13 posts tagged with "AI"

Plan mode is a read-only autonomy posture for AI sessions, picked from the chat footer. The assistant can only investigate; anything that writes, runs or deploys is refused until it hands over a plan and you approve it, which restores the posture you were in before. Session chats only, and never persisted.

New features

Plan (read-only) joins Ask permission and Yolo in the autonomy picker of a session chat.
Only read-only tools run while planning; writes, job runs and deploys are refused with a "Blocked in plan mode" row.
You enter plan mode from the picker, or by accepting the "Start planning?" card the assistant raises on open-ended tasks.
The proposed plan is shown for approval with "Approve and implement" or "Keep planning"; approving restores the exact posture that preceded plan mode.
One versioned plan document per session, pinned to the top of the artifacts list, with the approved version marked and unapproved revisions labeled as drafts.

AI sessions (beta)

AIv1.766.0

Docs

AI sessions are now in beta and enabled by default. Each session ties an AI chat to a workspace or fork with a live preview panel; the AI builds scripts, flows and apps as drafts you review in a unified diff drawer and deploy. Run many sessions in parallel; a banner switches back to the legacy chat.

New features

Enabled by default during the beta: a banner under the session chat switches back to the legacy side panel chat (stored per browser), and a matching banner in the legacy chat reactivates sessions. Operators keep the legacy chat.
Each session is tied to a workspace or fork, picked before the first message; forks are only created when the conversation actually starts.
The preview panel opens edited items as live editors and workspace pages (runs, schedules, settings) as tabs, refreshing only what the AI touched.
A chat-scoped edits bar and unified diff drawer track what the conversation changed, with per-item deploy and discard, and fork promotion via the compare page.
Jobs the AI runs detach to a background jobs tray with live status, cancel and approval actions, and survive page reloads.
The AI writes plans and specs as markdown artifacts that open in the preview panel with copy and download actions.
With a full-code app preview open, the AI debugs the running app: console logs, job runs, live DOM queries (search_dom, read_dom), screenshots, and a click-to-attach element inspector.
Slash commands: /compact summarizes the conversation in place, /clear starts a fresh chat, and workspace AI skills appear as commands.
Sessions are grouped by workspace family in the sidebar, with unread badges and parallel draft sessions persisted in the browser.

Attach text files to AI chat messages

AIAI Chatv1.765.0

Docs

Text files can now be attached to individual AI chat messages via picker, drag and drop, or paste. They show as chips in the input box and are sent with the next message. Content is read on demand by the AI through file tools instead of being inlined, so it is never resent with the conversation history.

New features

Attach text files to a chat message via the attach button, drag and drop, or paste
Up to 8 files per message, 1 MB per file
Files appear as chips in the input box, sent with the next message and cleared after send
The AI reads file content on demand via read_file and search_files tools, keeping the conversation history small
Linked folders remain session-wide assets, shown in the footer next to the mode picker

Web search sources in AI chat

AIv1.764.0

Docs

When the AI chat searches the web (OpenAI and Anthropic providers), the tool card now shows the search query as it is typed and expands with the list of consulted sources (favicon, page title and link) as each search completes. Sources are kept with the conversation history.

New features

The web search tool card shows the search query live in its label while the model types it.
The card auto-expands with the list of consulted pages (favicon, title and link) as soon as each search completes, while the model continues its answer.
Sources persist with the conversation history.

Reasoning effort for AI agent steps

Docs

AI agent steps can now use extended thinking, configured with a model-adaptive reasoning effort dropdown in the provider/model picker. Supported on Anthropic Claude, OpenAI o-series and GPT-5, Azure OpenAI, Google AI Gemini, AWS Bedrock Claude, OpenRouter, DeepSeek, and Mistral, with an explicit off option. Reasoning tokens are billed, and thinking is streamed into a collapsible Thinking box in chat mode.

New features

A reasoning effort dropdown appears below the model in the AI agent step provider/model picker, with model-adaptive levels (low / medium / high, plus minimal / xhigh / max on some models) and an explicit off option.
Supported on Anthropic Claude, OpenAI o-series and GPT-5, Azure OpenAI, Google AI Gemini, AWS Bedrock Claude, OpenRouter, DeepSeek, and Mistral; available levels depend on the specific model.
Reasoning tokens are billed and counted in the step token usage; thinking is streamed into a collapsible Thinking box in the flow chat interface during execution.

Azure AI Foundry support for Windmill AI

AIAI ChatAI agents

Docs

Azure AI Foundry is now natively supported as an AI provider in Windmill, exposing the broader Foundry catalog (OpenAI models plus Llama, DeepSeek, Mistral, Phi and more) over an OpenAI-compatible API in both AI chat and AI agent steps.

New features

Azure AI Foundry now available as a native AI provider via the azure_foundry resource type
Access to the broader Foundry model catalog beyond Azure OpenAI: Llama, DeepSeek, Mistral, Phi and more
Uses the same Azure authentication (api-key header) and base URL configuration as Azure OpenAI
Available in both AI chat (copilot) and AI agent flow steps

Custom headers on AI requests

Docs

The customai resource accepts a headers object to attach arbitrary headers to every AI request, and a new AI_HTTP_HEADERS env var applies headers globally across all providers (OpenAI, Anthropic, Azure, Bedrock).

New features

headers field on customai resources: map of header names to values, attached to every request made with that resource.
AI_HTTP_HEADERS env var on the server and workers attaches headers to every outgoing AI request across all providers (OpenAI, Anthropic, Azure, Bedrock, etc.).
Useful for upstream proxies, gateway authentication, and routing metadata.

PDF input for AI agents

Docs

AI agent steps can now accept PDF documents as input alongside images, with provider-specific encoding for Anthropic, OpenAI, and Google AI handled automatically. Requires a workspace-level S3 object storage.

New features

PDF documents can be passed as attachments to AI agent steps.
Provider-specific encoding handled automatically: Anthropic document blocks, OpenAI input_file blocks, Google AI inline data.
MIME type detected from the file extension; requires workspace-level S3 object storage.

Route Anthropic and Google AI through Vertex AI

Docs

The googleai and anthropic AI resources now expose a platform field to switch between the provider's standard API and Google Vertex AI, using a Google Cloud service account access token for authentication.

New features

platform: google_vertex_ai option on googleai and anthropic resources.
Set base_url to your Vertex endpoint and api_key to a Google Cloud OAuth2 access token (Vertex AI User role).
Run Claude models on Vertex AI from existing Anthropic-shaped scripts and AI agents.

Instance-level AI settings

AISelf-hosted

Docs

Superadmins can configure AI providers (OpenAI, Anthropic, Google AI, Azure, Bedrock, customai) at the instance level so every workspace inherits the same models, headers, and routing without per-workspace setup.

New features

Configure AI providers (OpenAI, Anthropic, Google AI, Azure, Bedrock, customai, etc.) once at the instance level.
Workspaces inherit instance defaults and can still override per-workspace.
Centralizes model selection, base URLs, and shared credentials for self-hosted deployments.