Release notes

This release brings together product updates since the previous public stable release, 0.1.23. It focuses on reading long conversations and agent work, sources and file previews, Missions and scheduled tasks, protection for edits, Expert Squad dispatch, and model configuration. Version 0.1.24 was an unpublished candidate; the complete changes in this stable upgrade are listed below.

Added

• Mission draft editing: edit a pending request after creation, discard edits, or save and review it. Saving checks the current draft; a concurrent update produces a conflict notice and preserves your input so an older page cannot overwrite a newer request.
• Copy individual sources: Sources entries can copy a complete web URL or file path and line range. Copy feedback remains visible as the streamed answer grows.
• Open source images and PDFs: image sources open in the image preview and PDF sources open in the paginated reader. PDFs support text selection and copying, and retain the reading position on return. Decoding errors provide a clear explanation and recovery action.
• Locate local search results: code-search matches can open their actual file and line. Counts, context and paths come from the real search results.
• Search loaded logs: find text in currently loaded log records. The filtered display and copy action use the same results while preserving original records and levels. This does not search the complete log history.
• Optional Parallel Search setup guide: English and Chinese Model Context Protocol (MCP) documentation adds anonymous search and page-fetch examples, including streamable-http, oauth: false, Fast mode and rate-limit handling. Real credential-free connections verified searching and then fetching a result page.

Changed

Expert Squads and task execution

• Preset workflows are optional references: squads retain explicit roles and capability boundaries. The coordinator uses the actual request and delivery evidence to select experts, divide work, change order, continue investigation or request review, without mechanically running every node in a preset graph.
• Search before generating a default squad: when no squad is selected manually, the native assistant and Missions search and inspect existing squads first. Built-in Dynamic organizes a squad for the request when no close match exists. Manual selection takes precedence, and existing Tasks retain their fixed package identity.
• Multiple independent executions by one expert: work is distinguished by its actual Task, Session and dispatch identity. Direct dispatch and continuing existing work are supported; completion depends on actual delivery and review evidence.
• Authorized general Skills are directly available: the general Skill loader is included in the resident tool set, reducing cases where a model knows a capability exists but cannot load it directly. Specialized extensions still use the existing capability catalog.

Conversations and workspace

• One long-text rendering path: Markdown parsing runs in a worker. Growing paragraphs, lists, tables and code blocks use the same parser; live events use streaming compression that can produce immediate output.
• Clearer source hierarchy: main conversations, side conversations and agent docks consistently show source titles, websites, sections, filenames and line ranges. Full excerpts open as readable body content, reducing cramped tooltips and repeated location labels.
• Activity follows actual execution: preparation of tool arguments, execution, stopping, failure and completion are distinguished. Multiple inputs received during one reply appear in their actual execution batches and retain completion duration.

Fixed

Long conversations, side conversations and agents

• Fixed verified cases where completed long Markdown remained at Rendering… after returning to a page, content height restored incorrectly, or expanding sources repeatedly rebuilt text and controls.
• Fixed jumps to the bottom or lost reading anchors after opening sources, reading upwards, switching Tasks, or reopening docks. Source expansion, excerpt and tool-body scroll positions, and keyboard focus are retained.
• Fixed missing history or reference relationships after reopening side conversations, and missing retry actions after lookup failures. Submission acceptance and panel closure settle against the actual request.
• Fixed lost text prefixes after reconnection, missed deltas and replay snapshots, and incorrect restoration of completed agent text, source membership or empty text states.
• Fixed finished inputs remaining Pending, unclear activity after stopping, and completion timestamps being overwritten by later rendering.
• Improved keyboard reading: tables, code sources, read-only files and diff views can receive focus. Home/End, list selection, menu closure and tool-tab switches retain the appropriate reading and focus positions.

Web sources, files and previews

• Fixed fetched addresses replacing actual page titles, missing section fragments, unreadable international URLs, long links hiding website identity, and repeated titles. Display names and actual navigation addresses are retained separately.
• Fixed Windows absolute paths, original attachment names, exact filenames, line positioning in mixed-line-ending files, and stale whole-file reference highlights. Ordinary code terms and bare filenames retain their literal meaning.
• Fixed placeholders remaining after empty text attachments loaded, missing diagnostics after file reload failures, and previews retaining old content. Read-only sources and editable file states are clearer.
• Fixed lost selections, focus or previews during live file-tree and diff updates. Initial directory loading, failure and recovery have explicit states.
• Fixed existing chart artifacts remaining constrained to their original width in fullscreen and Chinese fonts differing from the app. Compact and fullscreen views use the same actual artifact.

Editing, settings and model configuration

• Fixed unsaved file and artifact edits being overwritten by navigation, reloads, external changes or late responses. File-save shortcuts perform actual writes; version conflicts retain input and explain how to proceed.
• Fixed settings search destroying scheduled-task and network drafts when temporarily switching pages. Clearing search returns to the preserved editing state.
• Fixed changes to a Provider display name losing environment variables, model capabilities, limits or other settings. Full model IDs are retained: qwen3-next:80b is no longer truncated to qwen3-next.
• Fixed model-discovery responses writing old results into a new form after closure, reopening or a project switch. Saved models and reasoning options temporarily absent from the catalog remain identifiable and can explicitly return to the project default.
• Fixed unavailable model configuration preventing read-only access to historical conversations, events, files, Git information and worktrees. Configuration errors required for execution are separated from reading existing work data.
• Fixed connection-draft saving, service-address validation and recovery actions. Valid address paths and empty usernames are retained, and results belong to the actual connection and workspace.
• Improved access to inputs and controls in narrow desktop windows. Global search finds Missions, and settings navigation opens the correct conversation workspace. Interface language, editor text, sidebar width, browser tabs and external addresses are retained according to their actual state.

Missions, scheduled tasks and runtime recovery

• Fixed retrying an accepted first dispatch after it consumed a Mission draft being mistaken for a missing draft or affecting a later draft. Request replay refers to the execution already accepted.
• Fixed renaming scheduled tasks changing their start date, weekdays or next run. Start date, weekdays, time and timezone are saved explicitly; UTC calendars show their timezone and complex rules retain their original values.
• Fixed lost original requests and participant-evidence boundaries after context compaction, duplicate sources, and coordinators losing access to earlier worker text. Continuation uses actual persisted messages and sources.
• Fixed the old Windows service taking down a prepared replacement when it exited during a standalone-service restart. Native request timeouts, exit reasons and physical process settlement retain their actual ownership.
• Fixed squad installation scope, built-in identity, fixed Task selection and immutable package publication consistency. Embedded squads participate in localized search, and completed package publication settles promptly.

Tool results, errors and delivery

• Fixed generic conversion changing original tool argument types. Strings, numbers, booleans and other inputs retain the tool’s declared contract. File and Skill pagination reports actual read ranges and complete line counts.
• Fixed byte truncation splitting UTF-8 characters, damaged plain-text web structure, and incorrect HTML/XHTML content-type handling.
• Fixed tool summaries losing actual read names, paths or failure text. Search-service diagnostics, complete multiline errors and failed history are retained; error details support pointer and keyboard copying.
• Fixed streamed service errors showing only opaque internal information, missing conversation failure reasons after Sources, and empty delivery cards when there was no actual content.
• Fixed the website reading built-in Dynamic from the wrong directory. All five built-in squad archives and Registry data generation were restored. Packages can legally omit a static workflow, with corresponding production-build and database-lifecycle validation completed.

User impact and known limitations

• Back up runtime data before upgrading. Storage structure and squad-dispatch contracts changed. Incompatible databases refuse startup with SCHEMA_RESET_REQUIRED and require an explicit rebuild. There is no automatic historical migration; do not rebuild without retaining a copy.
• Check task-creation arguments in custom integrations. The low-level Software Development Kit (SDK), HTTP task.create and panel_create_task require the selected squad’s promptProfile explicitly. The command-line interface (CLI) selects a squad with --expert-squad. Automatic selection in the default assistant does not imply implicit selection in low-level calls. Custom squads may now omit preset workflows.
• Sources and long conversations have undergone multiple real streaming and page checks. Continuous rendering and switching in complex long conversations remain under investigation; not every flicker or jump is considered resolved.
• Parallel Search is an optional third-party service enabled by users. Anonymous access has low rate limits. Wait before retrying when limited; quotas and authentication follow the service provider’s rules.
• Version 0.1.24 stopped before public release because the website build failed. This version passed native builds and installer acceptance on all five platforms, and published its official artifacts and update channel. The combined pipeline exceeded its 40-minute limit at the website stage; a separate complete deployment subsequently delivered the site. Public delivery was verified, while the original timeout/cancellation result remains recorded.

Version scope

• These notes are based on actual source changes from v0.1.23 to v0.1.25. Validation, documentation and generation commits are not counted as separate features. Development after v0.1.25 is outside this release.