Skip to content

Verification evidence - 2026-11-01

Historical snapshot only. These results describe the recorded working tree and conditions, not current acceptance. See current verification.

Scratch workspaces and workspace deletion

A new GUI chat used to derive its workspace from the parent of the application directory, so a checkout at D:\LLM made every chat run in D:\. The workspace is the file tools' authority boundary, so that silently treated the whole drive as the project, and it also removed project identity from skills, memory, instruction discovery and coding subagents.

The desktop app no longer supplies an implicit project. A run created with an empty workspace_path gets a Ru-owned scratch workspace under the data root (workspaces/<session id>), written with a .ru-scratch.json marker; the run response returns the canonical workspace_path and workspace_kind, and the session listing returns workspace_kind too. The client adopts the returned path, groups every scratch chat under one Scratch sidebar heading, and approves the scratch trust request silently through the normal trust endpoint.

Deletion follows the session: the deletion coordinator removes the scratch folder after the session record (a locked folder reports ArtifactsPending and a retry finishes it), and a startup sweeper reclaims folders whose session no longer exists. The marker is removed last so an interrupted delete can retry, and a folder without a matching marker is never deleted. The workspace group's Delete workspace… action confirms once, deletes each chat through the single-session endpoint, reports deleted/partial/failed counts, and never deletes the project folder on disk. The CLI keeps its single-session agent session delete (which now also removes that session's scratch folder) because the CLI has no workspace grouping.

The scratch folder is ordinary workspace territory: trust, policy, tools and skills behave exactly as for a chosen project. See ADR 0041.

Evidence

  • ScratchWorkspaceStoreTests: allocation under the data root with a marker; invalid session ids rejected; scratch paths limited to direct children; idempotent delete; a foreign folder without a marker or with another session's marker left in place; a locked file keeps the marker and a retry succeeds; orphan sweep keeps live and foreign folders.
  • HarnessSessionDeletionCoordinatorTests: a scratch folder is deleted with its session; a retry after the session row is gone still deletes the folder; a folder Ru did not create is never deleted.
  • AgentRunIdempotencyTests.Create_without_a_workspace_allocates_a_marked_scratch_workspace: the HTTP create path returns workspace_kind: "scratch" and a marked folder under the data root.
  • GUI: AgentWorkspaceSelectionTests (no implicit project), SidebarOrganizationTests (all scratch sessions share one group), SessionSidebarViewModelTests (bulk delete targets only the group and reports counts, a failed delete does not abort the rest), and the updated source contracts for the request/trust paths.
  • Full .NET matrix on the final tree: Broker 1,416, CLI 2,063, Client 306, Harness.Core 914, Harness.Persistence 366, Hosting 42, Protocol 201, Runtime Adapters 321, Runtime Model 142, SessionInterchange 45, Updates 25, RuBackend 1,632, gui-winui3 2,007 - 9,484 passed, 0 failed (artifacts/scratch-workspaces-matrix.log).
  • Python: 268 passed, 1 skipped (artifacts/scratch-workspaces-python.log).

End-to-end against a managed Host

A build of RuBackend --managed-runtime was started with a throwaway data root and driven through its own Public Contract:

  • POST /sessions/new then POST /agent/v1/runs with workspace_path: "" returned workspace_kind: "scratch" and <data root>/workspaces/<session id>; the folder held the .ru-scratch.json marker naming the session.
  • GET /sessions reported the same workspace_kind and workspace_path.
  • DELETE /sessions/{id} removed the scratch folder.
  • The CLI against the same Host listed the scratch session (workspace_kind: "scratch") and agent session delete <id> --yes removed both the session and its folder.
  • After a Host restart, the startup sweeper removed an orphan folder whose marker named a missing session and kept an unmarked folder in workspaces/.

Two environment facts surfaced while doing this and are not product changes: a data root under D:\ fails DPAPI key hardening because its inherited ACL grants only Modify (owner change needs WRITE_OWNER), so the E2E run used a %LOCALAPPDATA% data root like the shipped app.

Test-run history, reported as required: the first full matrix hung for 18 minutes in the CLI project with no progress; the CLI project then passed on its own. The next two full runs each failed one timing-sensitive test under load (TgSquareWordmarkRasterTests.Frames_use_two_buffers... and UpdateMaintenanceTests.Idle_update_acknowledges...), and each passed alone. Because the sweeper was registered for every Host instance, including test factories, it was moved to managed-runtime hosts only; after that the full matrix passed. A third gap was real: the client's strict create-response parser initially rejected the two new fields (workspace_path, workspace_kind); the parser now allows them, and AgentRunIdempotencyTests plus the Client suite cover the envelope.