Workstream Canvas

Workstream canvas

Every workstream keeps one canvas the owner can open beside the chat to see where it stands, and what it needs from them, without reading the transcript. It is brought up to date as the work advances, not at the end.

Where

• ~/.cursor/projects/<workspace>/canvases/ws-<slug>.canvas.tsx: <workspace> is the Cursor project folder for this repository (find it from absolute paths in the environment, or list ~/.cursor/projects/); <slug> is a short kebab-case name for the workstream (ws-copy-paste-sync).
• One canvas per workstream. Edit only your own; another agent's workstream has its own canvas. If you continue someone else's workstream, keep its canvas.
• Read ~/.cursor/skills-cursor/canvas/SKILL.md once per session before the first write: its rules apply (theme tokens only, no empty states, label every chart).

Start a workstream

1. Copy template.tsx to the path above.
2. Replace every value in the workstream object, and delete the example rows that don't belong. Leave the rendering code alone; an empty list hides its section or tab.
3. Link the canvas in your reply: [Workstream title](/absolute/path/ws-slug.canvas.tsx).

Boards: features with milestones

When the work is a feature with milestones, each made of tracker rows (the library, LIB-01 to LIB-35, is one), start from board-template.tsx instead. It shows overall progress and a bar for each milestone, a Now panel (what's being built, what waits on the owner, what's next), Needs you, every row with its state and issue, the measurements against their budgets, the agents with a button to open each conversation, and the plan's steps, log and decisions folded underneath.

• Fill in BOARD, MILESTONES (each row with its tracker ID, issue and state), LATER, NOW, AGENTS, DOCS and workstream; leave the rendering alone.
• A row's state is done (merged where it ships, its done-when met), built (merged and tested, with a check at scale or a small part left, which its note names), doing, you (waiting on the owner) or todo. Never call a row done that the tracker doesn't.
• Everything below applies to boards too: the owner's marks first, one step per todo, updates in the same turn, and only what was verified.
• For a post or a report, capture.py renders a canvas outside Cursor with Cursor's own canvas runtime and captures its sections (its docstring has the commands); the blog post How I manage a large feature across AI agents was illustrated with it.

A plan's todos are its steps

When the work has a plan (plan mode's ~/.cursor/plans/<name>.plan.md, or one in docs/plans/), the canvas shows every todo:

• The canvas is the plan's first todo. Plan mode can't write it, so it's written as soon as the plan is approved, before any other todo starts.
• One step per todo, in the plan's order, with the todo's ID as the step's id, a short step and detail from its content, and doneWhen from the plan. Link the plan under links.
• A todo and its step change in the same turn: pending is not started, in_progress is in progress, completed is done and cancelled is dropped. blocked is the canvas's own, with the step's note saying what it waits on. A todo added or split gets its step at once, so the two lists stay one to one.

What it shows

• Overview: progress through the plan; Needs you, open items first (blocking ones on top, with a command to copy) and done items folded underneath with what came of them; the Ready and Blocked lanes; the latest log entries.
• The owner answers Needs you on the canvas: each open item has Done, Skip and Ask the agent. Done and Skip fold the item away at once (with Undo). Ask the agent opens a new chat with the canvas attached and a prompt naming the item, and marks it "With an agent".
• Plan: each step with what it involves, its "done when" criterion, its reference (tracker ID, issue or commit) and its status.
• Decisions, the full Log, and Measurements when there are any.