Trust model

Spark runs extensions inside the main window, with no sandbox and no iframe. An imported package is code you chose to trust. Everything below follows from that decision.

Tiers

TierExampleTrust
CoreTerminal, Markdown, DataShipped with Spark, reviewed with the binary.
OfficialOpenAlexA named integration, shipped with Spark.
CommunityImported or downloaded packageYour explicit trust.

certification: "spark" is reserved for packages published and signed by Spark.Today no cryptographic check is wired, so the badge on built-ins is declarative. Third-party certified organisations are not supported yet.

Before install

  • The v2 manifest is validated in Rust, not only in the renderer.
  • spark.* and official.* identifiers are reserved; the version must be x.y.z.
  • Relative paths only: .., backslashes and symbolic links are refused; .git and node_modules are skipped.
  • Size, file count and depth are bounded.
  • certification is refused inside an imported package.
  • HTTPS is required in release; localhost is tolerated in debug.
  • The zip's SHA-256 is checked before extraction, then the package is staged and renamed atomically.

After install

  • Activation is given 5 seconds; a failure is cleaned up rather than leaving partial contributions.
  • Each tile has an error boundary with Retry and Disable the extension.
  • Uninstalling asks for confirmation and removes the package and its state entry.
  • Local import is restricted to developer mode.
  • Registry sources are capped at 20 and their URLs are validated.

What the guardrails do not do

  • They do not sandbox the code.
  • They do not prevent an extension from calling the Tauri commands allowed for the window.
  • They do not isolate services and events per extension: identifiers are global.
  • They do not restrict an extension's network access.
  • They do not verify signatures, and they do not update anything automatically.
  • They do not replace a code review for a certified package.

Native surface

Every command allowed for the main window is reachable from code running in that window. Capabilities live in src-tauri/capabilities/; CSP and asset scope intauri.conf.json. The asset protocol only covers the PDF cache and$APPCONFIG/extensions/**.

Wording

When a feature touches trust, say what is verified and what is not. Never write "secure" where you mean "signed" or "checked locally". Distinguish a built-in from a community package, and document a limit instead of hiding it.