You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Make desktop behavior and package-cost claims for the agentic component library
provable. Contributors and agents should use one supported desktop-driver and
Storybook workflow, CI results must distinguish authoritative native evidence
from capability skips and simulated evidence, and component/theming bundle
boundaries must be reproducible.
Status
In progress. Four of nine direct work items are closed; five remain open.
PR #4306
completed the native-provider implementation represented by #4257:
Added source-built macOS and shared Windows/Win32 helpers behind the public,
Appium-independent Desktop Driver protocol.
Added verified native build/cache selection, exact launch and attach leases,
cancellation, input recovery, normalized accessibility snapshots, and native
window/element capture.
Changed stories-and-tests to attach to the exact owned Storybook
application and execute authored plans through the native provider rather
than the manifest-derived fake target.
Added a shared Desktop Driver setup action and explicit helper build,
verification, doctor, cache, and artifact steps to macOS, Windows Fabric, and
Win32 Paper pull-request jobs.
Documented the architecture, native providers, test integration, CI
authority model, security boundary, and genuinely unfinished qualification
and release work.
PR #4294
previously landed the public package, shared Storybook orchestration, Windows
and Win32 smoke jobs, native catalog traversal, and the first Button, Checkbox,
and Input plans. The retired windows-tests harness was also removed there.
Current evidence boundary
Authored plans now execute through real native providers on macOS, Windows
Fabric, and Win32 Paper. The hosted jobs deliberately disable capabilities
their desktop sessions cannot authoritatively provide: Windows and Win32
disable keyboard, physical click, and wheel; hosted macOS disables physical
click. Those tests report explicit capability skips. A green hosted job proves
the provider, semantic assertions, application identity, and lifecycle paths
that actually ran; it is not proof of interactive input, durable macOS TCC
authority, DPI coverage, or release readiness.
What remains
Promote bundle comparisons from advisory output to reviewed required
budgets after normal variance is established
(#4253).
Restore the eleven focus-crash/focus-movement regression cases removed with
the legacy harness as native portable plans wired to Windows Fabric CI
(#4254).
Expand the portable accessibility cohort beyond Button, Checkbox, and Input;
add accessible-name and keyboard-reachability coverage, reporting, switch
normalization, and authoritative macOS/Windows runs
(#4255).
Qualify and promote the native Storybook jobs: broaden the story cohort,
establish matching manifest evidence, measure flake rate, define retry and
quarantine policy, assign artifact retention/redaction and runner owners,
and require only capabilities available on authoritative runners
(#4256).
Qualify Windows Graphics Capture on unlocked Windows Fabric and Win32
sessions at 100%, 150%, and 200% DPI; add explicit provenance and
locked/non-interactive diagnostics
(#4260).
Complete the remaining Desktop Driver qualification and release hardening
recorded in packages/agentic/desktop-driver/PLAN.md,
including managed macOS authority, interactive Windows qualification,
compatibility, security, cache, performance, reliability, packaging, and
publishing gates.
Native role/state/value support and Button, Checkbox, and Input plans exist; broader name, reachability, reporting, cohort, and authoritative proof remain.
All three desktop PR jobs use native providers and retain diagnostics; coverage expansion, authoritative capabilities, policy, reliability evidence, and required promotion remain.
Native provider implementation and CI integration merged in PR #4306. Remaining production qualification is explicit in the package plan and this rollup.
The public Desktop Driver package is merged, built by the unified graph,
and covered by package tests.
Storybook exposes one shared CLI for manifest generation, native catalog
traversal, authored-plan execution, and ownership-safe process
lifecycles.
macOS, Windows Fabric, and Win32 Paper pull-request jobs explicitly build
and verify native helpers, run stories-and-tests, and retain
diagnostics.
At least one authored semantic plan executes through each real native
provider without a fake fallback.
Cross-platform component and Flex-theme bundle baselines and dependency
boundaries are checked and reported.
At least three components beyond Button have portable role, name, state,
value, and keyboard assertions validated on an authoritative real
desktop endpoint.
The replacement suite preserves all required focus-crash coverage from
the retired Windows harness.
Required capabilities have no false-green path: unavailable capabilities
skip explicitly, and authoritative jobs fail distinctly for product,
application, helper, authority, and infrastructure failures.
Production exit criteria
macOS, Windows Fabric, and Win32 Paper run the required authored cohort on
authoritative environments with auditable manifest and helper identity.
Retry, quarantine, expiry, artifact retention/redaction, escalation, and
runner-capacity policies have named owners.
Backend prerequisites and compatibility ranges are published and tested.
The driver package passes its security, performance, cache, reliability,
packaging, external-consumer, and publishing gates.
Bundle growth budgets are required gates and baseline changes require
reviewed rationale.
Flex-only consumers exclude legacy and optional capabilities unless those
capabilities are explicitly imported.
Windows Fabric and Win32 Paper publish qualified window-scoped visual
evidence with capture provenance and locked-session diagnostics; pixel
differences remain evidence, not a pass/fail gate.
Portable accessibility assertions are required for the initial component
cohort on macOS and Windows.
Agent context
The supported public contract is client-neutral and Appium-free.
A fake-host pass proves schema and orchestration determinism only.
Native story traversal proves renderability; authored native plans prove only
the capabilities and assertions that did not explicitly skip.
Deterministic gates and optional visual evidence remain separate.
Interactive Windows qualification requires a logged-in, unlocked desktop.
Authoritative macOS qualification requires a managed, auto-logged-in Aqua
session, stable signing identity, and explicit TCC/PPPC authority.
Windows Fabric still excludes Accordion and Callout because they fail-fast
the current RNW Fabric host during traversal. Platform exclusions remain
explicit in generated manifests and reports.
Goal
Make desktop behavior and package-cost claims for the agentic component library
provable. Contributors and agents should use one supported desktop-driver and
Storybook workflow, CI results must distinguish authoritative native evidence
from capability skips and simulated evidence, and component/theming bundle
boundaries must be reproducible.
Status
In progress. Four of nine direct work items are closed; five remain open.
PR #4306
completed the native-provider implementation represented by
#4257:
Appium-independent Desktop Driver protocol.
cancellation, input recovery, normalized accessibility snapshots, and native
window/element capture.
stories-and-teststo attach to the exact owned Storybookapplication and execute authored plans through the native provider rather
than the manifest-derived fake target.
verification, doctor, cache, and artifact steps to macOS, Windows Fabric, and
Win32 Paper pull-request jobs.
authority model, security boundary, and genuinely unfinished qualification
and release work.
PR #4294
previously landed the public package, shared Storybook orchestration, Windows
and Win32 smoke jobs, native catalog traversal, and the first Button, Checkbox,
and Input plans. The retired
windows-testsharness was also removed there.Current evidence boundary
Authored plans now execute through real native providers on macOS, Windows
Fabric, and Win32 Paper. The hosted jobs deliberately disable capabilities
their desktop sessions cannot authoritatively provide: Windows and Win32
disable keyboard, physical click, and wheel; hosted macOS disables physical
click. Those tests report explicit capability skips. A green hosted job proves
the provider, semantic assertions, application identity, and lifecycle paths
that actually ran; it is not proof of interactive input, durable macOS TCC
authority, DPI coverage, or release readiness.
What remains
budgets after normal variance is established
(#4253).
the legacy harness as native portable plans wired to Windows Fabric CI
(#4254).
add accessible-name and keyboard-reachability coverage, reporting, switch
normalization, and authoritative macOS/Windows runs
(#4255).
establish matching manifest evidence, measure flake rate, define retry and
quarantine policy, assign artifact retention/redaction and runner owners,
and require only capabilities available on authoritative runners
(#4256).
sessions at 100%, 150%, and 200% DPI; add explicit provenance and
locked/non-interactive diagnostics
(#4260).
recorded in
packages/agentic/desktop-driver/PLAN.md,including managed macOS authority, interactive Windows qualification,
compatibility, security, cache, performance, reliability, packaging, and
publishing gates.
Work items
Beta exit criteria
and covered by package tests.
traversal, authored-plan execution, and ownership-safe process
lifecycles.
and verify native helpers, run
stories-and-tests, and retaindiagnostics.
provider without a fake fallback.
boundaries are checked and reported.
value, and keyboard assertions validated on an authoritative real
desktop endpoint.
the retired Windows harness.
skip explicitly, and authoritative jobs fail distinctly for product,
application, helper, authority, and infrastructure failures.
Production exit criteria
authoritative environments with auditable manifest and helper identity.
runner-capacity policies have named owners.
packaging, external-consumer, and publishing gates.
reviewed rationale.
capabilities are explicitly imported.
evidence with capture provenance and locked-session diagnostics; pixel
differences remain evidence, not a pass/fail gate.
cohort on macOS and Windows.
Agent context
the capabilities and assertions that did not explicitly skip.
session, stable signing identity, and explicit TCC/PPPC authority.
the current RNW Fabric host during traversal. Platform exclusions remain
explicit in generated manifests and reports.
Merged evidence