agentgraft

Widget API malformed requests returned HTTP 400 instead of server errors.

API tenancy security denied access with HTTP 404 to prevent user enumeration.

Voice input supported Japanese and Chinese and detected accented letters correctly.

Webhooks returning non-object JSON no longer crashed turns; features stabilised.

Dashboard flash messages rendered correctly on every page following redirects.

Signup and account form fields enforced validators for username, password, and slugs.

Display formatting no longer incorrectly split paragraphs.

Message deletion counts correctly counted only messages, ignoring attachments.

Memory backfill command classified unclassified historical messages and updated user mood.

Added internationalisation in seven languages for site, docs, dashboard, and widget, with voice mode support.

Added language-prefixed URLs, redirects, and navigation switchers for multilingual support.

Updated widget to support language selection via settings and Accept-Language headers.

Fixed text selection inside the widget on Safari to avoid collapsed range issues.

Fixed workspace page to correctly count each tool once per agent.

Added SDK function `ag.graft(palette=...)` to configure widget accent colours like "jade" or "indigo" to match themes.

Added a route strip UI to confirm or override destinations before transmission.

Users could customize accent colours via SDK or a visual picker in settings.

Avatar face featured liquid simulation to visually represent resource availability.

Sidebar header stopped displaying agent name, and send button required valid input.

Eliminated the separate global composer input box in the sidebar to streamline interface.

Added energy level and mood speech bubble to the header face.

Updated sidebar search button to display visible text label.

Announced route changes via live region for screen readers.

Removed emotion tile strip from the landing page.

Deployed new winking-mark favicon and navigation icons.

Improved production stability by applying explicit resource limits and tuning worker configurations.

Preserved dashboard-managed API credentials during deployment to prevent accidental deletion.

Updated landing page with sitemap.xml and robots.txt for improved search indexing.

Redesigned Memories settings page with iOS-style grouping, topic map, and local date rendering.

Updated Memory decay logic to run daily, preserving visitor edits and adding a 14-day grace period.

Fixed hidden scrollbar in settings sheet and removed duplicate history entries for single memories.

Added Deepgram Flux TTS integration allowing streaming audio replies using Bree model.

Enhanced voice input with background barge-in listening and automatic follow-up dictation.

Added animated avatar face expressions synchronized with audio states like speaking and listening.

Implemented FIFO audio queue system to serialize chat replies and inject bridge audio.

Updated settings modal to allow configuration of voice commands, pause phrases, and language.

Added visual countdown timer overlay for spoken follow-up actions.

Added browser wake-word detection and customisable wake phrase settings for voice input activation.

Added speech engine selection controls in settings for Text-to-Speech and Speech-to-Text.

Changed read-aloud resource citations to render naturally instead of exposing raw database IDs.

Updated settings modal with new radio card interfaces for selecting speech providers and managing audio permissions.

Added Launcher face mood integration and read-along highlighting for speech engine.

Repositioned and simplified the widget usage meter while retaining accessible details.

Redesigned the public site, documentation, authentication, widget harness, and dashboard.

Turned the landing page into interactive demonstrations of key product capabilities.

Added branded recovery pages for common 400, 403, CSRF, 404, and 500 errors.

Added image attachments to chat, including paste, drag-and-drop and image-only messages.

Added upload previews, progress indicators, failure handling and controls for removing attachments.

Added GIF support with grouped frame previews and labelled image context for agents.

Added image galleries, full-screen viewing, reattachment and flexible deletion of images and descriptions.

Added privacy-preserving image support for private chats and image capability handling across agents.

Added SDK support for image attachments, including up to five images per message and GIF uploads.

Changed image-upload synchronisation to preserve server settings unless hosts explicitly enabled or disabled uploads.

Documented image-upload configuration, including Safari CSP requirements and production examples.

Added a paginated Uploads library in Settings for managing images across all conversations.

Clarified image deletion: message-level removal affects one occurrence; Uploads removal affects all uses.

Added a choice to remove an image and its description, or retain the generated description.

Improved the image viewer with zoom, pan, fit/close toggling, controls, and keyboard shortcuts.

Extended viewer navigation across all thread images and added GIF playback controls.

Turned assistant references such as “Image 2” into clickable image chips.

Added “Reading your image(s)” feedback while image descriptions were being prepared.

Updated Delete all my data to explicitly include images and attachment data.

Added a responsive drawer sidebar for narrow panels, with scrim, Escape dismissal and automatic closing after thread selection.

Applied the mobile layout to short landscape touch devices to prevent cramped desktop-style interfaces.

Prevented floating chat-header controls from overlapping by progressively condensing or hiding lower-priority controls.

Improved header and message layouts to accommodate floating controls, long banners and scrolling content.

Corrected panel and composer sizing so fullscreen and resized panels fit accurately without cumulative growth.

Simplified the automatic-routing composer with a sparkles icon and the clearer “Ask — picks right chat” placeholder.

Made quota status visible in text, including remaining capacity, limits and refill times.

Streamlined thread-list scanning with clearer highlight counts, roomier titles and less prominent section metadata.

Added touch-friendly controls for pinning, archiving, deleting, promoting and closing threads.

Prevented iOS focus zoom and made the mobile composer start at one line before expanding with text.

Added safe-area spacing for notched devices across the header, messages, sidebar and composer.

Contained overscroll in search results, thread lists, modals and other scrollable regions.

Allowed reference chips to wrap naturally and added keyboard activation while preserving modified-click behaviour.

Added a streaming event for embedded widgets to signal when the assistant started or finished processing.

Added streaming-status notifications for background conversations and private chats, even when the chat panel was closed.

Updated embed documentation with available widget events and a page-level assistant-busy example.

Embedded widgets notified host pages when agents started tool calls via new tool_call events.

Embedded widgets notified host pages when tracked tool calls completed via tool_result events.

Tool-activity events were unavailable for private chats to protect tool-call metadata.

Added per-profile @ag.on_tool_call hooks for auditing, authorisation, metrics, retries and custom responses.

Added the public ag.ToolCall object with structured invocation, user, agent and request details.

Added ag.current_tool_call() for accessing the active tool invocation during hooks and tool execution.

Added invoking agent attribution to tool webhooks, including support for shared and private-agent tools.

Changed hook execution to run after request validation and support deterministic nesting and explicit HTTP errors.

Fixed tool argument validation to return 400 before execution and correctly report in-tool TypeError failures as 500.

Updated SDK documentation with the tool-call hook lifecycle, agent attribution and revised webhook behaviour.

Changed the embedded chat widget to render inline and display TeX mathematics as MathML.

Changed widget mathematics rendering from KaTeX to Temml, with safer currency handling and fallbacks.

Removed the widget's post-render KaTeX enrichment pass.

Added support for Pydantic v2 models as Python SDK tool parameter annotations.

Added actionable HTTP 400 validation errors for invalid structured tool inputs.

Released SDK version 0.7.0 with documented Pydantic model parameter support.

Added settings to close the chat after sending and reopen it when a reply arrived.

Preserved typing context while the chat was closed and indicated activity through the launcher.

Made highlights across formatted inline content appear as one continuous highlight.

Made copying, quoting and hovering multi-fragment highlights apply to the complete passage.

Displayed tool activity names as more readable, capitalised labels with identifiers available on hover.

Removed the incomplete-generation status and “Generation stopped” indicator from loaded message history.

Added the ability to fork from specific messages, allowing the creation of parallel branches.

Enhanced message area with multi-column branch viewing and branch-aware actions.

Introduced a user interface for managing branches including archiving, deleting, promoting, detaching, and renaming.

Updated editing, truncation, and resend functionalities to target the correct branch lane.

Expanded thread list to show branch hierarchies and indicate active branches.

Restored last active thread on reload.

Smarter unread-tail indicator now updates during/after streaming.

Hover-preview for highlight chips added.

Improved text selection in message area.

Visible “stream stopped” recovery affordance added.

Enhanced code block presentation with header bar and copy/wrap controls.

Introduced soft line-wrapping toggle for code blocks.

Improved thread list visual state with clearer status markers.

Added recall last prompt in composer for easy message retrieval.

Simplified prompt buttons for consistent preset behaviour.

Adjusted enter-to-send for IME composition and reliability.