Agent Chat in Split-Screen

Circuitry's chat editor becomes an agent the moment you put it in a split-screen with another document. The agent sees what you have open in the other pane and can act on it — manipulate a workflow, edit a code file, run wizards, all turn-by-turn with your approval.

Pairing chat with a workflow or file

1. Open a chat editor (from the "+" menu or by clicking a Chat node).
2. Open a workflow, code file, or design document in the other pane.
3. The chat header shows the paired document. You're now in agent mode.

When paired, the chat agent automatically receives:

• The active document's path, name, and language.
• Your current text selection (up to 2000 chars).
• The containing function at the cursor (up to 4000 chars), so questions like "what does this do?" work without you having to copy-paste anything.
• The list of tools the agent can call — workflow manipulation, file read / write, wizard composition, etc.

Plan approval

Before the agent runs any tool that changes your workflow or files, it shows a plan — a structured list of actions. Click Approve to run them, or Reject to cancel. This is the safety net: nothing changes without you saying yes.

Diff approval for code edits

When the agent edits a code file, the changes appear as diff hunks in the editor — green for additions, red for deletions, exactly like a code review. Accept hunks individually or all at once. The same diff UI is shared with Cmd+K inline prompts, so the workflow is consistent.

Claude Code CLI awareness (new)

If you're using a CLI-tool model — Claude Code, for example — Circuitry now feeds the paired-document context (file path, language, selection, containing function) directly into the CLI's first user message. Previously this context was dropped on the way through the CLI's stdio protocol. Now selecting code in the right pane and asking "what does this do?" in a CLI-backed chat works the same as it does with cloud models.

This applies automatically — no setting to flip. If you've configured Claude Code to use Circuitry's MCP server for workflow tools, you still get the full agent-mode experience; the new behaviour just makes sure the file context shows up regardless of MCP configuration.

Models and how they affect chat

The chat panel has its own model selector. Cloud models (OpenAI, Anthropic) route through Circuitry's server-side key injection so your API key never leaves the server. Local models (LM Studio, Ollama) talk to your localhost endpoint directly — no Circuitry server hop. CLI models (Claude Code in persistent mode) keep a long-lived process alive across messages for sub-second turn-around.

Tips

• Selecting code before you ask is almost always faster than describing what you want the agent to find. The selection lands in the system prompt automatically.
• Plans are reviewable: if the agent wants to run workflow.deleteNode, you'll see exactly which nodes before approving.
• Diffs are atomic per hunk: you can accept the agent's logic but reject its variable rename, hunk-by-hunk.
• Switch models freely mid-chat: the chat agent will use whatever model is selected for the next message, so you can start with a cheap model and bump to a stronger one for the hard parts.