Repository navigation
Description
Current status - 2026-11-16
Blocked for the qualified support baseline by F15 #4441 (open, partial qualification). Evidence collection can continue; an open baseline gate does not prohibit parallel endpoint work.
Landed since the foundation: neutral macOS and Windows application ownership and containment (#4492, #4494), generic attach-only handoff with macOS and Windows native evidence (#4491), manifest provenance and credential isolation (#4490), macOS permission-enabled partial qualification with open findings (#4493), macOS physical-input fixes and an advisory physical CI lane (#4499, #4500), and Windows activation refusal plus Win32 secondary popup registration (#4498). #4445, #4446, #4447 and #4448 are closed. Each PR description records its exact revisions, commands and results; each package's references/current-status.md keeps the latest result per endpoint with a link. Historical results stay at their exact source revisions. No all-pass or merged-head native qualification claim is justified, and macOS Fabric, macOS Paper, Windows Fabric and Win32 Paper are reported separately.
Remaining consolidation includes:
- The per-endpoint failures and gaps enumerated in Infrastructure: Live-qualify the desktop driver and MCP foundation on macOS, Windows, and Win32 #4441, which also carries the remaining F1-F5 and F16 native evidence
- Application-ownership reliability cases that Add neutral macOS application ownership with exact workload identity #4492 and feat(desktop-test): qualify Windows application ownership and containment #4494 list as unqualified: actual forced OS PID recycling, arbitrary descendant discovery/containment (macOS launch kinds are explicitly uncontained), power-loss rename during claim publication, Metro/channel survivors, and repeated-run reliability
- Cross-service exact-application exclusion and input-mutex qualification (from F2 Infrastructure: Implement settled desktop lifecycle supervision and fenced ownership records #4442)
- A real non-React integrated application and AppKit fixture
- Independent supervisor, helper, app, Metro, bridge, runner and MCP faults
- A repeated-run reliability window
Standing gates carried over from the retired desktop-test checklist:
- Automatic reclaim of durable-uncertainty claims is Unsupported. Claim schema 3 records OS start identity; schema 2 and unverified claims inspect as unknown. List it as
unsupportedin the matrix. Automatic adoption/recovery is not implemented; recovery records imply neither reclaim nor retained-app handoff. - Arbitrary descendants on macOS remain unqualified (owners Infrastructure: Live-qualify the desktop driver and MCP foundation on macOS, Windows, and Win32 #4441 and this item).
- Physical-input suites must never run in parallel on one interactive desktop.
Coordinate release/CI windows with #4345/#4349 rather than claiming duplicate independent evidence. Managed macOS input lanes from F16 are tracked in #4349. Capture fidelity is #4348.
- Serialized
--input all(moved from F16 Infrastructure: Enforce session input policy and attribute delivered input mechanisms #4467): run the input lanes one after another and account for each lane separately. An interrupted run reports the remaining lane as Not run. Never run physical lanes in parallel on one interactive desktop.
Discovery/operation bytes and tokens, tool counts, cold/warm startup and latency, retries/recovery cost, retained resources and cleanup need exact revision/client/tokenizer/sample-count/artifact methods. Windows Fabric and Win32 Paper MCP measurements are recorded in #4498 and macOS Fabric and macOS Paper measurements in #4493 (revision 141b88b4b); see the Measurements section of packages/agentic/desktop-driver-mcp/references/current-status.md. A successful macOS click cost and mutation-cancellation costs remain Not run. Historical macOS numbers and green hosted jobs cannot fill those cells. Preserve every failed attempt and every explicit Unsupported/Skipped outcome.
Goal
Maintain a revision- and endpoint-specific support matrix for the desktop automation packages, covering injected lifecycle failures and agent-efficiency measurements. Each feature item produces its own native and consumer evidence when it lands; this item consolidates that evidence and fills the gaps that only show up across components.
Part of #4435 (plan item F14).
Scope
- Owner: cross-platform qualification; each feature owner remains responsible for its own evidence.
- Paths (under
packages/agentic/): driver native fixtures and contracts, harness and Storybook consumer tests, owning app CI and scripts, and thereferences/current-status.mdfiles. - Inject supervisor, helper, app, Metro, bridge, runner, and MCP failures independently. Cover same-port replacement, surviving children, stale descriptors, late renders, grant revocation, and uncertain cleanup.
- Include a non-React application, and add the AppKit fixture rather than assuming it exists.
- Report macOS Fabric, macOS Paper, Windows Fabric, and Win32 Paper separately.
- Measure identical tasks for schema overhead, tool calls and response bytes, cold and warm startup, retries, recovery cost, retained resources, and cleanup correctness, and record them with their method (per-run evidence in the pull request description or an issue; latest result per endpoint in
references/current-status.md).
Dependencies
- Blocked by: F15 Infrastructure: Live-qualify the desktop driver and MCP foundation on macOS, Windows, and Win32 #4441 (open), which establishes the baseline for the delivered foundation.
- Consolidates: each advertised feature only after its implementation and endpoint evidence land.
- Related: Infrastructure: Complete Desktop Driver production release qualification #4345 (driver production release qualification) and Infrastructure: Qualify and require native Storybook CI coverage #4349 (required native CI).
Acceptance
- The matrix lists each endpoint and feature with passed, failed, skipped, not-run, or unsupported status at an exact revision.
- Hosted physical-input or capture skips are recorded as skipped, not passed; a successful semantic
activateis not evidence that a physical click works. - Measurements record revision, client, platform, renderer, tokenizer, sample count, and artifact locator.
- Historical results are never reported as qualification of a later revision.
Validation
- Iterate with focused
yarn test <name>in each changed package, then run itsyarn format,yarn lint,yarn build, andyarn test. Finish withyarn clean-lage-cachethenyarn lage buildci; runyarn lage test-links --verbose --groupedafter markdown changes. - Add a patch changeset for each changed published package. Update the affected topic section and its status-overview line in the owning
references/current-status.md, and affected README/AGENTS guidance, in the same change. - Native, provider, or process-topology changes need matching-OS native contracts and the affected consumer lane. JavaScript tests and bundles do not qualify native behavior; hosted capability skips are not passes. Report each check as passed, failed, skipped, not run, or unsupported.
Evidence
- Latest per-endpoint results: "Reliability and concurrency" and "Application ownership" in
packages/agentic/desktop-test/references/current-status.md, plus the status overviews inpackages/agentic/desktop-driver/references/current-status.mdandpackages/agentic/desktop-driver-mcp/references/current-status.md. Per-run evidence lives in pull request descriptions and issues; each package'sreferences/current-status.mdkeeps a status overview, stable topic sections, and only the latest result per native endpoint with a link. - Retired plan:
driver-next.mdat0569e1b86, sections 8 (F14) and 11.
Activity
JasonVMo commentedon Nov 16, 2026
Status update: still blocked for the qualified baseline by #4441, which stays open with partial qualification. The body now lists landed PRs (#4490-#4494, #4498-#4500) and adds residuals from the closed ownership issues: forced PID recycling, arbitrary descendants, power-loss rename, Metro/channel survivors and repeated-run reliability. It also records that automatic reclaim is Unsupported and that physical-input suites must not run in parallel on one desktop.