Skip to content

Ideas workspace

Lightweight workspace for capturing notes and brainstorming with an agent. Two panes only.

Sidebar label: Ideas · icon: lightbulb.

Available panes

  • Notes (⌘⇧1) — Note list on the left (within the pane), markdown-style body on the right. Backed by SwiftData (IdeaNote model) with autosave.
  • Agent (⌘⇧2) — Same agent pane as Prompt; useful for thinking through an idea without spinning up a terminal.

Notes structure

Each note has:

  • A title (first line of the body, auto-derived).
  • A body (plain text, soft-wrapped, monospaced).
  • Last-edited timestamp surfaced as edited <relative time>.

Notes are not workspace-scoped — every Ideas workspace shares the same note pool, but the active selection is per-workspace.

Agent integration

The Agent pane in Ideas works the same as in Prompt:

  • Default provider: Claude Code via OAuth.
  • Local LLMs work too — point at Ollama or any OpenAI-compatible endpoint via Settings → Providers.
  • cwd is the workspace folder if set, otherwise nil.
  • Every turn carries a workspace snapshot — workspace name, project folder, CLAUDE.md/AGENTS.md/GUIDE.md/README.md from that folder, the active tab's body, and short excerpts of sibling tabs — so asking for ideas is grounded in the project, not generic. See Agents → Workspace context block for the full payload shape.

Save-to-tab proposal cards

When the agent's response in an Ideas workspace looks like a list (numbered or bulleted, three or more items), an inline confirmation card appears under the response:

  • Title: N ideas for "<active tab name>".
  • Each item rendered as a checkbox row — uncheck items you don't want.
  • Mode toggle: Create as new tabs (one new note per item) or Append to active tab (bullets added to the current note's body).
  • Add N ideas commits the selected items via SwiftData; Dismiss collapses the card.

The card is gated by the active workspace kind. Other kinds (Prompt, Review) ignore proposals and render the response as plain text.

For Anthropic-direct providers, items are produced via the propose_items tool. For Claude Code subprocess and local HTTP providers, a fallback list parser extracts items from the finalized response — phrasing requests like "give me 8 ideas for X" reliably triggers the card.

When to use Ideas vs Prompt

Use Ideas when… Use Prompt when…
Drafting copy, planning, journaling Running commands, editing files
Bouncing rough ideas off a model Driving an agent against a real codebase
You don't need a terminal or kanban You need full cockpit panes