Design for mounting a Streamable HTTP MCP server inside the host-agent
process so Hermes Agent (or any MCP client) can drive devices directly.
Reuses the existing console FastAPI + uvicorn on port 8765, adds bearer-
token auth, per-device session-level locks with 60s TTL, and cloud
coordination via a new heartbeat field. Cloud scheduler skips devices
reported as MCP-busy.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The root tests/conftest.py fixture disables APEX_HUMANIZE_ENABLED by default for deterministic assertions, but apps/device-host-agent/tests/ was outside its scope. test_e2e.py::test_public_sdk_reports_fake_device_success_and_runtime_failure failed because tap coordinates were jittered (2.67..., 3.37...) instead of exact (2, 3).
Add the same autouse fixture to apps/device-host-agent/tests/conftest.py so all tests in that directory inherit the deterministic behavior.
PaddleOCR itself returns no color info, only text/bounds/confidence.
Add pixel-level post-processing in perception/ocr.py: crop the
screenshot to each OCR box, split pixels into two luminance clusters
via Otsu threshold, and treat the minority cluster as the text stroke
(foreground) and the majority as the background. New
SceneElement.foreground_color/background_color fields ("#rrggbb",
None when not OCR-sourced or sampling fails) round-trip through
to_dict/from_dict alongside the existing accessibility-state fields.
Planner system prompt documents the new fields as a secondary signal.
pillow is promoted from an implicit paddleocr transitive dependency to
an explicit direct dependency since perception/ocr.py now imports PIL
directly; uv.lock re-resolved with no version change (already locked
at 12.3.0).
Adds long_press/double_tap atomic gestures, a centralized humanize layer
(coordinate jitter, curved W3C-Actions swipe, timing jitter) gated by
APEX_HUMANIZE_ENABLED, and planner integration. 651 non-integration tests
pass on the branch.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- test_cancellation_full_path_queued_immediate_and_dispatched_collaborative
exercises the full public-API cancellation path: immediate cancel of a
queued task, collaborative cancel of a dispatched task surfaced through
lease renewal and a cancelled terminal report, and visibility of the
cancelled status via both the get and list endpoints.
- Full backend suite (869 passed, 50 skipped) and cloud-console frontend
suite (27 passed) + typecheck show no regressions; the only failures are
4 pre-existing live-LLM integration tests unrelated to this change.
- New internal API route POST /internal/v1/hosts/{host_id}/tasks/{task_id}/cancel,
authenticated via the host's own bearer credential (authorize_host) with an
ownership check, since host tokens carry no scopes and cannot reach the
public SDK's tasks:submit-scoped cancel endpoint.
- HostAgentClient.cancel_task() calls the new internal route directly.
- create_console_app() gains a cancel_task callable with automatic default
wiring from host_client, so production app.py needs no changes.
- Local console: POST /tasks/{task_id}/cancel route resolves the local
execution id to its Cloud source_task_id before cancelling, and the task
detail page/template show a Cancel button plus notice/error banners.
- Tests across all three layers: internal API route, Jinja2 template
rendering, and FastAPI console route behavior.
- Widen TaskStatus to include "cancelled"; add it to TasksView's
STATUSES filter dropdown.
- Add TaskCancellationResponse type and cancelTask(taskId) to api.ts.
- Add a Cancel button to TasksView's task detail panel, gated on
tasks:submit and a non-terminal task status; updates the displayed
status on success and surfaces errors via the existing error path.
- Extract the cancellability rule into a pure taskCancellation.ts
module (mirroring taskProgress.ts/plannerHistory.ts) with unit
tests, since the project has no Vue component-mounting test setup.
Task 6/9 of task-cancellation change.
- POST /v1/tasks/{task_id}/cancel: tasks:submit scoped, 200 for
immediate/idempotent cancellation, 202 for newly recorded pending
cancellation, 404 for unknown task, 409 for terminal task.
- TaskCancellationResponse{task_id, status} model.
- Widen list_tasks status_filter Literal to include "cancelled".
- CloudClient.cancel_task(task_id).
- SDK-level tests covering queued/assigned/idempotent/404/409/scope
cases for both the router and CloudClient.
Task 5/9 of task-cancellation change.
- LeaseGuard gains an is_cancellation convenience property
- ActiveAssignmentRunner marks the lease lost with a cancellation
reason when a renewal response reports cancel_requested
- AssignmentExecutor threads stop_reason through to TaskRunner/
WorkflowRunner and maps a cancellation-flavored stop to
AssignmentExecutionResult.status = "cancelled" instead of "failed"
- AssignmentProcessor forwards a three-way done/cancelled/failed
status when reporting the terminal result
- Add/extend tests across lease, assignment, processor, and client
9-task TDD plan: humanize module, Driver.long_press/swipe_path/double_tap
on WDA+Android (+ all Fake subclasses), tool wrappers, tap/swipe hooks,
and planner integration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Design for long_press/double_tap atomic gestures and a centralized
humanize layer (coordinate jitter, curved W3C-Actions swipe, timing
jitter) gated by APEX_HUMANIZE_ENABLED.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- LeaseRenewalResponse gains cancel_requested (populated from the
repository's renew_lease result)
- TerminalResultRequest.status widened to accept "cancelled"
- Add internal API tests for a renewal surfacing cancel_requested=True
and a cancelled terminal report being accepted/idempotent
SceneElement gains enabled/clickable/selected/checked/focused (bool | None),
populated from the literal attributes Appium's XCUITest and UiAutomator2
page_source already emit (iOS: enabled only; Android: all five). None means
"not reported by this platform", not false. to_dict() omits unset fields to
keep the LLM-facing scene JSON compact; planner_prompts.py documents the new
fields so the AI planner knows how to use them (e.g. don't tap disabled
elements, use selected/checked to judge whether a toggle already matches the
goal).
- Add nullable cancel_requested_at column (migration 0012)
- Widen ScheduledTaskStatus/TerminalTaskStatus to include cancelled
- Add CancellationRequestStatus + request_task_cancellation() to
CloudRepository protocol and SQLAlchemy implementation
- renew_lease() now returns LeaseRenewalResult, surfacing whether
cancellation is pending, instead of a bare status string
- reap_expired_leases() resolves pending-cancellation tasks to
cancelled instead of requeuing/failing them
- record_task_result() accepts cancelled and clears
cancel_requested_at on any terminal write
Note: internal_api/api.py's renew_assignment route still compares
renew_lease()'s return value against a bare string; it will be
updated in the next task (Internal Host<->Cloud protocol) to consume
LeaseRenewalResult and populate the new cancel_requested wire field.
- TaskRunner.run() and WorkflowRunner.run()/resume() accept an optional
stop_reason callable alongside should_stop, distinguishing a genuine
cancellation from other stop conditions (e.g. lost lease).
- is_cancellation_reason() shared helper added to runtime/task.py.
- WorkflowRunner._stop_status() now branches cancelled/failed based on
stop_reason, correcting a prior blanket cancelled-on-any-stop behavior
that conflicted with the host-agent-protocol spec's requirement to
distinguish cancellation from lease-loss stops.
- Default behavior (stop_reason=None) is preserved exactly for both
runners so existing callers/tests are unaffected.
- Task 1 of openspec change task-cancellation.
PaddleOCR's OCR.yaml pipeline defaults to use_doc_orientation_classify
and use_doc_unwarping enabled, which are meant for photographed paper
documents. Applied to a flat, upright device screenshot, UVDoc
geometrically warps the image before detection, and returns box
coordinates in that warped space with no inverse mapping back to the
original image.
Verified on a real screenshot: with unwarping on, the same detected
element ("新项目") shifts from y=158 to y=71 versus the original image,
and 2 boxes near the top edge (status bar time/battery) are dropped
entirely. Disabling both flags by default (still overridable via
explicit kwargs) makes detected boxes match the original screenshot.
paddlepaddle 3.3.1 silently corrupts non-ASCII (CJK) recognized text
into literal U+FFFD replacement characters during rec postprocessing,
while leaving confidence scores high and ASCII/digit text unaffected.
The same release also breaks CPU oneDNN inference on Windows entirely
(NotImplementedError in onednn_instruction.cc). Verified on a real
task screenshot that downgrading to 3.2.x eliminates the corruption
with no other environment changes (same GBK-locale machine).
Host-agent console showed OCR/UI-tree overlay boxes misaligned with the
displayed screenshot. Two independent causes, both confirmed with real
task data and pixel-level measurement of a user-provided screenshot:
1. perception/ui_parser.py parses XCUITest UI-tree bounds as iOS logical
points, while scene_builder.py's Scene.width/height (via infer_png_size)
and OCR bounds are in screenshot pixels, never reconciled (2.0x on
Retina devices). build_scene() now detects the scale from the first
x==0,y==0 UI element and rescales OCR bounds down to points-space,
reporting Scene.width/height in points too. No-op for Android, where
UiAutomator2 bounds already match pixels 1:1. This also fixes tap()
landing at the wrong location for OCR-matched text, and lets the IOU
fusion between UI-tree and OCR elements actually fire on iOS.
2. runtime/task.py captured `scene` (OCR/UI-tree data) before the LLM
planning call, but re-captured `before_screenshot` for each step
afterward - a real time gap during which on-screen content (e.g. a
keyboard) could shift, producing a directional drift between the
overlay and the displayed image. The first step of each plan batch
now reuses the screenshot already taken for planning instead of
capturing a new one; later steps in a multi-step batch still take a
fresh capture (left unresolved, scoped out by request).
Regression tests added for both the scale reconciliation (using real
828x1792 vs 414x896 numbers) and the screenshot reuse behavior.
Remove the 20-step execution limit that was causing "max steps exceeded" errors for long-running tasks. Increase the default max_steps to 999999 in all configurations, effectively removing the practical limit while maintaining the safety mechanism.
Changes:
- runtime/task.py: TaskRunnerConfig.max_steps 20 → 999999
- agents/collab_runner.py: CollaborativeTaskRunnerConfig.max_steps 20 → 999999
- storage/device_config.py: DEFAULT_MAX_STEPS 20 → 999999
Fixes issue 3: the host-agent console showed OCR results but never real
UI-tree data, because _ui_tree_nodes() checked for a get_ui_tree/ui_tree
tool action that has never existed anywhere in the codebase.
- storage/timeline.py: add a ui_tree_results field to TimelineRecord and
Timeline.append(), mirroring the existing ocr_results field.
- runtime/task.py: _append_timeline() now extracts scene.elements with
source == "ui" into ui_tree_results (scene_builder.build_scene() already
preserved these; they were just never persisted).
- host_agent/web/app.py: _ui_tree_nodes() reads the new field directly
instead of the dead tool-action check. New _overlay_payload() exposes
each step's scene dimensions and fused element list for client-side
rendering.
- task_detail.html: adds a toggle to overlay OCR (orange) and UI-tree
(blue) bounding boxes on the before-action screenshot, plus a visual
marker for the actually executed action (tap circle, or an animated
swipe path) using an SVG viewBox so no manual coordinate-scaling JS is
needed. Legacy/incomplete records degrade to no overlay, never an error.
Also corrects openspec/specs/runtime-task-evidence and
host-agent-console-task-pages, which had encoded the same nonexistent-tool
assumption, via the new host-agent-console-visual-evidence change.
600 tests passing; ruff/compileall/openspec validate all clean.
Forced tool_choice ("any"/"required") makes both Anthropic and OpenAI
skip any text/thinking block before the tool call, which silently made
rationale and thinking always None despite the planner-reflection-history
change's capture code being correct. Switch the primary call to
tool_choice="auto" (Anthropic: type=auto, disable_parallel_tool_use=true;
OpenAI: "auto") so the model can emit its reflection text, and add a
one-time forced retry (Anthropic "any", OpenAI "required", thinking
disabled) if the model responds without a tool call, guaranteeing a step
never stalls. Also add OpenAI text_output capture from message.content,
which was never extracted before (Anthropic-only gap).
Update planner-reflection-history design.md/tasks.md to document the bug
found during the pending manual smoke test (task 8.5) and the fix (new
section 9).
driver.tree() failures (WDA/Appium session errors) previously raised
uncaught, killing describe_screen() before OCR ever ran. Malformed
tree content (invalid XML) had the same problem inside
parse_ui_tree(). Both are now caught and logged, falling back to an
empty ui_elements list so the scene degrades to OCR-only, mirroring
the existing OCR-failure fallback in run_ocr().
Documents Skill Management in CLOUD_DEPLOYMENT.md (cloud-skill store,
per-host entitlement, incremental sync, local authoring/override,
inventory report, skills:admin scope) and applies ruff check/format to
all touched modules. All tasks complete; full non-integration suite
green (593 passed) and openspec validate --strict passes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
q792602257andClaude Opus 4.6