xg-gh-25/SwarmAI

SwarmAI is a desktop-based agentic operating system that orchestrates autonomous software development and content production through a modular skill architecture. It integrates a governed context engine, code intelligence graph, and self-evolution tracking to manage project knowledge and automated workflows. The system provides a unified interface for chat, file management, and external channel integration, supported by a local SQLite backend and a suite of specialized skills for document generation, web automation, and media processing.

A new EvolutionBadge component has been added to the Swarm Radar panel to display the count of successful evolution sessions. The badge shows the total number of evolutions and provides a breakdown by trigger type (reactive, proactive, and stuck) via a hover tooltip, helping users monitor the system's self-evolution activity directly within the chat interface.

desktop/src/pages/chat/components/radar · high confidence

The s\_project-manager skill now includes a new verify\_ddd\_complete gate that checks DDD project completeness based on the specific assets a brain governs. This gate requires code-intel.json only for brains governing a code-repo asset; data-agent and pure-knowledge brains pass without it. It also supports both the new six-section numbered layout and the legacy bare layout for canonical documentation paths.

_backend/skills/s\project-manager/scripts · high confidence

The Briefing Hub now includes a new 'Working' section that displays actionable items sourced from email, Slack, and calendar events. This feature presents items with priority indicators and source labels, allowing users to click an item to automatically populate the chat input with relevant context. The implementation introduces a shared utility for building this context and a reusable component that is integrated into both the Welcome Screen and the Radar Sidebar.

desktop/src/pages/chat/components/briefing · high confidence

The RightSidebar's history view has been upgraded to allow searching through the actual content of past conversations, not just their titles. When a user types a query, the system performs a backend full-text search and displays the matching sessions in a flat list, replacing the previous time-grouped view. Additionally, clicking a session in this new history overlay opens a read-only preview of the conversation in-place, rather than navigating away to a new tab. The component also retains a legacy client-side title filter for uncontrolled usage.

desktop/src/pages/chat/components/RightSidebar · high confidence

The DOCX skill has been updated with new document style standards. It now includes an 'Amazon Narrative Format' for creating 6-pagers and LT review docs, featuring dense prose, bold lead assertions, and specific appendix structures. Additionally, an 'XG document style standard' has been added, enforcing defaults like Calibri 10.5pt body text, 0.7-inch margins, and KaiTi for Chinese text. The skill also includes comprehensive instructions for using the docx-js library, OOXML schema validation, and redlining workflows.

_backend/skills/s\docx · high confidence

The repository is initialized with core documentation files (AGENTS.md, AI\_CONTEXT.md, CHANGELOG.md, CODE\_OF\_CONDUCT.md) and a comprehensive .gitignore that excludes private project data, internal skills, test artifacts, and environment files. A .secrets.baseline file is added to track known secret fingerprints for the security gate, establishing the initial security posture for the codebase.

(repo-wide) · high confidence

This change introduces the backend channel subsystem, enabling the agent to connect to external messaging platforms like Slack via a new gateway and adapter architecture. To ensure channels remain active 24/7, a macOS launchd daemon is installed to keep the backend running independently of the desktop app, featuring a guardian watchdog for self-healing restarts and automatic log rotation. The system includes native Slack streaming with reaction controls, per-channel model overrides, and a Slack-native owner approval UI for managing trusted senders. Security is strengthened with an egress redactor that automatically strips credentials and exfiltration URLs from outbound messages, and strict per-channel session isolation with TTL-based eviction.

backend/channels · high confidence

The app now includes a CodeGraph component that renders a force-directed visualization of the code intelligence graph, showing top connected symbols as nodes colored by module. This graph is exposed via a new 'View Graph' button in the BottomBar's Code Intelligence popover. To support embedding the graph within the Brain Hub detail pane without breaking existing fullscreen callers, the CodeGraph component now accepts an additive \inline\ prop that switches its outer container from a fixed fullscreen overlay to a relative, parent-filling layout.

desktop/src/components/layout · high confidence

The backend now includes a new evolution subsystem that tracks correction classes (e.g., CLASS\_A) to monitor and close self-evolution loops. This change introduces a \CorrectionClassTracker\ to record correction events, monitor gate effectiveness, and auto-resolve classes after 30 days of silence. It adds a \canonical\_class\_key\ normalizer to fix key-drift issues where the same logical class was split across different entries. A \closed\_loop\ module provides a Goodhart guard to distinguish genuine evolution (fewer mistakes) from metric gaming (logging less), and a meta-test to verify the integrity of the feedback circuit. Additionally, a \governance\_router\ routes classifications to the tracker or a pending queue, and a \gate\_scaffold\ generates fail-open gate stubs for human review.

backend/core/evolution · high confidence

Adds the initial backend implementation for SwarmAI Hive, enabling the provisioning, lifecycle management (deploy, stop, start, update, delete), and update operations for cloud instances on AWS EC2 and CloudFront. The new code includes the HiveProvisioner for managing AWS resources (IAM, EC2, S3, CloudFront), user-data scripts for automated instance bootstrapping, and core module initialization.

backend/hive · high confidence

This change introduces two new state management modules in the desktop application. The MessageStore replaces the previous fragmented message state handling with a centralized, per-tab source of truth that enforces phase-gated operations (idle vs. streaming) to prevent race conditions and reconciliation errors during chat sessions. It includes a watchdog timer to handle stream stalls and optimizes rendering performance by gating updates for background tabs. Additionally, the TerminalStore provides a module-level registry for terminal tabs, managing the lifecycle of PTY processes (spawn/kill) independently of React's rendering cycle to ensure stability across tab switches and React StrictMode double-mounts. These stores are exposed via a new index and a React hook (useMessageStore) for reactive UI updates.

desktop/src/stores · high confidence

SwarmAI now ships with a standardized, 11-file context system (backend/context/) that defines the agent's identity, operational rules, security standards, and memory model. This includes AGENT.md for execution directives and mandatory coding gates (R1–R32), SOUL.md for core principles, and a new SECURITY-BASELINE.md enforcing secure-coding rules (A1–B8) on every change. A new CONTEXT.md provides a ubiquitous language glossary to ensure consistent terminology across all 69 skills, while MEMORY.md, EVOLUTION.md, and KNOWLEDGE.md establish a structured, agent-owned memory and correction registry. This replaces ad-hoc context management with a governed, cross-file contract that standardizes how the agent reasons, codes, and remembers.

backend/context · high confidence

Added md\_freeze.py, a new script that enables safe translation of Markdown documents by freezing fenced code blocks behind sentinels so the LLM only translates prose, then stitching the original blocks back in to preserve byte-identical content. The tool provides freeze, stitch, and verify commands to split, reassemble, and structurally verify documents, ensuring no silent drift in code, JSON, or file trees during translation.

_backend/skills/s\translate/scripts · high confidence

The backend now includes a project-scoped code intelligence platform that builds a dependency graph for each project, enabling structural decomposition and business-rule extraction. This change introduces language-agnostic graph clustering to identify structural communities and extraction candidates, alongside a topology analysis layer that detects high-degree 'god nodes' and surprising cross-domain connections. It also adds a deterministic business-rule manifest generator that derives per-domain rules from graph edges and source-level reconciliation calls. A new graded incremental re-analysis system allows the graph to be updated efficiently by distinguishing between cosmetic, structural, and full updates based on file signatures. Additionally, a PreToolUse hook injects dependency context (symbols, callers, routes, risk scores) into agent sessions, while a codebase map provides a concise summary of the project's structure, routes, and hot files at session start.

_backend/core/code\intel · high confidence

A new \backend/hooks\ package has been added to execute automated tasks when a session closes, decoupling these operations from the critical chat path. The package includes \WorkspaceAutoCommitHook\ for smart, conventional git commits with code-review gating, \DailyActivityExtractionHook\ for summarizing conversations, \DistillationTriggerHook\ for archiving and curating memory entries, \ContextHealthHook\ for maintaining knowledge indexes, \EvolutionMaintenanceHook\ for pruning stale evolution entries, and \ImprovementWritebackHook\ for capturing lessons into project documentation.

backend/hooks · high confidence

Adds several new React hooks and utility modules to the desktop frontend: AudioKeepAlive to prevent macOS audio session teardown in WKWebView, useAudioPlayer for Web Audio API playback, and useAttentionQueue for the unified 'Need You' channel. It also introduces streaming-guards and streaming-machine to enforce send-queuing logic and manage the chat lifecycle via a state machine, alongside fileChangedBroker and railSsot to centralize file-change event dispatching and path-matching logic for the Canvas rail.

desktop/src/hooks · high confidence

The narrative-writing skill (v1.3.1) now provides a structured authoring workflow for seven document types—six-pager, PR/FAQ, HLD, LLD, ADR, SOP, and project plan—each with a dedicated template and guided questions. It enforces a five-criteria quality bar, requires reader testing and stakeholder simulation via sub-agents, and mandates weasel-word and AI-ism checks using a provided script and reference lists before delivery.

_backend/skills/s\narrative-writing · high confidence

A new skill named 'repo-to-ddd' has been added to the backend, allowing users to transform any codebase into a structured, AI-ready format. The skill produces a \.ai-context/\ directory containing DDD-structured artifacts (PRODUCT.md, TECH.md, IMPROVEMENT.md, PROJECT.md, code-intel.json) and an AGENTS.md entry point, designed to help AI agents navigate, understand, and safely modify the project. It supports multi-language repos, monorepo fan-out, and optional signal ingestion (docs, wikis, Slack) to achieve higher understanding levels. An installer script is included to deploy these artifacts into various IDEs (Claude Code, Cursor, VS Code, etc.).

_backend/skills/s\repo-to-ddd · high confidence

The desktop/public area now includes a new Brain Hub demo page (brain-hub-demo.html) that serves as a left-nav overlay mockup for the phase-1 Brain Hub design, featuring a wizard for creating new brains, build progress tracking, and code-intel reuse visualization. Additionally, Chinese-language versions of the evaluation architecture (eval-architecture-zh.svg) and sequence (eval-sequence-zh.svg) diagrams have been added to support multilingual documentation, mirroring the existing English diagrams with localized text and font stacks.

desktop/public · high confidence

Added a new 'Browser Agent' skill that enables DOM-based browser automation using Playwright. This skill allows users to navigate websites, read compressed page content, click elements, fill forms, extract data, and take screenshots. It supports multi-tab workflows, persistent sessions via Chrome DevTools Protocol (CDP), and integrates with MCP tools for simple tasks while falling back to the script for complex interactions. The skill includes a self-contained setup that installs Playwright locally and provides commands for launching, navigating, interacting with elements by index, and managing browser state.

_backend/skills/s\browser-agent · high confidence

This change introduces a suite of new scripts in the \scripts/\ directory to enforce quality and security gates. \check\_commit\_trailers.py\ enforces the \Co-Authored-By: Swarm\ identity in CI, bypassing shadowed local git hooks. \lint\_doc\_frontmatter.py\ validates required metadata fields in design docs. \lint\_skills.py\ ensures the published skill corpus contains no unresolvable private identifiers or hardcoded paths. \security\_scan.py\ and its shell wrapper provide a decoupled security gate that checks for high-severity bandit findings, hardcoded secrets, and wildcard CORS regressions using fingerprinted baselines. Finally, \smoke\_e2e.py\ adds post-deploy end-to-end verification, including checks for stuck streaming sessions and mid-stream disconnect recovery.

scripts · high confidence

A new DDD skill, s\_repo-to-ddd, has been added to provide a self-contained mechanism for regenerating the code-intel.json projection from a bound repository's source code. This capability ships a portable CLI (refresh\_code\_intel.py) and helper scripts that operate using only Python standard library and git, ensuring no dependency on the SwarmAI backend. The tool scans real import graphs to produce a v2 code-intel.json file, strictly adhering to a narrow refresh mode that never modifies the core DDD documentation. It resolves output paths dynamically based on the $SWARM\_WORKSPACE environment variable or falls back to a repo-adjacent location, ensuring portability across different host environments.

_backend/templates/ddd-skills/s\repo-to-ddd · high confidence

A new \s\_ddd-distribute\ skill and its \scripts/distribute.py\ orchestrator have been added to package a grown DDD into distributable capability packages (AIM capabilities or Open-Plugins). The tool enforces a strict reach model: it reads the DDD's declared targets from \aim.json\, allows emitting only a subset of those declared targets, and refuses to widen reach or publish externally unless visibility is explicitly set to \external\. It includes a \--check-stale\ mode to detect when source changes have not been re-emitted, preventing silent staleness, and a \--with-enablement\ flag to optionally ship class-A enablement skills for bare foreign hosts. All packaging logic is delegated to core modules, with this script handling argument parsing, policy validation, and human-in-the-loop confirmation.

_backend/skills/s\ddd-distribute · high confidence

Introduces the \s\_ddd-persist\ skill, which allows agents to sediment knowledge into a DDD's documentation with strict routing discipline. The skill includes a bundled \locked\_write.py\ engine that ensures safe, concurrent read-modify-write operations on Markdown files using file locks, preventing data corruption. It enforces a decision-tree for persistence (admission gate, governance check, content routing) and provides mechanisms for retiring or moving entries safely, ensuring that only stable, decision-relevant knowledge is stored additively.

_backend/templates/ddd-skills/s\ddd-persist · high confidence

A new 'html-artifact' skill has been added to generate single-file HTML artifacts for reports, code reviews, comparisons, and scorecards. The skill enforces a dual-consumer protocol: agents always receive a complete inline markdown summary in chat, while the HTML file serves as a supplementary, shareable artifact for human consumption. It includes a design system (base.css) with a 'print magazine' aesthetic, four complexity levels (static to interactive), and specific templates for reports, code reviews, comparisons, and scorecards. Additionally, a Playwright-based HTML-to-PDF conversion script is provided as the sole sanctioned method for PDF generation, replacing the previously unreliable Chrome headless CLI approach.

_backend/skills/s\html-artifact · high confidence

The Pollinate skill now generates presentations as self-contained HTML decks using a new build layer. This adds a reusable \<deck-stage\> web component for slide navigation, keyboard/touch controls, and auto-scaling, along with a Playwright-based export script that produces PDFs with preserved clickable hyperlinks. The engine ships with 34 design systems (including 8-Bit Orbit, Biennale Yellow, BlockFrame, Blue Professional, and Bold Poster) that define typography, colors, and components, and includes shared animation patterns and a base HTML template to ensure consistent, zero-network-render output.

_backend/skills/s\pollinate/templates/html-deck · high confidence

Added the s\_library skill, providing a CLI interface to mount external directories as pointers (without copying data) and search the resulting index. The skill distinguishes between code directories, which are indexed into a symbol graph, and documentation directories, which are chunked and indexed into the shared FTS5 knowledge base at mount time, making them immediately recallable. It supports listing registered mounts and searching across library domains, acting as the agent-facing entry point to the core library mount engines.

_backend/skills/s\library · high confidence

The PDF skill now includes a suite of new scripts to handle form data and document conversion. For PDF forms, the system can detect fillable fields, extract field metadata, and populate forms either by updating native PDF form fields or by adding text annotations for non-fillable documents. It also provides utilities to convert PDF pages to images and generate visual validation overlays for bounding boxes. Additionally, a new \md2pdf.sh\ script enables converting Markdown to PDF using either a Tectonic (LaTeX) or WeasyPrint (HTML) engine, supporting professional and minimal styling templates, table of contents generation, and batch processing.

_backend/skills/s\pdf/scripts · high confidence

A new PPTX skill has been added to the backend, enabling the system to create, edit, and analyze PowerPoint (.pptx) files. This skill provides a comprehensive set of tools for building presentations from scratch using an HTML-to-PPTX workflow, modifying existing slides, restyling content to match specific themes, and extracting text or generating visual thumbnails. It includes detailed instructions for handling Office Open XML structures, validating documents against ISO-IEC 29500-4 schemas, and managing design elements like typography, color palettes, and layout patterns.

_backend/skills/s\pptx · high confidence

Added two new PreToolUse hooks to improve tool usage safety and reliability. The pytest safety hook blocks full test suite runs unless explicitly requested via the SWARMAI\_SUITE=1 environment variable or specific test file filters, and sanitizes commands by removing unsafe piping (tail/head) and normalizing Python paths. The SPA domain hook prevents WebFetch from returning degraded content on known single-page application domains (e.g., xiaohongshu.com, weibo.com) by blocking the request and suggesting a curl alternative with a mobile user-agent.

.claude · high confidence

The application now includes dedicated React context providers to manage distinct UI and system states, each accompanied by comprehensive test coverage. LayoutContext centralizes workspace scope, modal management, and session metadata, persisting the workspace scope to localStorage. OverlayContext establishes a single-source-of-truth for fullscreen surfaces, bridging legacy window events to a unified overlay registry while ensuring mutual exclusion. TerminalContext manages PTY tab lifecycles and panel visibility, persisting panel state to localStorage. HealthContext exposes backend health status and triggers on-demand checks. ThemeContext handles system preference detection, with specific tests ensuring resilience against missing matchMedia APIs. These contexts replace scattered state management, providing a structured foundation for the redesigned layout and overlay subsystems.

desktop/src/contexts · high confidence

The Pollinate Studio backend now includes a suite of Remotion templates for generating video content, introducing support for both horizontal (16:9) and vertical (9:16) aspect ratios. These templates provide a visual editing interface via Remotion Studio, allowing users to customize colors, typography, and layout through a defined schema. The package includes specialized compositions for standard videos, short-form vertical content (optimized for platforms like Bilibili), and platform-specific thumbnails (16:9, 4:3, 3:4, 9:16). It also features a library of reusable UI components for rendering data visualizations, code blocks, animated diagrams, and audio waveforms, along with robust error handling and localization support for Chinese and English content.

_backend/skills/s\pollinate/templates/remotion · high confidence

The Settings page now includes several new tabs to manage core functionality. The AI & Models tab allows users to configure AWS authentication (SSO, Ada, API Key, Bedrock), manage available models, and adjust thinking mode settings. A new Channels tab provides controls to set up, edit, and disconnect Slack integrations. Users can now manage data safety via a Backup tab that displays status, configures repository URLs, and triggers manual backups or restores. The Capabilities tab exposes runtime module availability (e.g., Vector Search, Slack Bot), while the Core Engine tab offers a dashboard of growth metrics, memory effectiveness, and learning state. Additionally, the About tab has been updated to show version info, platform details, and a built-in update checker with download and restart capabilities.

desktop/src/components/settings · high confidence

A new 'Web Design Review' skill has been added to the backend, enabling the system to audit frontend code against accessibility (WCAG 2.1 AA), semantic HTML, responsive design, performance, and design quality standards. The skill integrates a local BM25 search engine backed by a 99-rule UX guidelines database (derived from UI UX Pro Max) to provide evidence-based, specific Do/Don't guidance during audits. It also incorporates a shared 'design-judgment' skeleton from the frontend-design skill to evaluate info-density and surface patterns. Users can trigger this review via commands like 'review my UI' or 'check accessibility', receiving structured reports with file:line references and severity levels (Critical, Warning, Info).

_backend/skills/s\web-design-review · high confidence

A new XLSX skill has been added to the backend, enabling the agent to create, edit, and analyze Excel files (.xlsx, .xlsm, .csv, .tsv). The skill provides detailed instructions for financial modeling standards, such as color coding and number formatting, and guides the agent to use Excel formulas rather than hardcoded values to ensure dynamic spreadsheets. It includes a new \recalc.py\ script that leverages LibreOffice to recalculate formulas and detect common Excel errors (e.g., \#DIV/0!, \#REF!), and supports data analysis via DuckDB for SQL-based exploration and pandas for data manipulation.

_backend/skills/s\xlsx · high confidence

The assets directory now includes a comprehensive set of new visual and documentation resources to clarify the system's architecture. A new HTML design document (SwarmAI-Architecture-Design-Doc.html) provides a high-level overview of the agentic OS. Several new SVG diagrams have been added to illustrate specific components: the autonomous pipeline (aidlc-autonomous-pipeline-v4.svg) details the 9-stage, 3-gate workflow; the context engineering diagram (context-engineering.svg) explains the 11-file priority chain and token budget management; the DDD three-layer stack (ddd-three-layer-stack.svg) visualizes the interface, intelligence, and orchestration layers; and the eval architecture (eval-architecture.svg) outlines the decoupled evaluation system. These assets serve as updated reference materials for the platform's engineering and operational structure.

assets · high confidence

The \.kiro/hooks\ directory now includes four new user-triggered hooks that automatically analyze and fix code changes. The 'Code Quality Scan & Auto-Fix' hook scans for code smells, anti-patterns, and best-practice violations, automatically applying fixes for High and Medium severity issues. The 'Docs Update Scan & Fix' hook detects when documentation (README, AGENT.md, specs) needs updating based on code changes and applies those updates. The 'PE Design Review, Revise & Propagate' hook reviews newly created design.md files for architectural, security, and correctness issues, revising the design and propagating High/Medium fixes to related requirements and tasks. The 'Security Leak Scan & Auto-Fix' hook scans for hardcoded secrets, credentials, and other security vulnerabilities, automatically fixing Critical severity issues by replacing them with environment variables or config references.

.kiro/hooks · high confidence

The backend exposes a suite of new REST endpoints to power the Radar, Library, and Project overlays. The attention router aggregates the unified 'Need You' queue from multiple sources, while the library router provides live filesystem reads for the Native store and mount points. Pipeline and Pollinate routers serve real-time run state and content-asset gallery data to the Radar panel, and the projects router handles full CRUD for workspace projects. Additional routers introduce DDD cultivation proposal approvals, autonomous job status monitoring, and recall metrics visibility, with all heavy I/O offloaded to worker threads to prevent event-loop blocking.

backend/routers · high confidence

The backend now includes a dedicated set of scheduled job handlers in \backend/jobs/handlers\ to automate code intelligence, documentation maintenance, and system health. \code\_intel\_reindex.py\ manages incremental and full re-indexing of project code graphs with ownership guards and delegation for heavy tasks. \ddd\_refresh.py\ and \ddd\_self\_audit.py\ autonomously detect stale DDD documentation and perform semantic drift reviews, while \ddd\_weekly\_report.py\ aggregates cultivation activity into weekly summaries. \conversation\_digest.py\ processes channel messages into DDD proposals on an opt-in basis. System health is maintained by \memory\_health.py\ (deterministic integrity checks and LLM-powered pruning), \eval\_scheduled.py\ (biweekly evaluation runs with drift alerting), and \library\_freshness.py\/\library\_health.py\ (periodic health checks for mounted libraries and knowledge files).

backend/jobs/handlers · high confidence

This change introduces four new scripts in the backend/scripts directory to enhance pipeline integrity and data correctness. artifact\_cli.py provides a standalone CLI for managing the artifact registry and pipeline runs, including logic to detect untrackable (gitignored) files and enforce completion gates. backfill\_cjk\_escapes.py is a safe, dry-run-first migration tool that re-encodes JSON columns containing literal \\uXXXX escapes into raw UTF-8, ensuring Chinese characters are correctly indexed in FTS search tables. ci\_eval\_gate.py acts as a quality gate for pushes, verifying that the latest eval report is fresh (matching the current code digest), green (BVT passed), and free of red-line violations. Finally, check\_migration\_class\_test.py adds comprehensive tests for the goal-run class-completeness gate, validating enumeration commands and migration status reconciliation.

backend/scripts · high confidence

The backend now includes a new \utils\ package providing shared utilities for core operations. \bundle\_paths\ centralizes logic for locating resource files across development and Tauri production environments, ensuring the daemon can find its configuration and data files regardless of how the binary is deployed. \file\_lock\ introduces cross-platform file locking (supporting both Unix and Windows) to prevent race conditions during concurrent access to shared files. \jsonl\_rotation\ adds automatic rotation for JSONL log files to manage disk usage. Finally, \mcp\_validation\ extracts and standardizes validation logic for MCP configuration entries, including security checks to prevent environment variables from pointing to the system database.

backend/utils · high confidence

This change introduces a suite of new React components for the chat interface, including AttachedFileChips for displaying and removing attached files, ChatDropZone for handling drag-and-drop file attachments, ChatErrorMessage for structured error display with retry and rate-limit countdowns, EvolutionMessage for rendering self-evolution events, ScreenshotButton for capturing the current screen, SubAgentProgressBanner for tiered sub-agent progress visibility, and VoiceConversationIndicator for voice mode status. It also adds comprehensive property-based tests for AttachedFileChips, ChatContext logic, and SwarmAgent invariants, while removing the legacy PermissionRequestModal.

desktop/src/components/chat · high confidence

This change introduces several new components to the chat interface. The ActivityFeed component provides a collapsible summary of tool actions (file creation, modification, commands run) per assistant message. The AlertsPill replaces the previous notification indicator with a 'Needs You' entry in the left sidebar that opens a fullscreen overlay for pending items. The BrainHomeView redesigns the Welcome screen landing with a two-tier hierarchy: 'Needs your decision' for brains requiring input and a 'Brain pulse' strip showing health status. Additionally, the AssistantHeader adds a branded SwarmAI label and timestamp to assistant messages, while the ChatHeader now displays a Canvas outputs pill when the canvas is closed and has pending outputs.

desktop/src/pages/chat/components · high confidence

This change introduces a suite of new shared UI components to the desktop application, including a \BackendStartupOverlay\ that manages the backend initialization phase with a wall-clock ceiling to prevent infinite loading, a \CredentialBanner\ that provides method-aware remediation for expired authentication (e.g., AWS SSO, IAM roles), and a \CommentPopover\ for inline review comments. It also adds a \FileEditorCore\ component that implements a large-file performance guard to prevent UI freezing by skipping synchronous syntax highlighting and diffing for files over 100,000 characters. These components are accompanied by extensive property-based and integration tests to ensure correct state transitions, accessibility, and behavioral correctness.

desktop/src/components/common · high confidence

Added the \s\_context-hygiene\ skill, which provides a methodology and a read-only scanner (\scripts/scan.py\) to clean and compress SwarmAI's 12 context files. The skill distinguishes between system files (requiring source edits and rebuilds), runtime files (editable directly), and auto-generated sections (which should not be hand-edited). The scanner mechanically surfaces candidates for removal—such as echoed titles, volatile drift numbers, and dated pointer fragments—so the agent can read and judge each item before deletion, ensuring that judgment kernels and principles are preserved while noise is removed.

_backend/skills/s\context-hygiene · high confidence

The frontend-design skill now includes a comprehensive set of curated design data assets to guide UI/UX generation. This includes 12 slide presets, a 34-template bold index, animation mood mappings, and anti-AI-slop rules to prevent generic outputs. Additionally, it provides extensive color palettes for 161 product types, chart selection guidelines for 23 data visualization types, and a design-judgment framework to improve layout decisions. All assets are properly attributed to their open-source sources (UI UX Pro Max, frontend-slides, beautiful-html-templates) under MIT licenses.

_backend/skills/s\frontend-design/data · high confidence

Added a suite of shell scripts in \desktop/scripts/\ to streamline local development and cross-platform builds. \dev-with-backend.sh\ automates starting the Python backend and Vite dev server for standalone Tauri development. \build.sh\ and \pre-build.sh\ handle frontend dependency installation and asset synchronization, with \pre-build.sh\ specifically addressing Windows \cmd.exe\ compatibility for Tauri's \beforeBuildCommand\. \tauri-build.sh\ is a new wrapper that conditionally disables updater artifact signing for local developers (who lack the signing key) while preserving it for CI, and forces the \CI\ environment variable to bypass a Finder-related DMG detachment bug on macOS. \build-backend.sh\ was updated to support Windows targets and use \uv\ for dependency management. A test script (\test\_tauri\_build\_wrapper.sh\) validates the signing-key logic.

desktop/scripts · high confidence

The \backend/skills/s\_frontend-design/scripts\ area now contains the core logic for a new frontend-design skill. This includes \core.py\, which implements a BM25 search engine over local CSV data for UI/UX style guides, colors, charts, and other design domains; \design\_system.py\, which aggregates these search results to generate comprehensive design system recommendations; and \search.py\, a CLI entry point that exposes these capabilities via command-line arguments, including options for domain-specific search, stack-specific guidelines, and persistent design system output generation.

_backend/skills/s\frontend-design/scripts · high confidence

The Pollinate skill now includes a dedicated, self-contained Text-to-Speech (TTS) module that supports multiple synthesis backends (Azure, Amazon Polly, Edge, CosyVoice, Doubao, ElevenLabs, OpenAI, and Google). Users can select their preferred provider via environment variables or user preferences, with Edge TTS available as a free, keyless default. The engine features advanced SSML processing for improved pronunciation, including Chinese polyphone disambiguation, abbreviation expansion, and automatic language switching for English terms within Chinese narration. It also provides precise word-boundary tracking for accurate subtitle (SRT) generation and section timing synchronization for video production.

_backend/skills/s\pollinate/scripts/tts · high confidence

The file viewer now supports previewing a wider range of file types through dedicated renderers. Audio and video files play natively using browser controls, while images support zoom and pan. PDFs are rendered with page navigation and zoom controls, and CSV/TSV files are displayed as sortable, virtually-scrolled tables. HTML files are shown in a sandboxed iframe with a source-view toggle and a fit-width mode for wide reports. Unsupported file types display a unified info card with context-aware actions like opening in the system app or attaching to chat.

desktop/src/components/file-viewer/renderers · high confidence

Added the \s\_notify\ skill, enabling users to send messages to nine notification channels (Feishu, DingTalk, WeCom, Telegram, Email, ntfy, Bark, Slack, and generic webhooks) via a unified \send\_notification\ function. The skill reads configuration from \\~/.swarm-ai/notify-channels.yaml\, automatically converts markdown messages to channel-specific formats, and can be triggered by users or invoked by other skills and scheduled jobs.

_backend/skills/s\notify · high confidence

The desktop application introduces a comprehensive set of new service modules in \desktop/src/services\ to support the redesigned Radar, Community, and Jobs overlays. The \attention.ts\ service replaces the previous fragmented frontend merge of paused pipelines, jobs, and waiting tabs with a single backend aggregation layer (\GET /api/attention\) for the unified 'Need You' channel. The \community.ts\ service provides read-only access to the Community overlay's feed, sources, and engagement metrics, ensuring honest truncation flags are surfaced to the UI. The \jobs.ts\ service exposes scheduled job statuses, including consecutive failures and error reasons, to power the Radar's attention queue and Jobs & Runs section. Additionally, new services for \hive.ts\ (cloud instance management), \codeIntel.ts\ (codebase health and graph visualization), \ddd.ts\ (Brain Hub health and review data), \channels.ts\ (external channel CRUD), \evolution.ts\ (SSE event types), and \logForwarder.ts\ (frontend log persistence) are added, alongside constants for explorer-to-chat communication in \explorerEvents.ts\.

desktop/src/services · high confidence

This change introduces a new signal feed adapter system in \backend/jobs/adapters\ to populate the product's signal pipeline with diverse, real-time data. New adapters include \eastmoney\_market\ for Chinese A-share market movers, \trending\ for Chinese social media hot searches (Weibo, Zhihu, etc.), \weibo\_trending\ for keyword-matched Weibo posts, \github\_trending\ for trending repositories, \github\_releases\ for tracked project updates, \github\_community\ for scanning specific repos for engagement opportunities, \hacker\_news\ for AI-filtered stories, and \rss\ for standard feed parsing. To support these external fetches securely, a shared \http\_client\ module is added, featuring an SSRF egress guard that validates resolved IPs to block private, loopback, and metadata addresses, and enforces HTTPS-only connections. The implementation also includes retry logic for transient network errors and disables environment proxy variables to ensure predictable, thread-safe outbound traffic.

backend/jobs/adapters · high confidence

A new \todo\_db.py\ script has been added to the \s\_radar-todo\ skill, enabling direct read/write access to the SwarmAI \data.db\ SQLite database. This tool bypasses the backend API, allowing agents in sandboxed environments to manage todo items (add, list, get, update, status, delete) without needing network connectivity to a dynamic port. It resolves the database path using the \SWARM\_DATA\_DIR\ environment variable (with a deprecated fallback to \SWARM\_APP\_DATA\_DIR\ or \\~/.swarm-ai\) to ensure it writes to the same store the backend reads, and outputs structured 'work packets' containing context like next steps, acceptance criteria, and related files for immediate agent execution.

_backend/skills/s\radar-todo/scripts · high confidence

A new Python script has been added to the backend audit tools to detect mismatches between command-line arguments defined in skill scripts and those documented in INSTRUCTIONS or SKILL markdown files. This tool scans skills for argparse flags and reports which ones are missing from the documentation, helping maintain consistency between code and user-facing docs.

backend/scripts/audits · high confidence

The workspace settings interface now includes dedicated tabs for configuring skills, Model Context Protocol (MCP) integrations, and knowledge bases. The Skills tab allows users to enable or disable workspace-specific skills, with safeguards for privileged capabilities. The MCP settings panel consolidates management into two sections: toggling catalog integrations and managing personal/dev MCP servers. The Knowledgebases tab provides a full CRUD interface for adding, editing, and deleting knowledge sources (local files, URLs, indexed documents, etc.). These components replace the previous scattered configuration UIs and are backed by new service methods for workspace configuration.

desktop/src/components/workspace-settings · high confidence

The Outlook Assistant skill now includes three new shell scripts to manage email lifecycle and user configuration. The deletion-log.sh script tracks deleted emails in a local JSON file, supporting viewing, searching, and clearing of deletion history, while also migrating legacy data from the old config location. The preferences.sh script introduces a structured markdown-based system for storing user preferences (such as important senders, folder rules, and behavioral settings), with commands to view, edit, and set these preferences. The restore.sh script leverages the deletion log to help users identify and restore accidentally deleted emails by providing instructions for the Outlook MCP move\_email tool. These changes enhance the assistant's ability to manage email triage and organization with user-configurable rules and recovery capabilities.

_backend/skills/s\outlook-assistant/scripts · high confidence

Added a suite of Python scripts in backend/skills/s\_pollinate/scripts to support the Pollinate content production engine. These include brand\_chart.py for applying direction tokens to Excel charts, check\_prereqs.py for validating environment dependencies, check\_rpv.py and check\_specs.py for automated quality and platform specification validation of video assets, convergence\_gate.py for enforcing poster HTML quality layers, cross\_format\_check.py for verifying consistency across multi-track outputs, deck\_notes\_injector.py for processing PowerPoint speaker notes, font\_link\_backfill.py for managing Google Fonts CDN links, and format\_recommend.py for recommending production tracks based on audience and context.

_backend/skills/s\pollinate/scripts · high confidence

The desktop application is rebranded from Owork to SwarmAI (identifier com.swarmai.desktop, version 2.1.0). This release introduces two new capabilities: an integrated terminal using a vendored PTY implementation for interactive shell sessions, and a one-tap screen capture feature that grabs the display under the mouse cursor for chat attachments. To support these features, macOS entitlements and usage descriptions are added for microphone and screen recording access. The app also now includes an auto-updater configured for GitHub releases and resolves PATH detection on macOS by sourcing the user's login shell.

desktop/src-tauri · high confidence

Introduces a new Slack channel adapter that supports both Socket Mode (WebSocket) and HTTP polling fallback. This allows the system to connect to Slack without requiring a public webhook URL, and automatically switches to polling if the WebSocket connection is blocked by corporate proxies or VPNs. The adapter also includes logic for handling Slack API limits, authentication errors, and message formatting.

backend/channels/adapters · high confidence

Introduces the operational foundation for deploying SwarmAI on EC2, including a Caddy reverse-proxy configuration that terminates traffic from CloudFront and enforces basic authentication, a systemd service unit with circuit-breaker restart limits, and a release packaging script that bundles the backend, pre-built frontend, and a seed dataset sourced from the live SwarmAI workspace. The entry also includes a package verification script to ensure release integrity and setup/update scripts that are now deprecated in favor of the automated desktop provisioner and API.

hive · high confidence

The FileViewer now uses a dedicated \useFileViewerTabs\ hook to manage tab state, introducing a strict dirty-guard: closing a tab with unsaved changes is a no-op until the caller explicitly clears the dirty flag, ensuring users are not forced to confirm before discarding edits. The implementation supports up to 10 tabs, automatically evicting the oldest non-dirty tab when the limit is reached, and preserves scroll positions across tab switches. A new test suite verifies this close contract to prevent regression.

desktop/src/components/file-viewer/hooks · high confidence

The terminal is now a persistent, VSCode-style bottom panel that auto-opens a single shell when first shown and supports numbered tabs with a + button and per-tab close. The panel collapses to display:none rather than unmounting, so scrollback and live PTYs survive collapse/reopen without re-spawning. Mouse-selection drift is eliminated by preloading the JetBrains Mono font before xterm measures character cells, and a focus race is fixed so keystrokes work immediately on first open. The layout is denser (font size 11, line height 1.0) with a full 16-color ANSI palette for legible output, and an Attach to Chat button lets you send the active terminal’s recent output to the chat.

desktop/src/components/terminal · high confidence

The previous split between the FileEditorPanel and BinaryPreviewModal has been replaced by a single, unified FileViewer component that supports multiple file types (images, PDFs, HTML, video, audio, CSV) via lazy-loaded renderers. The viewer now features a tabbed interface (FileViewerTabBar) allowing users to open and switch between multiple files within the same panel, with content caching to prevent re-fetching on tab switches. A new status bar (FileViewerStatusBar) displays file metadata and provides copy-path and attach-to-chat actions. The panel (FileViewerPanel) now uses a responsive default width based on viewport fraction, with smooth resize animations and per-tab state isolation to prevent cross-tab bleeding.

desktop/src/components/file-viewer · high confidence

The Workspace Explorer has been redesigned with a new SwarmWS-branded header that includes a search input, collapse/expand all, and sort toggle (default, name, git-first). The file tree now uses a 3-tier visual hierarchy with section headers for Knowledge (yellow accent) and Projects (blue accent), while System files are dimmed and collapsed by default. A new collapsible Working Files section at the top displays only user-content git changes (modified, added, untracked) from Knowledge and Projects directories. The explorer also features a new toolbar for creating new files/folders and uploading files, a context menu with rename, delete, copy path, attach to chat, and ask-swarm options, and improved visual properties like depth-based indentation and file-type icons.

desktop/src/components/workspace-explorer · high confidence

Introduces a new pure-ASGI authentication middleware (HiveAuthMiddleware) that enforces HTTP Basic authentication on all API routes in Hive mode, providing a defense-in-depth layer alongside the existing Caddy edge gate. This change ensures that privileged endpoints (such as workspace and job management) require valid credentials even if the network edge is bypassed, while remaining a transparent pass-through in non-Hive modes to avoid locking out local users. The implementation is designed to be fail-closed and compatible with streaming responses, preventing potential security risks identified by CloudSec.

backend/middleware · high confidence

The backend now enforces a hierarchical permission model (OWNER, TRUSTED, PUBLIC) for all channel messages, ensuring the agent respects sender capabilities. A new SenderIdentity data class provides the single source of truth for authorization, replacing any inference from message content. Additionally, outbound messages are automatically redacted for credentials and exfiltration URLs via the egress redactor before being sent, closing a privacy gap.

python · high confidence

A change to existing behaviour in backend/schemas/\_\pycache\\_ — 2 commits, 11 files.

(repo-wide), backend/schemas/\\pycache\\_ · medium confidence · unverified_

The desktop application's icon has been replaced with a new design featuring a stylized bee and hexagonal grid pattern, reflecting the updated SwarmAI brand identity. This change updates the visual appearance of the app in the system tray and application launcher.

desktop/src-tauri/icons · high confidence

The autonomous pipeline scripts have been split from a monolith into distinct, self-contained modules to improve maintainability and introduce new capabilities. The pipeline now includes a confidence scoring system (confidence\_score.py) that evaluates delivery readiness based on acceptance criteria verification, review findings, and TDD status, alongside a GoalMetrics module (goal\_metrics.py) for tracking cycle efficiency and velocity. A new auto-PR creation tool (pipeline\_pr.py) automatically generates pull requests with detailed delivery stats, while a WTF gate (wtf\_gate.py) assesses fix risk to prevent overly broad changes. Additionally, the pipeline enforces bounded sub-agent prompts via a spawn budget contract (spawn\_budget.py, spawn\_prompt.py) to prevent infinite loops, and resolves git default branches dynamically to support non-standard branch names like 'master'.

_backend/skills/s\autonomous-pipeline/scripts · high confidence

The backend test suite is now memory-safe, splitting execution into batches via a new safe\_test\_runner.py to prevent 9GB+ RSS spikes, with Hypothesis profiles standardized in conftest.py. Model configuration is centralized in a new model\_registry.py, replacing five drifted hardcoded copies with a single authority that makes claude-opus-4-8 the default flagship. Legacy AWS Bedrock tool-use demos and the old .env configuration have been removed, and the .env.example has been updated to reflect the new desktop-focused SQLite and Tauri settings.

backend · high confidence

The chat page logic has been restructured into pure, testable utility modules to fix session-leak bugs and unify tab-landing behavior. \landPromptInTab.ts\ and \resumeTarget.ts\ now provide a single, shared decision engine for where prompts and resumed sessions land, preventing the silent wiping of history-bearing idle tabs during dispatch. \sessionBinding.ts\ and \dispatchBackfill.ts\ enforce strict per-tab session ownership, stopping background tabs from inheriting a foreign live session (which caused perpetual thinking and stuck queues) and ensuring dispatched ToDo records are correctly backfilled with their session IDs. Additionally, \utils.ts\ introduces \mergeOlderMessages\ to fix pagination seams where assistant responses were split across bubbles, and \toDisplayMessage\ now preserves the \client\_id\ to prevent duplicate message bubbles during initial loads.

desktop/src/pages/chat · high confidence

The desktop application now includes a dedicated ESLint configuration (\eslint.contract.config.js\) that actively enforces the URL path contract, preventing double \/api\ prefixes in API calls by running a structural AST-level check in CI. This is accompanied by updates to the main \eslint.config.js\ to include this rule and ignore the \src-tauri\ directory. Additionally, the app has been rebranded to SwarmAI, reflected in \index.html\ (title, icon, FOUC prevention script) and \tailwind.config.js\ (font changes), while stale documentation (\BUILD\_GUIDE.md\, \backend.env.example\) and build artifacts have been removed.

desktop · high confidence

The desktop app now features a comprehensive light and dark theme system with CSS variables, including a warmer dark mode and improved sidebar icon contrast. Startup behavior is more robust: a new idempotency guard in the entry point prevents duplicate React roots on Tauri reloads, and the onboarding gate now correctly routes new users to the wizard even if the backend is not yet fully initialized. For developers, a new test setup file destroys MessageStores after every test to prevent Vitest hangs, and suppresses known JSDOM undici errors.

desktop/src · high confidence

The modal system has been restructured to support a unified Workspace Settings interface and new user-facing confirmation flows. WorkspaceSettingsModal now consolidates Skills, MCP, and Knowledgebases configuration into a single tabbed modal, replacing the previous separate Skills and MCP settings modals. Additionally, three new modals have been introduced: ConvertToTaskModal simplifies the task conversion workflow by removing multi-workspace selection in favor of a static SwarmWS indicator; PrivilegedCapabilityModal provides a standardized warning dialog for enabling privileged skills or MCPs; and RefreshContextModal offers a clear, non-technical explanation of the context refresh process to manage user expectations. The legacy SkillsModal and MCPSettingsModal files have been removed as their functionality is now integrated into the new workspace settings.

desktop/src/components/modals · high confidence

The backend database layer has been simplified to use SQLite exclusively, removing all DynamoDB integration code and configuration. The \dynamodb.py\ client module has been deleted, and the database initialization logic in \\_\init\\_.py\ no longer checks for a \DATABASE\_TYPE\ environment variable to switch between providers; it now always instantiates the \SQLiteDatabase\. Consequently, the abstract base class \BaseDatabase\ has been updated to remove the \skills\ and \skill\_versions\ table properties, while adding new properties for channel-related tables (\channels\, \channel\_sessions\, \channel\_messages\, \channel\_user\_identities\). This change eliminates the need for AWS credentials and cloud infrastructure for database storage, consolidating all data persistence to the local SQLite engine.

backend/database · high confidence

The shared type definitions in desktop/src/types have been significantly restructured to align with backend changes and new features. The Agent interface no longer includes per-agent sandbox configuration (sandbox is now app-level), and skill references are now named allowedSkills. The Skill type has been simplified to a filesystem-based model, removing version control and database-specific fields, and adding health status tracking. New type files have been introduced for chat threads (ChatThread, ChatMessage, ThreadSummary) and the ToDo domain (ToDo, ToDoCreateRequest, ToDoUpdateRequest, ToDoAttachment), providing a single source of truth for these entities. Additionally, workspace configuration types have been added for skills, MCP servers, knowledge bases, audit logs, and policy violations.

desktop/src/types · high confidence

The s\_steeringify skill is now deprecated and redirects users to the PROMOTE operation within the s\_self-evolution skill. While the original skill extracted recurring patterns from EVOLUTION.md corrections into STEERING.md rules using a cross-reference graph, the new approach unifies correction capture, pattern detection, and rule promotion into a single lifecycle manager. Users should now use s\_self-evolution PROMOTE, which adds intake classification, budget enforcement, bias-class awareness, and a candidate queue for sub-threshold proposals, replacing the previous standalone steeringify workflow.

_backend/skills/s\steeringify · high confidence

The \s\_persist\ skill introduces a unified routing system for saving knowledge, replacing the previous \s\_save-memory\ behavior that wrote everything to \MEMORY.md\. It now routes content based on type and project context to specific destinations: DDD docs (PRODUCT/TECH/IMPROVEMENT/PROJECT) for project-specific knowledge, \MEMORY.md\ for cross-project cognitive knowledge, \EVOLUTION.md\ for self-corrections, and \Knowledge/Library/\ for searchable reference facts. The skill includes an admission gate to prevent storing transient or drifting data, enforces file-layer ownership rules (e.g., governance changes require approval paths rather than direct writes), and provides a \ddd-retire\ CLI for proper entry retirement and archiving to maintain recall integrity.

_backend/skills/s\persist · high confidence

The macOS desktop application now includes a guardian watchdog that automatically recovers the backend daemon if it crashes or becomes deregistered, preventing extended outages. This is supported by a new launchd agent configuration and a Python-based guard module that uses intent sentinels to distinguish between accidental crashes and intentional stops. Additionally, unbounded log file growth is resolved through periodic log rotation in both the backend wrapper and the guardian scripts, and the daemon now runs with self-heal enabled by default.

desktop/resources/daemon · high confidence

Resolves a configuration error where a null \context\_token\_budget\ value caused failures during agent initialization. The \build\_agent\_config\ function now correctly filters out \None\ values, ensuring that only explicitly set budget parameters are passed to the agent, preventing crashes or invalid state when the budget is not defined.

backend/core · high confidence

The background job infrastructure has been consolidated into a new \backend/jobs\ package, replacing the previous workspace-level layout. The scheduler now runs in-process within the daemon's asyncio loop, eliminating the need for external \launchd\ plists and their associated authentication failures. A new \model\_resolve\ module serves as the single source of truth for job model selection, ensuring all headless tasks use the correct flagship model with the 1M context window suffix, which fixes the recurring '400 Input is too long' errors. Additionally, the system introduces a robust credential resolution strategy for AWS Bedrock that pre-resolves tokens to handle launchd contexts, and adds a \channel\_monitor\_fallback\ to maintain Slack monitoring via bot tokens when the primary MCP-based auth expires.

backend/jobs · high confidence

The i18n initialization module now explicitly registers the Chinese locale (zh.json) alongside English. This change restores Chinese translations for the entire application, correcting a regression introduced in the v1.0.0 rebrand where the Chinese locale was dropped, causing all Chinese translations to silently fall back to English.

desktop/src/i18n · high confidence

The estimate-tokens skill no longer relies on the previous shell-based word-count heuristic, which significantly under-counted CJK characters and used an incorrect 200,000-token window. It now delegates to the same calibrated estimator used by the prompt assembly, ensuring reported token counts match actual context usage. The default context window has been adjusted to 91,000 tokens (reflecting the effective budget for 1M models), and the script includes robust discovery logic to locate the canonical estimator in various deployment environments, failing loudly if the source cannot be found.

_backend/skills/s\estimate-tokens/scripts · high confidence

The workspace components now use CSS variables for colors, ensuring consistent theming across the FileBrowser, CodePreview, and the new FolderPickerModal. File preview reliability is improved by switching HTML rendering to a data: URL to fix blank frames in the Tauri WKWebView, and by unifying unsupported file handling with the shared UnsupportedRenderer. Additionally, the Tauri environment detection is corrected to use the canonical isDesktop() utility, and code copying is routed through a dedicated clipboard utility.

desktop/src/components/workspace · high confidence

Added E2E evaluation fixtures for autonomous pipeline, outlook assistant, radar todo, and memory skills; Added comprehensive test coverage for MessageStore lifecycle and race conditions; Added comprehensive test coverage for chat rendering, TSCC, and context UI; Added comprehensive test coverage for critical chat, streaming, and UI stability fixes; Added comprehensive test coverage for the Code Intelligence platform; Added contract tests for chat, session, and system API endpoints; Added runtime i18n integration tests; Added service-layer unit and property-based tests for API contract correctness; Added static analysis tests for graceful shutdown logic; Added streaming test utilities for multi-tab isolation; Added test coverage for ChatPage layout containment, send keying, and EvalDashboard features; Added test coverage for legacy code cleanup safety invariants; Added tests for AuthConfigPanel authentication method selection and verification; Added tests for FileViewer and FileViewerPanel behavior; Added tests for HTML deck track and style selection logic; Added tests for chat display bug conditions and preservation invariants; Added tests for chat streaming deduplication, error detection, and UI panel behavior; Added tests for chat tab dispatch, landing, and session binding logic; Added tests for keep-mounted per-tab chat views and cross-tab isolation; Added validation tests for frontend-design slide assets and previews; Expanded test coverage for backend components; Expanded test coverage for desktop hooks and state management; New evaluation dashboard and onboarding flow with comprehensive test coverage; New utility modules and comprehensive test coverage for desktop core features.

The application has been rebranded from 'Claude Agent Platform' to 'SwarmAI' across the backend, desktop, and root manifests. This release upgrades the backend to use \claude-agent-sdk\ version 0.2.109 and introduces new capabilities including Slack integration (\slack-bolt\), voice transcription (\amazon-transcribe\), and code intelligence tools (\tree-sitter\, \watchfiles\). The desktop application now supports automatic updates via the Tauri updater plugin, features a unified file viewer with PDF support, and includes a new integrated terminal using \portable-pty\. Additionally, the Python test suite now defaults to serial execution to prevent deadlocks, and the backend version has been bumped to 2.1.0.

(dependencies) · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• First survey — no prior run to compare against. CAI 56.

• Code Health 65
• Architecture 87
• Maturity 81
• Readiness 56
• Security 68
• Accessibility 49

• 300 commits — 202 feature/other, 98 fixes

• backend/core — 73 commits
• backend/tests — 49 commits
• backend/skills — 34 commits
• backend/context — 32 commits
• desktop/src — 20 commits
• (root) — 14 commits
• backend/hooks — 13 commits
• backend/scripts — 12 commits
• Projects/SwarmAI — 11 commits
• desktop/src-tauri — 9 commits
• docs/CODEBASE_METRICS.md — 7 commits
• backend/jobs — 5 commits
• .github/workflows — 4 commits
• backend/templates — 3 commits
• backend/hive — 2 commits
• backend/routers — 2 commits
• desktop/resources — 2 commits
• desktop/scripts — 2 commits
• docs/decks — 2 commits
• scripts/check_commit_trailers.py — 2 commits

• fix: Close R32 self-enforcement gap: add A4 wildcard-CORS regression check to
• fix: Diagnose and fix per-turn latency regression: user reports every tool-ca
• fix: FIX size-valve entry-splitting bug + restore gutted corrections. The val
• fix: Fix Hive S3 bucket-naming drift + supply closure (direction Y)
• fix: Fix a false-positive in the loop_age starvation probe (run_69198f8c, jus
• fix: Fix completion_gate.completion_surface_verdict so a run in a gitignored
• fix: Fix ddd_packager.py AIM capabilities-package emit-layer non-compliance (
• fix: Fix host-specific hard-coded output path + doc/impl drift in DDD-native
• fix: Fix open_vec_db self-contradiction (HIGH-3, XG). vec_db.py open_vec_db:
• fix: Fix stale architecture content in the TSCC Flow tab + OWNER map (TSCCMod
• fix: Fix streaming long-turn false-timeout P0: retry_manager marks a live-but
• fix: Fix stubborn orphan Memory Index block in MEMORY.md: route all 5 distill
• fix: Fix the Capabilities > Connections panel to honestly reflect what it IS
• fix: Fix the large-file-crashes-Canvas causal chain (B+C+A closed loop): B=DD
• fix: Redesign C&M overlay Memory+Evolution tabs (backend+frontend). THREE fix
• fix: Root-fix cold-spawn serialization latency: delete MCP_CONNECTION_NONBLOC
• fix: bugfix(pipeline): test-stage auto-aggregate 对称补齐 — 修 cross_boundary_e2e completion gate 不对称坑
• fix: bugfix: Tests cross-poison each other via un-restored asyncio event loop
• fix: bugfix: judge-telemetry.jsonl has no rotation/size-cap — it grows monoto
• fix: bugfix: memory distillation silent failure — distillation_hook skips [UN
• …and 280 more

• 0 containers · 1 bounded contexts · 0 dependency edges (baseline)

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

xg-gh-25/SwarmAI was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.