goodbuddy/FEATURES.md

GoodBuddy Feature Matrix and Roadmap

English | 简体中文

This document records the capabilities available in GoodBuddy and its roadmap. Unchecked items are not included in the current release unless the item says otherwise.

Status

• Completed
• To implement

Status describes implementation, not publication or test coverage. Validation records are listed separately and do not introduce another feature status.

Feature Matrix

Desktop foundation, workspaces, and context

• Cross-platform desktop application: Supports Windows, macOS, and Linux release targets on x64 and arm64.
• Chinese hardware and operating-system coverage: Standard Linux packages target compatible x64 and arm64 environments, including UOS, Kylin, Hygon, Zhaoxin, Kunpeng, and Phytium systems. This is not vendor certification. LoongArch has a separate experimental loong64 preview outside standard releases and automatic updates; it is built only on explicit request. See the preview boundaries.
• Configurable global shortcut: Enable, disable, or record an Electron accelerator under Platform Features / General. The default remains CommandOrControl+Shift+Space; if registration conflicts or saving fails, GoodBuddy keeps the previously registered shortcuts and shows an actionable status.
• Projects, independent conversations, and conversation branches: Isolates context by project and manages conversations, attachments, and Git workspace changes. A stable local conversation can be copied into an independent branch; the branch retains a source badge but does not copy Tasks, queues, or artifact ownership. The project selector distinguishes local, managed SSH, and remote messaging-channel projects. Managed SSH projects are grouped by Host, with the real Agent connection state on the Host heading and only the remote path on each project row.
• Cross-project conversation activity: A shared activity entry lists running conversations and conversations needing attention across projects, including older conversations with live work. Anchored project-to-conversation menus provide keyboard navigation and exact conversation opening without a full activity dialog. The summary hides when idle, long conversation submenus scroll independently, and project selection continues to show activity counts. Successful background conversations retain a completion notice during the current client session until their chat is visible and loaded, not across restarts.
• Pinned conversations and complete saved history: Pin common conversations and retain their position after restart. Lists and search no longer stop at 100 conversations, and consecutive sends no longer truncate saved history beyond 500 messages. Model context limits still apply; previously lost messages cannot be restored.
• On-demand contextual help: Supplementary page and settings explanations use shared title- or field-adjacent help controls with mouse and keyboard access. Important operation consequences and errors remain visible.
• Compact conversation controls: Composer options share a compact settings panel while Runtime, mode, sending, and queue controls remain available. Conversation search has an inline clear action that restores the scoped list and returns focus to the input.
• On-demand conversation history: Lists load lightweight summaries rather than every conversation's process metadata. Opened and active conversations retain full details; search, copy, export, continued requests, and remote pending questions remain available. Unchanged image reconciliation does not resend saved messages, and idle acknowledged details follow the existing view-cache lifetime. See the history loading and retention rules.
• File, screenshot, window, and clipboard context: Added to model context only after explicit user selection. Paste one or more supported local files into the chat composer with Ctrl+V to add attachments, using the same parsing progress and limits as the attachment button. Text and screenshot paste remain available. See attachment rules.
• Text recognition in image-based PPTX files: With local OCR configured and an OCR-enabled parsing workflow, embedded PNG, JPEG, and WebP images are recognized alongside native text, retaining slide locations. OCR releases model memory after 60 seconds without active or queued work and reloads on demand; fast text/index modes still skip OCR. See the parsing behavior and real-file evidence.
• HTTP OCR and persistent attachment previews: Configure HTTP PaddleOCR-VL under Settings > Document Parsing, shared by chat documents and knowledge imports. Saves originals and extracted images and provides text/image/detail previews, explicit image-to-text actions, and selecting document images into a named conversation draft before sending. Ordinary chat images use image input by default, subject to the effective model and Runtime. This adds no Application Center entry; Local Inference Monitor manages local services only. Real validation covers Windows, direct visual input, and managed OpenCode on the shared Linux Host. Deployment-specific advanced OCR models and other Runtime combinations need separate verification. See scope and evidence.
• Per-file workspace diffs: Changed files expose separate staged and unstaged diffs, including deleted, renamed, and untracked files. Refresh reloads the selected diff, browsed directory, and expanded directories; more than 50 changes remain accessible through incremental loading.
• Workspace file and Git management: Create, rename, move, and delete files or directories, inspect properties, search and switch local or remote branches, create branches, explicitly Fetch, and browse commit history and per-file commit diffs. File and Git controls are grouped by view, branch search uses the shared compact form style, and subdirectories retain text breadcrumbs without an isolated root icon. Commit history stays docked at the bottom, initially expands to half the Git content area, and supports resizing with independently scrolling panes and one shared refresh action. Project switches automatically detect Git without reusing another project's status. Moves stay inside the workspace, deletion requires confirmation, and branch conflicts never trigger automatic stash or discard. Managed SSH operations require Agent 0.11.23.
• Rich responses: Supports GitHub Flavored Markdown, LaTeX math, constrained Mermaid diagrams, and static in-conversation HTML previews. Mermaid's large-diagram viewer keeps all edges reachable when zoomed and exports the complete diagram as PNG, independently of zoom and pan. Complete HTML and HTML code blocks can be previewed in place after an Agent reply finishes, with icon actions to reveal the source or open a full-screen preview. The feature is on by default and can be disabled under Platform Features / General; previews cannot run scripts, access the network, submit forms, or open windows.
• AI response and full-conversation copy: Completed AI responses expose a bottom action that copies Markdown source without reasoning, tool logs, or citation metadata. Full-conversation copy uses the same validated Preload/Main clipboard path.
• Assistant workbar, multiple terminals, and resizable layouts: The toggle sits beside the theme switch. The right workbar uses a persistent “+” capability catalog and application tabs. Task Center and Workspace are permanent singleton tabs; browser tabs are independent instances. Task Center can show the current project or all projects, and retains the selected scope with the layout. Tasks without a project appear under all projects with an “Unbound project” label. Users can open multiple independent terminals for the current local or managed SSH project with bounded output, resizing, termination, and explicit reconnection. Closing a terminal tab ends its Shell; restarting the app restores only an ended tab description and never restarts the Shell. The main sidebar and Magic Notes editor's AI pane support pointer and keyboard resizing. The AI pane remembers its visibility and width, and stacks below the editor at narrow widths.
• Project Agent Space (planned): Unifies roles, knowledge, Skills/MCP, models, approval policy, budgets, and timeouts in a Project, with reusable templates.
• Application center and navigation: Clicking the bottom App Center opens a lightweight upward anchored popup with enabled apps, regardless of pinning or opening history, and no modal backdrop. Clicking an app row closes the popup; Local Inference Monitor opens a separate modal preserving the workspace, while other apps open their main content pages. Manage Apps opens searchable cards with settings details; layout and interaction rules are defined in the UI design. All four application cards support arrow and drag reordering regardless of enablement or pinning. Cards, the launcher, and the filtered sidebar share one persisted order. Knowledge is always enabled and pinned, with Open and reorder controls. Supervisor, Notes, and Local Inference Monitor have enablement and pinning settings. Supervisor uses the existing Heartbeat identity and defaults to off when no preference is saved; explicit saved choices are preserved. Disabling it retains plans and history, blocks new reviews, and hides sidebar feedback; in-flight work may finish. Enabling the app does not create automatic plans. Older application orders retain their relative order when missing defaults are added. Durable settings saves, including configuration-tool writes, synchronize through change events; reopening the center locks edits until refresh completes, and newer snapshots supersede stale reads. The local-inference modal shows a single service list without task history or external connections. Embedding shows its independent process CPU and working-set memory; ASR/OCR explain shared-process attribution limits. ASR readiness is unknown rather than inferred from active requests. Supported service controls retain impact confirmation; failed operations refresh the list and require a new confirmation. TTS remains unavailable. Cancelling inference remains active in service-stop impact until worker acknowledgement or exit. Implementation is connected to production App; acceptance records distinguish full-App checks, component fixtures, real local-engine checks, and remaining platform/package coverage. Historical failures are not the current full-suite result. See the validation progress.
• Privately deployable application marketplace (planned): Extends application management with an organization-owned catalog and application distribution. Current management covers built-in apps only, not marketplace browsing or installation. See FR-15.
• Additional assistant workbar and execution-space capabilities (planned): Builds on the current workbar and multiple terminals with broader execution supervision, unified Runtime monitoring, managed processes, target-pinnable workspace/browser/artifact instances, bottom docking, and separate windows. Task Center remains the singleton Task index; attachments and knowledge remain in the conversation composer, while memory, when implemented, and historical execution context belong to the associated Task. See the Feature PRD.
• ShareServer Office co-editing (design): Uses an optional ShareServer integration with ONLYOFFICE Docs to open multiple filename-labelled document tabs in the assistant workbar. The planned scope covers DOCX, XLSX, and PPTX manual editing, save and undo, source-file conflict protection, and later AI selection edits. The selected ShareServer and editor process document content; Desktop neither bundles nor starts Document Server and does not claim offline editing when no service is configured. See the design.

Agent Runtimes and model connections

• Compact subagent history without lost details: Local and managed SSH subagent events store changes instead of repeated full progress snapshots. The first upgrade automatically converts existing history and reclaims disk space, with visible progress and quit/retry support. Chats, execution details, results, and remote event deduplication are preserved; older clients cannot reopen the upgraded database. The schema-35 repair released in 0.13.2 also handles repeated tool blocks written after upgrading to 0.13.1, without deleting events. Terminal tasks release their write caches; duplicate remote replay does not rebuild a released cache. See storage and upgrade behavior.

• Direct model Runtime: Supports question answering, knowledge synthesis, controlled tool execution, image generation, and reference-image editing through providers supporting the OpenAI-compatible image-editing endpoint. Editing sends attached images as multipart data and still requires validated inline image output rather than fetching provider-returned URLs. Follow-up image requests reuse the latest successful image and conversation text, including after reopening a saved conversation; explicit attachments take priority. Missing editing support does not block text-based generation: the affected reply shows a quiet footer explaining that reference images were not used. See image generation.

• Image tools in the current conversation: Enable conversation access for an image model to let tool-capable chat models generate images and edit uploads or earlier artifacts in Execute, without a separate conversation or manually switching to an image model. Direct text models, local OpenCode, Continue, DeepSeek Harness, and managed remote OpenCode use the shared service; Runtime assignment is derived from model settings. Requests may incur provider charges and editing requires provider support. See the implementation and validation records.

• Cross-project Runtime process reuse: Compatible local OpenCode and DeepSeek Harness configurations share heavyweight processes; managed OpenCode sessions share a process on the same Host. Project adapters and session workspaces, models, tools, questions, and cancellation stay independently routed. Remote pending questions and task status recover after reconnect or Desktop restart without resending accepted prompts. Requires Desktop 0.13.5 and Agent 0.13.0 together for the new remote behavior. See the design and measured validation.

• Direct model programming agent: Local direct text models can run the platform Shell in Execute mode and delegate one level of programming Subagent work while inheriting the parent request's mode, model, workspace, and capability scope. OpenCode, Continue, DeepSeek Harness, and managed SSH do not receive duplicate copies of these tools. The Windows local command path and a real-model edit, test, fix, and review loop have passed; native macOS and Linux command validation remains. Execute commands can select absolute, relative, or symbolic-link directories outside the workspace; the workspace stays the default and relative-path base. Ask remains read-only.

• OpenCode and Continue: Use isolated child processes, an environment variable allowlist, unified configuration, cancellation, startup and control-request deadlines, bounded streaming output, and activity records. Shared process cleanup preserves complete Windows process-tree termination and terminates POSIX process groups when a child uses an independent process group. Chat status checks use a one-shot Runtime probe that is immediately cleaned up rather than entering the model-and-project execution cache. For managed local OpenCode, this passive check validates the selected path and model credential without starting a throwaway server; the explicit settings test, native inventory still performs full startup and health checks, then shares that execution Runtime with the first real request instead of cold-starting another server. Execution Runtime adapters remain reusable per project, while local OpenCode configuration dependencies and content-addressed Skill snapshots are shared globally across projects. Session data, tool output, and request-scoped MCP state remain isolated. GoodBuddy-managed local OpenCode keeps requests in the same conversation ordered, while different conversations in the same project and different projects can run in parallel. Request-scoped dynamic MCP tools remain isolated through default wildcard disablement and explicit current-request enablement. Managed local and remote OpenCode skip unused automatic Git snapshots to avoid synchronous diff stalls; file tools, subagents, and workspace Git diffs remain available. Continue continuously drains unused utility-host stdout so console output cannot fill its pipe. These changes apply to newly started Runtimes, not external OpenCode Servers. OpenCode and Continue runs do not stop at a fixed tool-call or activity count, and every observed tool and Subagent activity remains in the local conversation instead of being discarded while the native task continues. Interactive questions are answered only by foreground conversations; scheduled tasks, remote channels, delegated work, and other background runs fail immediately with guidance to rerun in the foreground instead of waiting indefinitely. Local OpenCode startup uses the selected absolute path directly without checking the file or running --version first; invalid paths fail at actual launch. On managed Linux ARM Hosts, Runtime activation reuses the registry and manifest verified during setup instead of rehashing or rechecking the complete OpenCode binary.

• Managed SSH OpenCode and Continue: Controlled by the separate Remote Projects tab under Settings / Platform Features and disabled by default. Disabling it does not affect local projects, ordinary desktop capabilities, or desktop releases. When enabled, users can manage SSH Hosts with pinned Host Keys, browse bounded remote directories, and create Ask or Execute projects. Both modes start the signed Runtime directly without bubblewrap. Ask applies OpenCode's Ask permission configuration and permits only native read approvals at the Agent tool-dispatch boundary, while Execute explicitly allows native permissions and retains all permissions of the selected SSH account. The Agent owns accepted Prompts, provider/tool rounds, Runtime processes, a stable model ledger, and a bounded semantic transcript over a private Unix socket and ACP v5. Work continues on the Host after Desktop exit, network loss, or local-process termination. The current Agent may remain resident; a superseded Agent drains active work before exiting. Idle Runtime processes are reclaimed after tasks finish, while shared processes retain other active tasks, including pending questions. Subsequent turns recreate the Runtime and restore conversation history. Reconnection attaches only to the original controller, binding, and operation, then atomically merges provenance into the original conversation without resending the Prompt. Model profiles and API keys enter Agent memory only for the accepted operation and are not written to the Renderer, SSH arguments, remote environment, or disk. Remote components are not embedded in the desktop package; switching projects updates only local selection and does not connect or download. Settings reads a small signed catalog and reports the local version, latest compatible online version, and update state for Linux x64, Linux arm64, and macOS arm64. Intel Macs are not supported. Downloads occur only after a user action, and packages can also be imported or exported offline. Each compound .gbagent package contains the Agent, fixed Node, and compatible OpenCode and Continue Runtimes maintained by the desktop source. Online sources follow the GitHub/Beijing OSS choice in About and Updates. The cumulative signed catalog binds minimum Desktop version, Agent protocol, platform, architecture, size, SHA-256, and fixed URL. All three targets share one catalog. Users must upgrade Desktop before the first mixed-platform catalog is published because older Linux-only readers reject catalogs containing Darwin entries. The public-key registry accepts equivalent JSON whitespace and line endings while still strictly validating schema, Ed25519 keys, environment, and revocation; catalog, package, manifest, payload signatures, and streaming SHA-256 remain unchanged. A missing architecture package disables managed SSH only for that architecture. Projects store only Host, remote path, Runtime selection, and mode. Current Host identity is read when a Workspace/Runtime is first used, then reused by other projects in the same process. Managed SSH conversations expose supported OpenCode and Continue choices; Continue requires the updated compound package and matching Desktop. Multiple projects and conversations can run concurrently on one Host; Agent-owned prompts are not stopped at a fixed model-call count or prompt-wide output-token total, and Runtime output uses transport backpressure instead of cancellation when GoodBuddy buffers fill; recovery separately reports network, Agent, Runtime, committed-event cursor, completion, or failure/retry and blocks only the affected project. A real Linux x64 Host has covered short and long detach, forced local harness exit, concurrency, SSH relay loss, cancellation, definite and uncertain provider failure, Agent SIGKILL/restart, and recovery from a reopened Desktop SQLite database. Successful tool START/END events appear exactly once, with no Prompt, provider, or tool replay observed. The current Agent source lock is 0.13.3, while the current Desktop release candidate is 0.13.12; formal publication status follows the separate Agent and Desktop release channels. Previous macOS validation covered native package installation, detached lifecycle, Attach, real Ask/Execute, and cancellation of tools in separate process groups on a real Host; this does not imply publication. Agent 0.11.23 adds remote workspace management and optional model limits; its bundled Runtime no longer imposes a fixed ten-minute Prompt deadline. Unlimited request duration retains a separate connection timeout. Agent 0.11.23 requires Desktop 0.12.11 or later. Agent 0.11.24 requires Desktop 0.13.0 and fixes native question routing and sending another message after cancelling a pending question.

• Manual SSH Host environment provisioning source path: After Host Key, authentication, and system probes succeed, GoodBuddy saves the Host and read-only probes the shared Agent/Runtime. Saving a Host or opening a project never installs automatically. A Host card has one primary action based on version facts, Install Remote Environment, Update Remote Environment, or Reinstall, plus a secondary SegmentedControl for Auto, Host Download, or GoodBuddy Transfer. The selection defaults to Auto and is not persisted; a Version Matched badge does not mean the environment is healthy. Auto probes only before operation preparation and chooses one acquisition route. Explicit choices remain in effect, and prepare, commit, or adoption failures do not silently fall back across acquisition routes. Both routes deliver the same signed compound .gbagent into fixed staging and share control-plane prepare, commit, Agent activation/health, Runtime activation, finalize, and explicit cleanup. The GoodBuddy route can download and validate a missing candidate during the same operation, cache it with a lease, and stream one archive plus its verified bootstrap Node over bounded SFTP without loading the roughly 294 MiB package into a Main-process Buffer; the Host performs one complete payload verification while extracting. An unfinished operation stores only the operation ID needed to clean staging. A later update makes a best-effort cleanup and starts prepare again, without persisting a remote metadata copy or allowing cleanup failure to block a new update or roll back a healthy environment. Existing projects resolve current Host identity on demand and run a fixed attach-or-bootstrap; registered health, capabilities, and prompt startup do not scan the full payload. See the design.

• SSH Host environment provisioning validation records: The current source passed isolated Linux x64 package installation, Ask/Execute, native subagent external writes, reconnection, and stop/bootstrap. The complete Host-card acquisition matrix across GitHub, Beijing mirror, Linux x64/arm64, cancellation, and offline GoodBuddy transfer remains. Candidate CI/native packaging and system sleep/wake are not yet verified; this partial development evidence does not imply publication.

• DeepSeek Harness (preview): Uses the fixed GoodBuddy Host and an OpenAI-compatible model connection. It prefers an administrator-provided connection, otherwise follows the compatible default model or first compatible connection without requiring a duplicate selection. Settings displays the actual administrator or fallback model source. Ask permits only real Host-registered read and skill tools plus Main-managed Web Search/Fetch proxies, and rejects plugin impersonation of those names. Execute allows all enabled built-in and plugin tools with the current user's permissions. Image input follows the selected model connection's declared capability: text models reject images before Host or model invocation, while image-capable models receive bounded inline JPEG/PNG content through a temporary Attachment Store. On Windows, Host startup and ACP sessions share one canonical workspace path, so equivalent path spellings reuse the same working Runtime instead of failing session creation.

• DSH npm plugin marketplace: Disabled by default and searches public npm dsh-plugin packages only after explicit user enablement. It uses the bundled npm to install exact versions with ordinary lifecycle scripts and supports enable/disable, JSON configuration, removal, automatic disablement after startup failure, and offline management of installed plugins. Disabling the marketplace hides only the catalog and management interface; it does not change the enabled state of installed plugins, and third-party code is not subject to Ask initialization isolation.

• Ask and Execute work modes: Ask remains read-only. Execute is the user's authorization for all tools, processes, network access, and writable paths available to the current local or SSH account, including paths outside the workspace and native subagent work.

• Efficient direct-model workspace tools: Direct models use bundled ripgrep with native arguments for file discovery and content search, read large UTF-8 files by line, and apply multi-file patches in Execute. Ask searches stay read-only within the workspace; Execute searches use current-account permissions. No system ripgrep install is required. Search preserves native output and exit codes, with complete large results available through output_read; actionable argument errors let the model correct its request and continue.

• Long context and paged tool output: Message history, long replies, and parsed document text no longer use the previous fixed truncation limits. Direct models can continue reading stored search, command, and Subagent output in pages; the selected model's context window and bounded transports still apply.

• Native Runtime interaction routing: OpenCode and Continue questions support choices, yes/no, free-text answers, and skipping through the existing question card. Local and managed SSH OpenCode also route questions from owned child sessions. Managed SSH requires Desktop 0.13.0, Agent 0.11.24, and a newly started managed Runtime; this does not extend question support to arbitrary ACP services. Successful answers and skips retain the original questions and answers at their original position across rounds and local conversation reloads; answers discarded by older versions cannot be recovered. Concurrent questions wait in order, duplicate events preserve drafts, and failed submissions remain retryable. Cancelling a pending managed SSH question allows another message in the same conversation. Execute permission confirmations are handled automatically rather than waiting for another approval. See the interaction boundaries.

• Native execution checklists: OpenCode and Continue update a read-only checklist above the conversation, with progress saved in history. Explicit clears and remote replay retain request ownership; cancellation does not mark unfinished items complete. Remote delivery requires the matching Agent package. See the checklist contract.

• Experts and Subagents: Supports explicit experts, team analysis, and up to three experts running in parallel. Experts inherit the parent Ask/Execute mode and can use enabled local direct-model tools; Ask remains read-only. They are not remote OpenCode child sessions. Chat shows each expandable full expert response first and the main Agent's synthesis below, and persists both with the conversation.

• OpenCode child progress and final results: Child cards show ordered text, reasoning, and tool progress separately from the final result. Remote live progress requires Agent 0.11.20; older packages still expose final results without reconstructing missing progress.

• OpenCode event-stream cleanup: Chat requests and native context compaction close their own subscriptions before ending event iteration, including completion, failure, cancellation, and early consumer exit, without cancelling parallel conversations.

• Accurate message-footer status: Distinguishes request preparation, retry waiting, retry dispatch, tool activity, pending answers, and terminal states. Local OpenCode reports native retry attempts and scheduled times; direct models report their own backoff phase. Footer dots are static, and unsupported Runtime retry details are not invented. Conversation-list indicators and send/stop behavior remain unchanged.

• Simplified Runtime selection: Conversation and project menus no longer enumerate every Runtime/model combination. Configure Runtime models in system settings; saved fixed project selections remain usable and explicitly display their fixed model.

• Role-bound model connections: Each role can inherit the default model or select an independent text-model connection. Invalid connections safely fall back to the default; the synthesis role always inherits the default.

• Multiple model protocols: Supports Anthropic Messages, OpenAI Responses, OpenAI Chat Completions, OpenAI Images, and unauthenticated local models. New users receive a local Ollama-compatible default and no preconfigured third-party cloud provider. During upgrades, GoodBuddy replaces a historical default only when it has never received credentials and still exactly matches the old built-in value; explicitly saved connections and encrypted credentials remain unchanged. Existing deployments that supply credentials only through compatible environment variables retain their historical connection parameters, while generic model environment variables still take precedence. Save and Test Model sends a bounded real text or image-generation request and validates the generated result instead of testing HTTP reachability alone, so it may incur a small provider usage charge.

• Custom model request fields: Each LLM connection accepts bounded JSON objects for additional request headers and top-level body fields. Direct requests and Continue support both; local OpenCode and DeepSeek Harness use only their natively supported headers, while the managed SSH OpenCode model gateway supports both. Runtime, protocol, and authentication fields take precedence, and these ordinary connection settings must not be used to store API keys or other secrets. See model request customization.

• Context usage and automatic compaction: Direct models update usage from each successful provider call. Images and tool rounds use the same accounting, with estimation only when the provider omits usage. The UI distinguishes This Model Call from Post-Compaction Conversation Estimate, and the compaction line is recalculated from current settings and the selected model window rather than storing old configuration per conversation. Before/after compaction markers use one estimation method, while activity history retains provider usage for each call. Conversations and multi-round tool Agents can compact repeatedly after a completed call crosses the threshold, reserving budget first for fixed prompts, tool definitions, and summaries. One response separately preserves compaction markers for Agent tool context and conversation history, and summaries remain reusable after app restart or when older messages leave the local history window.

• Persistent credential protection: Main encrypts API keys with secure system storage and never exposes them to the Renderer. A managed SSH accepted Prompt places the current profile and key in Agent memory only, never in SSH arguments, the remote environment, or disk; other paths keep credentials inside Main. Credentials remain attached to their model connection when the service URL changes or authentication is temporarily disabled, and are removed only when the user explicitly clears them or deletes the connection.

• Bounded failure diagnostics: Desktop stores a fixed allowlist of startup, Runtime, and remote-connection failure stages in the user-data directory, rotating at four 256 KiB files. Each GoodBuddy Agent rotates daemon, connection, recovery, and Runtime lifecycle diagnostics at three 64 KiB files in its private installation state and exposes them through the fixed diagnostics --installation-id command. Both sides store only allowlisted stages, stable error codes/types, and fixed short messages, never Prompts, credentials, file contents, paths, environments, SSH arguments, or raw provider responses. Diagnostic-write failure does not alter normal operation.

• OpenCode Runtime customization: GoodBuddy-managed OpenCode can discover native Agents, Tools, Commands, LSPs, Formatters, MCP, Skills, Prompts, and Resources. Tools show read, file-edit, command, network, Agent orchestration, and other types, source, and Ask/Execute availability while hiding OpenCode's internal invalid tool and temporary GoodBuddy MCP tools. Users can save a default Agent, override it per request, run Commands through the native SDK, view context usage, and invoke native Compact with a total timeout. Questions from concurrent external Server conversations are mapped by public request ID so answers cannot cross conversations. External OpenCode Servers report connection state only and do not claim native catalog access. Arbitrary plugin installation, Session Share, automatic Worktree, and OpenCode native-session persistence remain unavailable.

• Continue Runtime customization: Exposes native Rules, prompt templates, and MCP from static configuration plus editable GoodBuddy Rules and Prompt presets. Chat can select a preset per request and insert an editable Prompt. The current Continue Host has no trustworthy static native Tool discovery API and runs with an isolated CONTINUE_GLOBAL_DIR, so the UI explicitly marks static Tool discovery unsupported and does not present workspace or user Skills that the Host will not load as native capabilities. GoodBuddy-assigned Skills still execute through per-request staging. The temporary Continue Host does not reuse native session compaction; manual compaction uses the GoodBuddy summary model and verifies persisted summary coverage. Agent questions become unified question cards. Resources, Hooks, background Jobs, and Continue native-session management remain deferred.

• Native Runtime catalog semantics: Native capabilities are displayed across 11 tabs: Agents, Tools, Commands, Skills, MCP, Rules, Prompts, Resources, LSP, Formatters, and Context. Catalog state is independent from Runtime connectivity and distinguishes complete, partial, unavailable, connection-only, and unsupported. DeepSeek Harness enumerates bounded built-in/plugin Tools and Skills through the Host Registry, reports real Ask/Execute boundaries, and excludes per-request GoodBuddy Skills and Web/MCP proxies.

• Runtime supervision section (planned): A fixed section in the application-level assistant workbar for Task-level delegation, background execution, Workflows/Hooks, long-running work, and native-session supervision across OpenCode, Continue, and DeepSeek Harness. Users select only Conversation or Task; Job/Run remains internal rather than becoming a tree or independent object.

• Advanced Subagent supervision (planned): Task-level aggregation in the workbar's fixed Runtime section, configurable nesting, parallelism, budgets, and lifecycle controls. Basic expert-mode inheritance and single-level direct-model programming delegation are already available.

Skills, MCP, and knowledge

• On-demand Skills: Assignable to direct models, OpenCode, Continue, and DeepSeek Harness with bounded resources and controlled Runtime boundaries.
• Local tool-execution environment source path: Under Capabilities and Tools / Tool Environment, users can select GoodBuddy-managed Node.js, on-demand managed Python, or a genuinely validated custom interpreter for local Skills and stdio MCP. The page provides an independent Native/OSS source choice, diagnostics, installation progress, cancellation, and removal. New local Runtimes and stdio MCP processes receive immutable PATH snapshots without changing ordinary terminals, the system environment, or remote Hosts. Managed Node and native-source Python have passed real installation validation on Windows x64. Managed Python archive validation follows the target filesystem: Linux preserves valid case-distinct paths, while every target still rejects exact duplicates and unsafe entries.
• Local tool-execution environment validation records: All six platform/architecture OSS mirror objects have passed public byte, size, and SHA-256 verification. Managed Python remains an on-demand download and does not add license files to the Desktop package. Each standard package job uses a native target-architecture Runner and performs a real managed-Python install, SSL, pip, and venv probe before packaging. The immutable v0.12.0 attempt passed Windows and macOS but exposed Linux TAR case handling before publication. The immutable v0.12.1 attempt was cancelled before native packaging while synchronizing the Agent release. Published v0.12.2 subsequently passed all six native package jobs and their managed-Python install probes. Real Skill/MCP execution, custom interpreter execution, and coordination with affected running processes remain acceptance work. Six-platform release acceptance must not be claimed before the applicable checks pass.
• On-demand built-in MCP: Knowledge, Magic Notes, GoodBuddy configuration, and built-in browser MCP can be enabled independently and assigned to direct models, GoodBuddy-managed OpenCode, and Continue. Settings explicitly marks DeepSeek Harness unsupported. Built-in MCP uses short-lived local authority for the current request, and user configuration cannot loosen Ask/Execute read/write boundaries.
• MCP Tools: Explicitly enabled custom MCP can be assigned to direct models, GoodBuddy-managed OpenCode, Continue Agent Execute, and DeepSeek Harness, and loads only in Execute. Agent child processes receive only per-request local-loopback authority; MCP addresses, commands, and credentials remain in Main. Dynamic tools pass through discovery and the existing activity and permission boundaries; healthy direct-model and Harness calls reuse the discovered request catalog instead of rediscovering it before every call.
• MCP Prompts and Resource metadata: MCP testing discovers bounded Prompt, parameter, and Resource metadata only when the Server declares the corresponding capability and does not read Resource content. Supported Prompts can be inserted into the chat draft for further editing. OpenCode can report an experimental Resource catalog; the current Continue version explicitly does not support Resources.
• Local knowledge bases: Supports file, directory, and web imports, SQLite FTS5 retrieval, and source tracing. Knowledge selection persists per conversation and starts empty for new conversations. Advanced creation and retrieval parameters are collapsed by default. Documents distinguish ready, processing, and failed states, with source opening, retry, and confirmation before source removal.
• External knowledge-base connections: Manage Dify, FastGPT, and RAGFlow instances from the Knowledge page, then discover remote knowledge bases or enter their IDs and verify bindings with provider-specific retrieval settings and citations. Conversations support on-demand and required pre-answer retrieval. External systems provide retrieval only: GoodBuddy does not use their App, Chat, Workflow, or Agent APIs, and does not bulk-sync, locally index, or modify remote content. Bounded cited snippets are retained locally with their conversations. See the external knowledge-base PRD.
• External knowledge-base validation records: All three providers have passed real service, local HTTP MCP, and production-IPC short-answer checks; Dify additionally passed the complete App composer-to-answer-and-citation path. FastGPT/RAGFlow desktop-specific parameters, complex multi-library answers, and remote pre-answer retrieval remain unverified. See the validation evidence and remaining work.
• Knowledge graph: Supports rule-based, model-based, and hybrid extraction plus entity, relationship, alias, and evidence maintenance. Model extraction adds no fixed 8192-token output cap: OpenAI follows provider limits, and Anthropic uses the model connection's output setting.
• Embedding configuration and retrieval: Configures compatible Embeddings endpoints and uses them for semantic retrieval.
• Embedding diagnostics and indexing jobs: Provides real embedding diagnostics, per-document rebuild progress, cancellation, failure state, and result recovery after restart. Each successful document becomes immediately searchable.
• Hybrid retrieval workbench: Diagnoses full-text, Chinese phrase, vector, and graph channels with configurable Top K, thresholds, weights, local or learned reranking, and context budgets.
• Chunking, maintenance, and evaluation: Supports fixed, structured, and parent-child chunking, chunk maintenance, cancellable rebuilds, and bilingual retrieval evaluation.
• Controlled knowledge ontology: Each knowledge base can define entities, relationships, aliases, and endpoint constraints while retaining evidence offsets, confidence, and extraction source, with explicit guidance to rebuild the graph.
• Required retrieval and cited context: Conversations can retrieve on demand or before every response, show zero-result, degraded, failed, and cancelled states, and expose cited context or safely open its source.
• Magic Notes: A local-first notes and todo workbench with scope management, editing, filtering, and controlled AI comments. Notes open from a responsive card overview. Detail displays a continuous stream with independently collapsible left index and right AI pane. The index defaults to 168px and supports pointer and keyboard resizing from 140px to 320px, with saved width preferences; narrow layouts retain the fixed-width drawer without a resize handle. Real first-page thumbnails scroll to records without replacing editors. Successful saves establish the editor's normalized baseline; only unsaved changes require confirmation on leaving. Opening a note shows a blank text/canvas composer above the history, and creation resets it for the next record. Explicit editing updates a saved record. The separate To-dos view offers search, status filters, grouped tasks and direct completion, with instructions, source entries and AI comments in a right detail pane beside a stable task list. Narrow views offer on-demand detail with an explicit return to the list. A content-sized switch sits beside New note in the header; status filters follow search. Source-note navigation returns to the same task context, without showing the previous item's source after a selection change. Failed reads use application notifications with inline retry, and missing sources have an explicit fallback. The left navigation can show the incomplete-todo count, and create, save, and comment results use application-wide notifications. Agent/MCP note writes automatically refresh the open workbench while preserving selection and unsaved drafts.
• Paged canvas notes: The integrated PeopleLib Fabric + Quill editor combines flowing body text with pen/highlighter, object selection/transforms, floating text, images and paper templates. Editors and read-only viewers support 25%-300% view zoom, 100% reset and responsive Fit width, without changing saved content or export dimensions. Import PDFs as page backgrounds with native text extraction, then export a raster PDF of text and annotations; exported PDFs do not retain a searchable text layer. Manual saves store entry bodies and binary assets in local note files; SQLite retains metadata, indexes, revisions, todos and comments. Backups must coordinate SQLite and the notes directory; there is no one-click backup action. Undo/redo is per annotation page or Quill mode, not global document history; there is no infinite canvas or PNG download button. Drafts, saved entries and canvas-source todos use the default model's image-input capability for page images plus text, or an explicitly labelled text-only fallback. Canvas pages to send selects the first 1-8 pages in current order (default 1) for both text and images; saving and exporting still support 50 pages. Purely visual entries/drafts require an image-capable model. Immediate-comment mode offers manual canvas analysis; after-save automatic analysis failures do not block or undo saving. Text-only comments survive unchanged extracted text; visual comments are invalidated by layout changes and reanalyzed in after-save-auto mode. MCP rejects plain-text replacement of canvas entries. Implementation is complete; final validation status is tracked in the feature progress record. The first upgrade migrates stored bodies into files and raises the database schema to 39 together with attachment changes. Older clients cannot reopen that database; rollback requires a pre-upgrade backup.
• MCP Server Control Plane (planned): Unified MCP lifecycle, health checks, reconnection, schema cache, isolation, approval, and audit.
• Traceable note excerpts and AI editing (planned): Collects sourced excerpts from conversations, knowledge, and the web and provides confirmation-based summarization, rewriting, and organization.

Work management, long-term collaboration, and workflows

• Tasks, activity, and artifacts: Centrally manages task state, audit activity, and independent artifact files. Ordinary chat responses remain only in the conversation and are no longer copied into Artifacts; existing duplicate chat Markdown is hidden from the artifact list but not physically deleted. Run history is stored in Main SQLite without the previous 500-item, 4,000-character, or 2 MB Renderer limits; the page renders it in batches, without a persistent legacy-cache truncation warning; previously lost history is not restored. Token usage is grouped by Runtime and model and normalizes the different OpenAI-compatible and Anthropic Messages cache-reporting semantics when showing cache hit rate. Project and conversation usage groups show collapsed totals by default and expand to Runtime/model child rows; model grouping remains flat. Activity is grouped by conversation and collapsed by default so long histories do not fill the page.
• Compact conversation tool records: Expand individual tool records to read or copy results, errors, and input parameters. Conversation and child-task progress use concise rows without repeated Runtime summaries. OpenCode, Continue, DeepSeek Harness, and direct models show available file paths, commands, or search summaries and retain them through completion and conversation reloads.
• Task and custom-task experience: Each product-level Task belongs to one Conversation, while one Conversation can contain multiple Tasks. The left conversation list exposes Task children through a leading expand button; children share status dots, the parent does not repeat task badges, and the UI stops at Task rather than exposing Job/Run levels. A new custom Task can use the current or a new Conversation, defaults to Execute, and preserves Runtime, tool, and approval boundaries. Repeated triggers reuse the same Task; text returns to the Conversation, while independent files and images remain artifacts. Ordinary messages and due Scheduled Tasks share a persistent Conversation queue that runs one item at a time. Users can keep sending while a response is active; queued items continue in order and can be removed or inserted immediately by interrupting the current item. Task Center remains the complete index rather than creating a separate Automation Center. Current schedules support one-time, daily, and weekly triggers; advanced time zones, Cron, event triggers, and retry governance remain incremental PRD work. Scheduled messages use the conversation's current history, Runtime, work mode, and saved knowledge retrieval settings; task details show the actual mode used, and an active occurrence blocks duplicate Run Now actions. See the Task Center PRD and Scheduled Task PRD. Status counts and filters identify running tasks and those needing attention; task approvals remain actionable in their task cards.
• Memory and Smart Heartbeat: Provides periodic review, suggested memories, insights, follow-up tasks, and auditable run history.
• Automatic supervision settings and scope: Supervisor / Automatic supervision is the sole authoritative plan configuration, automatic report, and suggestion entry. Daily or weekly plans can be created and edited for Global or one or more selected Projects. Legacy single-project settings migrate without loss, and project-level memory and action output must explicitly target a Project in scope. Task Center and Settings no longer duplicate the form. Partition-aware review, candidate generation, and recall triggers remain to be designed. See the Smart Heartbeat PRD.
• Supervisor reviews, story graphs, and activity: Review a selected scope and period, read dated supervision results and history in Work review, inspect saved result graphs and sources, confirm or revise entities, and preview local knowledge entity writes. Sidebar feedback follows or pins a Conversation or Task and opens its matching result. Activity combines heartbeat and downstream review stages, including failures and links to older results. Automatic stages process only new, changed, or unprocessed source portions; no-change checks skip model calls. Manual review reprocesses bounded history. Calls remain read-only without tools and may incur model charges. See the implementation and evidence.
• Further supervision capabilities (planned): Full event-by-event replay, Experiment targets, user cancellation, chronological manual-edit audits, event triggers, and broader execution observation remain incomplete. Current automatic review respects input budgets and the rolling window; deleting reports does not reset processing progress or rebuild graphs. See the Conversation Supervision PRD.
• Batch runs and comparison lab (planned): Compares model, Prompt, role, and workflow configurations in batches and summarizes quality, duration, tokens, cost, failure rate, and artifact differences.
• Temporal memory and fact-conflict detection (planned): Adds validity periods, current/expired/conflicting fact detection, fact checking, and evidence tracing to memory and the knowledge graph.
• Visual controlled workflows (planned): Versioned DAGs, conditional branches, approvals, cancellation, and recovery, with execution still passing through Main Runtime boundaries.
• Unified Run Graph and replay (planned): Connects Tasks, Subagents, models, knowledge, tool approvals, usage, and artifacts for failure diagnosis, retry, and redacted export.

Browser, communication, voice, and application maintenance

• Shared built-in browser for runtimes: Uses GoodBuddy's isolated Chromium and never controls a browser installed by the user. A separate master switch decides whether assigned runtimes receive the capability in Execute, with no per-use prompt after enablement. Turning Agent access off does not disable manual browser workbar actions. The Browser workbar and Agent share the same Conversation-owned session and serialized operation path, with Back, Refresh/Stop Loading, address entry, Go, Interaction, and Close actions. Stop Loading does not close the session, and user navigation changes the page the Agent sees next. Multiple tabs in one conversation share login state while keeping separate pages and navigation. Each model request retains its starting tab binding; unused request reservations do not create browser resources. Released tabs can be reopened, explicit screenshots remain available, and ordinary operations no longer trigger unused automatic screenshot capture. Tabs can be closed while an AI request uses them without cancelling the whole request. A later explicit navigation creates a separate replacement tab; blank pages cannot reload, and switching or closing panels clears stale action errors.
• Client-computer control tools: Managed separately from the built-in browser with scope, cancellation, timeout, output, and activity boundaries.
• Remote messaging-channel projects: WeChat ClawBot, WeCom, and DingTalk each have a system-managed project, independent remote conversations, working directory, processing backend, default Ask/Execute mode, and task/activity ownership. Each channel controls complete-response length and segmentation according to platform capabilities rather than relying on a shared service truncation.
• WeChat ClawBot QR login and media: A separate Sidecar handles local QR scanning, verification codes, encrypted credentials, and text messaging. Personal WeChat chats support images and files, with at most four attachments and 12 MB total after decryption per message.
• Safe WeChat responses: Can return images generated by the current task, or generate a Markdown attachment from the current final text after an explicit user request. It never automatically reads or sends existing workspace files.
• WeCom and DingTalk connections: Main-only encrypted settings, read-only environment-variable overrides, connection tests, dynamic enable/disable, sender scope, and status diagnostics.
• Managed local-model download source: Platform Features / General selects ModelScope (default) or Hugging Face globally for future speech-input and OCR model downloads. A source without a complete verified file is explicitly unavailable and never silently falls back or combines files.
• Optional local speech-model management: Model weights are not bundled. Provides verified downloads, progress and cancellation, source links, ZIP or local-directory import, switching, and removal.
• Local recording and offline transcription: Captures microphone audio and transcribes with the selected local model, with stop, cancel, state feedback, and resource cleanup.
• Version checks and mirror source: About and Updates selects GitHub (default) or the mirror. Manual checks, startup checks, and the download page use the same choice and read only fixed trusted release indexes; GoodBuddy does not automatically download or install updates.
• Window recovery controls: Rendering errors show a localized Reload action and renderer crashes offer a native recovery confirmation. Reload may lose unsaved input. See window recovery.
• In-app feedback: About and Updates can submit problems, suggestions, or experience feedback with an optional email and one screenshot. Diagnostics are not uploaded by default; users can explicitly attach a bounded recent Desktop diagnostic summary. Failed submission preserves the draft, diagnostic choice, and request ID. Conversations, Prompts, credentials, file contents, paths, raw provider responses, and remote Agent logs are never attached.
• Private-network compatibility: Permits in-app HTTP plus invalid, self-signed, or expired HTTPS certificates. Fixed platform services such as feedback and WeChat credential or media endpoints retain strict certificate validation.

Open source, builds, and releases

• Current source candidates are Desktop 0.13.12 and Agent 0.13.3, with OpenCode pinned to 1.18.29 and Continue to 1.5.47. Publication status follows the independent Desktop and Agent release channels.
• Desktop 0.13.12 adds Supervisor reviews, graphs, activity, and automatic incremental processing, reduces streaming activity writes, and improves Mermaid, usage totals, conversation duration, Git, notes, and notifications. Database schema 46 requires a complete pre-upgrade backup for rollback to older clients.
• Agent 0.13.3 reclaims idle Runtimes and retires displaced Agents after existing work drains. Its packages require Desktop 0.13.12; update Desktop first, then the Host environment. Node remains 24.19.0.
• Release preparation uses focused regression tests, notes verification, typecheck, and lint. The full local suite is not rerun at the user's request; full main-branch CI, its production build, native packaging, and publication verification remain pending. No local production build or packaging is part of this preparation. Existing Linux x64 lifecycle evidence used eight real provider requests and incremental-review evidence used three; neither covers the complete native matrix.
• Validation records remain separate from implementation status. Current-source storage validation covers real local and Linux x64 Host tool workloads, concurrent projects/conversations, cancellation, and lossless database migration. Runtime process reuse has real Windows full-App and Linux x64 Host evidence; PPTX OCR and idle-memory release have real Windows file evidence. Image tools have real provider generation/editing and Host transport evidence. Full UI-driven natural-language invocation and local/remote switching are not covered by those records. The candidate still requires main CI and native release packaging. See the Runtime evidence and image-tool evidence.
• Before release preparation, 9836a4c passed 4,858 local tests with 67 skipped, full typecheck and lint; real-provider canvas analysis remains unverified. Its three native Agent CI builds passed. Desktop CI run 35485936807 passed 4,875 tests but failed eight Electron tests before UI assertions because the Linux sandbox helper was not configured. The candidate configures the helper and runs tests under Xvfb without disabling sandboxing. Candidate-local validation, main-branch CI and native tag packaging are tracked separately; no local production build or package is part of release preparation.
• 0BSD open-source license: Original code can be freely used, copied, modified, distributed, and commercialized. Third-party components and resources retain their own licenses.
• Reproducible dependency installation and source builds: Uses locked dependencies, Node.js 24, and unified test, type-check, lint, and production build commands.
• Six-target native release matrix: Native runners build Windows, macOS, and Linux x64 and arm64 targets with release manifests and SHA-256 hashes. Windows provides NSIS and portable ZIP, macOS DMG only, and Linux AppImage, DEB, and RPM, for 12 installers and 20 published assets.
• Desktop Runtime package verification: Offline OpenCode dependencies are copied without electron-builder's root node_modules exclusion; DSH session projection is bundled into the unpacked Host. Final package checks compare the offline dependency tree and reject external DeepSeek imports. Windows x64 CI additionally imports the packaged OpenCode plugin and checks the real DSH UtilityProcess handshake before publishing.
• DMG-only update manifests: Current Desktop and website readers accept DMG-only and historical DMG/ZIP manifests. Older macOS GitHub readers and all older mirror readers require a manual upgrade from the download page.

Open interfaces, team collaboration, and remote execution

• Remote Continue delivery: Current source supports a managed CN 1.5.47 package, Agent HTTP-to-ACP adaptation, model routing, project selection, and native checklists. Linux x64 development packages passed real model, cancellation and reconnection checks. Default packages include OC/CN; Windows desktop-to-Linux checks cover installation, checklist updates, long lists, cancellation and restart recovery. Execute session MCP delivery passed real text-model Host checks with a substituted image service, not real image generation. Candidate Agent 0.13.3 requires Desktop 0.13.12; publication and native platform acceptance remain separate from Linux development evidence. See the validation record.

• Remote task delegation: Enabled only after the user explicitly configures an endpoint and token, uses HTTP(S) according to the global private-network compatibility setting, and writes results to a persistent outbox.

• Headless Runtime API (planned): A local-first Task, event, state, and artifact API with scoped, expiring, rate-limited, and revocable tokens.

• GoodBuddy Team Hub (planned): An optional service for organizations, RBAC, project sharing, remote Agents, policy distribution, and tenant audit.

• SSH Host and remote execution-space validation records: Host CRUD, Host Key, encrypted credentials, Project UI, Workspace, OpenCode ACP v5, Agent-owned Prompt/gateway/transcript, read-only Ask, full-account Execute, cancellation, exact detached-Agent reconnection, and release-only dual-architecture resource verification are wired. Current Linux x64 source has passed a real-model and tool matrix covering detach, process exit, relay loss, concurrency, cancellation, provider failures, Agent restart, and Desktop SQLite recovery. A current-source, test-signed Agent 0.11.17 package on Linux x64 additionally passed two concurrent Ask reads, an Execute write/read, and six real managed-gateway model requests with the configured custom headers and body fields. No provider request was replayed, and scanning the isolated Agent state found no API key or custom request value. The public signing-key registry is provisioned. Direct control-plane source does not wait for a new installer-bearing package, but the current formally signed Linux x64/arm64 artifact matrix still requires public verification, along with GitHub/Beijing mirror, dual-architecture, and offline GoodBuddy transfer release acceptance. Until these gates pass, this path must not be described as published or as having passed formal release acceptance.

• Multi-cloud remote sandbox Agents (planned): Manages dedicated Linux sandboxes through cloud-provider APIs and SSH Agents. Credentials remain in Main, and high-risk control-plane operations receive separate confirmation.

Roadmap Principles

Planned workflows, Subagents, MCP, remote APIs, and sandbox capabilities must not bypass existing Main Runtime, Ask/Execute, permission, cancellation, timeout, or audit boundaries.