asystant_ai 0.12.0

An embeddable Flutter assistant with local tools and reactive_notifier.

0.12.0

• The round limit is optional. AsystantTurnLimits.unlimited() lets a turn run until the model answers without tool calls (or the person stops it); AsystantTurnLimits.maxRounds is now int?, null meaning no limit. Breaking for code that reads maxRounds as an int.
• AsystantAI.setTurnLimits(limits) changes the limits from the next turn on without reconnecting, so a host can let the person turn the limit on or off. A turn already running keeps its own.

0.11.0

• Mermaid flowcharts in messages are drawn as diagrams instead of code. A ```mermaid block whose source is a flowchart or graph (TD, TB, BT, LR, RL) renders as a card sized to the diagram, scaled to the message width (never below 60 %, the rest clipped with a fade, at most AsystantTheme.diagramMaxHeight tall). A tap opens it across the screen with pan, pinch zoom and a "View code" action. Everything is parsed, laid out and painted on the device: no web view and no network service.
  • Supported: every classic node shape, quoted labels and <br>; links -->, ---, -.->, ==>, ~~~, <-->, --o, --x and their longer forms, labelled with |text| or -- text -->; chains, & groups, ; separators, nested subgraph … end drawn as titled frames, %% comments. classDef, class, style, linkStyle and click lines are ignored; colors follow the app's theme in light and dark.
  • Layered layout: back edges are routed around the flow, every link label gets its own space, and the same source always gives the same drawing.
  • Other diagram types, statements outside that subset and charts over 150 nodes keep showing as a code block, exactly as before.
• AsystantMermaidDiagram(source: …) is public, for drawing a flowchart anywhere in a host app.
• AsystantMarkdownText is public, so hosts can render agent Markdown, diagrams included, outside the chat (task history, decisions, activity logs). Its constructor takes text, style (merged over the paragraph style), selectable (wraps it in its own SelectionArea), onOpenLink and strings; inside a chat it keeps using the chat's link handling and labels.
• AsystantTheme.diagramMaxHeight (420 by default) and the AsystantStrings labels diagram, openDiagram, showDiagramCode, showDiagram and diagramSummary(steps); host string bundles can override them.
• Depends on markdown directly (it was already a transitive dependency).

0.10.0

• Business tools: BusinessToolkit registers, corrects, signs in to and operates business APIs described by asystant_core 0.6.0 contracts. Return toolkit.tools from AsystantAI.tools and toolkit.contextPrompt() from AsystantAI.contextPrompts.
  • Tools: RegisterBusinessTool, UpdateBusinessTool, ConnectBusinessTool, OperateBusinessTool, ConfigureBusinessEnvironmentTool, ConfigureBusinessPrivateHeaderTool, EnterBusinessSecretsTool and ReadBusinessDocsTool, named by BusinessToolNames.
  • Register and update keep the source document (document_attachment_id for an attached text file such as the endpoints .md, or document as text) in BusinessToolkit.documentStore; read_business_docs returns the sections that match a query, capped in size, with every heading. JsonBusinessDocumentStore persists documents through the host's key/value callbacks.
  • Sign-in values, private headers and secret fields are typed in secure forms by tools separate from OperateBusinessTool, so the chat never withholds an operation's answer from the model.
  • BusinessApprovalPolicy: only destructive operations (DELETE by default) ask for approval.
  • BusinessContextPrompt tells the model every registered operation and, per environment and profile, how it signs in and whether a session is saved.
• SecureCredentialStore: a BusinessCredentialStore over flutter_secure_storage, keyed by account, business, environment and profile under a configurable prefix.
• AsystantStrings adds the business labels (businessRegisterTitle, businessInputLabel, businessParameterLabel, businessCodeSentTitle, businessDocsTitle, ...); host string bundles can override them.

0.9.6

• A finished turn's activity is folded into one compact line, such as "3 steps · completed" (or "with issues", with the warning icon, when a step failed, was declined or was stopped), and its steps, including the images a tool returned, open only when the person taps it; the line is at least one action height tall. The line now goes under the turn's answer instead of above it. The turn in progress still shows its steps open and live.
• An activity line the person opened stays open only for its own turn: switching conversations no longer passes it to the turn drawn in the same position.
• AsystantStrings adds activitySteps(count) (singular and plural), activityFinished and activityFinishedWithIssues for that line; host string bundles can override them.

0.9.5

• ChatPrivateInputCard pre-fills a field from PrivateInputField.initialValue (asystant_core 0.5.2) instead of always starting empty. Only text and email fields are pre-filled: secret, password, code and totp always start empty, whatever value the tool sends.
• PrivateInputKind.email fields are shown in clear with an email keyboard and the email autofill hint, and the form cannot be sent until the value looks like an email address.

0.9.4

• Fix raw JSON text appearing in chat when a model (observed with gpt-oss) writes a registered presentation tool's arguments (e.g. present_choices) in content instead of tool_calls; it now shows the same card a real call would.

0.9.3

• Let hosts put model selection in the chat header and show context usage as a thin progress line.
• Keep model choices per provider and allow hosts to open an OpenRouter model catalog from settings.
• Reuse a host-managed OpenRouter credential without asking for a second key, honoring its allowed models.
• Let hosts decorate chat transports for free-first routing and supply OpenRouter app attribution.
• Fix transport disposal during deferred initialization.

0.9.2

• Default all library chat surfaces to English; hosts can opt in to built-in Spanish or supply an AsystantStrings subclass for every label, including provider settings and private input.
• Let hosts register strings for any language or region in AsystantStrings.forLocale, with English fallback.
• Use English defaults for the optional dashboard and choice presentation.
• Document local API keys and provider-neutral server-issued credentials without requiring a specific backend.

0.9.1

• Show a secure inline card for an arbitrary number of private tool fields, including password and TOTP, after approval. The host can disable it with enableInlinePrivateInput: false.
• Keep entered values out of conversation state and withhold tool output after a private-input request.

0.9.0

• Add OpenAI/GPT as a built-in local key source. Apps can hide the entire provider settings menu or allow only selected provider kinds.
• Let hosts place attachment and private-value buttons inside or outside the composer independently, or hide either button. The Keel Bot chat uses the external attachment button.
• Fix bottom navigation overlap in provider and private-value sheets. A custom private value now asks for a name shown to the assistant while its value remains private.
• Support host-defined attachment MIME limits, filename allow/deny regexes and text-content deny regexes alongside extension, size and count limits.

0.8.0

• Add opt-in provider settings in the library's chat menu. A host backend remains the default; the owner can choose a local OpenRouter, Gemini, Claude API, or HTTPS OpenAI-compatible key and a primary model for either local or backend inference.
• Keep local keys separate from messages and conversations in device secure storage, scoped by the host's account namespace. The settings sheet masks key entry and never displays a saved key.
• Reconfigure the chat transport when provider settings change, retaining saved conversations under their provider identity.

0.7.4

• Replace $value text with an opaque, conversation-scoped reference before messages are displayed, persisted or sent to the model. Resolve references only for local tool fields declared with acceptsSecret: true, after approval, and withhold secret-bearing tool output.
• Add an opt-in key icon beside file attachment in both composer layouts. Its bottom sheet collects API keys, tokens, URLs, passwords or other values without putting them into the draft; the model receives a labeled reference. $value remains available independently.
• Allow AsystantAttachmentPolicy(rejectLikelySecrets: true) to reject text attachments with likely credentials at the chat boundary, including custom picker and programmatic attachments, with a localized explanation.

0.7.3

• Add showAsystantSecretPrompt for one-time values collected on the device after an action's approval. The dialog returns values only to its caller and never adds them to conversation state.
• Keep the control-center chat, persistent conversations, native presentations and per-invocation sensitivity policy in the reusable package.

0.6.0

• Add AsystantActionPolicy and AsystantSensitivity so tools or the host assistant can classify each validated invocation by a host-defined name and color and decide whether it requires approval.
• Show sensitivity on approval cards and in the collapsible activity panel.
• Add Approve all for this session to eligible approval cards. It applies only to the current assistant login, never skips required selections, and is not persisted.

0.5.0

• Register any number of AsystantPresentation definitions on an assistant. Each definition provides its tool schema, input validation, trusted data resolution and completed-card widget.
• Persist presentation identity and data in conversation snapshots, so native cards survive app restarts; unknown presentations retain a generic card fallback.
• Include AsystantChoicesPresentation for questions with tappable answers.

0.4.3

• AsystantChat.identityIcon lets the host replace the default sparkle in the header with its own identity mark.
• AsystantChat.welcomeContent lets the host provide the empty conversation view while keeping the built-in timeline, activity steps, composer and permission flow.

0.4.2

• AsystantChat.conversationActionsStyle: AsystantConversationActionsStyle.inline (default, unchanged) or .menu, which puts the conversation list, new conversation and delete behind one icon as a dropdown, each entry with its own icon. conversationMenuIcon replaces the trigger icon.
• New dependency: multiselect_field ^2.5.0, used for the dropdown.

0.4.1

• AsystantChat.headerActions: host buttons in the header row, before the conversation actions, for hosts whose main screen is the chat itself (navigation, sessions, notifications).

0.4.0

• ChatViewModel.reloadConversations() keeps the conversation on screen in the store, then reads the list again and replaces the one shown with exactly what the store answers, newest first. For hosts whose store answers per context of their own (one list per open document): after switching the context, call it, then openConversation or newConversation. See the integration guide, "Host content".
• AsystantAI.requiresConfirmation(tool) decides whether a call waits for the person's approval. It is asked on every call, after the preview and before the tool runs, and defaults to the tool's own requiresConfirmation; a host overrides it to answer from its settings, such as a switch that approves everything. ChatViewModel.configure takes it as confirmation (AsystantConfirmationPolicy). It never skips requiresSelection.
• The images a tool returns (ToolOutcome.images) are also kept on its completed step (AssistantStep.images), and the built-in chat shows them as thumbnails under the step (AsystantTheme.stepImageHeight, 120 by default).
• Host content inside AsystantChat: headerContent (a widget under the header), hostCards (AsystantHostCards pinned after the conversation, with AsystantCardAction buttons for decisions the host's workflow waits for) and managesConversations (off hides the conversation list, New and Delete, for hosts that keep one conversation per context). See the integration guide, "Host content".
• AsystantStrings.failureMessage(failure) words a failure in the chat; by default failure(code), and a subclass can add the provider's detail.
• GenUiCard takes actions, drawn under the card's content.
• Progress of long tools: what a tool reports with ToolContext.reportProgress shows on its running step as AssistantStep.progress (0 to 1) and AssistantStep.progressLabel, with AssistantStep.showsProgress. ChatViewModel shows the first report at once and then at most one every 100 ms, always the latest, and ignores reports once the step ended or the turn was canceled. The built-in chat draws a thin bar under the running step. Canceling the turn still reaches the running tool through ToolContext.isCanceled, and its step ends as canceled.
• Local knowledge (RAG): pass KnowledgeSearchTool(knowledge: AsystantKnowledge(...)) in builtInTools and the assistant searches the app's own documents, locally and without an embeddings model, with every provider (see asystant_core and the integration guide, "Local knowledge (RAG)").
• AsystantAI.init(turnLimits: AsystantTurnLimits(...)) configures the limits of each turn, which were fixed: maxRounds inference rounds (8 by default, up to 64) and maxCallsPerResponse tool calls in one response (16 by default, up to 16, the built-in transports' own bound). Defaults and what happens at a limit are unchanged; an out-of-range value throws a RangeError from init. See the integration guide, "Turn limits".
• The images a tool returns (ToolOutcome.images) are kept on its result message and reach the model with every provider (see asystant_core).
• AsystantTool.isAvailable is read before every model call, not only at initialization: the model is offered the tools available at that moment, so a host can scope them to what it shows.
• A tool's failure detail, and the reason a call was rejected, reach the model with the tool result, so it can correct the call.
• ToolOutcome.endsTurn ends the turn after that tool (see asystant_core).
• AssistantStep records the toolName, when it startedAt, the detail of a failure and the tool's data; its title becomes the tool's summary when it completes. A conversation store keeps them with the entries.
• ChatViewModel is exported, for hosts that draw their own chat over assistant.conversation.
• Fix: pending permission cards no longer render host card content next to Authorize/Decline, as the card-content contract requires.
• AsystantAI.init(provider: ...) is the way to connect an assistant: OpenRouterProvider, ClaudeCodeProvider or any later AsystantProvider. Tools, prompts, per-request context, attachments, the model picker and cards work the same with each. init(transport: ...) remains for a custom AssistantTransport (test doubles, proxies); pass exactly one of the two.
• With ClaudeCodeProvider and no models, the chat offers the models the installed Claude Code CLI declares.
• The example answers through ClaudeCodeProvider with --dart-define=CLAUDE_CODE=true on a desktop, and through a simulated transport otherwise.
• Documentation describes the providers and how to add one; the gateway session API no longer exists.

0.3.2

• Align the embedded chat with Aula-AI: compact header, inline composer and icon-only send/stop controls.
• Isolate input borders from host themes and support Enter to send / Shift+Enter for a new line.
• Pulse the working icon while keeping status text readable; respect reduced-motion preferences.
• Refine bordered messages and discreet inline failure feedback.

0.3.1

• Render assistant controls, status indicators, and cards with bundled SVG icons, independent of Material icon font downloads and web font caches.
• Refine the chat header, message typography, composer, and failure feedback.
• Add English chat previews at mobile and desktop widths with no icon fonts loaded.

0.3.0

• Add host-owned content for completed GenUI cards through cardContentBuilder.
• Keep custom result widgets separate from SDK permission controls.

0.2.1

• Selectable conversation and card text with clickable links and host-controlled navigation.
• Batched streaming updates reduce UI notifications without interrupting hidden conversations.
• Regression coverage for closing/reopening during inference, pending approvals, and account changes.
• Documented session-owned conversation lifetime and explicit cancellation.

0.2.0

• Deferred Flutter initialization: init() stores configuration; the mounted chat connects after its first frame. Headless clients await ensureInitialized() explicitly.
• Injectable application/context prompts and baseline safety guidance in SDK and gateway.
• Typed genUI bar/line charts, bounded numeric tool arguments and English golden previews.
• Initialization retry, concurrent setup deduplication and lifecycle cleanup.

0.1.0

• Initial MIT-licensed release.
• Typed local tools, explicit permissions and cancellation-aware execution.
• Session-backed gateway authentication and server-assigned model policies.
• English examples and integration guidance for the reference Rust API.
• Embeddable Flutter chat, adaptive layouts, genUI cards and reactive_notifier state.

0.7.0

• Define the five tool sensitivity levels none (Nulo), low (Bajo), medium (Medio), high (Alto) and admin (Admin). Host tools assign a level for each validated invocation; the library provides its name, color and approval behavior.
• Run none and low actions immediately. Require approval for medium, high and admin, with the existing optional approval for the current session. Required selections always wait for the person.
• Move the tappable context usage meter into the chat header, leaving the composer free for input.
• Preserve the legacy requiresConfirmation behavior for tools that have not been classified yet.

0.7.1

• Remove the empty composer footer after moving context usage into the header.

0.7.2

• Show sensitivity as a compact colored ticket icon beside each action in the activity row; keep the localized level name in its tooltip and accessibility label.
• Place the compact context meter beside the chat header controls, with a full tooltip and tappable detail.
• Keep sensitivity identifiers and stored names in English while localizing their visible labels.