* feat: add data model foundation for progressive skeleton traversal
Add children_count, root_ref, skeleton fields to core types. Add
get_subtree() to PlatformAdapter trait. Add remove_by_root_ref() and
write-side size check to RefMap. No behavioral changes yet.
* feat: implement progressive skeleton traversal with ref-rooted drill-down
Add --skeleton and --root flags to snapshot command for token-efficient
accessibility tree exploration. Skeleton mode clamps depth to 3 levels
and annotates truncated containers with children_count, allowing AI
agents to discover regions before drilling into them. Named containers
at skeleton boundaries (via name or description) receive refs as
drill-down targets. The --root flag starts traversal from a previous
ref with scoped invalidation — only refs from that drill-down are
replaced on re-drill.
Key changes:
- New ref_alloc.rs: shared ref helpers (INTERACTIVE_ROLES, actions_for_role,
ref_entry_from_node, is_collapsible) extracted from snapshot.rs
- New snapshot_ref.rs: drill-down logic with DrillDownConfig, scoped
invalidation via root_ref tagging on RefEntry
- macOS count_children() uses raw CFArrayGetCount without materializing
AXElement wrappers for performance at skeleton boundaries
- RefMap write-side size check prevents >1MB files
- Skeleton anchors consider both name and description for Electron compat
* fix: mention --skeleton in STALE_REF error suggestion
* docs: document --skeleton and --root flags in skill reference
* docs: update phases.md and CLAUDE.md for progressive skeleton traversal
Add skeleton traversal as Phase 1 objective P1-O10. Document --skeleton
and --root flags, get_subtree() trait method, new core modules, and
platform-agnostic notes for Phase 2/3. Update risk mitigations and
performance optimizations table.
* docs: make progressive skeleton traversal the default agent workflow
Update SKILL.md observe-act loop to skeleton-first approach. Add
progressive skeleton traversal as the primary workflow pattern. Update
anti-patterns, key principles, and command quick reference to lead
with --skeleton + --root. Full snapshot remains documented as fallback
for simple apps.
* fix: preserve skeleton drill-down anchors
* fix: preserve drill-down refs across skeleton re-snapshots
Skeleton snapshots now load the existing refmap and remove only
skeleton-level refs (root_ref: None), preserving drill-down refs
accumulated via --root. Previously, build() always created a fresh
RefMap, breaking the skeleton → drill → act → skeleton(verify) workflow
by wiping all drill-down refs on the verify step.
* fix: preserve drill-down depth for root snapshots
* fix: verify AX action effect on Electron elements before trusting success
Chromium's AX implementation returns kAXErrorSuccess for AXPress,
AXConfirm, etc. but only toggles ARIA state without firing DOM event
handlers. This caused the click chain to short-circuit on false
positives, preventing CGClick from ever being reached.
The fix detects web elements via AXWebArea ancestor walk, then verifies
each AX action actually had a DOM effect by comparing focused element
pointers before and after. When all AX methods produce no real effect,
CGClick fires as the genuine last resort.
Native elements are unaffected — the original verified_press behavior
is preserved in a separate code path.
* chore: ignore .context/ for local tooling artifacts
* style: collapse web_action_had_effect signature to single line
* test: cover RefMap save oversize rejection
Extract the serialize+size-check step into a private helper so the 1MB
write rejection path can be unit-tested without filesystem I/O. Also
moves the size check above the directory creation in save() so an
oversized refmap no longer creates ~/.agent-desktop on rejection.
* test: assert stale-ref suggestion mentions --skeleton
Locks in the Phase 4 polish requirement that STALE_REF errors guide
agents back to a skeleton refresh, not just a plain snapshot.
* test: cover skeleton-to-drill-down counter continuity
Asserts that skeleton refs (@e1..@e10) survive a scoped invalidation
of @e3, that drill-down refs allocated after the skeleton continue from
@e11 instead of resetting, and that remove_by_root_ref drops only the
drill-down children.
* test: cover --root + --surface rejection at execute() boundary
Adds a NoopAdapter that uses every PlatformAdapter trait default to
exercise execute()'s validation guards without standing up a real
adapter. Verifies that combining --root with a non-Window surface
returns INVALID_ARGS, and that the Window surface does not trigger the
guard.
* test: add filesystem-redirected integration tests for run_from_ref
Adds a thread-local HOME override + RAII HomeGuard so RefMap::save and
RefMap::load can be exercised against an isolated temp directory without
env-var racing across parallel tests. Also adds a StubAdapter that
implements PlatformAdapter with a canned subtree response so the full
run_from_ref drill-down flow can be unit-tested without standing up
macOS AX state.
Closes seven Phase 3 plan acceptance items in one pass:
- save+load roundtrip with HOME override
- oversize save rejection preserves previous file on disk
- run_from_ref returns subtree and persists drill refs
- stale root ref → STALE_REF with skeleton suggestion
- re-drill replaces drill refs only, counter continues
- multiple drill-downs from @e1 and @e2 coexist
- empty subtree drill-down produces no new refs
* test: add golden fixtures for skeleton output and drill-down refmap
Adds two committed JSON fixtures under tests/fixtures/ plus matching
unit tests that build the same input trees, run them through
allocate_refs / run_from_ref, and assert the produced ref ids,
parent-child layout, and root_ref tagging match the golden expectations.
Locks in the JSON shape so future serialization changes have to be
deliberate.
* fix: bound build_subtree recursion on Electron wrapper chains
Anonymous AXGroup/AXGenericElement wrappers do not advance the semantic
depth counter (web-wrapper depth-skip), so a long chain of nested
wrappers could recurse arbitrarily deep without ever hitting either
max_depth or ABSOLUTE_MAX_DEPTH. The previous code checked
`depth >= ABSOLUTE_MAX_DEPTH` against the semantic depth, which stayed
at 0 throughout a wrapper chain. On pathological Electron trees this
risked stack exhaustion before any cap engaged.
Add a separate `raw_depth` parameter that always increments and use it
for the absolute cap. Semantic `depth` still drives `max_depth` and the
skeleton boundary so the wrapper-flattening behavior is preserved on
normal trees.
Also drops the long-unused `_include_bounds` parameter; bounds
filtering happens in `allocate_refs` in core, not here.
* chore: add pre-commit hook running fmt + clippy + tests
Mirrors the CI quality gates (cargo fmt --check, cargo clippy
--all-targets -- -D warnings, cargo test --lib --workspace) so a
failing commit never reaches origin. Skips automatically when no
Rust/TOML files are staged. Bypass with --no-verify or SKIP_PRECOMMIT=1
when a non-Rust hotfix is genuinely needed.
Hook lives under .githooks/ and is opt-in per clone:
git config core.hooksPath .githooks
Setup instructions added to CLAUDE.md.
* fix: prevent orphaned drill-down refs leaking across skeleton refresh
remove_skeleton_refs() previously dropped every entry whose root_ref
was None and kept every entry whose root_ref was Some. After a series
of skeleton refreshes that pattern leaked drill-down refs whose root
anchor had already been removed: they could never be cleaned up by a
future remove_by_root_ref(target) because target referred to a
non-existent anchor, and they accumulated until the 1MB write guard
fired.
Restructure the function to keep only:
- skeleton anchors that have at least one drill-down pointing at them
- drill-down refs whose root anchor is among the kept anchors
Unreferenced anchors are still dropped (the original intent), and
orphaned drill-downs are dropped at the same time so the refmap can
no longer accumulate dead entries.
Updated existing test to assert the new keep-when-referenced semantics
and added a regression test for the orphan-drilldown leak path.
* fix: align resolve traversal with snapshot child-attribute set
find_element_recursive previously walked AXChildren first then fell
back to AXContents only, and resolve_element_name only derived its
fallback label from AXChildren. Snapshot building, however, uses the
full child_attributes() list (AXChildren, AXContents,
AXChildrenInNavigationOrder) via copy_children. That asymmetry meant
refs minted for containers exposed only through AXChildrenInNavigationOrder
appeared in the snapshot output but produced false STALE_REF errors on
drill-down or action commands because the resolver could not find them
again.
Route both call sites through child_attributes() so resolution sees
the same children as snapshotting.
* fix: bypass AXConfirm on web elements and add skeleton anchors to drill-down
* fix: compare AX elements via CFEqual in web_action_had_effect
AXUIElementCopyAttributeValue follows the CoreFoundation Create rule
and returns a freshly-allocated CF object on every call, so raw
pointer equality (`before.0 != after.0`) between two separate copies
of AXFocusedUIElement was ALWAYS true even when the focused element
was unchanged. That made Step 1 (AXPress) of activate_web_element
appear successful every time and left Steps 2-4 of the escalation
chain (AXConfirm bypass, child actions, CGClick fallback) dead code
on web elements.
Replace the pointer comparison with CFEqual, which compares
accessibility element identity rather than heap addresses.
* refactor: unify allocate_refs across snapshot and drill-down paths
allocate_refs in snapshot.rs and allocate_refs_with_root in
snapshot_ref.rs were near-exact copies that differed only in whether
each allocated ref carried a root_ref tag. The duplication meant any
bug fix or behavior change in one path risked silently drifting from
the other (bot flagged this after the first round of fixes added even
more duplicated skeleton-anchor logic to allocate_refs_with_root).
Move the shared logic into ref_alloc.rs behind a new RefAllocConfig
struct with Option<&str> for the root_ref_id, and delete the
snapshot_ref.rs copy. Both snapshot::build and snapshot_ref::run_from_ref
now route through the same function with their respective configs.
append_surface_refs and the existing unit tests in both modules are
updated to use the shared function + config.
* chore: drop with_root from drill test names after allocator unification
The snapshot_ref tests kept their original test_allocate_refs_with_root_*
names after the allocator dedupe even though allocate_refs_with_root no
longer exists. Rename them to test_drill_alloc_* so grep for
allocate_refs_with_root returns zero matches and nobody assumes a second
allocator path exists.
* docs: compound DRY ref-allocator dedupe into knowledge base
Adds docs/solutions/best-practices/deduplicate-ref-allocator-via-config-struct-2026-04-14.md
documenting the root cause, guidance, consequences, trigger conditions,
and before/after of the allocate_refs / allocate_refs_with_root
duplication that landed on PR #20 and was unified in d06a6c2.
Also:
- gitignore allows docs/solutions/** (was caught by docs/* rule)
- CLAUDE.md workspace tree surfaces docs/solutions/ so fresh agents
discover the knowledge store
- docs/phases.md drops two stale DrillDownConfig references (lines 62
and 226) — the symbol no longer exists in the codebase
* refactor: drop unused root_ref field from TreeOptions
TreeOptions.root_ref was populated from SnapshotArgs.root_ref by
tree_options() and then never read anywhere. The actual routing
lives in execute(): `if let Some(ref root) = args.root_ref` branches
to snapshot_ref::run_from_ref(adapter, &opts, root), passing the
root id as an explicit argument — opts.root_ref was never consulted
by any caller.
Delete the field from TreeOptions and its Default impl and stop
populating it in tree_options(). SnapshotArgs.root_ref remains as
the single source of truth for the drill-down target.
* fix: suppress skeleton flag when --root is set
tree_options() clamped depth based on skeleton && root_ref.is_none()
but still forwarded skeleton: args.skeleton unconditionally. When a
caller passed --skeleton --root @e3, opts.skeleton stayed true, so
adapter.get_subtree on macOS built a truncated skeleton tree instead
of the full drill-down subtree, AND ref_alloc::allocate_refs tagged
skeleton-anchor refs with the drill-down root_ref — both unintended.
Tie the skeleton flag to the same root_ref.is_none() guard the depth
clamp already uses. A drill-down is always a full subtree now, never
a skeleton view.
Adds test_tree_options_suppresses_skeleton_for_drill_down to lock the
behavior in and extends the existing clamp test to assert the flag
still propagates on non-drill paths.
* fix: detect web action effect via value + selected + focus change
The previous focus-change-only heuristic in web_action_had_effect was
wrong for non-Chromium AXWebArea elements (Safari WKWebView, Mail,
Notes): AXPress on a checkbox toggles AXValue correctly but does NOT
shift the focused UI element. web_action_had_effect returned false, the
chain escalated to CGClick, and CGClick toggled the checkbox a second
time — net zero visible change. The pointer-equality bug that ae78cbc
replaced was masking this by always returning true; CFEqual exposed
the real signal gap.
Capture element state into a PreActionState struct (focused, value,
selected) before any AX action runs, and consider the action to have
had an effect if ANY of those three fields differs afterward. Element
value covers WKWebView checkboxes, text fields, sliders. Selected
covers tabs, radio buttons, and list items. Focused still covers
Chromium's DOM-driven focus shifts. Triggers on whichever signal is
most relevant to the element type without needing role-specific code.
Also re-exports copy_bool_attr from crates/macos/src/tree for use by
the chain-steps module.
* fix: emit skeleton boundary when drill path is about to hit raw cap
In skeleton mode, a chain of anonymous web wrappers (AXGroup /
AXGenericElement with empty title and value) never advances the
semantic `depth` counter, so `child_depth > max_depth` never fires
no matter how deep the chain runs. The recursion then hits the
`raw_depth >= ABSOLUTE_MAX_DEPTH` top-of-function guard and silently
returns None, dropping the subtree without any children_count marker.
Agents see truncated output with no indication that anything was
dropped.
Extend the at_skeleton_boundary condition to also fire when
child_raw_depth is about to hit ABSOLUTE_MAX_DEPTH, so skeleton mode
always emits a visible truncation node for deep wrapper chains
instead of disappearing them. The old `raw_depth < ABSOLUTE_MAX_DEPTH`
half of the guard was redundant — the top-of-function check already
ensures the current frame has raw budget; the meaningful question is
whether the CHILD frames will.
* perf: compute is_in_webarea once per verified-press call
try_focus_then_verified_confirm_or_press called is_in_webarea(el) to
gate AXConfirm, then fell through to do_verified_press which called
is_in_webarea(el) again on its first line. Each call walks up to 20
AX parents via per-element IPC (AXUIElementCopyAttributeValue on
AXParent), so every web-area click through this path paid the cost
twice.
Extract a private dispatch_verified_press(el, caps, in_web) that
takes the web-area flag as an input. do_verified_press keeps its
existing (el, caps) signature for chain_defs compatibility and
computes is_in_webarea once, then delegates. The focus-then-confirm
path also computes it once and passes it through. The result is
identical on both paths with one AX traversal instead of two.
* fix: skeleton anchors in drill-downs must not inherit root_ref; restore AXBrowser AXContents fallback
Skeleton anchors discovered while processing a drill-down subtree were
being tagged with the drill root_ref, making them indistinguishable from
regular drill entries. re-drills would delete those anchors via
remove_by_root_ref, breaking sub-drilling. They now always carry
root_ref=None so they survive re-drills as stable targets.
AXBrowser child resolution dropped the AXContents fallback when
child_attributes() was unified — the resolver now uses
["AXColumns","AXContents"] matching the original traversal order.
* fix: suppress skeleton anchor creation in drill-down mode to prevent orphaned ref accumulation
* fix: bounds-based resolver pruning and CGClick fallback on chain timeout
Resolver was doing exhaustive DFS through entire AX tree (depth 50) to
find a single element. For large documents like a dense spreadsheet this
meant visiting tens of thousands of cells via AX IPC, causing multi-minute
hangs before the click chain even started.
Two changes:
- find_element_recursive now prunes subtrees whose spatial bounds do not
contain the target element's centre point, and aborts the whole search
after five seconds. Numbers table -> button not inside -> entire table
skipped. Resolution went from >1m49s to <50ms.
- execute_chain now sets a 1s app-level AX messaging timeout so calls to
children/parents fail fast instead of blocking for the system default 6s,
and when the 10s chain deadline fires it attempts the CGClick step before
returning failure so the coordinate fallback is always reached.
* refactor: resolve 9 code-review todos for progressive skeleton traversal
- skeleton refresh now starts from RefMap::new() (removes stale ref accumulation across refreshes)
- drill-down snapshot resolves real WindowInfo via PID lookup instead of synthetic empty id
- --root flag validates ref format before RefMap lookup (returns INVALID_ARGS not STALE_REF)
- ABSOLUTE_MAX_DEPTH truncation emits boundary node with children_count instead of silently dropping
- fix ref vs ref_id field name in commands-observation.md JSON examples
- fix phases.md TreeOptions listing root_ref (it lives in SnapshotArgs, not TreeOptions)
- add batch snapshot skeleton/root examples to commands-system.md
- split snapshot.rs and snapshot_ref.rs test modules into separate _tests.rs files (143/69 LOC)
- add integration tests for skeleton + drill-down workflow including invalid --root validation
* docs: compound progressive snapshot review hardening
* docs: tighten progressive snapshot compound write-up
* docs: update README with progressive skeleton traversal and fix command count to 50
* docs: correct command count to 53, add Notifications section and platform notes
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
74 KiB
agent-desktop — Phase Roadmap
Source of truth for the phased delivery plan. Derived from PRD v2.0 and the Skill Maintenance Addendum.
Phase Overview
| Phase | Name | Status | Platforms |
|---|---|---|---|
| 1 | Foundation + macOS MVP | Completed | macOS |
| 2 | Windows Adapter | Planned | macOS, Windows |
| 3 | Linux Adapter | Planned | macOS, Windows, Linux |
| 4 | MCP Server Mode | Planned | All |
| 5 | Production Hardening | Planned | All |
Each phase is strictly additive. Core engine, CLI parser, JSON contract, error types, snapshot engine, and command registry are never modified — only new PlatformAdapter implementations, new transports, and new modes are added.
Phase 1 — Foundation + macOS MVP
Status: Completed
Phase 1 is the load-bearing phase. It establishes every shared abstraction, every trait boundary, every output contract, every error type, the complete command trait and registry, and the full workspace structure. All subsequent phases build on top of this foundation without modifying core.
Objectives
| ID | Objective | Success Metric |
|---|---|---|
| P1-O1 | Working macOS snapshot CLI | snapshot --app Finder returns valid JSON with refs for all interactive elements |
| P1-O2 | Platform adapter trait | Trait compiles with mock adapter; macOS adapter satisfies all trait methods |
| P1-O3 | Ref-based interaction | click @e3 successfully invokes AXPress on the resolved element |
| P1-O4 | Context efficiency | Typical Finder snapshot < 500 tokens (measured via tiktoken) |
| P1-O5 | Typed JSON contract | Output validates against JSON Schema; schema is versioned |
| P1-O6 | Permission detection | Missing Accessibility permission prints specific macOS setup instructions |
| P1-O7 | Command extensibility | Adding a new command requires exactly 1 new file + 2 registration lines |
| P1-O8 | 50 working commands | All commands pass integration tests |
| P1-O9 | CI pipeline | GitHub Actions macOS runner executes full test suite on every PR |
| P1-O10 | Progressive skeleton traversal | Skeleton + drill-down workflow achieves 78%+ token savings on Electron apps |
Workspace Structure
agent-desktop/
├── Cargo.toml # workspace: members, shared deps
├── rust-toolchain.toml # pinned Rust version
├── clippy.toml # project-wide lint config
├── schemas/ # JSON Schema files for output validation
│ ├── snapshot_response.json
│ ├── action_response.json
│ └── error_response.json
├── crates/
│ ├── core/ # agent-desktop-core (platform-agnostic)
│ │ └── src/
│ │ ├── lib.rs # public re-exports only
│ │ ├── node.rs # AccessibilityNode, Rect, WindowInfo
│ │ ├── adapter.rs # PlatformAdapter trait
│ │ ├── action.rs # Action enum, ActionResult, InputEvent, WindowOp
│ │ ├── refs.rs # RefAllocator, RefMap, RefEntry
│ │ ├── ref_alloc.rs # Shared ref-allocation logic (RefAllocConfig, allocate_refs, INTERACTIVE_ROLES, is_collapsible)
│ │ ├── snapshot_ref.rs # Ref-rooted drill-down (run_from_ref) — delegates allocation to ref_alloc
│ │ ├── snapshot.rs # SnapshotEngine (filter, allocate, serialize)
│ │ ├── error.rs # ErrorCode enum, AdapterError, AppError
│ │ ├── output.rs # Response envelope, JSON formatting
│ │ ├── command.rs # Command trait + CommandRegistry
│ │ └── commands/ # one file per command
│ ├── macos/ # agent-desktop-macos (Phase 1)
│ ├── windows/ # agent-desktop-windows (stub → Phase 2)
│ └── linux/ # agent-desktop-linux (stub → Phase 3)
├── src/ # agent-desktop binary (entry point)
│ ├── main.rs
│ ├── cli.rs
│ ├── cli_args.rs
│ ├── dispatch.rs
│ └── batch_dispatch.rs
└── tests/
├── fixtures/
└── integration/
PlatformAdapter Trait
The single most important abstraction. Every platform-specific operation goes through this trait. Core never imports platform crates. 12 methods with default implementations returning not_supported():
pub trait PlatformAdapter: Send + Sync {
fn list_windows(&self, filter: &WindowFilter) -> Result<Vec<WindowInfo>>;
fn list_apps(&self) -> Result<Vec<AppInfo>>;
fn get_tree(&self, win: &WindowInfo, opts: &TreeOptions) -> Result<AccessibilityNode>;
fn execute_action(&self, handle: &NativeHandle, action: Action) -> Result<ActionResult>;
fn resolve_element(&self, entry: &RefEntry) -> Result<NativeHandle>;
fn check_permissions(&self) -> PermissionStatus;
fn focus_window(&self, win: &WindowInfo) -> Result<()>;
fn launch_app(&self, id: &str, wait: bool) -> Result<WindowInfo>;
fn close_app(&self, id: &str, force: bool) -> Result<()>;
fn screenshot(&self, target: ScreenshotTarget) -> Result<ImageBuffer>;
fn get_clipboard(&self) -> Result<String>;
fn set_clipboard(&self, text: &str) -> Result<()>;
fn synthesize_input(&self, input: InputEvent) -> Result<()>;
fn manage_window(&self, win: &WindowInfo, op: WindowOp) -> Result<()>;
fn get_subtree(&self, handle: &NativeHandle, opts: &TreeOptions) -> Result<AccessibilityNode>;
fn list_notifications(&self) -> Result<Vec<NotificationInfo>>;
fn dismiss_notification(&self, id: &str) -> Result<()>;
fn interact_notification(&self, id: &str, action_name: &str) -> Result<ActionResult>;
fn list_tray_items(&self) -> Result<Vec<TrayItemInfo>>;
fn interact_tray_item(&self, id: &str, action: Action) -> Result<ActionResult>;
}
Key Supporting Types
Action— Click, DoubleClick, RightClick, SetValue(String), SetFocus, Expand, Collapse, Select(String), Toggle, Scroll(Direction, Amount), PressKey(KeyCombo)InputEvent— MouseMove(x,y), MouseClick(x,y,button,count), MouseDown(button), MouseUp(button), MouseWheel(dy,dx), KeyDown(key), KeyUp(key), Drag(from,to)WindowOp— Resize(w,h), Move(x,y), Minimize, Maximize, Restore, CloseScreenshotTarget— Screen(index), Window(id), Element(NativeHandle), FullScreenNotificationInfo— id, app_name, title, body, timestamp, actions: Vec<String>, is_persistentTrayItemInfo— id, app_name, title, tooltip, has_menuTreeOptions—max_depth,include_bounds,interactive_only,compact,surface,skeleton(root is CLI-only viaSnapshotArgs.root_ref, not plumbed intoTreeOptions)
macOS Adapter Implementation
Located in crates/macos/src/ following the platform crate folder structure:
crates/macos/src/
├── lib.rs # mod declarations + re-exports only
├── adapter.rs # MacOSAdapter: PlatformAdapter impl
├── tree/
│ ├── mod.rs # re-exports
│ ├── element.rs # AXElement struct + attribute readers
│ ├── builder.rs # build_subtree, tree traversal
│ ├── roles.rs # AXRole string → unified role enum mapping
│ ├── resolve.rs # Element re-identification for ref resolution
│ └── surfaces.rs # Surface detection (menu, sheet, alert, popover)
├── actions/
│ ├── mod.rs # re-exports
│ ├── dispatch.rs # perform_action match arms
│ ├── activate.rs # Smart AX-first activation chain (15-step)
│ └── extras.rs # select_value, ax_scroll
├── input/
│ ├── mod.rs # re-exports
│ ├── keyboard.rs # CGEventCreateKeyboardEvent, key synthesis, text typing
│ ├── mouse.rs # CGEventCreateMouseEvent, mouse events
│ └── clipboard.rs # NSPasteboard.generalPasteboard read/write
├── notifications/
│ ├── mod.rs # re-exports
│ ├── list.rs # List notifications via Notification Center AX tree
│ ├── dismiss.rs # Dismiss individual or all notifications via AXPress
│ └── interact.rs # Click notification action buttons
├── tray/
│ ├── mod.rs # re-exports
│ ├── list.rs # List menu bar extras via SystemUIServer AX tree
│ └── interact.rs # Click menu bar extras, expand menus
└── system/
├── mod.rs # re-exports
├── app_ops.rs # launch, close, focus via NSWorkspace / AppleScript
├── window_ops.rs # window resize, move, minimize, maximize, restore
├── key_dispatch.rs # app-targeted key press
├── permissions.rs # AXIsProcessTrusted(), AXIsProcessTrustedWithOptions(prompt: true)
├── screenshot.rs # CGWindowListCreateImage
└── wait.rs # wait utilities
Tree traversal:
- Entry:
AXUIElementCreateApplication(pid)for app root - Children:
kAXChildrenAttributerecursively with ancestor-path set (not global visited set — macOS reuses AXUIElementRef pointers across sibling branches) - Batch fetch:
AXUIElementCopyMultipleAttributeValuesfor 3-5x faster attribute reads - Role mapping: AXRole strings → unified role enum in
tree/roles.rs - Max depth default: 10, configurable via
--max-depth - Name:
kAXTitleAttribute/kAXDescriptionAttribute. Value:kAXValueAttribute - Bounds:
kAXPositionAttribute+kAXSizeAttributecombined to Rect
Action execution:
- Click:
AXUIElementPerformAction(kAXPressAction) - SetValue:
AXUIElementSetAttributeValue(kAXValueAttribute, value) - SetFocus:
AXUIElementSetAttributeValue(kAXFocusedAttribute, true) - Expand/Collapse: Toggle
kAXExpandedAttribute - Select:
AXUIElementSetAttributeValue(kAXSelectedAttribute, true)on child - Keyboard/Mouse:
CGEventCreateKeyboardEvent/CGEventCreateMouseEventvia CoreGraphics - Clipboard:
NSPasteboard.generalPasteboardread/write via Cocoa FFI - Screenshot:
CGWindowListCreateImagefor window-specific or full-screen capture
Permission detection:
- Call
AXIsProcessTrusted()on startup - If false, return
PERM_DENIEDwith guidance: "Open System Settings > Privacy > Accessibility and add your terminal" - Optionally call
AXIsProcessTrustedWithOptions(prompt: true)to trigger system dialog
Notification management:
- Open Notification Center via AX: target the
NotificationCenterprocess (bundleId:com.apple.notificationcenterui) - List notifications: traverse the Notification Center AX tree — each notification is an
AXGroupwith title, subtitle, and action buttons - Dismiss: perform
AXPresson the notification's close button, orAXRemoveFromParentif supported - Interact: resolve action buttons within a notification group and perform
AXPress - Dismiss all:
AXPressthe "Clear All" button at the group level - Do Not Disturb detection: read Focus/DND state via
NSDoNotDisturbEnableduser defaults orCoreFoundationpreferences
System tray / Menu bar extras:
- Menu bar extras (status items) live under the
SystemUIServerprocess AX tree - List items: traverse
AXMenuBarItemchildren of the system menu bar - Click:
AXPresson the target menu bar extra element - Expand menus: after clicking a tray item, traverse the resulting
AXMenuas a surface - Control Center items: accessible via the
ControlCenterprocess (bundleId:com.apple.controlcenter)
AXElement safety:
- Inner field:
pub(crate)notpub(prevents double-free via raw pointer extraction) Cloneimpl must callCFRetainDropimpl must callCFRelease
Snapshot Engine and Ref Allocator
Platform-agnostic, lives in agent-desktop-core:
- Raw tree: Call
adapter.get_tree(window, opts) - Filter: Remove invisible/offscreen. Remove empty groups with no interactive descendants. Prune beyond max_depth
- Allocate refs: Depth-first. Interactive roles get
@e1,@e2, etc. Store in RefMap - Serialize: Omit null fields. Omit empty arrays. Omit bounds in compact mode
- Estimate tokens: Optionally warn if exceeding threshold
RefMap persisted at ~/.agent-desktop/last_refmap.json with 0o600 permissions, directory at 0o700. Each snapshot replaces the refmap file entirely (atomic write via temp + rename). Action commands use optimistic re-identification: (pid, role, name, bounds_hash). Return STALE_REF on mismatch.
Progressive Skeleton Traversal:
--skeletonflag clamps depth tomin(max_depth, 3), annotates truncated containers withchildren_countfor agent discovery--root <REF>flag starts traversal from a previously-discovered ref instead of window root- Named or described containers at skeleton boundary receive refs as drill-down targets (with empty
available_actions) - Scoped invalidation: re-drilling a ref replaces only that ref's subtree refs, preserving all others
- Core modules:
ref_alloc.rs(canonicalallocate_refs+RefAllocConfig),snapshot_ref.rs(drill-down flow that delegates allocation toref_alloc) - macOS:
count_children()uses rawCFArrayGetCountwithout materializingAXElementwrappers for performance - RefMap write-side size check prevents >1MB files
- Token savings: 78-96% reduction for dense Electron apps (Slack skeleton: ~3.6KB vs ~17.3KB full)
New Commands — Notification & System Tray (Post Phase 1)
Note: Notification management and system tray interaction were not part of the original Phase 1 delivery. These are new features to be implemented across all platforms as each platform adapter is built. The macOS implementations were added as a follow-up to Phase 1. Windows (Phase 2) and Linux (Phase 3) implementations follow the same pattern.
Notification Commands (macOS — Completed)
| Command | Description | Flags | Status |
|---|---|---|---|
list-notifications |
List current notifications with app, title, body, and available actions | --app (filter by app), --text (filter by text), --limit (max results) |
Completed |
dismiss-notification |
Dismiss a specific notification by 1-based index | <index>, --app (filter by app) |
Completed |
dismiss-all-notifications |
Clear all notifications, optionally filtered by app (single NC session, reports failures) | --app (filter by app) |
Completed |
notification-action |
Click an action button on a specific notification | <index> <action-name> |
Completed |
System Tray / Status Area Commands (New — Not Yet Implemented)
| Command | Description | Flags |
|---|---|---|
list-tray-items |
List all system tray / menu bar extra items with app name and tooltip | — |
click-tray-item |
Click a system tray item by ID or app name | <tray-item-id> |
open-tray-menu |
Click a tray item and snapshot its resulting menu for ref-based interaction | <tray-item-id> |
Wait Command Update (Notification — Completed, Menu — Completed)
The wait command has been extended with notification and menu support:
wait --notification— Wait for any new notification to appear (index-diff based detection)wait --notification --app Safari— Wait for a notification from a specific appwait --notification --text "Download complete"— Wait for a notification containing specific textwait --menu/wait --menu-closed— Wait for context menu open/close
Commands Shipped (57)
| Category | Commands | Count |
|---|---|---|
| App / Window | launch, close-app, list-windows, list-apps, focus-window, resize-window, move-window, minimize, maximize, restore |
10 |
| Observation | snapshot, screenshot, find, get (text, value, title, bounds, role, states, tree-stats), is (visible, enabled, checked, focused, expanded), list-surfaces |
6 |
| Interaction | click, double-click, triple-click, right-click, type, set-value, clear, focus, select, toggle, check, uncheck, expand, collapse |
14 |
| Scroll | scroll, scroll-to |
2 |
| Keyboard | press, key-down, key-up |
3 |
| Mouse | hover, drag, mouse-move, mouse-click, mouse-down, mouse-up |
6 |
| Clipboard | clipboard-get, clipboard-set, clipboard-clear |
3 |
| Notification | list-notifications, dismiss-notification, dismiss-all-notifications, notification-action |
4 |
| System Tray | list-tray-items, click-tray-item, open-tray-menu |
3 |
| Wait | wait (with --element, --window, --text, --menu, --notification flags) |
1 |
| System | status, permissions, version |
3 |
| Batch | batch |
1 |
JSON Output Contract
All commands produce a response envelope. Schema files versioned in schemas/.
Success:
{
"version": "1.0",
"ok": true,
"command": "snapshot",
"data": {
"app": "Finder",
"window": { "id": "w-4521", "title": "Documents" },
"ref_count": 14,
"tree": { ... }
}
}
Error:
{
"version": "1.0",
"ok": false,
"command": "click",
"error": {
"code": "STALE_REF",
"message": "RefMap is from a previous snapshot",
"suggestion": "Run 'snapshot' to refresh, then retry with updated ref"
}
}
Serialization rules: omit null/None fields (skip_serializing_if), omit empty arrays, omit bounds in compact mode, ref_count and tree inside data.
Error Taxonomy
| Code | Category | Example | Recovery Suggestion |
|---|---|---|---|
PERM_DENIED |
Permission | Accessibility not granted | Open System Settings > Privacy > Accessibility and add your terminal |
ELEMENT_NOT_FOUND |
Ref | @e12 not in current RefMap | Run 'snapshot' to refresh, then retry with updated ref |
APP_NOT_FOUND |
Application | --app 'Photoshop' not running | Launch the application first with 'launch Photoshop' |
ACTION_FAILED |
Execution | AXPress returned error on disabled button | Element may be disabled. Check states before acting |
ACTION_NOT_SUPPORTED |
Execution | Expand on a button element | This element does not support the requested action |
TREE_TIMEOUT |
Performance | Traversal exceeded 5s | Try --max-depth 3 or target a specific window |
STALE_REF |
Ref | RefMap is from a previous snapshot | UI may have changed. Run 'snapshot' again |
WINDOW_NOT_FOUND |
Window | --window w-999 does not exist | Run 'list-windows' to see available windows |
PLATFORM_UNSUPPORTED |
Platform | Linux adapter not yet shipped | This platform ships in Phase 3. Currently macOS only |
CLIPBOARD_EMPTY |
Clipboard | clipboard get but clipboard is empty | No text content in clipboard. Copy something first |
TIMEOUT |
Wait | wait --element exceeded timeout | Element did not appear within timeout. Increase --timeout or check app state |
NOTIFICATION_NOT_FOUND |
Notification | Notification ID not found | Notification may have been dismissed or expired. Run 'list-notifications' to see current notifications |
NOTIFICATION_UNSUPPORTED |
Notification | Notification daemon does not support listing | This notification daemon does not expose a history API. Consider using 'wait --notification' to catch notifications in real-time |
TRAY_NOT_FOUND |
System Tray | Tray item not found | Tray item may have been removed. Run 'list-tray-items' to see current items |
TRAY_UNSUPPORTED |
System Tray | No system tray available | System tray not available on this desktop environment. On GNOME, install the AppIndicator extension |
Exit codes: 0 success, 1 structured error (JSON on stdout), 2 argument/parse error.
Testing
Unit tests (core):
- AccessibilityNode ser/de roundtrips
- Ref allocator only assigns interactive roles
- SnapshotEngine filtering
- Error serialization
- JSON schema validation
- MockAdapter: in-memory PlatformAdapter returning hardcoded trees
Unit tests (macos):
- Role mapping coverage
- Permission check with mocks
- Tree traversal cycle detection
Integration tests (macOS CI):
- Snapshot Finder, TextEdit, System Settings — non-empty trees with refs
- Click button in test app — verify action succeeded
- Type text into TextEdit via ref — verify content changed
- Clipboard get/set roundtrip
- Wait for window
- Launch + close app lifecycle
- Permission denied scenario — correct error code and guidance
- Large tree (Xcode) snapshot in under 2 seconds
- List notifications — returns non-empty list when Notification Center has entries
- Dismiss notification — verify notification removed from Notification Center AX tree
- List tray items — returns known menu bar extras (Wi-Fi, Bluetooth, Clock)
- Click tray item — verify menu bar extra menu opens
Golden fixtures (tests/fixtures/):
- Real snapshots from Finder, TextEdit, etc. checked into repo
- Regression-test serialization format changes
CI
- GitHub Actions macOS runner on every PR
cargo clippy --all-targets -- -D warnings(zero warnings)cargo test --workspacecargo tree -p agent-desktop-coremust not contain platform crate names- Binary size check: fail if release binary exceeds 15MB
Dependencies
| Crate | Version | Purpose |
|---|---|---|
clap |
4.x | CLI parsing with derive macros |
serde + serde_json |
1.x | JSON serialization |
thiserror |
2.x | Error derive macros |
tracing |
0.1+ | Structured logging |
base64 |
0.22+ | Screenshot encoding |
accessibility-sys |
0.1+ | macOS AXUIElement FFI |
core-foundation |
0.10+ | macOS CF types |
core-graphics |
0.24+ | macOS CG types |
Documentation Delivered
- README with installation (npm + source), core workflow, command reference, JSON output, ref system, platform support table
- PRD v2.0
- Architecture diagram
- Claude Code skills:
.claude/skills/agent-desktop/(core, platform-agnostic) +.claude/skills/agent-desktop-macos/(macOS-specific) - Quick reference slash command:
.claude/commands/desktop.md
Phase 2 — Windows Adapter
Status: Planned
Phase 2 brings agent-desktop to Windows. Core engine, CLI parser, JSON contract, error types, snapshot engine, and command registry are untouched. Only the new WindowsAdapter implementation is added inside crates/windows/. The existing stub is replaced with a full implementation.
Objectives
| ID | Objective | Metric |
|---|---|---|
| P2-O1 | Windows adapter | snapshot on Windows returns valid tree for Explorer, Notepad, Settings |
| P2-O2 | All existing commands cross-platform | Identical JSON schema output on macOS and Windows for every command |
| P2-O3 | Windows input synthesis | click, type, press, all mouse commands working via UIA + SendInput |
| P2-O4 | Windows screenshot | screenshot produces PNG via BitBlt / PrintWindow or xcap crate |
| P2-O5 | Windows clipboard | clipboard-get / clipboard-set / clipboard-clear working via Win32 Clipboard API |
| P2-O6 | Windows CI | GitHub Actions Windows runner executes full test suite on every PR |
| P2-O7 | Windows binary release | Prebuilt .exe published via GitHub Releases and npm |
Windows Adapter Implementation
Full WindowsAdapter in crates/windows/src/ following the identical platform crate folder structure:
crates/windows/src/
├── lib.rs # mod declarations + re-exports only
├── adapter.rs # WindowsAdapter: PlatformAdapter impl
├── tree/
│ ├── mod.rs # re-exports
│ ├── element.rs # UIA element wrapper + attribute readers
│ ├── builder.rs # IUIAutomationTreeWalker traversal with CacheRequest
│ ├── roles.rs # UIA ControlType → unified role enum mapping
│ ├── resolve.rs # Element re-identification for ref resolution
│ └── surfaces.rs # Surface detection (menus, dialogs, flyouts)
├── actions/
│ ├── mod.rs # re-exports
│ ├── dispatch.rs # perform_action match arms via UIA patterns
│ ├── activate.rs # Smart activation chain (InvokePattern → Toggle → coordinate)
│ └── extras.rs # SelectionPattern, ScrollPattern helpers
├── input/
│ ├── mod.rs # re-exports
│ ├── keyboard.rs # SendInput keyboard synthesis
│ ├── mouse.rs # SendInput mouse events
│ └── clipboard.rs # OpenClipboard / GetClipboardData / SetClipboardData Win32 APIs
├── notifications/
│ ├── mod.rs # re-exports
│ ├── list.rs # List toast/Action Center notifications via UIA
│ ├── dismiss.rs # Dismiss individual or all notifications
│ └── interact.rs # Click notification action buttons
├── tray/
│ ├── mod.rs # re-exports
│ ├── list.rs # List system tray items via Shell_TrayWnd UIA tree
│ └── interact.rs # Click tray items, open tray menus
└── system/
├── mod.rs # re-exports
├── app_ops.rs # Process launch via CreateProcess, close via TerminateProcess
├── window_ops.rs # SetWindowPos, ShowWindow for resize/move/minimize/maximize/restore
├── key_dispatch.rs # App-targeted key press via SetForegroundWindow + SendInput
├── permissions.rs # COM security check, UAC elevation detection
├── screenshot.rs # BitBlt / PrintWindow or xcap crate
└── wait.rs # wait utilities (polling UIA element existence)
Windows API Mapping
| Capability | Technology | Details |
|---|---|---|
| Tree root | IUIAutomation.ElementFromHandle() |
Via uiautomation crate (v0.24+) wrapping UIA COM APIs via windows crate |
| Children | IUIAutomationTreeWalker.GetFirstChild / GetNextSibling |
With CacheRequest for batch attribute retrieval (3-5x faster) |
| Role mapping | UIA ControlType integers |
Map to unified role enum in tree/roles.rs — e.g. UIA_ButtonControlTypeId → button |
| Click | InvokePattern.Invoke() |
Pattern-based, falls back to TogglePattern.Toggle(), then coordinate click via SendInput |
| Set text | ValuePattern.SetValue() |
Falls back to SelectAll + SendInput keystroke sequence |
| Expand/Collapse | ExpandCollapsePattern.Expand() / .Collapse() |
Native UIA pattern |
| Select | SelectionItemPattern.Select() |
For combobox, listbox, tab items |
| Toggle | TogglePattern.Toggle() |
For checkboxes, switches |
| Scroll | ScrollPattern.Scroll() / ScrollPattern.SetScrollPercent() |
Native UIA scroll, falls back to mouse wheel |
| Keyboard | SendInput API |
INPUT_KEYBOARD structs with virtual key codes and scan codes |
| Mouse | SendInput API |
INPUT_MOUSE structs with MOUSEEVENTF_* flags |
| Clipboard | OpenClipboard / GetClipboardData / SetClipboardData |
Win32 APIs, handle CF_UNICODETEXT format |
| Screenshot | BitBlt / PrintWindow |
Window capture; or xcap crate for cross-platform consistency |
| App launch | CreateProcess / ShellExecuteEx |
Launch by name or path, wait for main window |
| App close | WM_CLOSE / TerminateProcess |
Graceful close first, force kill with --force |
| Window ops | SetWindowPos / ShowWindow |
Resize, move, minimize (SW_MINIMIZE), maximize (SW_MAXIMIZE), restore (SW_RESTORE) |
| Permissions | COM security / UAC | Detect elevation requirements; return PERM_DENIED if UIA access blocked |
| Notifications | UIA + Action Center | Toast notifications accessible via UIA tree of Windows.UI.Notifications.Manager. List via IUIAutomationElement traversal of Action Center pane. Dismiss via InvokePattern on close button. Interact via InvokePattern on action buttons. Do Not Disturb (Focus Assist) state via WNF_SHEL_QUIETHOURS_ACTIVE_PROFILE_CHANGED or registry query |
| System tray | UIA + Shell_TrayWnd | System tray items accessible via UIA tree of Shell_TrayWnd class. Overflow items in NotifyIconOverflowWindow. List via IUIAutomationTreeWalker on tray area. Click via InvokePattern or coordinate-based SendInput. Expand overflow via click on chevron button |
Notification Management (New Feature — Windows Implementation)
Windows notification management must be implemented from scratch as part of Phase 2. The macOS notification implementation (completed as a follow-up to Phase 1) serves as the reference pattern — same PlatformAdapter trait methods (list_notifications, dismiss_notification, dismiss_all_notifications, notification_action), same JSON output contract, same 1-based indexing.
Implementation approach:
- List notifications: Open Action Center via UIA (
Windows.UI.Notifications). Traverse the notification list — each toast is a UIA element withName(title),FullDescription(body), app info, and child action buttons - Dismiss:
InvokePattern.Invoke()on the notification's dismiss/close button. For "dismiss all", invoke the "Clear all" button in Action Center - Interact with actions: Resolve action buttons within a toast element tree, invoke via
InvokePattern - Focus Assist / Do Not Disturb: Query via
WNF_SHEL_QUIETHOURS_ACTIVE_PROFILE_CHANGEDstate notification or registry keyHKCU\Software\Microsoft\Windows\CurrentVersion\CloudStore\Store\DefaultAccount\Current\default$windows.data.notifications.quiethourssettings - Edge case: Some notifications may be transient (disappear after timeout). The
wait --notificationcommand should monitor for new toasts via UIA event subscription (UIA_Notification_EventId)
System Tray (New Feature — Windows Implementation)
System tray interaction must be implemented from scratch as part of Phase 2.
Implementation approach:
- List items: Access the system tray via UIA tree of
Shell_TrayWndwindow class. Tray items are children of the notification area. Overflow items live inNotifyIconOverflowWindow - Click:
InvokePatternon tray items, falling back to coordinate-basedSendInputfor items that don't expose UIA patterns - Open menu: After clicking a tray item, detect the resulting popup menu via UIA focus-changed events and expose it for ref-based interaction
Web/Electron App Compatibility
Chromium-based apps (Electron, Chrome, Edge, VS Code) expose deep, noisy accessibility trees where every HTML <div> becomes a UIA Group element. The macOS adapter solved this with three patterns that must be replicated identically on Windows.
Chromium detection:
- Detect Chromium-based windows via UIA process name or
Chrome_WidgetWin_1window class matching - If tree is empty or minimal for a Chromium window, warn: "This appears to be a Chromium app. Run the app with
--force-renderer-accessibilityto expose the accessibility tree" - Include this guidance in the
platform_detailfield of the error response
Web-aware tree traversal (depth-skip):
- Non-semantic wrapper elements (
UIA_GroupControlTypeId/UIA_CustomControlTypeId) with emptyNameAND emptyValueproperties do NOT consume depth budget during tree traversal - This matches the macOS pattern where
AXGroup/AXGenericElementwrappers are skipped - Without this, default
--max-depth 10finds ~3 refs in Slack; with it, finds 100+ refs - Implement in
crates/windows/src/tree/builder.rswith the sameis_web_wrapperlogic
Resolver depth:
- Element re-identification must search up to
ABSOLUTE_MAX_DEPTH(50), not a lower hardcoded limit - Electron elements commonly sit at depth 25+ in the raw tree; a shallow resolver cap causes
STALE_REFerrors - Implement in
crates/windows/src/tree/resolve.rsmatching the macOS pattern
Surface detection for Electron:
- When an Electron app opens a modal (file picker, dialog), UIA may report the dialog as the focused window itself rather than a child of the parent window
- Surface detection (
list-surfaces,--surface sheet/alert) must check if the focused window IS the target surface, not only search its children - Check both
ControlTypeandLocalizedControlType/ UIA patterns (analogous to macOS checking both AXRole and AXSubrole) - Implement in
crates/windows/src/tree/surfaces.rs
Progressive skeleton traversal works identically on Windows — --skeleton and --root flags are platform-agnostic, handled entirely by core. The Windows adapter only needs to implement get_subtree() (which delegates to the same build_subtree() as get_tree()). Token savings for Electron apps (VS Code, Slack) apply equally.
Minimum OS Requirements
- Windows 10 1809+ (October 2018 update)
- UIA COM interfaces available since Windows 7, but modern patterns require 10+
New Dependency
| Crate | Version | Purpose | License |
|---|---|---|---|
uiautomation |
0.24+ | Windows UIA wrapper via windows crate |
Apache-2.0 |
Added to Cargo.toml as target-gated dependency:
[target.'cfg(target_os = "windows")'.dependencies]
agent-desktop-windows = { path = "crates/windows" }
Testing
Unit tests (windows):
- UIA ControlType → role mapping coverage for all control types
- Permission check with mocks (COM security state)
- CacheRequest attribute batching correctness
- Element resolution round-trip (pid, role, name, bounds_hash)
Integration tests (Windows CI):
- Snapshot Explorer — non-empty tree with refs, buttons, text fields
- Snapshot Notepad — text area with value, menu items
- Snapshot Settings — modern WinUI controls
- Click button in test app — verify action succeeded
- Type text into Notepad via ref — verify content changed
- Set-value on a text field — verify value set via UIA
- Clipboard get/set/clear roundtrip
- Wait for window title pattern
- Launch + close app lifecycle (Notepad: launch, type, close)
- Resize, move, minimize, maximize, restore window operations
- Screenshot produces valid PNG
- Large tree snapshot performance validation
- Chromium detection — verify warning when tree is empty
- Electron app snapshot (VS Code) — default depth finds 50+ refs via web-aware depth-skip
- Electron surface detection — file picker dialog detected as sheet surface
- List notifications — returns non-empty list when notifications exist
- Dismiss notification — verify notification removed from Action Center
- Notification action — click action button on a test toast notification
- List tray items — returns known system tray entries (volume, network, clock)
- Click tray item — verify tray menu opens
Cross-platform validation:
- Same snapshot of a cross-platform app (e.g., VS Code) produces structurally identical JSON on macOS and Windows
- All error codes produce identical JSON envelope format
CI
- Add GitHub Actions Windows runner alongside existing macOS runner
- Both runners execute:
cargo clippy --all-targets -- -D warnings,cargo test --workspace cargo tree -p agent-desktop-corecontinues to contain zero platform crate names- Binary size check: Windows
.exemust be under 15MB
Release
- Prebuilt Windows
.exebinary published to GitHub Releases viacargo-dist - npm package updated to include Windows binary (platform-specific download)
- GitHub Release notes document Windows support and installation
Skill Update
Per Skill Maintenance Addendum:
- Create
.claude/skills/agent-desktop-windows/SKILL.md:- UIA permission model and UAC handling
- Windows-specific behaviors (UIA patterns, WinUI3 quirks, COM initialization)
- Chromium/Electron compatibility: depth-skip, resolver depth, surface detection patterns
--force-renderer-accessibilityguidance for empty trees- Windows error codes and
platform_detailexamples (HRESULT codes) - Troubleshooting guide (empty trees, COM errors, elevation failures)
- Update core
SKILL.md:- Add Windows platform skill to skill graph table
- Update platform support section
- Update
workflows.md:- Add cross-platform patterns noting Windows-specific differences
- Add Windows-specific workflow examples (e.g., navigating UWP apps)
README Update
- Update Platform Support table: Windows column → Yes
- Add Windows installation instructions:
- npm (same command, auto-detects platform)
- Direct
.exedownload from GitHub Releases - From source:
cargo build --releaseon Windows (note: requires MSVC toolchain)
- Add Windows permissions section:
- UIA works without special permissions for most apps
- UAC elevation may be required for elevated processes
- Chromium apps need
--force-renderer-accessibility
- Update "From source" section with Windows build requirements (Rust + MSVC)
Phase 3 — Linux Adapter
Status: Planned
Phase 3 brings agent-desktop to Linux, completing the three-platform story. Same additive approach — only the LinuxAdapter implementation is added inside crates/linux/. The existing stub is replaced with a full implementation.
Objectives
| ID | Objective | Metric |
|---|---|---|
| P3-O1 | Linux adapter | snapshot on Ubuntu GNOME returns valid tree for Files, Terminal, Settings |
| P3-O2 | All commands cross-platform | Identical JSON schema output on all 3 platforms for every command |
| P3-O3 | Linux input synthesis | click, type, press, all mouse commands via AT-SPI actions + xdotool/ydotool |
| P3-O4 | Linux screenshot | screenshot produces PNG via PipeWire ScreenCast (Wayland) / XGetImage (X11) |
| P3-O5 | Linux clipboard | clipboard-get / clipboard-set / clipboard-clear via wl-clipboard (Wayland) / xclip (X11) |
| P3-O6 | Cross-platform CI | GitHub Actions matrix: macOS + Windows + Ubuntu |
| P3-O7 | Linux binary release | Prebuilt binary published via GitHub Releases and npm |
Linux Adapter Implementation
Full LinuxAdapter in crates/linux/src/ following the identical platform crate folder structure:
crates/linux/src/
├── lib.rs # mod declarations + re-exports only
├── adapter.rs # LinuxAdapter: PlatformAdapter impl
├── tree/
│ ├── mod.rs # re-exports
│ ├── element.rs # AT-SPI Accessible wrapper + attribute readers
│ ├── builder.rs # D-Bus tree traversal via GetChildren
│ ├── roles.rs # AT-SPI Role enum → unified role enum mapping
│ ├── resolve.rs # Element re-identification for ref resolution
│ └── surfaces.rs # Surface detection (menus, dialogs, popovers)
├── actions/
│ ├── mod.rs # re-exports
│ ├── dispatch.rs # perform_action via AT-SPI Action interface
│ ├── activate.rs # Smart activation chain (DoAction → coordinate fallback)
│ └── extras.rs # Text.InsertText, Selection helpers
├── input/
│ ├── mod.rs # re-exports
│ ├── keyboard.rs # xdotool (X11) / ydotool (Wayland) keyboard synthesis
│ ├── mouse.rs # xdotool (X11) / ydotool (Wayland) mouse events
│ └── clipboard.rs # wl-clipboard (Wayland) / xclip (X11) clipboard ops
├── notifications/
│ ├── mod.rs # re-exports
│ ├── list.rs # List notifications via D-Bus org.freedesktop.Notifications or daemon-specific API
│ ├── dismiss.rs # Dismiss/close notifications via CloseNotification D-Bus method
│ └── interact.rs # Invoke notification actions via ActionInvoked D-Bus signal
├── tray/
│ ├── mod.rs # re-exports
│ ├── list.rs # List tray items via StatusNotifierItem D-Bus interface or AT-SPI
│ └── interact.rs # Activate/context-menu tray items via D-Bus methods
└── system/
├── mod.rs # re-exports
├── app_ops.rs # App launch via xdg-open / process spawn, close via SIGTERM/SIGKILL
├── window_ops.rs # xdotool / wmctrl for resize/move/minimize/maximize/restore
├── key_dispatch.rs # App-targeted key press via window focus + input synthesis
├── permissions.rs # AT-SPI2 bus availability check, DBUS_SESSION_BUS_ADDRESS detection
├── screenshot.rs # PipeWire ScreenCast portal (Wayland) / XGetImage (X11) / xcap crate
└── wait.rs # wait utilities (polling AT-SPI element existence)
Linux API Mapping
| Capability | Technology | Details |
|---|---|---|
| Tree root | atspi Accessible on bus |
Via atspi crate (v0.28+) + zbus (5.x) — pure Rust, no libatspi/GLib dependency |
| Children | org.a11y.atspi.Accessible.GetChildren |
Async D-Bus calls to AT-SPI2 registry daemon |
| Role mapping | AT-SPI Role enum |
Map to unified role enum in tree/roles.rs — e.g. Role::PushButton → button |
| Click | org.a11y.atspi.Action.DoAction(0) |
AT-SPI actions preferred over coordinate-based input |
| Set text | org.a11y.atspi.Text.InsertText |
AT-SPI text interface; falls back to clipboard paste |
| Expand/Collapse | Action.DoAction("expand") / Action.DoAction("collapse") |
Action name-based dispatch |
| Select | org.a11y.atspi.Selection.SelectChild |
For combobox, listbox, tab items |
| Toggle | Action.DoAction("toggle") or Action.DoAction("click") |
For checkboxes, switches |
| Scroll | Coordinate-based scroll events via xdotool/ydotool | AT-SPI has no native scroll pattern |
| Keyboard | xdotool key (X11) / ydotool key (Wayland) |
Shelling out for input synthesis |
| Mouse | xdotool mousemove/click (X11) / ydotool mousemove/click (Wayland) |
Display server detected at runtime |
| Clipboard | wl-copy / wl-paste (Wayland) / xclip (X11) |
Shelling out; display server detected at runtime |
| Screenshot | PipeWire ScreenCast portal (Wayland) / XGetImage (X11) |
Or xcap crate for consistency |
| App launch | xdg-open / direct process spawn |
Launch by .desktop file or command name |
| App close | SIGTERM / SIGKILL |
Graceful close first, force with --force |
| Window ops | xdotool / wmctrl |
Window resize, move, minimize, maximize, restore |
| Permissions | AT-SPI2 bus availability | Check for org.a11y.Bus on D-Bus session bus. Return PLATFORM_UNSUPPORTED with enable instructions if missing |
| Notifications | D-Bus org.freedesktop.Notifications |
List via GetServerInformation + monitoring Notify signals. History varies by daemon: GNOME uses org.gnome.Shell.Notifications, KDE uses org.freedesktop.Notifications with GetNotifications. Dismiss via CloseNotification(id). Interact via ActionInvoked signal. Do Not Disturb: GNOME org.gnome.desktop.notifications.show-banners, KDE org.kde.notificationmanager |
| System tray | D-Bus org.kde.StatusNotifierWatcher |
SNI (StatusNotifierItem) protocol for modern tray items. Legacy XEmbed tray items via AT-SPI tree of the tray window. List via RegisteredStatusNotifierItems property. Activate via Activate(x, y) method. Context menu via ContextMenu(x, y) method. Fallback: coordinate-based click for XEmbed items |
Notification Management (New Feature — Linux Implementation)
Linux notification management must be implemented from scratch as part of Phase 3. The macOS implementation (completed) and Windows implementation (Phase 2) serve as reference patterns — same trait methods, same JSON output contract, same 1-based indexing.
Implementation approach:
- List notifications: The standard
org.freedesktop.NotificationsD-Bus interface does NOT provide a "list current notifications" method. Approach varies by desktop environment:- GNOME:
org.gnome.Shellexposesorg.gnome.Shell.Notificationsinterface withGetNotifications()method (returns array of notification dicts) - KDE Plasma:
org.freedesktop.NotificationswithGetNotifications()extension, ororg.kde.notificationmanagerD-Bus interface - Other DEs: Monitor
NotifyD-Bus signals to maintain an in-memory notification history within the daemon session
- GNOME:
- Dismiss:
org.freedesktop.Notifications.CloseNotification(id)D-Bus method call. Works across all notification daemons - Interact with actions: Listen for user-triggered actions or programmatically invoke via
ActionInvokedsignal. Note: the D-Bus spec does not define a method to programmatically trigger actions — coordinate-based click on the notification popup via AT-SPI may be needed as a fallback - Do Not Disturb:
- GNOME:
gsettings get org.gnome.desktop.notifications show-banners(boolean) - KDE:
org.kde.notificationmanagerD-Bus interface,inhibitedproperty
- GNOME:
- Edge case: Notification daemon varies by DE — detect via
GetServerInformation()D-Bus method. ReturnPLATFORM_UNSUPPORTEDwith daemon-specific guidance if the notification interface is unreachable
System Tray (New Feature — Linux Implementation)
System tray interaction must be implemented from scratch as part of Phase 3.
Implementation approach:
- Modern tray (SNI): Most modern Linux apps use the
StatusNotifierItem(SNI) D-Bus protocol. Discover items viaorg.kde.StatusNotifierWatcher.RegisteredStatusNotifierItemsproperty - Legacy tray (XEmbed): Older apps use XEmbed protocol. Access via AT-SPI tree of the tray window, or coordinate-based interaction
- List items: Query
StatusNotifierWatcherfor registered items. Each item exposesTitle,IconName,ToolTip,Menu(D-Bus menu path) properties - Activate: Call
Activate(x, y)method on theStatusNotifierItemD-Bus interface - Context menu: Call
ContextMenu(x, y)method, or read theMenuproperty to get thecom.canonical.dbusmenupath and traverse the menu tree - Edge case: GNOME does not natively support SNI (requires
AppIndicatorextension). Detect and report via error suggestion if no tray is available
Display Server Detection
Runtime detection required for input, clipboard, and screenshot since Linux runs either X11 or Wayland:
- Check
$WAYLAND_DISPLAYenvironment variable — if set, use Wayland path - Check
$DISPLAYenvironment variable — if set and no Wayland, use X11 path - If neither, return
PLATFORM_UNSUPPORTEDwith guidance to check display server configuration - Input tools: verify
xdotool(X11) orydotool(Wayland) is installed; error with install instructions if missing - Clipboard tools: verify
xclip(X11) orwl-clipboard(Wayland) is installed; error with install instructions if missing
Web/Electron App Compatibility
Same Chromium/Electron compatibility patterns as Phase 2 (Windows), adapted for AT-SPI2. These patterns ensure default --max-depth 10 works with Electron apps like Slack, VS Code, and Chrome.
Web-aware tree traversal (depth-skip):
- Non-semantic wrapper elements with AT-SPI roles
ROLE_PANEL,ROLE_SECTION, orROLE_FILLERthat have emptyNameAND emptyValuedo NOT consume depth budget during tree traversal - This is the AT-SPI equivalent of macOS
AXGroup/AXGenericElementand WindowsUIA_GroupControlTypeIdskipping - Implement in
crates/linux/src/tree/builder.rswith the sameis_web_wrapperlogic
Resolver depth:
- Element re-identification must search up to
ABSOLUTE_MAX_DEPTH(50), not a lower hardcoded limit - Electron elements commonly sit at depth 25+ in the raw AT-SPI tree
- Implement in
crates/linux/src/tree/resolve.rsmatching the macOS/Windows pattern
Surface detection for Electron:
- When an Electron app opens a modal (file picker, dialog), AT-SPI may report the dialog as the active window itself rather than a child of the parent window
- Surface detection must check if the focused window IS the target surface, not only search its children
- Check both
RoleandRelationSet/RELATION_EMBEDSfor dialog detection (analogous to macOS AXRole + AXSubrole) - Implement in
crates/linux/src/tree/surfaces.rs
Chromium detection:
- Detect Chromium-based apps via process name matching (electron, chrome, chromium, code)
- If AT-SPI tree is empty for a Chromium app, warn about
--force-renderer-accessibility - On Linux, Chromium respects
ACCESSIBILITY_ENABLED=1environment variable as an alternative
Progressive skeleton traversal works identically on Linux — --skeleton and --root flags are platform-agnostic, handled entirely by core. The Linux adapter only needs to implement get_subtree() (which delegates to the same async tree walker). Token savings for Electron apps apply equally.
AT-SPI2 Bus Detection
- Check for
org.a11y.Buspresence on the D-Bus session bus - If bus is not running, return
PLATFORM_UNSUPPORTEDwith instructions:- GNOME: "AT-SPI2 should be enabled by default. Check
gsettings get org.gnome.desktop.interface toolkit-accessibility" - Other DEs: "Install
at-spi2-coreand ensureat-spi-bus-launcheris running" - Flatpak/Snap: "Ensure the app has
--talk-name=org.a11y.Buspermission"
- GNOME: "AT-SPI2 should be enabled by default. Check
Minimum OS Requirements
- Ubuntu 22.04+ / Fedora 38+
- GNOME 42+ (primary target), KDE Plasma 5.24+ (secondary)
at-spi2-corepackage installed (default on GNOME)- X11:
xdotoolinstalled. Wayland:ydotoolinstalled
Key Risks and Mitigations
| Risk | Mitigation |
|---|---|
| Wayland a11y gaps | Focus on GNOME (best AT-SPI2 support). Prefer AT-SPI actions over coordinate input. Document known gaps clearly in skill and README. |
| AT-SPI2 bus not running | Detect on first command. Return clear enable instructions specific to the detected distro/DE. |
| Display server fragmentation | Runtime detection (X11 vs Wayland). Separate code paths for input/clipboard/screenshot. Test both. |
| Rust a11y crate maintenance stalls | Pin atspi and zbus versions. atspi crate backed by Odilia accessibility project. Maintain patches if upstream stalls. |
| Input tool availability | Check for xdotool/ydotool on first use. Provide package manager install commands in error suggestion. |
New Dependencies
| Crate | Version | Purpose | License |
|---|---|---|---|
atspi |
0.28+ | Linux AT-SPI2 client | MIT/Apache-2.0 |
zbus |
5.x | D-Bus connection | MIT/Apache-2.0 |
tokio |
1.x | Async runtime (required by atspi/zbus for async D-Bus) | MIT |
Added to Cargo.toml as target-gated dependency:
[target.'cfg(target_os = "linux")'.dependencies]
agent-desktop-linux = { path = "crates/linux" }
Note: tokio is introduced here for the first time. Phases 1-2 are fully synchronous. The Linux adapter requires async D-Bus calls via zbus.
Testing
Unit tests (linux):
- AT-SPI Role → role mapping coverage for all role types
- Bus availability detection (mock D-Bus responses)
- Display server detection logic (Wayland vs X11 env vars)
- Element resolution round-trip (pid, role, name, bounds_hash)
Integration tests (Ubuntu CI):
- Snapshot GNOME Files — non-empty tree with refs, buttons, text fields
- Snapshot GNOME Terminal — text area, menu items
- Snapshot GNOME Settings — modern GTK4 controls
- Click button in test app — verify action succeeded
- Type text into GNOME Text Editor via ref — verify content changed
- Clipboard get/set/clear roundtrip (test both X11 and Wayland if CI supports)
- Wait for window title pattern
- Launch + close app lifecycle
- Resize, move, minimize, maximize, restore window operations
- Screenshot produces valid PNG
- AT-SPI2 bus not running — correct error code and guidance
- Electron app snapshot (VS Code) — default depth finds 50+ refs via web-aware depth-skip
- Electron surface detection — file picker dialog detected as sheet surface
- List notifications — returns non-empty list when notifications exist (GNOME)
- Dismiss notification — verify notification dismissed via D-Bus
CloseNotification - List tray items — returns known SNI items (if running under KDE or with AppIndicator extension)
- Click tray item — verify tray menu opens via
ActivateD-Bus method - Notification daemon detection — correct
GetServerInformationresult
Cross-platform validation:
- Same snapshot of a cross-platform app (e.g., VS Code) produces structurally identical JSON on all 3 platforms
- All error codes produce identical JSON envelope format on all 3 platforms
- Notification commands return identical JSON envelope structure across all 3 platforms (list, dismiss, action)
- Tray commands return identical JSON envelope structure across all 3 platforms
CI
- GitHub Actions matrix: macOS + Windows + Ubuntu (all three on every PR)
- All runners execute:
cargo clippy --all-targets -- -D warnings,cargo test --workspace cargo tree -p agent-desktop-corecontinues to contain zero platform crate names- Binary size check: all platform binaries must be under 15MB
Release
- Prebuilt Linux binary published to GitHub Releases via
cargo-dist - npm package updated to include Linux binary (platform-specific download)
- GitHub Release notes document Linux support, requirements, and installation
Skill Update
Per Skill Maintenance Addendum:
- Create
.claude/skills/agent-desktop-linux/SKILL.md:- AT-SPI2/D-Bus setup and bus detection
- Wayland vs X11 differences (input via xdotool/ydotool, clipboard via wl-clipboard/xclip, screenshot via PipeWire/XGetImage)
- Required system tools:
xdotoolorydotool,xcliporwl-clipboard - Linux error codes and
platform_detailexamples (D-Bus errors, bus not found) - Troubleshooting guide (bus not running, empty trees, missing tools, Flatpak/Snap permissions)
- Update core
SKILL.md:- Add Linux platform skill to skill graph table
- Update platform support section to show all 3 platforms
- Update
workflows.md:- Add cross-platform patterns noting Linux-specific differences
- Add Linux-specific workflow examples (e.g., GNOME app automation)
- Document display server detection behavior
README Update
- Update Platform Support table: Linux column → Yes
- Add Linux installation instructions:
- npm (same command, auto-detects platform)
- Direct binary download from GitHub Releases
- From source:
cargo build --releaseon Linux (note: requirespkg-config,libdbus-1-dev)
- Add Linux permissions section:
- AT-SPI2 bus must be running (default on GNOME, may need enabling on other DEs)
- Required tools:
xdotool(X11) orydotool(Wayland) for input synthesis - Required tools:
xclip(X11) orwl-clipboard(Wayland) for clipboard - How to check:
busctl --user list | grep a11y
- Update minimum OS versions: Ubuntu 22.04+ / Fedora 38+
- Update "From source" section with Linux build requirements
Phase 4 — MCP Server Mode
Status: Planned
Phase 4 adds a new I/O layer. Core engine and all three platform adapters are unchanged. The MCP server wraps existing command logic in JSON-RPC tool definitions, enabling agent-desktop to work as an MCP server for Claude Desktop, Cursor, and other MCP-compatible hosts.
Objectives
| ID | Objective | Metric |
|---|---|---|
| P4-O1 | MCP server mode via --mcp |
Responds to MCP initialize handshake, reports capabilities |
| P4-O2 | All commands as MCP tools | tools/list returns all tools with JSON Schema specs |
| P4-O3 | Claude Desktop validated | Claude Desktop invokes tools to control desktop apps end-to-end on all platforms |
| P4-O4 | Tool annotations | readOnlyHint, destructiveHint, idempotentHint on every tool |
Entry Point
The binary crate's main.rs detects mode:
- If invoked with
--mcpor stdin is a pipe: enter MCP server mode - Otherwise: parse CLI arguments, execute command, print JSON to stdout
This is the invariant: every MCP tool maps 1:1 to a CLI command. agent-desktop snapshot --app Finder is identical to invoking the MCP desktop_snapshot tool. Testing, debugging, and documentation are never fragmented.
New Crate: agent-desktop-mcp
crates/mcp/src/
├── lib.rs # mod declarations + re-exports
├── server.rs # MCP server bootstrap, initialize handler, capabilities reporting
├── tools.rs # Tool definitions with #[tool] macro, parameter JSON Schemas
└── transport.rs # Stdio transport (primary), optional HTTP+SSE
MCP Tool Surface
Each MCP tool maps 1:1 to a CLI command. Tool names are prefixed with desktop_ to avoid collision with other MCP servers.
| MCP Tool | CLI Equivalent | readOnly | destructive |
|---|---|---|---|
desktop_snapshot |
snapshot |
true | false |
desktop_click |
click <ref> |
false | false |
desktop_type_text |
type <ref> <text> |
false | false |
desktop_set_value |
set-value <ref> <text> |
false | false |
desktop_press_key |
press <keys> |
false | false |
desktop_find |
find <query> |
true | false |
desktop_list_windows |
list-windows |
true | false |
desktop_focus_window |
focus-window |
false | false |
desktop_launch_app |
launch <app> |
false | false |
desktop_close_app |
close-app <app> |
false | true |
desktop_screenshot |
screenshot |
true | false |
desktop_scroll |
scroll <dir> |
false | false |
desktop_drag |
drag <from> <to> |
false | false |
desktop_select |
select <ref> <val> |
false | false |
desktop_toggle |
toggle <ref> |
false | false |
desktop_clipboard_get |
clipboard get |
true | false |
desktop_clipboard_set |
clipboard set <text> |
false | false |
desktop_wait |
wait |
true | false |
desktop_get |
get <prop> <ref> |
true | false |
desktop_is |
is <state> <ref> |
true | false |
desktop_list_notifications |
list-notifications |
true | false |
desktop_dismiss_notification |
dismiss-notification <id> |
false | true |
desktop_dismiss_all_notifications |
dismiss-all-notifications |
false | true |
desktop_notification_action |
notification-action <id> <action> |
false | false |
desktop_list_tray_items |
list-tray-items |
true | false |
desktop_click_tray_item |
click-tray-item <id> |
false | false |
desktop_open_tray_menu |
open-tray-menu <id> |
false | false |
Transport
- Stdio (primary): MCP host spawns
agent-desktop --mcpas child process. JSON-RPC over stdin/stdout. This is the only required transport. - HTTP+SSE (optional, stretch goal): For remote scenarios. Additive, non-blocking for core milestone.
- Session: On MCP
initialize, detect platform, check accessibility permissions, report capabilities. RefMap is session-scoped (held in memory, not persisted to disk like CLI mode).
Initialize Handler
On receiving MCP initialize:
- Detect platform (macOS / Windows / Linux)
- Check accessibility permissions (
check_permissions()) - Report capabilities: list of available tools, platform, permission status
- If permissions not granted, include guidance in capabilities response
New Dependencies
| Crate | Version | Purpose | License |
|---|---|---|---|
rmcp |
0.15.0+ | Official MCP Rust SDK — #[tool] macro, JSON-RPC handling, transport |
MIT/Apache-2.0 |
schemars |
0.8+ | JSON Schema generation for tool parameter definitions | MIT/Apache-2.0 |
tokio |
1.x | Async runtime (required by rmcp for MCP server event loop) | MIT |
Note: If tokio was already introduced in Phase 3 (Linux), it is already available. Otherwise, it is introduced here.
Binary Crate Changes
src/main.rs— Add--mcpflag detection, route to MCP server modeCargo.toml— Addagent-desktop-mcpdependency (non-platform-gated, available on all platforms)- No changes to
dispatch.rs,cli.rs, or any command files — MCP tools call the sameexecute()functions
Testing
Unit tests (mcp):
- Tool definition schema validation — every tool's JSON Schema is valid
- Tool invocation round-trip — call tool, verify response matches CLI output
- Initialize handler — correct capabilities, platform detection, permission status
Integration tests:
- Full MCP protocol compliance — initialize, tools/list, tool invocation, error responses
- Claude Desktop end-to-end: launch app → snapshot → click button → verify action
- Cursor end-to-end: same workflow
- Session isolation: RefMap is session-scoped, not shared across sessions
- Protocol edge cases: malformed requests, unknown tools, invalid parameters
Cross-platform:
- MCP server works identically on macOS, Windows, and Linux
- Same tool invocations produce same JSON structure on all platforms
MCP Config Examples
Provide ready-to-use config snippets for:
Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"agent-desktop": {
"command": "agent-desktop",
"args": ["--mcp"]
}
}
}
Cursor (.cursor/mcp.json):
{
"mcpServers": {
"agent-desktop": {
"command": "agent-desktop",
"args": ["--mcp"]
}
}
}
Skill Update
Per Skill Maintenance Addendum:
- Create
.claude/skills/agent-desktop-mcp/SKILL.md:- MCP tool surface documentation (all tools, parameters, annotations)
- Transport configuration (stdio setup, optional SSE)
- Session management (RefMap scoping, initialize flow)
- Tool-to-CLI mapping reference
- MCP-specific error handling
- Update core
SKILL.md:- Add MCP mode section
- Add MCP skill to skill graph table
- Update
workflows.md:- Add MCP workflow patterns (tool invocation from Claude Desktop, Cursor)
- Add session lifecycle patterns
README Update
- Add "MCP Server" section:
- How to start:
agent-desktop --mcp - What it does: wraps all CLI commands as MCP tools
- Session behavior: RefMap scoped per session
- How to start:
- Add Claude Desktop configuration snippet
- Add Cursor configuration snippet
- Document
--mcpflag in CLI reference - Add note: every MCP tool maps 1:1 to a CLI command
Phase 5 — Production Hardening
Status: Planned
Phase 5 transforms agent-desktop from functional to enterprise-grade. Persistent daemon process, session isolation for concurrent agents, comprehensive quality gates, and distribution via native package managers.
Objectives
| ID | Objective | Metric |
|---|---|---|
| P5-O1 | Persistent daemon | Warm snapshot completes in <50ms (vs 200ms+ cold start) |
| P5-O2 | Session isolation | Two agents hold independent RefMaps without interference |
| P5-O3 | Enterprise quality gates | All gates in quality gates table pass |
| P5-O4 | Package manager distribution | Available via brew (macOS), winget/scoop (Windows), snap/apt (Linux) |
Daemon Architecture
The daemon is a long-running process that maintains state between CLI/MCP invocations, dramatically reducing startup latency.
Auto-start:
- CLI detects if daemon is running by checking for socket file (
~/.agent-desktop/daemon.sockon Unix, named pipe on Windows) - If not running, spawns daemon as background process
- Daemon listens on the socket for incoming commands
Auto-stop:
- Daemon exits after configurable idle timeout (default 5 minutes)
- No active sessions = idle timer starts
- Any new connection resets the idle timer
Session multiplexing:
- Each CLI invocation or MCP session gets a unique session ID
- Each session has its own RefMap (held in memory by daemon, not on disk)
- Sessions are isolated: agent A's refs never collide with agent B's refs
- Session destroyed on disconnect or explicit
session kill
Health check:
agent-desktop statusreturns: daemon PID, uptime, active session count, platform, permission status
New Commands
| Command | Description |
|---|---|
session list |
List active daemon sessions with IDs, creation time, last activity |
session kill <id> |
Terminate a specific daemon session, release its RefMap |
CLI-to-Daemon Migration
When daemon is running:
- CLI command parses arguments as usual
- Instead of directly calling the adapter, CLI connects to daemon socket
- Sends serialized command to daemon
- Daemon executes command in the caller's session context
- Returns JSON response to CLI
- CLI prints response to stdout
When daemon is not running, CLI falls back to direct execution (same as Phases 1-4). Daemon is purely an optimization, never a requirement.
Enterprise Quality Gates
| Gate | Requirement |
|---|---|
| Security | No arbitrary code execution. No privilege escalation. All actions allowlisted via Action enum. No network access (daemon communicates only via local socket). |
| Performance | Cold start <200ms. Warm snapshot <50ms via daemon. Tree traversal timeout 5s default, configurable. |
| Reliability | Zero panics in non-test code. Graceful daemon recovery on crash. Stale socket cleanup on startup. |
| Observability | Structured logging via tracing crate. --verbose flag for debug output. Timing metrics per operation logged at debug level. |
| Compatibility | Tested against target app matrix: Finder, TextEdit, Xcode, VS Code, Chrome (macOS); Explorer, Notepad, Settings, VS Code (Windows); Nautilus, Terminal, Firefox (Linux). |
| Distribution | Single binary per platform. No runtime dependencies. Reproducible builds. SHA256 checksums for every release artifact. |
| Documentation | README, CLI reference, MCP reference, per-platform setup guides, troubleshooting. |
Performance Optimizations
| Optimization | Platform | Details |
|---|---|---|
| CacheRequest batching | Windows | Batch UIA attribute fetches via CacheRequest — reduces COM round-trips |
| Async tree walking | Linux | Parallel D-Bus calls for tree traversal — concurrent child fetching |
| Cached subtrees | All (daemon) | Reuse unchanged subtrees between snapshots in same session — skip re-traversal of stable UI regions |
| Warm adapter | All (daemon) | Adapter stays initialized between commands — skip COM init (Win), D-Bus connect (Linux), AX bootstrap (macOS) |
| Progressive skeleton drill | All | Skeleton overview + targeted drill-down reduces token consumption 78-96% for dense apps — fewer tokens per snapshot means more budget for actions |
Package Manager Distribution
| Platform | Package Manager | Format | Install Command |
|---|---|---|---|
| macOS | Homebrew | Formula | brew install agent-desktop |
| Windows | winget | Manifest | winget install agent-desktop |
| Windows | scoop | Manifest | scoop install agent-desktop |
| Linux | snap | Snap package | snap install agent-desktop |
| Linux | apt | .deb package | apt install agent-desktop |
Each package manager distribution includes:
- Prebuilt binary for the target platform
- SHA256 checksum verification
- Automatic PATH setup
- Uninstall support
Testing
Daemon tests:
- Daemon starts on first CLI command when not running
- Daemon stops after idle timeout with no active sessions
- Multiple concurrent sessions have isolated RefMaps
- Session list returns correct session metadata
- Session kill terminates session and releases resources
- Stale socket cleaned up on daemon restart
- Daemon crash recovery — CLI falls back to direct execution
- Warm snapshot performance: <50ms after initial cold start
Quality gate tests:
- Security: verify Action enum is exhaustive, no shell injection vectors
- Performance: benchmark cold start (<200ms) and warm snapshot (<50ms)
- Reliability: stress test with concurrent sessions, verify zero panics
- Compatibility: snapshot + click workflow on each app in target matrix
Package tests:
- brew formula installs and runs on macOS
- winget/scoop manifest installs and runs on Windows
- snap package installs and runs on Ubuntu
- All packages produce correct
versionoutput - All packages handle permissions correctly on their platform
Skill Update
Per Skill Maintenance Addendum:
- Update
commands-system.md:- Add
session listcommand documentation - Add
session kill <id>command documentation - Update
statuscommand to document daemon-specific fields (PID, uptime, sessions)
- Add
- Update
workflows.md:- Add daemon lifecycle patterns (auto-start, idle timeout, health checks)
- Add concurrent agent patterns (session isolation, multi-agent coordination)
- Add performance optimization patterns (warm snapshot, cached subtrees)
- Update platform skills:
- Document enterprise quality gates in each platform skill
- Add daemon-specific troubleshooting (stale socket, port conflicts)
README Update
- Add "Daemon Mode" section:
- How it works: auto-start, auto-stop, session isolation
- Configuration: idle timeout, socket location
- Health check:
agent-desktop status
- Add package manager installation methods:
brew install agent-desktop(macOS)winget install agent-desktop(Windows)snap install agent-desktop(Linux)
- Add "Performance" section:
- Cold start vs warm snapshot benchmarks
- Daemon mode benefits
- Update installation section with all distribution channels (npm, brew, winget, scoop, snap, apt, source)
- Final polish:
- Complete CLI reference for all commands including
session listandsession kill - Comprehensive troubleshooting guide covering all platforms
- Per-platform setup guides linked from main README
- Complete CLI reference for all commands including
Cross-Phase Requirements
README Update Schedule
The README is updated at the end of each phase to reflect the current state:
| Phase | README Changes |
|---|---|
| Phase 1 | Initial README: npm + source installation, core workflow, all 50 commands, JSON output, ref system, error codes, platform support table (macOS only) |
| Phase 2 | Add Windows: .exe installation, Windows permissions, update platform table, Windows build instructions |
| Phase 3 | Add Linux: binary installation, AT-SPI2 setup, update platform table, Linux build instructions, minimum OS versions |
| Phase 4 | Add MCP Server: --mcp usage, Claude Desktop config, Cursor config, tool-to-CLI mapping |
| Phase 5 | Add daemon mode, package managers (brew/winget/snap), performance benchmarks, final troubleshooting guide |
Skill Maintenance Rules
Per the Skill Maintenance Addendum:
- Every new command must be added to the appropriate
commands-*.mdfile - Every new platform gets its own skill directory under
.claude/skills/agent-desktop-{platform}/ - Every new mode (MCP, daemon) gets its own skill file
- Breaking changes to JSON output or CLI flags must update all affected skill files
- Skill files are reviewed as part of the PR checklist for any command-surface change
CI Matrix Evolution
| Phase | CI Runners |
|---|---|
| Phase 1 | macOS |
| Phase 2 | macOS + Windows |
| Phase 3 | macOS + Windows + Ubuntu |
| Phase 4 | macOS + Windows + Ubuntu (+ MCP protocol tests) |
| Phase 5 | macOS + Windows + Ubuntu (+ daemon tests, package build verification) |
All runners enforce: cargo clippy --all-targets -- -D warnings, cargo test --workspace, cargo tree -p agent-desktop-core contains zero platform crate names, binary size <15MB.
Dependency Introduction Schedule
| Dependency | Introduced In | Purpose |
|---|---|---|
clap 4.x, serde 1.x, thiserror 2.x, tracing 0.1+, base64 0.22+ |
Phase 1 | Core: CLI, JSON, errors, logging, encoding |
accessibility-sys 0.1+, core-foundation 0.10+, core-graphics 0.24+ |
Phase 1 | macOS AX API FFI |
uiautomation 0.24+ |
Phase 2 | Windows UIA wrapper |
atspi 0.28+ + zbus 5.x |
Phase 3 | Linux AT-SPI2 client via D-Bus |
tokio 1.x |
Phase 3 | Async runtime (required by atspi/zbus) |
rmcp 0.15.0+ |
Phase 4 | Official MCP Rust SDK |
schemars 0.8+ |
Phase 4 | JSON Schema generation for MCP tool parameters |
Platform API Quick Reference
| Capability | macOS | Windows | Linux |
|---|---|---|---|
| Tree root | AXUIElementCreateApp(pid) |
IUIAutomation.ElementFromHandle() |
atspi Accessible on bus |
| Children | kAXChildrenAttribute |
TreeWalker.GetFirstChild |
GetChildren D-Bus |
| Click | AXPress |
InvokePattern.Invoke() |
Action.DoAction(0) |
| Set text | AXValue = val |
ValuePattern.SetValue() |
Text.InsertText |
| Keyboard | CGEventCreateKeyboard |
SendInput |
xdotool / ydotool |
| Clipboard | NSPasteboard |
Win32 Clipboard API | wl-clipboard / xclip |
| Screenshot | CGWindowListCreateImage |
BitBlt / PrintWindow |
PipeWire / XGetImage |
| Permissions | AXIsProcessTrusted() |
COM security / UAC | Bus availability |
| Notifications | Notification Center AX tree (com.apple.notificationcenterui) |
UIA tree of Action Center / Toast Manager | D-Bus org.freedesktop.Notifications + daemon-specific history |
| System tray | SystemUIServer AX tree + ControlCenter AX tree |
UIA tree of Shell_TrayWnd + overflow window |
D-Bus StatusNotifierWatcher + XEmbed fallback |
Risk Register
| ID | Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|---|
| R1 | macOS TCC friction deters adoption | High | High | Clear first-run guidance. Detect before any op. One-command setup: permissions --request. |
| R2 | Electron/Chrome no a11y tree by default | High | Medium | Detect Chromium windows. Print --force-renderer-accessibility guidance in error response. |
| R3 | Custom-rendered UIs invisible to a11y | Medium | High | Phase 5 stretch: vision fallback. Short-term: document limitation in README and skills. |
| R4 | Wayland a11y gaps | Medium | Medium | Focus on GNOME (best AT-SPI2 support). Prefer AT-SPI actions over coordinate input. Document gaps. |
| R5 | Rust a11y crate maintenance stalls | Low | High | Pin versions, maintain patches. atspi backed by Odilia project. Fork-ready. |
| R6 | MCP spec changes break compat | Low | Medium | Pin rmcp version. Monitor spec under Linux Foundation governance. |
| R7 | Tree traversal too slow (>5s) | Medium | Medium | Depth limiting via --max-depth. Focused-window-only. Cached subtrees in Phase 5 daemon. Progressive skeleton traversal (--skeleton + --root) reduces token consumption 78-96% for dense apps. |
| R8 | Ref instability confuses agents | Medium | High | Clear docs: refs are snapshot-scoped. STALE_REF error with recovery hint. Stable hashing in Phase 5. Progressive skeleton traversal with scoped invalidation provides a stable drill-down workflow for navigating complex UIs. |