Kandev canvas authoring

SkillProductivity

Author a Kandev task canvas as a self-contained web application.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Kandev canvas authoring skill

What this skill tells your AI

The instructions your AI receives, as published by kdlbs/kandev in apps/backend/internal/mcp/canvasskill/files/SKILL.md and read by ahel’s review.

Use one read_canvas_authoring_skill_kandev call without path when you need the authoring contract. That response is the complete core bundle. It includes this workflow, the manifest and browser protocol summary, appearance rules, the minimal scaffold, and the exact supporting-file inventory. Do not read the core bundle again during the same authoring task.

Required workflow

  1. Call create_canvas_kandev with a short title and an application summary. It creates an inactive task canvas and returns its source directory, manifest scaffold, initial permission policy, and exact scaffold inventory.
  2. Use native file tools in that returned directory. The initial files are manifest.yaml, index.html, appearance.js, script.js, and styles.css. Replace or extend them in the same directory.
  3. Keep every source path relative to the returned directory. Do not write outside it. Bundle executable dependencies. Node, a package manager, and a network build step are not available at runtime.
  4. Run local checks, then call publish_canvas_kandev with the returned canvas ID and source path. Read validation diagnostics and correct rejected source before publishing again.

The first valid release of a new owner-created task canvas uses the returned initial permission policy. It can activate without a second approval for its declared supported task-scoped data, event, state, and exact HTTPS-origin permissions. A later permission increase, imported package, or workspace promotion still requires human review. Do not add a trust flag to the manifest, and do not request permissions outside the policy.

Core application contract

  • Include <meta name="viewport" content="width=device-width, initial-scale=1">.
  • Use relative ./_kandev/v1 paths for Kandev data, state, actions, and events.
  • Treat Kandev domain data as the source of truth. Derive filters and summaries in memory instead of storing a second copy of domain records.
  • Store only small application-specific shared values in instance state. Keep temporary input in memory and use conditional revisions for writes.
  • The canvas has an opaque origin. Do not use browser storage, service workers, origin-wide cookies, host URLs, or authorization headers.
  • Avoid secrets in source, URLs, query strings, logs, and client state.
  • Render loading, empty, error, and retry states. Keep destructive actions explicit and explain their result.
  • Use accessible labels, keyboard operation, visible focus, and touch targets.

Kandev injects a reserved startup bootstrap into the entry document before authored scripts. It reports early document errors and checks the relative context route after document load. The host reveals the frame only after a versioned acknowledgement for the current attempt. A missing acknowledgement or context failure becomes recoverable after 15 seconds. Keep the entry valid HTML and render loading, empty, error, and retry states in the app.

Minimal manifest

Use the returned manifest_scaffold as the starting point. New manifests use api_version: 2, one lowercase web-app key, a package-relative entry, and at least one task-canvas or workspace-canvas placement. Declare only the api_read, api_write, events, state, and network_origins permissions that the application needs. The owner-authorized first release can receive only these supported task-scoped grants. The entry and all relative assets must be in the published package.

Browser protocol summary

Resolve all routes from the application document with ./_kandev/v1. Use context, the paginated data routes, state/{key} with If-Match, and the bounded events stream as documented by the optional references. Events are hints that invalidate a read. Refetch authoritative data after an event, reconnect with Last-Event-ID, and perform a full refetch after runtime.resync_required.

Appearance protocol

The host may send the public presentation-only message kandev.web_app.appearance with version: 1, mode: light|dark, and exactly these color tokens: background, foreground, card, cardForeground, muted, mutedForeground, border, primary, primaryForeground, accent, accentForeground, destructive, destructiveForeground, and ring. Accept it only when event.source === window.parent, the type and version match, the keys are exact, and each serialized color is bounded. Map the tokens to the same-name kebab-case CSS variables. Keep light and dark fallbacks so the app remains usable before the first message. The message has no identity, capability, data, storage, navigation, or action fields.

The generated appearance.js implements this listener. It is optional, but copy its pattern when replacing the scaffold.

Read a supporting reference only when its topic is needed:

  • references/browser-api.md for detailed browser routes and errors.
  • references/manifest.md for the full manifest shape and validation rules.
  • references/data-and-state.md for domain data and instance state.
  • references/events-and-recovery.md for events, reconnect, and retries.
  • references/security.md for opaque-origin and source safety rules.
  • references/ui-patterns.md for responsive and accessible UI patterns.

Signals

GitHub stars
771
Forks
114
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
kandev-canvas-authoring
Source
github.com/kdlbs/kandev