Skip to content

Latest commit

 

History

History
792 lines (697 loc) · 56.7 KB

2026-11.md

File metadata and controls

792 lines (697 loc) · 56.7 KB

2026-11 update log

Index and query commands: README.md. New entries go at the end.

U-20261101-01 · 2026-11-01 · Give the shared managed-task helper twenty seconds to finish · #incident #agent #ci #test

  • Run. Dev CI 41214973954 on fe649360: Linux 3.14 and Windows 3.14 passed. Windows 3.10 had 1 failure out of 16,604 tests, in a suite of 58:37. The transfer fix of U-20261031-05 held: no transfer test failed. macOS: see U-20261101-02.
  • The failure. test_agent_dag_continuation.py::test_approval_transaction_rolls_back_all_bindings_and_adoptions expected "database operation failed". It got a 409: "DAG continuation evidence or decisions changed; review again".
    • In the fixture, node a logged "DAG managed task requires review". The 2-second wait_seconds of the shared task() helper in test_dag_managed_tasks.py had expired before the task finished. The task then completed in the background, between the review and the approval, so the approval's snapshot no longer matched.
    • This is the second test the 2-second wait failed on Windows. The first was the retention test, which U-20261031-03 gave its own 20 seconds.
  • Fix. The shared helper waits 20 seconds, the value the other managed-task fixtures already use. The retention test's local override is removed again. Tests that need a task not to finish still pass their own wait_seconds (0.01 for the timeout case).
  • Cost, measured. The 19 test files that use the helper: 235 tests in 113.0 s with 2 seconds, 99.3 s with 20 seconds. No test waits longer, because a healthy task returns as soon as it finishes.
  • Open items: none.

U-20261101-02 · 2026-11-01 · Move every macOS square from macos-14 to macos-15 · #decision #ci #macos

  • Why. The macOS 3.10 job of Dev CI 41214973954 failed with no step run. The annotation: "This job failed because of a scheduled macOS 14 runner image brownout. macOS 14 runner images will be retired on November 2, 2026" (actions/runner-images#13518). The cancelled macOS job of 41208376103 was the same.
  • Decided by the maintainer: move every workflow at once, rather than only dev.yml or rerunning around the brownouts.
    • dev.yml: the macOS 3.10 square.
    • quality.yml: the macOS 3.10 and 3.14 squares.
    • platform-smoke.yml: its matrix entry and the macos-capabilities job.
    • docs/CAPABILITY_MATRIX.md: the arm64 sentence that names the runner.
  • Not changed. Text that records the 2026-08-19 measurement on macos-14 (test/verify/macos_verify.py, the platform-smoke.yml comment, the WHATS_NEW translations, test_window_backend_macos_real.py) says what was measured on which image, so it keeps naming macos-14. EXPECTED in macos_verify.py is unchanged: the macos-capabilities gate on macos-15 is what checks that the newer image grants the same TCC permissions.
  • Open items. That gate runs only on pushes to main and pull requests to it: see Progress.md.

U-20261101-03 · 2026-11-01 · Bind this package's HTTP servers without a reverse lookup of their address · #incident #ci #macos #server

  • Run. Dev CI 41256321097 on e779af0c, the first on macos-15. The macOS 3.10 square had 7 failures out of 16,405 tests, in a suite of 27:41. None was a TCC-gated test.

    • Six were child-process fixtures (_transfer_process.py) reporting "owned transfer startup exceeded its original budget" (10 seconds), then a failed shutdown.
    • One was the process-death test of test_agent_dag_background_crash.py.
    • Each took about 20.4 seconds: the startup budget plus the 10-second stop wait. In the parent process, test_admin_client.py::test_add_host_round_trip took 35.6 seconds on macos-15 and 1.0 on macos-14. The tests after it in the file, starting the same two servers, took about 1 second each.
  • Measured. A throwaway branch (diag/macos15-startup, deleted after the run) timed it on both images (run 41258738137, Python 3.10):

    macos-15 macos-14
    first socket.getfqdn('127.0.0.1') in a process 35.03 s 0.005 s
    second one in the same process 0.002 s 0.000 s
    a fresh child process doing the same 35.04 s, 35.01 s 0.11 s total

    HTTPServer.server_bind sets server_name = socket.getfqdn(host). On macos-15 every process therefore paid 35 seconds the first time any of its HTTP servers bound.

  • Fix. New LocalThreadingHTTPServer in utils/http_headers.py. Its server_bind records the bound host as server_name instead of resolving it; nothing in the package reads server_name. Every HTTP server in the package now uses it:

    • RestApiServer
    • _MCPHttpServer
    • PrometheusExporter
    • HttpChallengeServer
    • WebhookTriggerServer

    That also helps a deployment whose reverse DNS times out.

  • Guard. test_http_server_bind.py starts the REST, exporter, challenge and webhook servers with socket.getfqdn patched to raise, and checks that the MCP server class is a LocalThreadingHTTPServer. An AST scan fails on any other module that constructs or subclasses a stock HTTPServer/ThreadingHTTPServer. It flags the pre-fix rest_server.py, http_transport.py and exporter.py.

  • Evidence. 1,161 tests pass: the 71 REST, MCP, webhook, exporter, ACME, admin and trigger test files, plus the doc, complexity and cross-project contract gates. ruff reports nothing. import je_auto_control still loads no PySide6.

  • Open items: none.

U-20261101-04 · 2026-11-01 · Widen the presence reader test's safety gate · #incident #agent #ci #test

  • Run. Dev CI 41259682516 on 6faab853. macOS 15 3.10 passed: the fix of U-20261101-03 held, and the TCC-gated tests pass on macos-15. Linux 3.14 and Windows 3.10 passed too. Windows 3.14 had 1 failure out of 16,617 tests, in a suite of 52:10.
  • The failure. test_agent_presence_reader.py::test_first_bind_blocks_are_bounded_and_late_replies_never_authorize[receive] expected close() to refuse while the blocked receive was still in progress, and close() succeeded.
    • The harness's blocked stage waits on its gate for at most 2 seconds and then raises, which ends the blocked operation.
    • The test measures ready() (bounded at 0.7 s), a notice (0.3 s) and close() (0.7 s): 1.7 s in all, 0.3 s under the gate. A slower Windows runner went past 2 seconds, the gate released itself, and the operation had already ended when close() checked.
  • Fix. The gate waits 10 seconds. It is only a safety net: the fixture's cleanup always opens it, so no test waits longer. The file's 34 tests run in 8.1 s as before.
  • Open items: none.

U-20261101-05 · 2026-11-01 · Report each managed-task area's readiness passively · #done #agent #diagnostics

  • What. The P3/P6/P7 diagnostics item in Progress.md (item 7, first bullet), apart from measured. New utils/agent_runtime/readiness.py: agent_runtime_readiness() reports each area as one of unsupported, missing_dependency, unauthorized, not_probed, not_recorded or measured. The areas: desktop input, desktop capture, presence source, Android transport, Android managed control, iOS managed control, browser and local model.
  • Passive by construction.
    • It reads the environment, PATH (shutil.which) and installed packages (importlib.util.find_spec).
    • It never imports an optional dependency, contacts a device, launches a helper or asks for a permission.
    • So no area is measured, every area carries native_evidence: not_recorded, and an installed tool counts only as not_probed.
    • A wrapper/capabilities observation maps without promotion: available becomes not_recorded; needs_permission, paused and revoked become unauthorized.
  • Honest about gaps.
    • Android and iOS managed control are unsupported: no device-side provider and no WebDriverAgent adapter exist yet, even when adb is installed.
    • The presence source is unsupported everywhere except Windows.
    • The local model is not_probed, because each task names its own model.
  • Surfaces. The facade and __all__ (1446 names), AC_agent_runtime_readiness (857 commands), the MCP tool ac_agent_runtime_readiness (761 tools, drive_input), the CLI agent-task readiness, the Agent Tasks view's Actions menu ("Show what each area still needs", in four languages), actions.pyi, the three READMEs, docs/AGENT_TASKS.md, the English and Chinese Sphinx pages, and architecture_explore.md.
  • Evidence. The 14 tests of test_agent_readiness.py pass. 1,225 surface and catalogue gate tests pass (Actions menu, language catalogues, RBAC, MCP, stubs, facade, CLI, Script Builder). The doc-count gates pass after re-measuring.
  • Open items. measured needs the protected native acceptance records, which wait until native acceptance resumes; Progress.md item 7 says so.

U-20261101-06 · 2026-11-01 · Add macOS and X11 physical-activity sources · #snapshot #agent #presence #macos #x11

  • What. The first step of Progress.md item 3's macOS/X11 bullet: a trusted physical-activity source for each platform, behind the same native API contract as Win32PresenceAPI. That contract is:

    • open(physical) and poll(timeout) on one owner thread;
    • poll waits at most 50 ms, drains at most 1,024 events (more is AgentControlBusy) and returns (interactive, held);
    • close releases what the source owns.

    The existing WindowsPresenceSource loop and the sixty-second idle PresenceState run any of them, through the new presence_platform.py.

  • X11 (presence_x11.py).

    • It selects XInput2 raw key, button and motion events for every device on the root window, the only global XInput2 view, plus HierarchyChanged.
    • It parses the raw header itself (=HIIHH): python-Xlib 0.33 registers parsers only for device events, which reach focused windows, not for raw events.
    • Events from XTEST slaves never count as physical. A source that is not a slave device refreshes the device list and is then ignored.
    • held is tracked from physical press/release events, or-ed with the core keymap and pointer buttons, which errs toward held.
    • The seat is usable only when logind reports Active=yes and LockedHint=no for XDG_SESSION_ID. It is read through loginctl with a fixed argv, a 1-second timeout and a 1-second cache. Without logind or a session id, or on any failure, the seat counts as locked.
  • macOS (presence_quartz.py).

    • A listen-only kCGSessionEventTap. Without the Accessibility permission CGEventTapCreate returns nothing and the source refuses to start.
    • Mistaking a person for an injection is the unsafe direction, so an event counts as injected only when kCGEventSourceUnixProcessID is positive and its source state is not the HID system state.
    • held adds HID-state key and button queries, for anything already down before the tap started.
    • A disabled tap (timeout or user input) fails the source.
    • The console is usable when CGSessionCopyCurrentDictionary is on console and not locked.
  • Wayland stays without a source. presence_backend returns unsupported whenever WAYLAND_DISPLAY or XDG_SESSION_TYPE=wayland is set, because a client sees no global input and cannot tell a person from libei or uinput injection.

  • Not changed. Non-Windows shared-desktop input stays refused: the presence worker still only has the Windows pipe, and helper launch, endpoint transfer and challenge/lease integration are still to come. Readiness now reports presence_source as not_probed for macOS and X11, with that refusal in its detail.

  • Evidence.

    • 66 tests pass in the three new test files and the readiness tests, all with fakes.
    • Two native tests run only where they can: under the Linux CI's Xvfb (XTEST key and motion injection must not count as physical), and on the macOS runner. There a posted mouse move's PID and source-state fields are recorded and the classification checked for consistency.
    • The existing 34 Windows source and server tests pass.
  • Open items. See Progress.md item 3:

    • the non-Windows worker and channel wiring;
    • target identity, window-scoped input, helper isolation and the dialog gate;
    • native acceptance of physical versus injected input on real hardware;
    • a trusted Wayland source.

U-20261101-07 · 2026-11-01 · Write the deployment and rollback plan for the agent runtime · #docs #agent #deployment

  • What. The maintainer asked for a written plan rather than an executed deployment for P4 (production browser integration) and P8 (multi-host). docs/superpowers/plans/2026-11-01-deployment-rollback.md covers:

    • the current state;
    • the checks before every deployment;
    • deploying and rolling back the main working tree;
    • the browser helper and managed Chromium;
    • multi-host rollout.

    Every decision it needs is marked "待決定" (to be decided).

  • Facts it rests on.

    • Jeffrey_RPA loads the main working tree through an editable install, so a change to that tree is a deployment.
    • The main tree's local dev is behind the remote dev, which is pushed from the feature worktree.
    • Another session holds uncommitted changes in the main tree, and git stash is shared by every worktree.
    • The managed task DAG never deploys a remote service.
  • Nothing was executed. No environment, service, .venv or browser was changed.

  • Open items. Progress items 1 and 6 now point at the plan, and item 1 holds the DECIDE on its open choices.

U-20261101-08 · 2026-11-01 · Hold a task's first action until a fresh observation re-establishes its state · #done #agent #recovery

  • What. The first bullet of Progress.md item 5: a safe recovery strategy for login, half-filled field and window state. AgentTaskSpec.resume_conditions lists at most 32 clauses, limited to the kinds a fresh observation decides: screen, browser, file_exists, file_sha256, file_text and file_manifest. result_equals and actions_complete are refused there, because receipts are history and cannot prove that a session is still signed in.
  • How it holds. New utils/agent_runtime/resume_gate.py.
    • ResumeGate.record evaluates the conditions on every observation while the task still holds its lease. That is the same verify_clauses that completion uses, now split out of verify, including the target's verify_browser.
    • ResumeGate.require admits an action only when its own observation_id is the last one on which every condition passed. An older passing observation does not admit an action on a newer one.
    • A failed or unknown condition puts the task in waiting_user with the reports and raises AgentHumanReview; nothing is dispatched.
    • Once the task has any action receipt, the gate is spent.
  • Compatibility.
    • An empty list is left out of to_dict, so every existing spec and request digest is unchanged.
    • Submission (validate_task_conditions) and the DAG task contract both validate the field.
    • A recovery reviewer may add conditions to a successor: _check_scope no longer compares that one key, and every other source field must still match.
    • runtime.py stays within its limit at exactly 750 lines: the logic is in the new module, and the runtime gained four hook lines.
  • Evidence.
    • The 11 tests of test_agent_resume_gate.py pass.
    • A full run of every test_agent_*/test_dag_* file plus flow control passes: 3,415 passed, 70 skipped, 0 failed in 34:54.
  • Open items. Progress.md item 5 keeps the real-application acceptance. Operators still have to write suitable conditions for each flow: the gate never infers login or window state by itself.

U-20261101-09 · 2026-11-01 · Read the XInput opcode through the public Display API · #incident #ci #x11

  • Run. Dev CI 41270009945 on 8e1e8368.
    • The new native X11 test (U-20261101-06) failed under Xvfb with "X11 physical presence observer could not start". The cause was AttributeError: get_extension_major: that method is on python-Xlib's protocol object, not on the public Display. The fake display had it, so the headless tests could not catch it; this is what the native test is for.
    • The same run also showed four stale line counts on Linux and macOS: the versioned-API fix had not re-measured the map.
    • The macOS native tap test passed there: its classification of an event the process posts itself is consistent.
  • Fix. The opcode comes from display.query_extension('XInputExtension') .major_opcode, and the fake display now has the same shape. Every other call was checked against python-Xlib 0.33's public Display: has_extension, screen, flush, fileno, pending_events, next_event, query_keymap and close. The xinput_* methods come from the extension. The map is re-measured.
  • Evidence. The headless X11 tests pass. Whether the native test passes under Xvfb is for the next Linux CI run: Docker is not running here, and starting it would change the machine.
  • Open items: none.

U-20261101-10 · 2026-11-01 · Audit the planner, executor and verifier boundaries against injected page text · #done #agent #security

  • What. The single-host part of Progress.md item 6's first bullet. docs/AGENT_TASKS.md gains a "Role boundaries" table, and each boundary names the code that enforces it and the tests that hold it:

    • a planner only proposes;
    • the executor admits a proposal only inside the reviewed spec and current authority;
    • the verifier decides completion from fresh evidence;
    • another node never carries this host's authority.
  • New regressions (test_agent_role_boundaries.py). The target's visible text carries an instruction aimed at the planner, and an obedient model follows it. The tests check four outcomes:

    • A tool outside the spec. The task goes to waiting_user and nothing is dispatched.
    • allowed_tools, roots, spec or authorization smuggled into a proposal. They never reach the stored spec or the target, because _grounded_action keeps tool, arguments and reference alone.
    • A path outside roots. It is refused before dispatch.
    • A planner's "task complete". It is not evidence: no report passes.

    A drafted workflow also contacts neither the task service nor a database.

  • A gap the stricter tests found. A refused first action (U-20261101-08) did not keep its resume-condition reports. The runtime's AgentHumanReview handler rewrote waiting_user with an empty verification list. AgentHumanReview now has reports (empty by default), and the handler records it. ResumeStateRequired carries the reports, so the refused task's verification shows screen: failed.

  • Lint. The readiness change had left four lines over 120 characters in the export lists of __init__.py, api/core.py and utils/agent_runtime/__init__.py. They are wrapped, and ruff check je_auto_control/ is clean.

  • Evidence. A full rerun of every test_agent_*/test_dag_* file, the two new files and flow control passes: 3,420 passed, 70 skipped, 0 failed in 36:03.

  • Open items. Cross-node role evidence on a real multi-host deployment stays in Progress.md item 6, with the deployment work.

U-20261101-11 · 2026-11-01 · Start the Android device-side helper with its fence ledger · #snapshot #agent #android

  • Why on the device. AndroidCommandLease requires an old writer to stop after a takeover, expiry or cancellation. Each command must be settled by its own request id, boot and epoch, with input released and nothing left queued. A host lock, an adb shell return code or a local process exit proves none of that. Per the maintainer's choice, the helper is built here and built and tested in CI, since this machine has no JDK or Android SDK.

  • Stage 1 of 4.

    • native/android_helper/README.md specifies protocol version 1: the hello, fence, begin and settle requests over the abstract socket je_auto_control_helper through adb forward, one bounded JSON object per line.
    • core/ holds FenceLedger, the pure-Kotlin state machine every request goes through.
    • A command runs only under the epoch it was begun with.
    • fence moves the epoch forward for good: queued older commands become fenced, and an injecting one is told to stop between events. A command a fence overtook can never become completed.
    • It refuses a stale epoch, a reused id and a full queue. Only one command injects at a time. settle answers from the ledger alone, and an unknown id settles as rejected.
  • CI. .github/workflows/android-helper.yml runs the JVM tests (JDK 17, Gradle 8.11.1, Kotlin 2.0.21, JUnit 4.13.2) when native/android_helper/** changes on dev or main, or in a pull request to main. It holds no secret and publishes nothing.

  • Not done.

    • Stage 2: the instrumentation app (socket server and UiAutomation injector).
    • Stage 3: the host provider implementing AndroidControlProvider, with getevent as the physical-input source, since injected events never reach /dev/input.
    • Stage 4: emulator tests in CI.

    Nothing is wired into the runtime. Real-device acceptance stays deferred.

  • Open items. Progress.md item 4.

U-20261101-12 · 2026-11-01 · Build the Android helper's protocol handler and instrumentation app · #snapshot #agent #android

  • Stage 2 of 4 (U-20261101-11 was stage 1).
  • A ledger fix the first JVM run found. A fence reply said queue_empty while a command was still injecting, which settle never did. It now counts the injecting command as pending, so the two agree.
  • ProtocolHandler (core/, JVM-tested).
    • It turns one decoded request into one reply.
    • It accepts only the closed tap, swipe and key shapes, with exactly their fields and bounded integers (coordinates up to 100,000, a duration up to 10 s, a keycode up to 400).
    • A malformed request gets an error reply and never reaches the ledger.
  • The instrumentation app (app/).
    • HelperServer serves the abstract socket je_auto_control_helper. An abstract socket has no file permissions, so it serves only a peer running as shell (how adb forward connections arrive) or root, and closes anything else before reading a line. A line over 64 KiB ends the connection.
    • Injector is the only writer. It asks the ledger between events whether it may go on. Once a fence overtakes it, it sends ACTION_CANCEL or the key-up for what it pressed and reports the command as not completed.
    • LocalServerSocket is closed explicitly: it only implements Closeable on newer API levels than minSdk 26.
  • CI. The Android workflow now also runs :app:assembleDebug (AGP 8.7.3, compileSdk 35) and keeps the debug APK for the emulator stage. Run 41275815574: fence-ledger and instrumentation-app both passed.
  • Open items. Progress.md item 4 now holds a DECIDE. Today the managed Android adapter dispatches input from the host through the ADB session (dispatch_checked), while this helper injects on the device. Routing managed dispatch through the helper's begin/settle would change a security-critical path, and that is for the maintainer to approve before stage 3 (host provider and getevent source) and stage 4 (emulator) proceed.

U-20261101-13 · 2026-11-01 · Run the Linux presence broker, worker and server over the owned Linux channel · #snapshot #agent #presence #linux

  • Why. The X11 source (U-20261101-06) had no way to reach the runtime: the presence broker, the worker and its server were hardwired to the Windows named pipe and Windows peer security. Linux already had an authenticated packet channel (presence_channel_linux.py). Its endpoint, though, is a fresh private socket that a client may trust only when the endpoint came through a trusted launch.
  • What. New presence_pipe_linux.py, shaped like the Windows pipe and security:
    • LinuxPresencePipe binds a fresh endpoint under XDG_RUNTIME_DIR and has no lookup name. connect_if_present is empty until use(endpoint) hands it an endpoint.
    • LinuxBytePipe and LinuxPresenceListener wait with select before every bounded receive or accept. LinuxChannel revokes itself when a bounded operation runs out of time, so a blocking wait inside one would kill the channel. Close cancels a wait first.
    • endpoint_to_wire and endpoint_from_wire carry the endpoint with closed fields, bounded integers, an absolute directory and a bounded socket name.
    • LinuxPresenceSecurity takes the peer from the connected socket's pidfd (retain_connected_peer), so the existing presence challenge compares process births.
  • The trusted launch path.
    • PresenceServer.start adds the listener's endpoint to its owned bootstrap reply.
    • The broker accepts that reply shape, adopts the endpoint only from its own SourceProcess, and connects. LinuxChannel.connect then pins the actual listener to the process the endpoint names before the presence handshake.
    • presence_platform picks the transport, Linux or Windows. Windows construction is unchanged.
    • Two small additions to presence_channel_linux.py: LinuxChannel.fileno() and LinuxListener.fileno(), used for readiness waits and peer evidence.
  • Not changed. presence_reader still uses only the Windows pipe, macOS has no presence transport, and the runtime still refuses non-Windows shared desktop tasks. Unlike Windows' first-instance sharing, each Linux broker owns its own source.
  • Evidence.
    • On Windows, all 405 presence tests pass.
    • test_agent_presence_service_linux.py covers endpoint parsing and the no-launch case. On Linux CI, a real broker process launches a real worker process: the endpoint travels in the bootstrap reply, and bootstrap/acquire/check/inspect run over real sockets.
  • Open items. Progress.md item 3.

U-20261101-14 · 2026-11-01 · Carry the listener's pid namespace in the Linux endpoint · #incident #ci #linux

  • Run. Dev CI 41279647369 on 1d6e3f0e, the first run of the Linux presence chain on real sockets (U-20261101-13). Two tests failed on Linux 3.14; every other square passed.
    • The endpoint round trip raised TypeError: LinuxPeerIdentity.__init__() missing 1 required positional argument: 'pid_namespace'. The wire form named five identity fields, and LinuxPeerIdentity has six.
    • The broker-and-worker integration test failed for the same reason. The broker could not rebuild the endpoint its worker reported, so bootstrap came back as "native helper protocol failed". The chain refused rather than connect to an endpoint it could not verify.
  • Fix. The wire form carries pid_namespace alongside user_namespace, and both must be nonempty strings. A new test compares the wire form's field lists with LinuxPeerIdentity, FileIdentity and LinuxEndpoint on every platform. A field added to them now fails on Windows too, instead of only under a real Linux listener.
  • Open items: none.

U-20261101-15 · 2026-11-01 · Observe only the maintenance call's own writes in the receipt test · #incident #ci #test

  • Run. Dev CI 41281523936 on f38053dc.
    • Linux passed, including the real broker-and-worker presence chain (test_an_owned_worker_hands_its_endpoint_to_the_broker_and_serves_presence). The Linux presence chain of U-20261101-13 now works over real sockets.
    • macOS 3.10 failed one existing test, test_agent_transfer_import_lifecycle.py::test_current_receipt_maintenance_does_not_reserve_an_idle_database_writer. It counted two write connections.
  • Cause. The test patches host.runtime.store.connection and counts every write connection opened during controller.maintain(), from any thread. It holds controller._mutex to keep other transfer maintenance out. The service's own lease watch, though, runs runtime.maintain() every 0.25 s without that lock. On a slow runner, one of its passes wrote inside the observation window.
  • Fix. Only writes on the calling thread count. TransferController. maintain runs synchronously on that thread, so every write the call itself makes is still observed. The test's 16 cases pass.
  • Open items: none.

U-20261101-16 · 2026-11-01 · Accept either refusal when a revocation races the lease watch · #incident #ci #test

  • Run. Dev CI 41283326131 on c0c02ea6.

    • macOS, Linux and Windows 3.10 passed. That includes the receipt test of U-20261101-15 and the Linux presence chain.
    • Windows 3.14 had 1 failure out of 16,712 tests: test_agent_recovery_ownership.py::test_revocation_during_grounding_cannot_reach_successor_input. It raised CancelledError where the test expected AuthorizationError.
  • Cause. The test revokes the reviewer's role inside the target's observe. The service's lease watch (runtime.maintain() every 0.25 s) can see that revocation first and cancel the task. _perform then stops at its cancellation checkpoint, one line before check_task_authority, which would have raised AuthorizationError.

    Either way the action was prepared and settled as failed, and the target was never driven. Only which refusal names it depends on timing.

  • Fix. The test accepts AuthorizationError or CancelledError. Its assertions on what matters are unchanged: target.count == 1 and the receipts ["failed"]. This is the third test on the service fixture this round that had ignored that thread (see U-20261031-04 and U-20261101-15).

  • Open items: none.

U-20261101-17 · 2026-11-01 · Patch the dependencies named in Dependabot security alerts · #bugfix #deps #security #ci

  • What: cleared the Dependabot security alerts that an update can fix today.
    • pyproject.toml, dev.toml: pillow==12.3.0 becomes pillow==12.3.0; python_version < '3.11' and pillow==13.0.0; python_version >= '3.11' (41 advisories, critical among them GHSA-v9pv-5gc8-w2m6 and GHSA-5x8w-8jj5-q52h). Pillow 13 needs Python ≥3.11, so 3.10 keeps the newest release it can install. The pdf extra's security floor moves from pypdf>=6.16.1 to pypdf>=6.21.4, the first release without the 6.20–6.21 advisories.
    • uv.lock: pillow 13.0.0 on 3.11+, pypdf 6.19.0 → 6.21.4, and dnspython 2.8.0 → 2.9.0 on 3.11+ (it needs 3.11 too). The dnspython floor is a [tool.uv] constraint-dependencies entry in pyproject.toml; it changes only the lock.
    • .github/requirements/publish.txt: setuptools 84.0.0 → 84.2.1, certifi 2026.7.22 → 2026.10.23; publish.in records the cut-off --exclude-newer 2026-10-25, still a 7-day wait.
    • Third_Party_License.md names the Pillow versions now pinned.
  • How: each lock regenerated with its recorded command plus -P for the patched packages only; every other pin is kept. pypdf 6.21.4 (2026-10-28) is the first fixed release and is newer than the cut-off.
  • Not changed: autocontrol-lsp/vscode/package-lock.json keeps minimatch 5.1.9 (GHSA-g3hv-75wg-m249, fixed only in 11.0.1). It comes from vscode-languageclient, and every release through 11.0.0 calls Minimatch#makeRe(), which minimatch 11 removed.
  • Test: test_arm64_dependency_markers.py read exactly one Pillow line and failed on the split. It now reads every line of a package and checks that none names the platform (sys_platform, platform_machine, platform_system, os_name); a Python-version marker passes. A new case checks that pattern, and putting the arm64 marker on Pillow still fails it.
  • Evidence: the full headless suite (python -m pytest, as in dev.yml) in a Python 3.12 environment with Pillow 13.0.0, pypdf 6.21.4 and dnspython 2.9.0: 16711 passed, 191 skipped, and the one failure was that test, fixed here (its file, the dev.toml parity and publish lock tests: 32 passed).
  • Downstream: TestPioneer and WebRunner lock Pillow through je_auto_control; they move to 13.0.0 once a release carries this pin.
  • Open items: Progress.md, "minimatch 11 在 VS Code 擴充套件". → corrected in U-20261112-02 (the advisory no longer covers 5.1.9).

U-20261102-01 · 2026-11-02 · Merge the release line into dev for 1.1.0 · #release #migration

  • What: main (1.0.x) merged into dev. In each subsystem both lines re-implemented while beta, dev's version is kept: the action journal, config sync, healing evaluation and template candidates, capability states, MCP tool modes and deferred role ownership. Everything else main added is kept: mobile sessions, stoppable runs, user administration, the result hook, the split remote-desktop panel, background runs in the Script and Script Builder tabs, Wayland physical recording, and PySide6 6.12.0.
  • Result / numbers: the facade has 1,504 names (143 added and 65 removed since v1.0.6). je_auto_control.api keeps all 7 v1.0.6 names. There are 876 executor commands (16 removed) and 820 MCP tools (15 removed). actions.pyi was regenerated (876 signatures). Four duplicate MCP registrations the merge produced were dropped (ac_probe_capabilities, ac_journal_read, ac_config_store_get / _commit). The version is declared as 1.1.0.
  • Files: CHANGELOG.md (1.1.0, with every removed name), docs/API_LIFECYCLE.md (1.1.0 section), pyproject.toml, uv.lock, je_auto_control/__init__.py, je_auto_control/_export_bindings.py. Main-only tests, examples (28, 29, 31-33) and Sphinx pages (config sync, Wayland capabilities, examples) were removed with the code they described.
  • Typing debt: the 334 modules the development line added were never held to the strict mypy tier. They are listed in test/verify/typing_not_yet_strict.txt, which may only shrink. typing_contract_verify.py passes in both modes (0 failing; 18 listed with --extras).
  • Behaviour: four differences follow the development line: the RBAC caller is a context variable, some reads need drive_input, MCP refusals are terser, and the state machine runs on the calling executor. CHANGELOG.md 1.1.0 lists them.
  • Open items: none.

U-20261102-02 · 2026-11-02 · Clear the SonarCloud findings the 1.1.0 release PR reported · #ci #test #quality

  • What: SonarCloud listed 2,087 open findings in the new code of PR #538 (Release 1.1.0). All are maintainability findings; the quality gate passed. 1,616 were in tests, 471 in the package, scripts, examples and the Android helper.
  • Result / numbers:
    • S9073 (1,208): every flagged assert a and b is one assert per condition, so a failure names the condition that broke.
    • S5778 (400): each pytest.raises block holds only the call under test; setup calls are hoisted, and branching bodies moved into a local function.
    • S1192 (292): scripts/generate_facade_exports.py emits one constant per owner module the generated bindings repeat (249 findings); the rest are named constants. A generated owner named like a credential carries a nosec B105 with its reason.
    • The remaining package findings are fixed in place: redundant classes in except tuples (checked against the real MRO), \d/\w regexes with re.ASCII so they still match ASCII only, nested conditionals and with statements, cognitive-complexity splits, empty blocks given a reason. The Script Builder table's 34 append runs became extend; its built spec list is byte-identical.
    • 33 findings are left as intended and closed on SonarCloud with a reason: 15 S5754 (BaseException handlers that re-raise through a helper or hand the error to another thread), 12 S5886 (dataclasses.replace inferred as DataclassInstance), 4 S7504 (list() snapshots of a dict the loop mutates), 1 S5655 (a writable ctypes buffer) and 1 kotlin S6624 (no version catalog for the Android helper; the version is locked by gradle.lockfile).
  • Verification: ruff, bandit, radon cc -nc, generate_facade_exports.py --check, typing_contract_verify.py (both modes) and the doc-count and budget tests pass; architecture_explore.md line counts re-measured.

U-20261102-03 · 2026-11-02 · Require Python 3.11; stop supporting 3.10 · #migration #deps #ci

  • Why: every package in the workspace moves its floor to Python 3.11 together. Security releases such as Pillow 13.0.0 need 3.11, and supporting 3.10 kept older, vulnerable versions in the locks.
  • What: requires-python is >=3.11 and the 3.10 classifier is gone; the CI matrices start at 3.11; the docs that gave 3.10 as the floor now give 3.11.
  • Locks:
    • uv.lock: the 3.10-only branches are gone; removed: exceptiongroup, tomli; one version left instead of a 3.10 one beside it: dnspython, numpy, pillow.
  • Also: The Pillow dependency is a single pillow==13.0.0, and the dnspython constraint applies on every interpreter. The Pillow and dnspython Dependabot alerts were open only because of the 3.10 entries. The CI cells that ran 3.10 to cover Windows, Ubuntu and macOS now run 3.11, and mypy targets 3.11. The version is declared as 1.2.0, and CHANGELOG.md says what changed.
  • Open items: none.

U-20261103-01 · 2026-11-03 · Hold the collector off around the presence reader's timed windows · #incident #agent #ci #test

  • Run. Quality run 41476383837 on the 1.2.0 release PR (#548). Ubuntu 22.04 3.14 failed twice, the first attempt and its rerun, on the same test and in the same place: test_agent_presence_reader.py::test_reader_loss_does_not_block_worker_owned_release_only_ack_cleanup. Its first harness.reader.ready() raised physical presence reader deadline exhausted. In the same run, the other 17,546 tests passed, and so did the other eight squares.
  • Why only there. The same commit passed the Ubuntu 3.14 square of Dev CI (run 41476340230). The two jobs differ only in coverage run, and on Linux both run under gdb (ci_native_diagnostics.py --native). ready() must finish connect, handshake and reply within 0.25 s. On the 1.1.0 release PR (#538) the same test passed at 0.455 s, against 0.254 s for its neighbours.
  • Cause (inferred, not observed). Allocation counts trigger the collector, so a fixed test order puts a collection in the same place on every run. After 17,000 tests under coverage and gdb, one collection of the heap takes longer than the 0.25 s window. #546 and #547 shifted the allocation pattern enough to move a collection into it. The CI log has no direct evidence of a collection. The reasons are that the failure repeats at one point, that only the slowest configuration fails, and that the reader's code did not change.
  • Fix. The harness fixture collects once, disables the collector for the test and always re-enables it, even when Harness() raises. The product deadline is unchanged. A collection pause in production still fails closed and requires an explicit resume. The file's 34 tests run in 8.3 s, as before.
  • Check. If the cause is something else, the quality run on #548 will fail again in the same place. It then needs a gc callback that logs long collections, not a wider deadline.
  • Open items: none.

U-20261110-01 · 2026-11-10 · Cooperative cancellation for Windows focus waits · #done #accessibility

  • What: Windows focus-event queue waits now poll the bound cancellation token with one monotonic deadline and release the subscription through the existing finally block.
  • Evidence: Regression failed before the fix with DID NOT RAISE CancelledError; the focused accessibility, thread and public API suite passed 98 tests with one skip.
  • Open items: Real COM acceptance and the legacy stop path without a cancellation token remain in Progress.md.

U-20261110-02 · 2026-11-10 · Executor stop reaches focus-event waits · #done #accessibility

  • What: Focus-event waits check the existing executor stop token between bounded queue reads. They retain the single deadline, remove their subscription on stop, and honor cleanup shielding until a second stop request.
  • Evidence: Both new stop regressions failed with focus wait kept polling after stop before the fix. Focus and public API tests pass after the fix.
  • Open items: Native COM acceptance remains in Progress.md.

U-20261110-03 · 2026-11-10 · GUI task cancellation reaches headless waits · #done #gui

  • What: Ordinary TabTask work binds a core cancellation scope linked to the GUI token. Stop, timeout and owner destruction reach cooperative waits; pre-cancelled work does not start, and completed work releases the scope. Script stop shielding stays unchanged.
  • Evidence: Five new regressions failed before the fix and passed afterward; delegate controller/executor suite: 43 passed; new tests plus GUI scripts: 22 passed. Mobile surface tests: 13 passed, confirming probe/run already execute asynchronously.
  • Open items: Noncooperative backend calls and native/device acceptance remain in Progress.md.

U-20261110-04 · 2026-11-10 · Reconcile USB wait and channel-error progress · #done #usb

  • What: Verified the existing viewer adds the bounded device timeout to its reply budget and channel-level errors reach their matching request. Removed the stale implementation gaps from Progress, retaining actual device/WebRTC/legacy-host acceptance.
  • Evidence: test_usb_client_wait_and_channel_errors.py: 31 passed. Source: viewer_client._round_trip, _client_requests.device_timeout_s, and channel host/client adapters.
  • Open items: Real USB, WebRTC and legacy-host acceptance remain in Progress.md.

U-20261110-05 · 2026-11-10 · Verify existing RBAC administration delivery · #done #rbac

  • What: Confirmed user administration already shares rbac.admin across AC commands, MCP, REST, CLI and the GUI panel. Removed the obsolete Python-only delivery gap from Progress.
  • Evidence: test_rbac_user_admin.py and test_rbac_users_gui.py: 34 passed. Tests cover admin-only requests, first-admin CLI bootstrap, last-admin protection, token disclosure/audit behavior and offscreen GUI actions.
  • Open items: Actual clients, TLS combinations and remaining capability coverage stay in Progress.md.

U-20261110-06 · 2026-11-10 · Give the Report page independent ownership · #done #gui

  • What: Migrated Report from a main-window mixin to a lazily created widget with its own controls, translations and Actions. Close releases local state; reopen creates a new page. Construction/show reads actual recorder status; closing does not disable shared recording.
  • Evidence: Delegate report/registry/exception/actions tests: 77 passed with offscreen Qt. All repository control callers migrated; old mixin path removed.
  • Open items: Remaining owner-built panels and report-generation responsiveness are tracked under GUI F1/F3 in Progress.md.

U-20261110-07 · 2026-11-10 · Carry owned Linux presence endpoints into helpers · #done #agent

  • What: Broker/pool retain a validated endpoint from their owned source launch. An exact live-reference grant supplies a fresh snapshot; host/channel/spec and worker configuration pin it. Reader uses the platform transport and refuses Linux startup without the trusted endpoint. Source lease wire fields and Windows desktop admission remain unchanged.
  • Evidence: Delegate focused suites: 78 passed; endpoint/Windows synthetic IPC: 33 passed, 3 Linux-only skipped; endpoint/runtime/worker binding: 34 passed. Mypy passed for the six modified production modules. Added Linux-only real-socket synthetic-source end-to-end coverage.
  • Open items: Linux socket execution is unverified because the existing WSL validation distro cannot mount its missing ext4.vhdx; native authority integration and acceptance remain in Progress.md.

U-20261110-08 · 2026-11-10 · Generate reports without blocking the GUI · #done #gui

  • What: Report generation now captures plain name/format inputs and runs through TabTask. It shows localized running status, refuses duplicate starts, and discards late results after close without touching a reopened page. Already-running writes may finish.
  • Evidence: Delegate report/registry/actions suite: 79 passed; running-feedback regression suite: 8 passed. Tests hold a fake writer blocked while a Qt timer ticks and verify name snapshots and close/reopen isolation. Mypy and Ruff passed.
  • Open items: Other GUI task/lifecycle gaps remain in Progress.md.

U-20261110-09 · 2026-11-10 · Bound internal semantic process-memory observations · #snapshot #agent

  • What: Added internal immutable identities and reviewed probes, bounded region reads, versioned semantic decoding, sequence checks and retained failed-handle cleanup. Provider failures are sanitized while checkpoint authority errors and interruption signals propagate.
  • Evidence: Injected-provider suite: 84 passed. Delegate strict mypy, Ruff and exception-policy tests passed. No native process memory was read.
  • Open items: Windows provider, trusted resolver, owned fixture, runtime grants/lifecycle, public surfaces and native acceptance remain in Progress.md; M0/M1 are incomplete.

U-20261110-10 · 2026-11-10 · Forward nested owner grants to presence sources · #done #agent

  • What: Linux broker reads inherited bootstrap once after parent-watch startup and passes the immutable grant through its private source launcher; each source owns a separate lease.
  • Evidence: Delegate suites: 95 passed, 2 skipped; final entry/factory file: 7 passed. Malformed grants and missing service/policy fail closed. Ruff and broker mypy passed.
  • Open items: Trusted Linux login/seat authority, live owner-service integration and native acceptance remain in Progress.md.

U-20261110-11 · 2026-11-10 · Own Screenshot lifecycle and run reads asynchronously · #done #agent

  • What: Screenshot is an independent lazy widget. Capture, pixel and size operations use TabTask with input snapshots, running status and close-discard; pixel errors survive retranslation.
  • Evidence: Delegate related suite: 90 passed; Ruff and targeted mypy passed. All native operations, dialogs and pickers were mocked before construction.
  • Open items: Remaining GUI panels and native acceptance stay in Progress.md. Already-entered capture writes may finish after close.

U-20261111-01 · 2026-11-11 · Prepare owned memory fixture buffers and bounded control · #snapshot #agent

  • What: Added a standalone child fixture with fixed payload and sequence buffers, strictly bounded JSON-lines control, validated state changes and clean EOF/stop. Bootstrap records explicitly contain self-reported identity. The nonce is stable for one fixture lifetime.
  • Evidence: Main-worktree fixture tests passed 25 cases, including hidden child communication, malformed input, exact frame limit and sequence exhaustion. Reader and documentation checks passed 119 cases. Ruff and package Bandit passed. Independent source review found no blocking issues.
  • Files: test/fixtures/process_memory_fixture.py, test/unit_test/headless/test_process_memory_fixture.py, docs/process_memory_contract.md.
  • Check limits: Fixture Pylint error checks passed. Whole-package Pylint reported existing findings and was stopped before completion. Package Radon reported five existing C-grade blocks; production package code was unchanged.
  • Open items: Native identity matching, provider, resolver, protection faults, runtime integration and native acceptance remain in Progress.md. No native cross-process reads or desktop operations ran.

U-20261111-02 · 2026-11-11 · Bound Windows memory reads behind reviewed attestation · #snapshot #agent

  • What: Added typed Win32 read/query bindings and owned handles with native process facts, bounded region reads, short-read rejection and failed-close retention. Application identity requires an explicit reviewed attestor. Decomposed existing memory validation and observation logic to meet the complexity limit.
  • Evidence: Independent review reran 147 reader, fixture and injected Win32 cases. Integration with document and file-budget checks passed 186 cases. Scoped three-target mypy, Ruff and Bandit passed; memory subsystem has no C-grade complexity blocks.
  • Files: utils/process_memory/windows.py, windows_native.py, models.py, reader.py, test_process_memory_windows.py, docs/process_memory_contract.md.
  • Open items: Trusted application attestation, fixture bootstrap matching, resolver, runtime wiring and native acceptance remain in Progress.md. Injected functions do not prove native Windows execution.

U-20261111-03 · 2026-11-11 · Release Image Detection and cancel pending click work · #done #gui

  • What: Replaced the retained workspace mixin with a lazy Image Detection widget that owns its inputs, mouse selection, translations and background task. Stop and close cancel pending work; close discards late results and reopen creates fresh state. The headless locate-and-click path checks cancellation before detection, movement and clicking.
  • Evidence: Implementation suite passed 219 cases, independent offscreen review passed 95, and integration with document checks passed 50. Scoped three-platform mypy, real Qt type checks, Ruff and Bandit passed. All desktop and dialog backends were faked.
  • Files: gui/_image_detect_tab.py, main_widget.py, tab_registry.py, wrapper/auto_control_image.py, lifecycle/cancellation tests, three READMEs and English/Chinese Sphinx guidance.
  • Open items: Remaining retained panels and backend/native acceptance are in Progress.md. Entered native calls cannot be preempted; cropping remains interactive.

U-20261111-04 · 2026-11-11 · Validate media types and restore package complexity checks · #done #typing

  • What: Validate PyAV video contexts and streams, release failed recording containers, type the existing H264 encoder patch, and separate the screen frame-clock helper from aiortc's integer timestamp. Removed four media modules from the measured extras exemptions. Extracted the unchanged native-target presence predicate to restore the remaining complexity limit.
  • Evidence: Media suite passed 157 cases and independent software-codec checks passed 6 without skips. Native-target suite passed 28 injected cases with independent source review. Integrated media, whole-package complexity and documentation checks passed 69. Integrated three-target typing reported zero ordinary errors and 14 exempt Qt modules on every target. Ruff passed.
  • Files: utils/remote_desktop/{hw_codec,session_recorder,webrtc_audio,webrtc_transport}.py, utils/agent_runtime/native_target.py, media tests and test/verify/typing_extras_exempt.txt.
  • Open items: Remaining Qt exemptions, strict typing and full final validation are in Progress.md. No desktop, audio hardware or native-target acceptance ran.

U-20261111-05 · 2026-11-11 · Reconcile mobile and modernization progress with current evidence · #snapshot #progress

  • What: Replaced stale blanket statements that E/H/F4 had not started with their actual remaining requirements. Existing device contexts, mobile operations, shared commands, typing gates and benchmark scripts are present; their presence does not prove complete delivery.
  • Evidence: Current mobile context, input/capture, application lifecycle and shared-entry tests passed 58 cases against fake ADB/WDA backends. The integrated typing results are recorded in U-20261111-04. Benchmark scripts were inspected, not rerun as a performance claim.
  • Open items: Actual SDK contracts, iOS extensions, native device checks, strict typing, complete documentation, comparable GUI benchmarks and final acceptance remain in Progress.md.

U-20261111-06 · 2026-11-11 · Check navigation and editor widgets against real Qt types · #done #typing

  • What: Handle optional navigation items consistently, read/write script fields through concrete Qt editor types, and preserve typed trigger arguments with single-read optional table cells. Removed three measured extras exemptions without new casts or ignores.
  • Evidence: Implementation suite passed 108 offscreen cases. Independent review passed 21 editor/trigger cases and 8 selected navigation cases without skips. Integration with document/type-contract checks passed 63. Full extras measurement reported the same 11 remaining WebRTC modules for win32, Linux and Darwin; scoped ordinary and real Qt mypy, Ruff and Radon passed.
  • Files: gui/navigation.py, gui/script_builder/step_form_view.py, gui/triggers_tab.py, editor/trigger tests and test/verify/typing_extras_exempt.txt.
  • Open items: Remaining WebRTC typing debt and native GUI acceptance remain in Progress.md. Tests used offscreen Qt and fake trigger engines.

U-20261111-07 · 2026-11-11 · Release core panels while retaining shared recorder ownership · #done #gui

  • What: Auto Click, Record and Script now own lazy widgets and cancellable background tasks. Closing a view releases controls; reopening Record reads shared recorder state. macOS snapshots the thread reference and X11 status does not acquire the startup lock.
  • Evidence: Implementation passed 213 cases plus 17 input-hook checks. Independent review passed 114 cases and found the macOS race, now covered by a deterministic fake regression. Integrated verification passed 137 cases with 2 platform skips. All input backends were faked; native lifecycle acceptance remains in Progress.md. Document, Actions menu, complexity and file-budget gates passed 54 cases after migrating the menu test to the new widget-owned actions. Package Ruff and Bandit passed. Pylint cannot resolve Qt extension members and unavailable macOS/X11 libraries on this host; it was not clean.

U-20261111-08 · 2026-11-11 · Reconcile conditional MCP path checks with current evidence · #snapshot #mcp

  • What: Corrected the outstanding path-policy description: path-or-other checks absolute/existing paths and file URLs while retaining literals. Ambiguous relative paths, shell interpretation and nested free-form payloads remain outside its guarantee.
  • Evidence: Existing policy suite passed 54 cases against actual registry schemas and fake handlers. Platform symlink and real-client acceptance remain in Progress.md.

U-20261111-09 · 2026-11-11 · Split MCP executor adapters along strict domain boundaries · #done #mcp

  • What: Replaced the oversized bridge with six domain modules and migrated all private callers. Each module uses strict typing; the obsolete strict baseline row and size exemption were removed. Executor imports remain lazy.
  • Evidence: Independent review authenticated the original snapshot and verified 253 function bodies/defaults/original annotations, 506 fake forwarding calls and 820 unchanged tool descriptors. Implementation passed 240 tests; independent review passed 61; integration with document and whole-package budget gates passed 88. Scoped three-target mypy and Ruff passed. No desktop or device commands executed.
  • Open items: Remaining strict typing and full validation are tracked in Progress.md.

U-20261111-10 · 2026-11-11 · Isolate default EI input in an owned helper · #done #wayland

  • What: Added bounded EI messages, owned worker/client and default cache integration. Requests share a deadline, cancellation and retained failed-close ownership. Invalid ACKs retire before another caller may emit. The direct in-process API and half-open guard remain.
  • Evidence: Implementation regression suite passed 273 cases with 6 platform skips. Independent review passed 217 and reproduced an ACK race, fixed with three deterministic concurrent regressions. Final worker suite passed 34; integrated EI tests passed 220. Scoped mypy, Ruff, Bandit and Radon passed. All native backends were fake; real child tests used a synthetic fixture. Integration document/complexity gates passed 44 cases; full three-target ordinary typing reported zero errors. Scoped Pylint error checks passed.
  • Open items: Native context/fd recovery, compositor input release, consent and pause semantics remain in Progress.md. Helper retirement is not native release evidence.

U-20261111-11 · 2026-11-11 · Enroll EI helpers in the strict typing gate · #done #typing

  • What: Added all four new EI modules to the explicit strict override. This corrects the registration gap found by the full typing-tier contract test; the debt baseline did not grow.
  • Evidence: Strict checks passed on win32, Linux and Darwin; typing-tier and document-line gates passed 19 cases. The original 7cb7fde2 full-suite baseline completed with 17,879 passed, 160 skipped, 16 subtests and one existing complexity-gate failure; coverage was 87.66%. Those five complexity blocks have since passed the integrated package gate. This baseline is not a final-suite result for the changed tree.
  • Open items: Final fixed-source validation remains in Progress.md.