Orchestrate Using AI Assistants

Braidrun's AI assistant is not a plug-in toolbar - it directly understands the YAML, DAG, and running status of the current page, can change workflow, trigger execution, install skills, and switch languages. Here’s how to get started and use it daily.

Where To Find It

• Floating button in the lower right corner (available on all main pages)
• Shortcut keysAlt + G

Inside the drawer is a streaming conversation panel that supports Markdown/code highlighting (12 languages) and slash commands.

What can it do?

• Create/modify/verify/execute workflow
• Create, enable, and deactivate schedules
• Refresh or install skill
• Switch interface language
• Answer systemic questions such as "What does this system support/How to write a certain configuration/How to use a certain API"

Slash Command

/execute               # run the current workflow (alias: /run)
/validate              # validate YAML without running
/export                # download workflow YAML
/new                   # start a fresh workflow
/undo                  # roll back the last assistant change (up to 10)
/help                  # list all slash commands

Trigger them by starting your input with "/". Slash commands take priority over free text and call the platform's API directly, bypassing the model — so they don't count against the AI assistant's daily quota. On pages like the editor and the workflow list, matching shortcut buttons also appear above the input box.

Diff Preview and Undo Stack

All changes to resources such as workflow / schedule / credential will first give you a Diff preview:

• New steps/fields are highlighted green
• Modified fields are marked yellow
• Deleted steps/fields are highlighted in red

After confirmation, you can use /undo to roll back at any time, retaining up to the last 10 modifications. The session itself can also be exported to Markdown with one click.

Working With The Editor

• When you have the assistant build or modify a workflow in the editor, the canvas reveals each change in order: nodes are added and edges drawn one step at a time, so you can watch the workflow take shape. The changes still aren't persisted until you click Save.
• The editor's diagnostics panel can hand off a single issue or all of them with "Fix with AI," and the step properties panel has an entry that opens the assistant with the current step's context already loaded — no copy-pasting YAML yourself.
• Every step the assistant changes also enters the editor's own undo / redo history, so Ctrl / Cmd + Z rolls them back one at a time.

For a full walkthrough of each area of the editor, see Visual Editor.

Preparation: Configure Your AI Assistant

Go before using it for the first time Account Settings → AI Assistant Open and select a model provider (e.g. OpenRouter, OpenAI, Anthropic, DeepSeek), or configure the subscription model in the Claude Code / Codex area. The real secret is placed in Credential Management → My Credentials Maintenance, it will be automatically resolved according to the current account when running. Administrators do not need to pre-configure the AI assistant for regular users.

Claude Code / Codex Model

If you want the AI assistant to call Claude Code or Codex, choose the authentication method in Account Settings → AI Assistant → Claude Code / Codex:

• Claude Code: You can use Anthropic API Key or Claude Subscription Token (provider=claude_code_oauth).
• Codex: You can use OpenAI API Key, or you can use Codex to subscribe to auth.json (provider=codex_subscription).
• Subscription mode does not write token/auth.json into the AI assistant configuration; it only reads it from your personal credentials or team credentials.

Documentation Knowledge Base

To have the AI assistant answer a system documentation question (e.g. "What is the default timeout for manual_approval"), go to Settings → Platform Persistence Storage → AI Assistant Documentation Knowledge Base Confirm it is enabled. Built-in guides, project master documents, and selected reference docs are indexed by default, and automatic building at startup is enabled by default; you can view the status and manually click "Rebuild Knowledge Base" on the same page.

"Thinking Process" Reasoning Streaming

When the selected model supports reasoning (Claude Sonnet 4.5 / 4.6 / 4.7, DeepSeek-R1, GPT-o1 / o3 series, Gemini Thinking, etc.), the AI assistant drawer will first display a folded "AI Assistant's Thinking Process" card. The model first flows the reasoning summary here, and then generates the final answer:

• The reasoning_delta event is collapsed by the client and will not refresh the screen (avoiding re-rendering every frame)
• Complete reasoning is only rendered with the thinking subcategory when reasoning_complete
• Click on the thinking card to expand/collapse; it is collapsed by default to avoid occupying the field of view
• The thinking content will also be recorded in the undo stack along with /undo to facilitate review.

Models that do not support reasoning will not display any additional cards and behave as before.

Long-Term Memory

The AI assistant is independent per session by default. If you want it to "remember" your preferences ("I like Chinese answers + short markdown", "My ASA workflow always runs early in the morning"), you can turn on Long-Term Memory in the AI Assistant configuration:

• After enabling, user messages and assistant replies will be vectorized and stored in the cross-session ltm:<user_id>:* namespace.
• When the next round of dialogue starts, the top-K relevant memories will be automatically retrieved as context injection
• There is a "Clear Long-term Memory" button in the upper right corner of the sidebar of the AI Assistant. POST /api/global-agent/memory/clear clears all namespaces of the current user with one click; suitable for use when changing devices/switching teams.

Tips

• If you want the AI to generate YAML: explicitly ask it to output snake_case YAML
• Branch flow: give priority to generating the structure of classifier + condition
• Big change: use natural language to describe the intention, and then break it into multi-step Diff
• combine Debugger : After the AI modification is completed, enable debug and dry-run the workflow before going online.
• When selecting a model that supports reasoning, you can directly see the "thinking process", and the debugging prompt design is very useful.