Workflow Copilot
An AI assistant that edits, debugs, runs, and deploys the workflow you have open — always with your confirmation.
The Workflow Copilot is the Copilot tab in the workflow editor. Ask it to edit, debug, deploy, or explain the workflow you have open. It uses a real AI planner backed by BYOK: it proposes graph changes as a diff you review and confirm, reads your last run to explain failures, and can run or deploy the workflow on request — nothing is applied silently.
What the Copilot can do
The planner picks one mode per request:
Add, delete, connect, disconnect, rename, move, or reconfigure nodes — proposed as a diff.
Read the last run and its per-node errors, explain what failed, and suggest a fix.
Run the workflow or deploy it, with confirmation.
Ask a clarifying question instead of guessing when a request is ambiguous.
Editing the graph
The Copilot edits through seven typed graph tools. Each produces an immutable patch to
{ nodes, edges } — it never mutates the canvas directly — and every patch is validated before it
can be applied.
| Tool | Effect |
|---|---|
addNode | Create a node from the catalog and auto-connect it into the graph |
deleteNode | Remove a node and every edge touching it |
connectEdge | Connect two nodes |
disconnectEdge | Remove a connection |
updateNodeConfig | Merge new values into a node's configuration |
renameNode | Change a node's label |
moveNode | Reposition a node (coordinates are clamped to the canvas) |
A patch that would leave the graph invalid is rejected before it reaches you — for example deleting the only trigger or the output node, connecting a node to itself, connecting into a trigger, creating a cycle, duplicating an edge, or referencing a node that doesn't exist.
Propose, diff, accept
Edits, runs, and deploys are proposed, not auto-applied. A proposed edit is shown as a colored graph diff:
- Added nodes and edges in green ("Add node…", "Connect A → B").
- Removed nodes and edges in red ("Remove node…", "Disconnect A → B").
- Modified node fields in amber ("Update … (fields)").
You choose Accept or Reject. Accepting applies the patch to the canvas; rejecting leaves the workflow untouched. The buttons sit below the changes so you can review the proposal before choosing. While Copilot prepares it, the chat shows the current action and recent activity from the runtime stream. These are task updates, not the model's private reasoning; unknown tools receive a general action label.
You can specify that a rename is the only change, for example “Rename only node Output to Result” or “Renombra únicamente el nodo Output como Resultado”. The proposed diff must still name the exact node and requested label.
Debugging a run
Ask "why did the last run fail?" and the Copilot builds a debug context from the most recent run: the per-node run states, the nodes that failed with their errors, and error-level logs. It uses that to explain the real failure and propose a concrete fix (which it can offer as an edit). If there are no runs yet, it says so instead of guessing.
Running and deploying from the Copilot
When you ask it to run or deploy, the Copilot surfaces a Run workflow or Deploy workflow action. Each button uses the short-lived confirmation issued for that proposal and its current graph. Confirming a run opens the persisted execution in the workflow's execution panel; confirming a deploy updates the canvas only after publication succeeds. To start from a particular block, use the canvas execution controls.
If you then run the candidate manually, the panel follows the newer run when you return to the editor. Selecting another execution also takes precedence while you inspect it. The earlier Copilot run remains in Recent executions. If its retained details cannot be retrieved, a separate Copilot notice offers Retry. Its lookup error or pending state does not change the selected manual execution or mark that execution as failed.
If the graph changes or the confirmation expires, the action is refused and its error appears in the conversation. Ask for a fresh proposal after reviewing the current graph. An interrupted response is not retried automatically: check Executions or the deployment state before requesting another action. An unsuccessful action never displays Run started or Deploy started.
The chat panel
The Copilot keeps a real message history. You can start a New chat (the current thread is archived) and reopen past threads from Chat history. Attach workspace resources as structured context, and dictate a prompt with voice input. When no AI provider is configured, the Copilot falls back to the deterministic keyword behavior so the panel still works.
New chat and history wait while the Copilot is planning or applying a confirmed proposal, so an unfinished action remains in its current conversation.
Voice recordings are limited to 60 seconds and 10 MB. Confluye verifies the audio container on the
server, resolves the active workspace's speech-to-text credential without exposing it to the browser,
and uses gpt-4o-mini-transcribe for OpenAI. Connect an OpenAI, Azure OpenAI, Groq, or compatible
custom provider in Settings before using the microphone.