus the
existing unique index.
Release note: 外部知识库对话检索不再执行仅用于保存校验的能力预检，RAGFlow 图谱绑定每次
提问少一次请求；停用实例时不再把它错误标记为连接失败。
Connect Dify, FastGPT and RAGFlow through read-only bindings, encrypted local credentials, shared retrieval and persistent citations. Configuration and removal affect GoodBuddy only.
Show active conversations across projects, retain older live work, and improve project dialog scrolling. Queue concurrent questions and route managed SSH OpenCode questions through the Agent. Preserve recoverable read/search errors and partial search results.
Document the Electron UI automation standard, real model and external knowledge validation, and remaining acceptance limits.
Validation: earlier full regression passed 4081 tests with 66 skipped; subsequent focused remote-question checks and real Linux evidence are recorded in feature progress. Precommit typecheck and lint passed; latest project dialog and shared UI suites passed 58 tests. The full suite was not rerun for the final dialog changes.
Release note: Use external knowledge in conversations, find active work across projects, and answer concurrent local or managed SSH OpenCode questions. Managed SSH questions require the updated Agent adapter.
Add read-only Dify, FastGPT, and RAGFlow probes, a redacted live capability baseline, and focused adapter tests. Document version-specific API and UI behavior, including Dify 0.x/1.x differences and RAGFlow GraphRAG evidence.
Use real in-memory SQLite for the remote-message ordering and active-task
retention query tests. Both exceeded the CI five-second limit while seeding
502 independently committed records on disk.
Keep the record counts, public database calls, and assertions unchanged.
Disk persistence, transactional migration, and reopen tests remain disk-backed.
Production database behavior is unchanged.
Knowledge selection was shared globally, document readiness overstated usable content, and important import and source-management paths exposed technical details or lacked safe recovery actions. Knowledge scope now persists per conversation, readiness reflects real processing channels, and the workspace guides users from creation and import into cited chat with controlled source actions.
Release note: 知识库现按对话保存使用范围，准确区分可用、处理中和失败文档，并提供更简洁的创建、检索测试、来源打开、删除确认和失败重试流程。
The complete-snapshot boundary test performs 501 synchronous SQLite source and document writes. It passed functionally but could exceed its 10-second case timeout on the serial Linux CI runner, so give this I/O-heavy case a bounded 30-second allowance without changing product behavior or global test limits.
Long document imports could send oversized graph prompts, surface empty model
responses without useful diagnostics, and fail the entire new document even
when its full-text index was usable. Graph extraction now uses bounded token
batches with retry and timeout handling, preserves successful partial results,
and reports graph failures independently from new-document indexing.
Chinese graph retrieval now recognizes bounded CJK query terms, while graph
search keeps one-hop relationship context visible around matched entities.
Release note: 修复长文档知识图谱抽取易因空响应或超时导致整篇文档导入失败的问题；现在图谱按 Token 分批重试，失败时仍保留新文档全文检索，并改进中文图谱问答和关系上下文展示。
Add managed SSH Hosts and projects with signed side-by-side GoodBuddy Agent and OpenCode Runtime installation, Ask and Execute workspace access, a Main-owned model bridge, detached Agent lifecycle, durable ACP output recovery, and project activation UI.
Use protocol v2 with bounded FIFO frames over SSH instead of custom credit, priority, batching, or model-message chunking. Mismatched Host Agents now upgrade automatically, an exactly verified Node runtime is reused on-host, and local Portable builds can embed strictly verified Agent artifacts.
Release note: 新增托管 SSH 远程项目，可自动安装或升级签名 Agent/OpenCode Runtime，并在 SSH 断线后继续 Agent 工作和同步未确认输出。
GoodBuddy previously required an external OpenAI-compatible endpoint for vector retrieval. It can now install, verify, import, remove, and run the built-in Granite multilingual embedding model in an isolated local inference process, while retaining custom embedding connections and migrating existing settings.
The model connection UI now manages built-in and custom vector connections consistently, reports the actual retrieval path, preserves credentials across failed saves, and renders the vector tab synchronously. Related settings copy, switch styling, metadata typography, and renderer bundle guardrails are aligned with the resulting workflow.
Release note: 新增可管理的内置 Granite 多语言向量模型，支持本地下载、离线导入、连接切换与诊断；同时优化向量设置及相关界面的文案、开关和状态显示。
mesalogo