Design Skill: workflows for interface, website, visual-system and user-flow diagnosis, design exploration, specifications, responsive layouts, accessibility checks and implementation.
Office Skill: create, read, edit and check Word, Excel, PPTX and PDF files, including document layout, spreadsheet formulas and charts, presentations and document review.
Presentation workflows: improve slide layouts, presentation narratives and speaker notes, and turn images into editable slides; reopen deliverables to check content and structure.
Personal Skills across execution modes: install once and find, reference and use them with @ in engineering, lightweight and orchestration modes. Built-in descriptions are in Chinese; personal Skills retain their authors’ descriptions.
On-demand Skill references: selected Skills provide text, references and paginated content to the current task; access is revoked when the task ends, is cancelled or updated.
Agent building: shared build and upgrade entry points for the GUI, CLI, SDK and API, linked to package identity, platform and version, with configuration versioning, administrator authentication and upgrade audits.
One-click updates: update checking is on by default. After a new-version notification, click Auto Update to check, download, verify the signature, install and restart; disabling automatic checks is remembered.
Anonymous diagnostic copy: copy stop, reload and completion lifecycle diagnostics from the debug panel without including message text or credentials.
Improved 10
Three execution modes: update engineering, lightweight and orchestration capabilities, improving continued execution, model routing, stopping and completion receipts.
Long reasoning and tasks: extend the execution window while real reasoning or tool progress continues; improve bounded recovery after temporary stream interruptions while retaining progress and unfinished status.
Context compression and recovery: improve summaries and checkpoint recovery, preserving the original goal, history and completed tool results to reduce repeated execution.
Model switching and continuation: preserve conversations and tool receipts when switching remote-authorized or separately configured models; keep the current selection on errors instead of reverting to an old or retired default.
Approvals and time limits: start tool execution timing after approval, so waiting for confirmation does not consume execution time; cancellation and duplicate-call protection remain effective.
Settings organization: group accounts and workspaces, experience, AI and extensions, and system maintenance; scroll navigation and content separately and improve About, updates and error displays.
Narrow windows and high zoom: move secondary title-bar actions into menus while keeping project information accessible; put attachments, expansion and voice in More for narrow composers, retaining Send and Stop.
Skill management: improve installation dialogs, retain unsubmitted input, and adjust table columns and narrow-window scrolling so long descriptions do not crowd categories, permissions or actions.
Offline office runtime: the Mac Apple Silicon package includes its interpreter, format libraries, converters and Chinese fonts; basic office processing uses bundled resources rather than a development environment.
Accounts and permissions: open platform login independently of a workspace and distinguish it from Agent server accounts; unknown roles, loading and failed reads are no longer presented as confirmed states.
Fixed 21
Persistent login and logout: fix success notices before credentials are saved and reuse of stale rotated tokens; request login for expired tokens and improve account persistence and logout after reopening, overwrite installation and upgrades.
Remote-model protocol compatibility: fix rejection or interruption caused by incompatible metadata, reasoning history and completion formats; retry only clearly recoverable current requests, preserving original history and tool results.
Local models without API keys: fix successful connection checks followed by key requirements at startup; use consistent rules for saved connections, selection, health checks and execution, and update the UI when health recovers.
Actual model selection: fix ignored orchestration settings, incompatible protocol choices after refresh and incorrect model identity, reducing differences between the selected and executed model.
Long-session recovery: fix scroll jumps when loading history, missing sidebar entries from oversized previews and duplicate restored transfer states; improve ownership coordination when multiple instances read an active session while preserving chat and tool receipts.
Tool and Skill calls: fix premature revocation of selected Skills, models receiving only Skill names and tool receipts sent to the wrong channel; deleted Skills no longer reappear from stale discovery and design Skill counts are not duplicated.
Task completion status: restore final replies after truncated tool output; stop treating progress text as completion or normal compression notices as failed answers.
Confirmation and protection states: require explicit confirmation before deleting personal Skills or restoring old sessions; fix failed audit or approval reads being shown as disabled protection or a healthy queue.
Menus and shortcuts: fix unresponsive native New Conversation and Check for Updates actions, and some Mac shortcuts failing after saving; add Chinese accessibility labels for code-diff options.
Popovers and layout: fix native context-menu positioning after zoom, misaligned window and sidebar-collapse buttons, narrow account popovers, clipped plugin dialogs and Escape closing both a plugin and settings.
Main window and workspace: improve showing hidden windows and rebuilding unresponsive ones; prevent historical run monitoring from replacing file or prompt panels, retaining task state during window replacement.
Office conversion and preview: fix bundled resource paths, conversion initialization in restricted environments and missing Chinese glyphs, preserving formulas and charts; damaged PDFs return clear diagnostics.
Page identity and fields: preserve valid channels when login identity is completed and isolate sessions after account switching; support dotted dates and fix multi-field planning and readback after writes.
Official desktop adaptation and installation: resolve changed execution entry points after official-program upgrades; back up and verify old launchers before atomic migration. Overwrite installation checks application and active-service identity and preserves running filter services and tasks.
Update installation: simultaneous update triggers in the main window and settings perform one installation; fix Mac archive metadata installation failures and use https://www.boltfusion.cn/downloads/boltfusion-client/latest.json as the built-in update source.
Mac platform login now recovers Keychain errors automatically and retries the current submission once; cancellation or another failure stops recovery.
Added redacted login diagnostics available without sign-in, including package identity, macOS version and recovery codes; passwords, tokens and private keys are excluded.
Improved private-model streaming: productive inference can exceed 180 seconds while permanent stalls remain bounded; complete JSON and tool receipts are preserved without replay.
Long-task compaction: verified new tool progress renews the allowance instead of imposing a lifetime task cap. Consecutive compactions without progress remain bounded. This Mac build passed three compactions and warm/cold continuation.
Image previews: PNG, JPG, WebP, GIF and SVG can be viewed directly. The enlarged view follows image proportions and window bounds, with original-size, zoom, pan and close controls.
Large files: logs, text, PDF, HTML, audio/video and Office files remain file links instead of loading full content or embedded previews automatically. Missing images show a clear explanation.
2026-11-06 manual QA increment for Mac 1.1.0, package 1d6088f18026. Exact installed identity/signature, image formats and five fixed window widths were verified. Productive tasks passed three compactions and warm/cold continuation without replaying tools; no-progress protection remains. Previous changelog entries are retained. Apple Developer ID, notarization and formal updater acceptance are not claimed. Twelve Mac points remain, including the viewer closing across responsive layout changes; customer-device recovery is not inferred from local tests.
Image viewer: images in conversations and reasoning logs show at a readable size. Open one to zoom, see the original size, pan or reveal it in its folder; press Esc to close.
Tasks survive restarts: when the app is reinstalled over itself or restarted, a running task reconnects and continues without repeating completed tool steps. If it cannot recover, the reason is shown and you can send "continue" manually.
Git branch display: the top-right corner shows the branch of the conversation's repository, a short commit when detached and a marker for uncommitted changes, or "Not bound to Git" outside a repository.
Implementation agents use their knowledge base: in all three modes the E9 implementation agent finds and calls its own knowledge base by business alias and cites evidence; when the knowledge base is unavailable it answers in degraded mode and says why.
Images in orchestration mode: orchestration mode can pass images to image-capable models; when the selected model does not accept images, you are asked to switch models or remove the image.
Design and office Skills: discover and reference integrated design, Word, Excel, PowerPoint and PDF workflows across the three execution modes.
Presentation workflows: includes reusable slide layout, speaker-note and image-to-editable-slide workflows. Optional plugins, OCR and image services require their own configuration.
On-demand office resources: retain locked manifests, launchers and licenses in the GUI; download and verify the platform runtime on first use, then reuse a verified cache. LibreOffice is no longer bundled as a large GUI payload. Offline users need the same verified cached recipe.
Personal Skills: discover an installed Skill across all execution modes and retain its enable, disable, update and removal state.
Account access: open platform sign-in independently of the workspace and distinguish platform authorization from Agent service accounts.
Lifecycle diagnostics: inspect bounded stop, reload, resume and completion diagnostics without recording credentials or message bodies.
Improved 18
Responsive layout: the interface scales with the window size; the input toolbar always stays on one line and folds into compact buttons with a settings panel in narrow windows; tablets keep the conversation list docked, and phones get a top bar and bottom navigation that match the main menu.
Conversation list: each project first shows 5 conversations, and "Load more" adds 5 at a time with the remaining count, hidden when there is nothing more; while starting, placeholders the size of a conversation row appear, and titles use the full row width.
Process cards: process and tool records start collapsed when a conversation opens, and expanding them no longer carries over to other conversations.
End-of-turn summary: every turn ends with the same card, green with proof of completion, neutral for an ordinary end, and highlighted for failed or unfinished turns.
@ reference list: arrow keys, Enter and Esc keep working in the list with a Chinese input method or while results load.
Home capability cards: one layout and icon style, shown in three columns, two columns or a compact list by width, without pushing the input box off small screens.
Model selection: the model switcher in settings lists only models the current gateway offers.
Install size and runtimes: server-side dependencies are merged into one trimmed copy, saving about 0.5 GB after installation; the lightweight mode runtime dependencies are updated to match.
Execution modes: improve model routing, continuous execution, cancellation and actual completion receipts across engineering, lightweight and orchestration modes.
Long-running tasks: extend bounded execution while real reasoning or tool progress arrives; preserve completed tool receipts during limited recovery.
Stall visibility and resume: retain stage diagnostics and checkpoints without replaying confirmed completed tools or treating idle waiting as progress.
Dynamic models: configure or discover model IDs, endpoints and protocols without rebuilding the GUI; validate the protocol actually used.
Context limits: honor explicit connection limits and available service capabilities; retain summarization, compression and task continuation without hardcoded customer limits.
Model switching: preserve the selected model, conversation and tool results instead of silently returning to a stale default.
Approval timing: start bounded tool execution after approval and retain cancellation and replay protection.
Settings and navigation: reorganize accounts, workspace, experience, extensions and system maintenance with independent scrolling.
Compact layouts: move secondary actions to menus in narrow or zoomed windows while keeping Send and Stop available.
Skill management: retain draft installer input and keep long descriptions from displacing permissions and action controls.
Fixed 39
With full access, starting Microsoft Store PowerShell from the command tool in terminal mode failed with "Access is denied".
Clicking "Collapse full content" on a long answer changed nothing on screen.
Reasoning logs were incomplete and could mix in other conversations; later turns of the same conversation now keep the knowledge bases it referenced (up to 4), and new conversations do not inherit them.
Text typed with a Chinese input method disappeared when the input already held an @ reference.
After a conversation switched to a model from another connection, requests still went to the original connection, and the same model could appear twice in the list.
A slow check on one connection made other healthy connections look timed out and greyed out.
When the conversation model list opened upwards, the top options sat under the title bar and could not be clicked.
Referencing a knowledge base or Skill reported "Failed to start reasoning".
Installing or upgrading stopped the ChatGPT context filter gateway and made open conversations time out.
The local authorization gateway could not re-authenticate after expiry, and messages were unclear when the upstream was unreachable.
The machine code exported by the client did not match the local Spec base environment, so one computer could be registered as two machines.
A task kept spinning when the server sent no final status; it now stops loading within a bounded time, keeps its checkpoint and tells you, without replaying the task.
Knowledge first installation and reinstallation: support both numeric-project and packageId ownership layouts. Prepare missing managed configuration through the real connector, then strictly validate TOML, ownership and project boundaries. Explicit local-only mode with commercial features disabled touches only the built-in target.
Authorization renewal: retain the verified new signed license in session memory until resource synchronization completes, then verify persisted material. Prevent new/old license mismatch loops and skipped synchronization when the capability set remains unchanged; preserve signature, identity and revision checks.
Selected GPT routing and compression: use current collaboration parameters for send and resume, retain the selected GPT through compaction and summarization, and prevent stale DeepSeek routing. Health checks no longer repeatedly overwrite global model configuration or compete with knowledge installation.
Short prompts and reference menus: keep actual user input in copy, reference and reopened history. Preserve asynchronous project-filter sessions when results are empty, so knowledge, keyword and project selection can continue.
Moved workspace history: match Windows-resolved directory identity for junctions and symbolic links without changing user paths, guessing drives or rewriting the history database.
Request diagnostics: use the actual application version and retain bounded HTTP status, phase, latency and sanitized request IDs. Limit token refresh to the original service, installation and account; diagnostics never block business execution.
WebView2 installation: include and select the bundled fixed runtime before creating the WebView; report incomplete or unsafe runtime paths clearly.
Startup verification: check the real installed executable and renderer, visible pages, saved history and exit/reopen behavior. This does not claim every customer GPU or OS issue is resolved.
Kernel identity: reject stale or mixed runtime versions using source, architecture and byte identity checks.
MCP migration: validate and atomically migrate conflicting legacy TOML structures while retaining user models, MCP entries, Skills and comments.
Spec output budgets: add bounded defaults to managed E9/E10 context and evidence tools while preserving explicit positive customer budgets.
Managed markers: repair uniquely identifiable historical MCP closing markers without moving global settings into Skill configuration.
Spec continuation and final answers: reduce line windows under the same evidence identity, retain image and continuation handles, and display the final answer after complete readback.
Completion handling: recover a final response after truncated tool output instead of treating progress text as completion.
Chat protocol compatibility: use the selected Chat Completions or Responses protocol for both testing and conversations; do not force all DeepSeek services to use Responses.
Thinking history: handle reasoning-content compatibility on the current request copy while preserving history, valid encrypted data, tool links and checkpoints.
Local models without API keys: use consistent save, selection, health and execution rules for compatible local services.
Orchestration model selection: load the supported model profile and return tool receipts to their owning runtime.
Sign-in persistence and sign-out: report success only after credentials are saved, refresh rotated tokens and prevent stale state from replacing the current session.
Windows atomic saves: retain the previous configuration on write rejection and support replacement under read locks.
Long conversation history: reduce preview and repeated-state payloads while preserving full chat, checkpoints and tool receipts.
Skill confirmation: retain selected Skill access until reading completes and require explicit confirmation for removal or session recovery.
Office paths and permissions: handle extended Windows paths and restricted tokens without broadening sandbox permissions or changing installed resources.
Menus and popovers: restore native New conversation, shortcuts and update checks; respect WebView scaling and close only the top dialog with Escape.
Workspace and authorization display: prevent old monitoring from replacing file or prompt panels and show actual pending or failed authorization reads.
Update controls: retain automatic-check preferences, avoid duplicate installs and preserve running tasks. Publishing this manual installer does not activate a Windows production update feed.
Long-task protection and gateway recovery: preserve active requests when gateway recovery times out, then restore the gateway automatically and continue the original conversation without forcibly terminating active tasks. Installed-build and final desktop validation are complete.
Windows x64 1.1.0 manual replacement installer, build ca354cb81fd6. Retains strict-by-default per-connection SSL verification with isolated persistent exceptions, startup/index recovery, tablet/narrow layouts and connection drafts. This update preserves cancellation signals through request wrappers so stopping also cancels remote generation. The installed build passed real remote Qwen cancellation and two explicit identical submissions, with matching model identities and request counts and unchanged manual VPN nodes. Partial output is retained without automatic replay. Installed-build validation is complete for preserving active tasks, gateway self-recovery and resuming the original conversation; final desktop validation also passed. This remains a QA manual installer; formal code signing and production automatic updates are not enabled. Exit normally before installation; user data is retained.
Added in this build: DeepSeek V4.1 reasoning compatibility, including custom model aliases. Generic minimal/medium/ultra map to low/high/max; valid xhigh/max and integer budgets are preserved. Installed binaries passed a strict isolated upstream, durable session resume and completed-tool no-replay checks; the customer-site receipt is pending. Immediate conversation naming, bottom text visibility and daily/test history separation passed Windows native acceptance. Formal signing and production updater limitations remain. Build d63805d020e4.
Additional verification: the installed Windows app completed a real DeepSeek V4.1 conversation and a same-thread continuation at medium reasoning, retaining context. Equivalent failure recovery retained completed tool results without replay. This republishes the same accepted 1.1.0 archive with its original build time and checksum. It remains a manual QA installer; formal signing and production auto-update limitations are unchanged.
Incremental update (Build 5f25e2016f85): project conversations start at five and load twenty more per click; a distant selection stays within the visible budget. Native checks cover isolation from the old sixty-item cache, 5→25→45→46, a cold restart at twenty-five, server cursors and five window widths. Version-tooltip sizing and wrapping, the Windows Ego browser route and long-task compaction/continuation safeguards are verified. Windows functional acceptance is 241/241; final desktop source gates passed with 5,288 tests passed, 15 skipped and no failures. All five local model checks passed after normal authorization was restored. This remains a manual QA installer with unchanged formal-signing and production auto-update limitations; Mac packages and customer production configuration are unchanged.
Increment (Build 2a8b52168f89): fixes repeated disconnect/retry loops after private DeepSeek Chat exhausts its wait budget, and premature cumulative-lifetime cutoffs despite real progress. Installed tests verify completion with205 seconds of progress, bounded stopping after180 seconds without progress, real disconnect/cold recovery and no replay of completed tools. JSON Object/Schema remains one complete document;18 installed cases pass. Resident sanitized Markdown retains the latest state and failure. Windows file-sharing races are bounded; diagnostic write failure never fails the task. Startup verifies the new gateway wait capability. Controlled simulation does not establish customer-site recovery. Manual QA installation and existing formal signing/automatic-update limitations remain.
Build 944819726940: Fixes explicit optional reasoning-effort rejection on private DeepSeek connections with one request-scoped retry, preserving thinking, checkpoints and completed tool results without changing saved settings. Verified ordinary/strict JSON, disconnect continuation, error boundaries and no tool replay. Startup checks the new gateway capability. Bounded waits and redacted Markdown diagnostics remain. Local simulation is not customer-site recovery; manual QA signing limitations remain.
Build 2da63258021f: bounds repetitive Chat continuation and reconciles stale history with live idle state; keeps reasoning/text/tool/non-progress counts and wait durations visible; records prolonged waits and observed anomalies even when the log page is closed; adds independent diagnostic export; repairs ordinary log download with source-byte cutoffs and SHA-256 manifests; adds Windows single-instance protection and read-only external-writer recovery, retaining history and unsent drafts until a real resume succeeds. Installed tests cover a124-second wait,601/601 recovered events,1239 exported source records and real writer/retry scenarios.5338 desktop and453 gateway tests pass. Bounded client observations do not prove upstream inference or entire conversation history. Customer-site recovery is unverified. Manual QA installer; formal signing/updater limits remain.
Build 18fc2b75827b (R37): Engineering tasks can resume automatically after an interrupted connection, preserving the original task and avoiding duplicate execution of completed tool actions. Large workspaces load file lists on demand, reducing background scans and preventing stale files after a switch. This is a manual Windows download update.
Build 2170ca320c63: Improves engineering-task recovery and stopping after 502 errors, disconnected streams or missing completion markers. Stopped tasks can continue in the original conversation without being mislabeled as failures. Large workspaces load file lists on demand and isolate late workspace results. Fixes a Windows replacement-install timeout when checking process ownership, retaining user-identity and process-start-time verification. Installation checks passed on the Windows host and VM; the VM verified real stop/resume and workspace behavior. This remains a manual QA installer. Formal code signing, production automatic updates and recovery at private customer gateways are not confirmed by this release.