Skip to content

LAUNCH_PLAN.md

Latest commit

 

History

History
73 lines (60 loc) · 3.56 KB

LAUNCH_PLAN.md

File metadata and controls

73 lines (60 loc) · 3.56 KB
Raw

OpenAdapt Launch Contract

Last reviewed: 2026-07-27

This is the durable public acceptance contract, not a chronological project plan. Current progress, metrics, and provider state belong in the private operating status. The product is not organized around a browser-only freeze: browser, native desktop, RDP, and Citrix/VDI are execution surfaces under one governed qualification contract.

Product Outcome

OpenAdapt turns a bounded demonstration into an inspectable workflow, executes healthy runs without model calls, verifies the declared business effect, and halts for review when authorization, identity, state, target, or effect cannot be proven.

The public launch path must let a user:

  1. Install the launcher or Desktop cockpit.
  2. Record a browser, native, RDP, or Citrix/VDI workflow using the canonical recorder for that surface.
  3. Compile and inspect the workflow without editing internal JSON or Python.
  4. Qualify risk, identity, effect strength, parameters, secrets, environment, and representative fault cases.
  5. Certify, seal, encrypt, version, and deploy the exact artifact.
  6. Execute it in managed or customer-controlled infrastructure with the required capabilities and data boundary.
  7. Receive VERIFIED, COMPLETED_UNVERIFIED, HALTED, FAILED, or ROLLED_BACK with evidence; Standard and Regulated profiles never treat COMPLETED_UNVERIFIED as success.
  8. Resolve attended decisions, validate repairs against the same qualification contract, promote or reject them, and roll back versions.
  9. Inspect evidence, usage, billing, and immutable artifact lineage.

Production mode must use real configured dependencies. Missing storage, runner, callback, verifier, billing, or authorization dependencies must fail closed and must never produce simulated success.

Surface Contract

  • Browser: Playwright-native recording retains DOM identity, source-time secret redaction, field geometry, and ordered frames.
  • Native desktop: openadapt-capture owns screen, input, timing, and window-scoped capture; live UIA, macOS Accessibility, or Linux AT-SPI evidence is retained at action time when available.
  • RDP and Citrix/VDI: customer-controlled external black-box operation uses the same window-scoped capture path with OCR, relational anchors, identity regions, fresh-frame resolution, and two-phase consequential actuation. No software inside the remote session is required.
  • Cloud: managed browser execution and the proprietary control plane support workflow, evidence, attended-operation, usage, and billing journeys.

Portable workflow intent remains separate from environment-specific bindings. Every runner advertises its observation, actuation, identity, verification, session, and security capabilities; admission refuses before actuation when the selected runner cannot satisfy the artifact.

Evidence Contract

  • Comparative claims require at least three trials per task and condition.
  • Results name the task, application, version, environment, oracle, run count, failure taxonomy, latency, model calls, silent incorrect success, and over-halt.
  • Mock, synthetic, public-application, customer-qualified, and production evidence remain distinguishable.
  • A screen banner or model judgment alone cannot produce VERIFIED for a consequential write.
  • Artifact and live-observation privacy boundaries are evaluated separately; compiling a recording does not make it safe to upload.

The public website, documentation, launcher, engine, Desktop, Cloud, repository descriptions, and organization profile must describe this same product.