Skip to content

L1_INTEGRATION.md

Latest commit

 

History

History
87 lines (73 loc) · 4.17 KB

L1_INTEGRATION.md

File metadata and controls

87 lines (73 loc) · 4.17 KB
Raw

Feeding a layered clinical-data platform (L1 integration)

Layered clinical research platforms separate acquisition (L1: get artifacts out of source systems) from standardization (L2: parse artifacts into a research-grade common data model) and federation (L3). The L1→L2 seam is deliberately thin: an on-disk file under a resolved extraction directory, plus manifest metadata ({file_number}_{date}_{doctype} filename or a CSV sidecar).

openadapt-flow targets the L1 role: instead of hand-building a bespoke acquisition harness per site (per-EMR navigation code, per-screen detection, custom recovery logic), a site operator records the workflow once and the compiler produces the deterministic, self-healing script that acquires the artifacts on every subsequent run — locally, with no per-run model calls on the happy path.

Current product path

  • Record → compile → replay → governed repair end-to-end, vision-only (PNG in, clicks/keys out), validated against a mock EMR-like app including four drift scenarios (theme, layout move, label rename, unexpected modal) in CI.
  • Postconditions per step derived from the recording (what actually changed on screen), so unattended runs verify progress instead of assuming it; semantic drift halts the run with an illustrated report rather than guessing.
  • Risk gate: steps tagged irreversible refuse to act on low-confidence resolutions.
  • L1 artifact emission (openadapt_flow.emit.l1_artifact.emit_l1_artifact): writes a workflow output file into an extraction directory under the canonical {file_number}_{date}_{doctype} name, appends a manifest.csv row, and drops a JSON provenance envelope (sha256, session, tool, version, captured_at) alongside — idempotent on identical re-emits, loud on content conflicts.
from openadapt_flow.emit.l1_artifact import emit_l1_artifact

ref = emit_l1_artifact(
    downloaded_pdf,                # produced by a replayed workflow step
    extraction_dir,                # the root the L2 layer watches
    file_number="P1",
    date="2026-07-06",
    doctype="referral",
    session_id=run_report.workflow_name,
)

Desktop and remote deployment

The compiler and governed runtime are shared across browser, Windows UIA, native macOS, and RDP. The substrate driver supplies observations and action delivery; OpenAdapt remains responsible for target uniqueness, identity, policy, postconditions, independent effects, repair, and audit.

  • Windows UIA: a fixed WinForms workflow passed 3/3 trials with independently confirmed SQLite effects. Native UIA actions produced 12 delivery receipts, while stale and ambiguous targets each refused 3/3.
  • Native macOS: a fixed TextEdit workflow passed 3/3 exact file-byte effect checks and refused a two-window ambiguous selector without modifying either file.
  • RDP: real Aardwolf RDP into Windows 11 passed 3/3 trials for a fixed remote-input task, with exact file readback through an independent guest-tools oracle and no model calls. See backends/RDP.md.

These accepted tasks establish working substrate paths. A clinical deployment qualifies the exact EMR, OS/session policy, display conditions, identity rules, and system-of-record effect oracle before supervised production writes. Citrix ICA/HDX follows the same adapter contract but requires qualification in the customer's published application; RDP evidence is not treated as Citrix evidence.

For artifact acquisition, downloaded files should remain the primary extraction path. When data must be read from a screen, use structured/native value access where available, with clipboard or OCR observations verified against the workflow's declared postcondition or external effect.

Trying it

pip install -e '.[dev]' && playwright install chromium
pytest -q                        # full suite incl. the drift/heal E2E matrix
openadapt-flow demo-record --out /tmp/rec
openadapt-flow compile /tmp/rec --out /tmp/bundle --name demo
openadapt-flow bench /tmp/bundle --n 3 --run-root /tmp/bench

Then read /tmp/bench/BENCH.md and the per-run REPORT.md — the same artifacts a site operator would review after an unattended run.