Expanded the global font stack to cover headings, body text, labels, code, and other text contexts
Added the following CJK fallback fonts:
PingFang SC
Microsoft YaHei
Noto Sans CJK SC
Ensured more consistent Chinese text rendering across macOS, Windows, and Linux
Eliminated the visual disjointedness caused by font jumps between different components

Fixed ripgrep unavailability on macOS (cc-haha#433)

Fixed an issue where launching the Tauri app from Finder / Dock lacked the PATH environment variable, preventing rg from being found
Added well-known install-path probing as a fallback
Currently supported probe paths:
/opt/homebrew/bin/rg
/usr/local/bin/rg
~/.cargo/bin/rg

Fixed Bun sidecar bundled-mode detection failure (cc-haha#208)

"isInBundledMode()" previously only checked "Bun.embeddedFiles"
Fixed a false negative for bundled mode in the Tauri sidecar scenario
Added heuristic detection logic based on "process.execPath"
Compatible with "bun build –compile" output

🚀 Improvements

Unified the Chinese font fallback strategy, improving cross-platform UI consistency
Strengthened macOS external command discovery, improving Finder / Dock launch stability
Improved bundled-mode detection logic, improving compatibility for Bun + Tauri sidecar scenarios

2026.05.17 · v0.2.1

🐞 Bug Fixes

Fixed abnormal file path display on Windows
Now correctly shows the file name instead of the full path
Affects tool calls, permission confirmations, file search, and other scenarios
Fixed incorrect project name parsing in the status bar
Fixed the project filter failing to correctly extract the project name from Windows paths
Fixed the file search breadcrumb not showing segments correctly on Windows
Fixed the directory picker failing to correctly expand levels for Windows paths

🚀 Improvements

This fix affects 10 frontend components in total
Covers the chat interface, sidebar, permission dialogs, file search, path display, and project detection modules

2026.05.18 · v0.2.2

📌 Release Overview This update mainly fixes issues with client streaming tasks and tool-call waiting, improving stability for long-running tasks and improving error feedback after abnormal interruptions.

🐞 Bug Fixes

Fixed the client streaming task wait time being too short
Tool call wait time extended to 10 minutes
Resolves tasks being prematurely aborted due to the previous 2-minute timeout
Fixed a per-chunk timeout issue in the stream parser
Both stream parsers now have a 5-minute no-data timeout mechanism
A clear error is now thrown when the upstream service is unresponsive for a long time, instead of waiting indefinitely
Fixed the UI showing no error feedback when a fetch was silently aborted
After an abort, the request now correctly ends its state and surfaces an error
Avoids the UI staying stuck on "Editing file…" indefinitely

🚀 Improvements

Improved stability for long-running tool calls and streaming response scenarios
Improved error visibility for network interruptions, no-data-returned, and other abnormal scenarios
Improved UI state recovery logic during file editing

2026.05.19 · v0.2.4

📌 Release Overview This update mainly fixes concurrency and state-race issues in the auto-update module, improving the stability and interaction consistency of the update flow.

🐞 Bug Fixes

Fixed the download progress bar jumping back and forth

Cause:
Missing concurrency protection
Multiple clicks would start multiple download tasks simultaneously
Multiple progress callbacks writing to state interleaved, causing the progress bar to regress
Fix:
Added an "isDownloading" state flag to prevent duplicate download triggers
Added a "peakPercent" monotonically increasing tracking mechanism to ensure the progress bar never regresses

Fixed the update dialog requiring two clicks to appear

Cause:
A state race between automatic update checks and manual checks
The state cycling through available → checking → available caused the dialog to disappear
Fix:
Added a "checkInProgress" promise de-duplication mechanism
Repeated calls to "checkForUpdates" are now skipped directly
Reduced the auto-check startup delay from 5 seconds to 1 second

🚀 Improvements

Improved state consistency and concurrency safety in the auto-update flow
Improved download progress feedback logic, reducing visual jitter
Improved the timing and interaction of the update dialog

2026.05.20 · v0.2.5

📌 Release Overview This update mainly completes the Apple notarization process for the Heicode macOS build.

🚀 Improvements

The Heicode macOS client has completed Apple Notarization
Improved installation and runtime trust on macOS
Reduced the likelihood of security warnings appearing on first run

2026.05.21 · v0.3.5

🔐 Comprehensive Security Upgrade

Conversation data is now encrypted end-to-end in transit, further strengthening privacy protection
Accounts are now bound to the device they log in from, improving account security
Added multi-device login management
View all logged-in devices from the web
Support one-click revocation of suspicious devices

✨ Feature & Experience Improvements

Improved display of account balance and resource usage, plus data decryption
The model list now supports automatic real-time syncing
Supports mainstream models including Claude, GPT, Gemini, Qwen, and DeepSeek
Improved overall stability of login-state detection

🐞 Bug Fixes

Fixed being unable to access a directory when using "@" to select a folder
Now correctly supports accessing projects at any path, including drive C:, D:, etc.
Fixed incomplete model list display and abnormal model call issues
Fixed an occasional stall where the app showed "not logged in" right after logging in
Cleaned up leftover issues from old-version upgrades
Upgrades now complete without needing to manually delete files
Fixed issues with task execution being interrupted or stalling

2026.05.22 · v0.3.6

📌 Release Overview This update mainly fixes client stability, permission-mode switching, and authentication-compatibility issues, while also unifying runtime behavior between the Windows and macOS clients.

🐞 Bug Fixes

Fixed the usage info box disappearing (Windows)

Fixed the balance / usage info below the chat box not showing after upgrading to v0.3.5
Improved balance-API compatibility logic with multi-path auto-detection
The info box is now retained with a "balance unavailable" notice even when the API is unavailable

Fixed a runtime error after switching permission mode

Fixed the following error appearing after switching permission mode mid-chat:
CLI exited during startup with code 1
Cause:
The old CLI process hadn't fully exited yet
The new CLI process started too early, causing a conflict
Fix:
The new process now waits for the old process to fully exit before starting
The old process is force-terminated automatically on timeout

Removed the "response took too long" auto-stop mechanism (macOS)

Fixed the occasional error during chat:
Error: response took too long, this turn has been automatically stopped
Removed the frontend/backend round-trip timeout watchdog
Now consistent with Windows client behavior
Improved stability for long sessions and streaming responses

Removed legacy key-compatibility logic

Server-side model calls now uniformly use device-signature authentication
No longer falls back to the legacy Authorization: Bearer
Improved overall security and authentication consistency

🚀 Improvements

Unified runtime behavior between the Windows and macOS clients
Improved CLI lifecycle-management stability
Strengthened stability for long-running tasks and streaming responses
Improved compatibility of balance and usage info display
Improved the client's overall error-recovery ability

2026.05.25 · v0.3.7 - v0.4.7

📌 Release Overview This phase mainly completes the switch to the V2 device-signature authentication system, and continues to strengthen error diagnostics, model management, and client stability.

🔐 Security Upgrade

Fully switched to V2 device-signature authentication

Server-side model calls have fully switched to V2 device-signature authentication
Each request is signed with an Ed25519 private key
The key is stored in the system keyring

Removed legacy Bearer Token compatibility logic

Removed the legacy:
Authorization: Bearer
The 30-day legacy compatibility window has officially ended

🚀 Improvements

Long-lived tokens are no longer transmitted back and forth with requests
Implemented device-level authentication isolation
Improved overall account and request security

🐞 Bug Fixes

Fixed incorrect balance display

Fixed abnormal balance display:
$0.0003 incorrectly displayed as $155.36
Cause:
The server already returned a USD float value
The client was applying the quota unit conversion a second time

🚀 Improvements

Improved balance and quota display logic
Improved accuracy of account data

✨ Update Dialog & Announcement System Improvements

Complete redesign of the update-notice dialog

Dialog width expanded from 24rem to 28rem
Notes area expanded from 10rem to 22rem

Announcement system improvements

The workflow now reads from RELEASE_NOTES.md
Unified as the single source of announcement content

UI & typography enhancements

Improved heading, button, and Markdown typography
Improved readability of long changelogs

🧠 Model Management & CLI Stability Fixes

Fixed the model picker showing unavailable models

The model picker no longer shows "built-in" models that aren't actually callable from the local providers.json
Avoids triggering 401 / 404 errors

Fixed server-side model fallback logic

Removed the hardcoded DEFAULT_MODELS fallback from the server-side /api/models

Fixed CLI startup failure when switching provider / model

Switching provider / model now automatically runs drainPendingPermissions
Avoids leftover unresolved tool_use entries in the transcript
Fixed the resulting CLI startup failure

Improved model status display

Now shows "Model Configuration" when no model is available
No longer incorrectly displays a stale historical model name

⚠️ Error Diagnostics & Messaging Upgrade

Enhanced startup-failure error classification

Now recognizes:
device-not-paired
401
not logged in
network error
encryption failure
server rejection, and other issues

Improved runtime CLI error messaging

Runtime CLI exits now include friendly categorized messages
No longer limited to the startup phase only

Fixed a false-positive on normal exit

A normal exit with code 0 no longer shows a red error banner

Improved abnormal-exit copy

Changed to a neutral three-step troubleshooting checklist
No longer defaults to attributing the issue to "device not paired"

🛠 Error Propagation & Recovery Improvements

Added a global `lastUpstreamError` register

Established a cross-module shared error-state mechanism

V2 local setup failures now show a specific reason

Now recognizes:
not_paired
server_pubkey_failed
encryption_failed

Manager server-side rejections now show detailed messages

Now parses:
401 / 403
5xx
other errors
Automatically shows a readable error message with recovery suggestions

Fixed loss of network-error detail

Fixed encryptedFetch swallowing exceptions due to a missing try/catch
DNS / TLS / timeout and similar exceptions are now uniformly classified as:
network_error

🚀 Improvements

Improved error propagation
Improved diagnosability of network and authentication issues
Provides clearer recovery guidance

⚙️ Underlying Stability Improvements

Improved the stderr drain mechanism

handleProcessExit is now async
stderr drain wait time now increases progressively:
500ms
1500ms
3000ms

Added Proxy debug logging

Added:
[Proxy] V2 dispatch start
[Proxy] V2 dispatch returned
To aid future troubleshooting and diagnostics

📌 Phase Summary This phase focused on two core efforts:

🔐 Authentication system upgrade

Fully switched model calls from Bearer Token to device-signature authentication
Improved security and device-isolation capability

🛠 User experience & diagnostics improvements

Gradually refined the previously vague:
"CLI exited with code 1" into specific error causes
Each error category now provides a corresponding recovery suggestion

🚀 Additional Improvements

Fixed incorrect balance display
Improved the update-announcement dialog experience
The model picker no longer shows unavailable models
Exiting normally with code 0 no longer triggers a false error report
Improved stability for long-running and streaming-response scenarios

2026.05.26 – 2026.06.08 · v0.4.8 – v0.5.0

📌 Release Overview

This phase strengthens login and account security, expands the available models, and unifies the macOS / Windows clients at v0.5.0.

Logging out now thoroughly clears local login state and device-binding information.
Login credentials are no longer stored in plaintext config files, making accounts more secure.
Fixed an issue where, in some cases, users couldn't use the app normally after logging in (model list and balance failed to load).

✨ Models & Experience

Added models including Gemini and GPT-5.x Codex.
Added a "reasoning effort" option to the input box.
Added a clear notice for network errors, making error messages easier to understand.

⚙️ Version

The macOS / Windows clients are unified at v0.5.0.

2026.06.15 · v1.0.1

🐞 Fixes & Improvements

Fixed device-pairing requests hitting the wrong server, which caused the app to be unusable after login (the pairing endpoint was pointed at the login service address instead of the gateway address, previously causing 404s / device not bound / model list and balance both failing to load)

✨ New Features

Added email Magic-Link login
Added client error telemetry (device-level, redacted, on by default, can be disabled in settings)

⚙ Engineering Changes

Client config directory moved to ~/.heicode (isolated from Claude Code, migrated automatically on upgrade)

2026.06.15 · v1.0.2

✨ New Features

Task Cockpit (Swarm visualization) launched: view and drive real Swarm runs — run list, state machine, task graph, event stream, artifacts, plus resource / budget / usage / audit details
"Launch Swarm" in the client: select a named configuration and fill in a goal to launch a Swarm task (named configurations are generated after binding resources in Manager)

🔧 Adjustments

Session "Agents" control: enabling Swarm and attaching a Cloud Agent are now mutually exclusive to avoid conflicts

🚀 Improvements

Reduced polling frequency for the Cloud Agent / session Agents list: drops to once per hour after the first successful fetch, significantly reducing background requests
Task results now distinguish between "no result yet" and "empty result", no longer mistakenly showing an empty result block

2026.06.15 · v1.0.3

✨ New Features

Added a "Documentation Center" entry to the user menu, for one-click access to the Heicode docs

🔧 Adjustments

Task Cockpit merged into Swarm: checking "Enable Swarm" in the session's "Agents" panel and selecting a Swarm configuration lets you launch a Swarm run just by sending a message (the message becomes the goal); run progress / pending approvals moved to the right-side "Swarm" panel (toggleable from the panel menu); the left-side "Cockpit" entry has been removed

🐞 Fixes & Improvements

Fixed occasional load failures in "Run Details" for a just-launched Swarm (gracefully degrades when the events / artifacts API is briefly unavailable, no longer causes a full-page error)

2026.06.16 · v1.0.4

🐞 Fixes & Improvements

Fixed a new install / upgrade from an older version going straight to the main UI without showing the login screen (user data is migrated but old credentials are not carried over; the first launch now goes through a proper login)
Fixed the download progress bar showing two overlapping / one short one long jittering progress indicators during a version update (downloads are no longer interrupted by periodic background checks, and are no longer duplicated)
Session switching is now faster and no longer gets stuck on "Loading session…"; fixed a blank screen after opening in certain cases

🚀 Improvements (Swarm)

Checking "Enable Swarm" from the initial page / a new session now creates a titled session and auto-navigates to it; Swarm results and deliverables are shown as replies within the session
The right-side "Swarm" panel can now be opened from any page (including Cloud Agent); Swarm event-stream labels are now human-readable (task created / running / retrying / completed…)
"Enable Swarm" is now a local toggle scoped to the current conversation (only shown once Swarm is enabled in Manager)

✨ New Features

Minimum supported client version: users are now prompted to upgrade if their version is too old

2026.06.17 · v1.0.5

🐞 Fixes & Improvements

Fixed the Files panel showing an "API error 403" (browsing the session's directory was blocked by the filesystem allowlist; the current session directory is now correctly allowed)
Fixed a Swarm session created earlier disappearing from the session list after restarting the client (Swarm sessions were mistakenly deleted by local CLI pre-warming as placeholder records; a dedicated container session is now always created fresh when launching)
Fixed the previous round's stale result being echoed when continuing to send commands in the same session after a Swarm task completed — subsequent commands after completion now start a new round, continuing on the workspace already updated by the previous round

🚀 Improvements (Swarm)

Opening the run panel now syncs to the final state immediately, no longer stuck showing "Running"; event-stream labels are more readable
"Pending Approvals" in the right-side "Swarm" panel now only follows the currently open task, no longer mixing in pending approvals from other tasks
The "Agents" control now correctly shows the "Enable Swarm" option once Swarm is enabled in Manager; removed the ineffective "Send command to this Run" input from the run panel

🚀 Improvements

Files / Terminal panels now open in the current session's directory (no longer the default directory)
Model configuration loads faster: the model list is now cached locally and shown immediately on launch, no longer briefly blank

2026.06.17 · v1.0.6

🐞 Fixes & Improvements

Fixed sending a slash command showing "CLI process failed to start / Provider not found" (when an old provider becomes invalid after login or a config-directory migration, it now automatically falls back to the currently logged-in account)
Fixed an occasional "An unexpected error occurred" when opening the "Session Inspector" via /status, /cost, or /context (read failures now degrade gracefully instead of causing a full-page error)
Fixed the usage panel always showing 0 when opening a historical session — it now backfills the session's real token consumption from session records
Fixed newly created sessions occasionally disappearing from the left session list, with the header not showing the project directory (list refreshes no longer discard the session currently in use)

🚀 Improvements

Added "Start dev" to the Preview panel: once a project directory is detected, run the dev script (npm/bun/pnpm/yarn) in the background with one click and auto-load the preview; "Stop" also ends the background process
Fixed the Preview panel blocking content with "this content has been blocked" when previewing a local dev service; no longer falsely reports "local service failed to start" when no service has been started, showing a friendly idle notice instead
Corrected the brand name in the Skills page copy (skills extend Heicode's capabilities)

2026.06.18 · v1.0.7

🐞 Fixes & Improvements

Fixed the model being unable to "see" or misreading images after sending them (including reading local images / screenshots) — images returned by tools are now delivered to the model in the correct multimodal format; also fixed a request failure caused by message-order issues when sending multiple images at once
Fixed PDF parsing: text-based PDFs now use pure-JS extraction (no longer depends on poppler, fixing parsing failures caused by "pdftoppm unavailable" on Windows); image-based / scanned PDFs are now handed directly to the model to read natively
Fixed an occasional "No such tool available" — handles the case where an upstream provider rewrites a tool name to Compat<name><random string>
Tightened spacing between paragraphs in assistant replies; the copy button now only shows at the end of the full reply

🚀 Improvements

Changed the default model from gpt-5.4 to qwen3.7-max
Improved model protocol routing: GPT / Qwen / Grok / DeepSeek now use the native OpenAI protocol, and Claude uses the native Anthropic protocol; fixed an issue where everything previously went through gateway conversion, causing attachment content like images / PDFs to be dropped
Corrected the brand name in the Skills page copy (skills extend Heicode's capabilities)

Images now display properly in conversations: images referenced with @, as well as images returned by tool reads / screenshots, now show thumbnails (no longer just a file path, and no longer a long block of encoded text); attachments and images are right-aligned in user messages

The session list no longer fills up with empty "Untitled Session" entries, and these empty placeholder files are now automatically cleaned up; the recent projects list loads faster

Added a fallback for UI crashes: a "reload" prompt is now shown on render errors, instead of a full blank screen

Update announcements now only show the current version's changes (no longer stacking historical versions)

🚀 Improvements

Improved the "needs your input" question cards: after answering, they auto-collapse into a one-line summary, the harsh white background is removed, and the overall look is more compact

Reduced misleading red exclamation marks: recoverable tool results and user-initiated cancellations are no longer marked red (downgraded to amber / neutral), with red reserved only for genuinely fatal errors

Faster startup: chart / diff components now load on demand; background polling is paused when the window is minimized, saving power and network usage

Updated execution permission-mode copy to "Request Permission" / "Accept Edits"

Images are now compressed by size to reduce unnecessary transfer (without reducing recognition quality)

2026.06.22 · v1.0.9

🐞 Fixes & Improvements

Fixed "Automation / Scheduled Tasks" not executing in the installed client — scheduled tasks previously only ran in the dev environment and did not actually run once packaged; they now execute on schedule as expected

Fixed a 400 error and a stuck, unrecoverable session caused by too many parallel tool calls in a single turn (exceeding the upstream limit of 128) when using models like GPT / Qwen / Grok / DeepSeek — the client now automatically splits these into compliant multi-part requests

Fixed the model configuration dropdown being empty in "Automation → New Task" — it now uses the same model list as the chat dialog

Fixed the "Skip Permissions" confirmation dialog: the popup is now centered (no longer offset); pressing Esc now only closes the confirmation dialog instead of accidentally closing the entire new-task form

🚀 Improvements

Simplified "Automation → New Task": removed the redundant required "Description" field, now only requires a task name and conversation content

Enhanced session management: supports archive / unarchive and duplicating a session (fork into a new session), plus new "Active / Archived / All" status filters

Added a "Preview" badge to the title bar and "Settings → About"

2026.06.23 · v1.1.0

✨ New Features

Added a "Context Window" section to the usage info panel: visualizes the current model's context usage (category breakdown + a color-coded progress bar graded by usage ratio), collapsible, and auto-updates when switching models
Clicking a file path in a conversation now shows the file's content directly in the right-side panel
Sessions now support full-text search by content (not just titles), and an entire conversation can be exported as Markdown (via right-click menu or /export)
Plan mode: when the AI produces an execution plan, it is now shown as a dedicated approval card, letting you approve and execute it in one click or continue refining it

🚀 Improvements

Hid the top session tab bar, unifying session switching via the sidebar for a cleaner UI

🐞 Fixes & Improvements

Fixed usage info sometimes showing 0: opening a historical session now shows the real cumulative token consumption (consistent with the Session Inspector)
Fixed occasional "CLI failed to start": transient startup failures are now automatically retried once, reducing the need for manual retries

2026.06.24 · v1.1.1

🐞 Fixes & Improvements

Usage info: fixed "input · output · total tokens" showing 0 or only the current turn after switching to a historical session / switching models — it now correctly shows the cumulative usage for the entire session; non-Claude (e.g. GPT) sessions no longer have their entire session usage zeroed out due to a pricing calculation error
Context window: fixed "used" showing 0 when restoring a session (now aligned with the category breakdown below), and changed to load in seconds (showing an estimate first, then refreshing with real-time data)
Fixed a 400 error when switching from GPT to Gemini in a session; fixed Gemini reply input tokens not being counted

🚀 Improvements

Added an edit menu to the input box's right-click menu (undo/redo/cut/copy/paste/select all); restored "copy" on right-click after selecting conversation content; the session right-click menu is now more compact
Added a warning before switching models across model families, reducing the risk of session corruption
You can now choose the save location when exporting a conversation as Markdown

2026.06.25 · v1.1.2

🐞 Fixes & Improvements

Fixed a lingering spinner (e.g. "Vibing…") remaining below the "option bar" (the prompt for you to choose an answer) in Plan mode after it appears
Fixed small-context-window models (e.g. qwen-max) having their context mistakenly "compressed" on the very first message, causing token consumption to nearly double — the compression threshold is now set sensibly based on the model's context window size, with no behavior change for large-window models
Fixed the question option bar sometimes not appearing in Plan mode
Windows: fixed the Bash tool reporting "No suitable shell found"; it now automatically detects the bash bundled with Git for Windows, and gives install guidance when it's missing

2026.06.26 · v1.1.3

✨ New Features

Added "Session Call Chain": in the usage info panel, you can now view each turn's model, duration, tool calls, and results for the current session, making troubleshooting more intuitive
Queued messages are now editable: a follow-up message queued while a reply is in progress can now be edited directly before sending
Swarm (requires the plan to be enabled): the Swarm feature panel is now fully wired up (sending/receiving tasks, approvals, and results via reverse proxy through HM); users without the Swarm plan enabled are unaffected and other features work as usual

🚀 Improvements

Long-session performance improvements: scrolling and rendering are noticeably smoother for long sessions with many messages (windowed virtual rendering); Markdown rendering and component re-rendering were also optimized
More thorough permission skipping: after choosing "Skip Permissions", the model no longer interrupts mid-task to switch into Plan mode asking for manual plan confirmation
Question cards auto-collapse: the "Heicode needs your input" card now collapses into a one-line summary after you answer (whether via an option or direct input), taking up less screen space

🐞 Fixes & Improvements

Stability & compatibility: fixed a 400 error ("tools array too long") when sending messages with OpenAI-compatible models (e.g. Qwen); slow or misbehaving providers no longer leave the conversation stuck indefinitely on "…"; the OpenAI-compatible path now uses a stable prompt-cache key for better cache hits

2026.07.01 · v1.2.0

✨ New Features

Swarm multi-turn conversation memory: you can now ask follow-up questions within the same Swarm without losing context
Added new output styles: built-in "Default / Explanatory / Learning" styles
Added LaTeX/KaTeX math formula rendering
New sessions can now be isolated by branch/worktree (running in a dedicated git worktree)
@-mentions now extend from files to skills/sub-agents

🚀 Improvements

Reworked the Swarm cockpit: the task panel now docks in the Dock, and approval cards are shown inline within the conversation (collapsing after a decision); task details are now split into "Overview / Execution / Subtasks" tabs; details and historical sessions load faster

🐞 Fixes & Improvements

Now shows a notice when device pairing hits its limit (previously failed silently)

2026.07.02 · v1.2.1

✨ New Features

Added "Suggested Next Steps": after each reply, suggested follow-up questions are shown above the input box — click to send immediately (Settings → General toggle, off by default)
Clicking the project name label above the input box now opens the working directory in the file manager, or copies the directory path with one click

🚀 Improvements

Unified the width and margins of new-session / in-session dialogs, so switching no longer causes layout jumps
Session cards (tool calls/thinking/tasks) now have a subtle shadow and clearer outline, for a softer visual look

🐞 Fixes & Improvements

Fixed "Suggested Next Steps" being incorrectly suppressed and never appearing under proxy billing setups

2026.07.08 · v1.3.0

✨ New Features

The window can now be freely resized horizontally, fixing an issue where the borderless window could only be resized vertically; the conversation area now fluidly widens with the window (max width raised to 1600px)
Added an "Output Style" selector (Default / Explanatory / Learning) to the "+" menu on the new-session page, so you can choose it before starting a session
Added /btw for a temporary side-thread: type /btw <question> to open an independent side thread that doesn't enter the main conversation context; also accessible from the "Panels" menu in the title bar
Support dragging files directly into the input box to add them as attachments
Added startup-time diagnostics (Settings → Diagnostics); shows a clear notice when a reply is truncated by the output limit

🚀 Improvements

Noticeably faster startup: improved session-tab restoration, cutting roughly 1 second off cold start

🐞 Fixes & Improvements

After restarting the client, attachments in historical messages are now restored as clean cards (path attachments/extracted content/images) instead of raw text or a large stacked block

2026.07.09 · v1.3.1

🐞 Fixes & Improvements

Fixed "update failed": after the update source's storage address was migrated, the updater now points to the new address, restoring auto-update
Fixed /btw temporary side-threads occasionally showing up in the session list
Fixed being unable to paste directly on macOS via right-click "Paste" without first granting system permission

2026.07.10 · v1.3.2

✨ New Features

Sessions are now isolated per logged-in account: you only see your own sessions; sessions from before the upgrade are automatically assigned to the current account and are not lost
The model picker now shows both official models and your custom models (BYOK) in a single dropdown: pick an official model to use the official channel, or a custom model to use your own key
Added "Custom Models" settings: add a third-party provider (with its own baseUrl/API key); once saved, the model automatically appears in the picker

🚀 Improvements

Session titles are now generated in the language you're using (a Chinese conversation gets a Chinese title); very short messages are used verbatim as the title instead of an English summary
After switching sessions, unsent drafts in the input box are preserved and still there when you switch back

🐞 Fixes & Improvements

Fixed request failures caused by a duplicated /v1 when the third-party model's address already included it
Fixed the 128-limit error (tools array/tool_calls) triggered by too many tools in long tasks — conversations no longer get stuck mid-task and no longer require a restart

2026.07.14 · v1.3.3

A stability and experience release — we recommend updating soon.

🚀 Improvements

Permission confirmation card: after you confirm, it auto-collapses into a compact bar and supports number keys 1 / 2 / 3 for quick selection (Reject / Always allow / Allow once)
Clearer authorization button wording: "Just this once" is now "Allow once"
Desktop notifications: after you confirm a permission in-app, the matching reminder in the notification center is cleared automatically (Windows)

🐞 Fixes & Improvements

Fixed intermittent errors when calling GPT-family and some third-party models, for more stable conversations
Fixed upload failures when pasting / selecting images that were caused by incorrect format detection
On a gateway timeout (e.g. 504), the client no longer pops up a long English error page — it now shows a concise, clear message
Fixed a possible interruption when using an Agent in a non-Git directory; it now falls back to the current directory and continues automatically

2026.07.20 · v1.4.0

The biggest update since 1.3.x, focused on being smarter, steadier, and more worry-free.

✨ New Features

"Auto" model in the main conversation: no need to agonize over which one to pick — it automatically chooses a suitable model based on your question (whether it includes an image, whether it needs deep thinking, and context length)
Built-in Agent browser: let the AI open web pages and look things up for you; the Preview panel shows in real time what it's browsing, and you can take over and click directly
Account menu: click your avatar to open settings or sign out
Automatic regional route selection: automatically picks the nearest mainland / overseas access point for faster, more stable connections
Enterprise managed policy: admins can centrally control the models / tools / network available to the team

🛡️ Stability Improvements

Model glitches no longer freeze the app: automatic retry when the upstream is unstable; auto-abort with one-click retry when there's no response for a long time; and one-click switch-model-and-retry — no more being stuck forever on "Thinking"
Automatic reconnection: clear prompts on network fluctuations and safe resumption after recovery; a Swarm task's badge no longer gets stuck on "In progress" after it finishes
More understandable errors: clear prompts for no seat / organization suspended / signed into the wrong organization, instead of just "Invalid token"
More accurate balance: shown by available credit (wallet + subscription), so no more "you have credit but it says you're out of money"

🐞 Fixes & Improvements

A batch of experience fixes: Ctrl+V to paste files directly, drag files into the input box, read Excel / Word / PPT, scroll within an extra-long input box, one-click scroll to bottom, permission-card number keys 1 / 2 / 3, and more

2026.07.21 · v1.4.1

A maintenance update after v1.4.0, adding one new panel plus a batch of experience improvements and fixes.

✨ New Features

Code review panel: a structured GUI that presents review findings (comparable to Bugbot / Codex review), making issues clear at a glance
Inline diff line comments: select a line in the diff to start a comment; added session review; when the selected model has no available channel, it automatically switches to an available one (single hop, no manual action needed)
Update announcements now support both Chinese and English, shown by your client language; a language selector was added to settings

🚀 Improvements

On startup, it goes straight to the new-session home page instead of automatically reopening the last session
The session list is sorted by time: an older session you re-open jumps to the top immediately
Region is now selected manually: mainland China / overseas (default mainland China); this only affects the chat access point and does not affect login
The stall watchdog was relaxed so it no longer misjudges tasks that are "slow but still running"

🐞 Fixes & Improvements

The skip-permissions dialog now shows the real working directory (no longer a literal "~")

2026.07.23 · v1.4.2

This release brings the desktop settings page, team / personal account switching, and a batch of model and preview fixes.

✨ New Features

Desktop settings page: a system-tray toggle and run-on-startup; branded installer, with run-on-startup on by default for new installs
Team / personal account switcher: switch context from the account menu; billing and available models follow the active context

🚀 Improvements

Context window shows its real size instantly (million-token models no longer flash 200k then jump)
"No available channel" errors now auto-switch to an available model (including in streaming) instead of showing raw JSON
When a site's anti-bot wall blocks WebFetch, the client suggests browser mode instead of failing silently
The model picker hides embedding models (not usable for chat)
Region-aware update source (China mainland via 21Vianet) for more reliable updates

🐞 Fixes & Improvements

Browse-preview fixes: an actionable hint when there's no session, correct per-session launch, and no more infinite spinner
Fixed request failures caused by the Opus-series thinking parameter
macOS: fixed chat request errors for some models (GLM / Qwen / DeepSeek / Grok)
UI cleanup: removed tool/agent card status icons; fixed model-name overflow in the toolbar when the right dock is open

2026.07.23 · v1.4.3

🐞 Fixes & Improvements

Hardened Content-Type handling for encrypted requests so a malformed header can't cause model chat failures (Windows was not affected — a preventive fix aligned with the macOS root-cause fix)

2026.07.28 · v1.4.4

This release focuses on never losing content to an interruption or an error, and completes the built-in browser and managed-policy work.

🚀 Improvements

Stop no longer discards what the model already produced; Retry keeps the interrupted answer (marked as such) and appends the new one below
Built-in browser polish: connect without a manual download, a settings toggle, per-session isolation, and no more accidental external-browser launches; switching to another browse-enabled session re-prepares its browser
Quitting now cleans up the sidecar process tree and leftover browser instances (no orphan processes or held ports), with no console flash or hang
The team/personal switcher is hidden when only one context is available
Managed policy: closed the outbound-domain gap while the policy is still loading; orgs can require high-risk operations to be refused when the policy is unavailable; auto mode / bypass-permissions mode can now be disabled centrally

🐞 Fixes & Improvements

Fixed mid-stream upstream errors being swallowed as a blank answer, and truncated response streams leaving the turn stuck on "thinking"
Fixed inflated +/- line counts on diff cards (a new file could show -1); "Copy path" now copies the plain path
Fixed the task panel picking up another session's tasks when switching quickly; the model list cache is cleared on logout so a different account can't briefly see the previous one's models

2026.08.12 · v1.4.5

This release is about runaway tool loops actually stopping and screenshots actually reaching you, plus solid Computer Use and error reporting.

🚀 Improvements

Rewrote the runaway-tool-call breaker: loops that rotate arguments on retry, or interleave other tool calls, are now caught too (it previously only covered strictly consecutive identical calls — the rarest shape a real loop takes)
Screenshots are visible and savable: images in tool results render inline, open full-screen, and can be saved to disk
Computer Use: system apps such as Finder / System Events are no longer reported as "not installed", and the authorization dialog no longer offers a button that has no effect when approved
Preview serves local pages over a local static server, so same-directory CSS / JS load correctly; the address stops working the moment the panel closes
"Retry with another model" hands the choice back to you instead of picking one and resending
Organization announcements and organization-level CLAUDE.md standards: bilingual with severity levels; injected standards merge with your own file instead of replacing it, and the UI names a contact when a standard is truncated
Longer prompt-cache retention, cutting rewrite cost substantially

🐞 Fixes & Improvements

Fixed hitting the turn limit being reported as an error: it is now a quiet notice, with no red error bubble, no useless "Retry" button, and no stacked notices across consecutive turns
Fixed sync failures showing only an error count: causes are now grouped and explained
Fixed upstream rate limits (429) delivered inside an HTTP 200 stream being shown as a generic failure
Fixed the synthetic "No response requested." placeholder appearing as an empty bubble, and session message counts disagreeing with the actual content
Fixed Opus 4.6 / 4.7 / 4.8 output limits being capped at 32k (the official limit is 128k), which truncated long answers
Fixed conversation context menus and floating layers being clipped at the bottom of the window; the queued-message list no longer pushes the conversation out of view
Fixed resumed sessions losing their attachments, leaving the model unable to see images / files it had sent
Fixed live token counts reflecting a single request instead of the whole turn; fixed context compaction receiving only part of the available space, and interrupt handling occasionally killing a freshly started turn
Fixed the "Retry with another model" button never having appeared at all (the model list it read was always empty)
macOS: fixed encrypted requests still using the legacy plaintext protocol on the chat-proxy and provider-test paths, which broke chat for some models (GLM / Qwen / DeepSeek / Grok)
macOS: fixed the Computer Use permission probe running in a virtual-env python subprocess, so permissions granted in System Settings still showed as missing