Skip to main content

S1-B2: Execution Plane

Intent​

Expand Seraph's ability to actually do work so the guardian thesis is backed by a more serious action layer.

Capabilities in scope​

  • real shell and process execution beyond Python-only snippets
  • stronger browser control and interactive web flows
  • background process/session handling
  • workflow engine direction for repeatable tool chains

Non-goals​

  • full desktop computer control across every app
  • unlimited shell access without policy controls
  • large-scale workflow marketplace

Required architectural changes​

  • split lightweight code execution from broader command/process execution
  • upgrade browser tooling from extract-only patterns toward interactive automation
  • define a process/session abstraction that can report status back to the UI
  • establish the first workflow representation and execution contract

Likely files/systems touched​

  • backend tools and tool registry
  • sandbox / execution environment
  • browser automation layer
  • chat and step streaming for longer-running actions

Acceptance criteria​

  • Seraph can run bounded shell/process tasks under policy control
  • browser automation supports richer interactive flows
  • longer-running tasks can surface intermediate state
  • at least one workflow path exists for repeatable multi-step execution

Dependencies on earlier batches​

Open risks​

  • execution breadth can outpace safety
  • UX may degrade if longer tasks are not surfaced clearly
  • browser automation can become flaky without careful constraints