Skip to content

feat(authoring): let an agent drive recording on native Windows, RDP, and Citrix - #527

Merged
abrichr merged 3 commits into
mainfrom
agent/native-agent-recording
Oct 17, 2026
Merged

abrichr merged 3 commits into
mainfrom
agent/native-agent-recording

Conversation

@abrichr

@abrichr abrichr commented Oct 17, 2026

Copy link
Copy Markdown
Member

AuthoringSession refused windows and rdp/citrix with CoachOnlyError (authoring.py:65-66, :428-430 on main), even though record_desktop_demo already records over WindowsBackend.

Changes:

  • All six backend kinds can be agent-driven. Windows goes through WindowsBackend (the in-guest agent's synthetic input, the same layer replay uses). RDP and Citrix go through the local client window and are screen-only: observe returns no tree, and screenshot() gives the frame.
  • Citrix is its own kind again, so the workflow surface reads citrix. Backend hints are limited to rdp_window / rdp_window_title / rdp_readiness_text.
  • Provenance: each events.jsonl line gets actor (agent/person). meta.json gets demonstrated_by, surface, authoring.input, and coordinate_space: window_pixels for remote displays.
  • The Citrix workspace qualification records through AuthoringSession and takes agent_authored.
  • Tests: mock-WAA Windows record → compile → replay (runs in the windows-mock job), rdp/citrix pixel recording with FakeClient, and an opt-in Docker stand-in e2e (OAFLOW_CITRIX_STANDIN_E2E=1, not run here because Docker isn't available).
  • README: honest scope.

Related: OpenAdaptAI/openadapt-types#47. None of this has run against real Windows, RDP, or Citrix yet; only the mock tests ran.

…trix

AuthoringSession refused windows, rdp, and citrix with COACH_ONLY, so an
agent could only drive the recording in a browser or a macOS or Linux
window. It now accepts every substrate Flow has a backend for. On native
Windows the session drives WindowsBackend, the same synthetic-input path
replay uses; RDP and Citrix are screen-only and drive the local client
window from pixels and keys.

The Recorder that sends the input writes the demonstration, as with
record_desktop_demo; Capture's observers drop injected input on purpose.
Each event now carries actor (agent or person), and meta.json carries
demonstrated_by, surface, and how the input reached the app. rdp/citrix
sessions stamp window_pixels and optional backend_hints for replay.

Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
… session

run_qualification takes agent_authored, which records the canvas stand-in
through AuthoringSession(backend_kind="citrix"), the path an agent drives,
instead of a bare Recorder. The opt-in stand-in e2e gains a test that runs
it and checks the recording says the agent did it. Also types the screen-
only observe branch for mypy.

Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr
abrichr enabled auto-merge (squash) October 17, 2026 03:38
@abrichr
abrichr merged commit 0f2e887 into main Oct 17, 2026
19 checks passed
@abrichr
abrichr deleted the agent/native-agent-recording branch October 17, 2026 04:29
@abrichr abrichr mentioned this pull request Oct 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant