Repository navigation
Conversation
The advertiser stop logged a pending or uncertain retirement at debug level and dropped the advertiser that still owned the name, and a recording whose finalising raised left Record disabled with no outcome shown. The advertiser now goes to the stop custody, and the recording says it may be incomplete.
The Browse LAN dialog joined zeroconf's thread on the GUI thread, swallowed a failed stop, and handed the browser a bound method that kept the dialog alive. The browser now reports through a weak lease, and its stop runs on an application-owned queue whose failures still reach the stop custody after the dialog is gone.
A cancellation that cuts a media cleanup short carries the original owners still holding resources, but the GUI kept only the outer stop. Each attached owner is now kept by the stop custody with its own stop.
Reading a media owner's live retirement status can wait on the lock its native cleanup holds. The custody now keeps the status the error captured when it was raised, and Diagnostics shows that instead.
A full stop waits on native work under the owner's lock, and a caller whose authority was revoked may never run it; until then the owner kept admitting audio, reports, sessions and input. close_admission() only stops admission, with no native call or cleanup lock, and StopQueue calls it on every owner before queuing the stop.
A GUI task's Stop or deadline cancelled its scope, but the HTTP helper and the signaling long-polls never looked at it, and stopping the WebRTC host left its publish loop waiting up to ten minutes for an answer. perform_call now checks the scope around each request, and the waits check their stop between polls.
A Qt connection holds its slot from C++, out of the collector's sight. The Script Builder connected a lambda over self to each of its 714 Add-step actions, so every tab built lived until exit, where Qt's teardown released the closure inside a destructor and deleted the tab a second time. Those slots, and the similar ones in four other tabs, are now methods or weak references.
The pin to 6.11.2 waited on finding what crashed the test process at interpreter exit under 6.12. That was tabs kept alive by slots holding themselves, fixed on dev; this branch runs the full matrix on 6.12.0 before the pin moves.
Every 3.12 to 3.14 square passed and exited cleanly on 6.12.0, but both 3.11 squares still abort: a None refcount underflow that is harmless where None is immortal. The pin now takes 6.12.0 only from 3.12 on.
The two assertions on the crash-dump step spelled LocalDumps\python.exe and autocontrol-native\dumps with a single backslash, which Python 3.11 reports as invalid escape sequences.
Trial PySide6 6.12.0 after the exit-crash fix
The trial ran clean on every 3.12+ square, so the decision is settled; what remains is unifying the pin once a 6.12.x runs clean on 3.11.
Dependabot alert 177 flags dnspython below 2.9.2 in uv.lock. The relock also records the gui extra's split PySide6 pin, which the lock still showed as a single 6.12.0.
817 commands had no entry in COMMAND_CAPABILITIES and relied on the submitting surface without anyone having decided that was right. Each was reviewed against the admin-only state the table protects; none needed raising, and a test now fails on a command in neither table. The review also found ac_journal_read's drive_input override sitting above the DATA_TOOLS spread, where it was silently overwritten.
The approver was whatever name the caller passed, so under RBAC one operator could file a request and approve it as somebody else, which defeats segregation of duties. Outside RBAC the names work as before.
They already passed disallow_untyped_defs and disallow_any_generics on all three target platforms, real Qt and aiortc types included, but the shrink-only lists left them free to regress.
Each had a single missing annotation or bare generic between it and the strict flags. Four whose correct annotation exposes a real type error, and the modules that alias ImageSource to Any, stay listed.
These had two or three missing annotations or bare generics each. Seven whose correct annotation exposes a real type error stay listed, as do the image modules that alias ImageSource to Any.
These had four to six missing annotations or bare generics each. The image modules that alias ImageSource to Any, and vlm_tab whose correct annotation exposes a real mismatch, stay listed.
These had seven to ten missing annotations or bare generics each. The unscoped ignores on optional backend imports go: the forced-Any override already covers those modules.
Each matching module spelt its own ImageSource = Any, which the typing contract allows only in a registered adapter. One adapter now owns the alias, so the modules that otherwise passed can join the tier.
Both shapes are what the code already narrows on, so the two locator modules can join the strict tier. MouseKeycode keeps its Any: typing it per platform exposes mismatched key tables in the wrapper layer.
Mostly agent_runtime. The cv2 error decorator now keeps the signatures of what it wraps, and the signaling routes declare the response shapes they already returned.
Annotating get_pixel let the contract sweep exercise this tool, and the sample it drew from the schema had one channel, which the handler has always refused. The schema now says so.
The last annotation stage: 17 more modules join the strict tier and the not-yet-strict list is empty. Typing the WebRTC control dispatch showed that a peer sending a list as the message type crashed the callback.
|
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
devis 58 commits ahead ofmain(2026-11-10 to 2026-11-12). This PR brings them tomainand runs the checks that only run on pull requests intomain, in particular themacos-capabilitiesjob ofplatform-smoke.yml. That job has not yet confirmed that macos-15 grants the same TCC permissionstest/verify/macos_verify.pyexpects (Progress.md› "macos-15 上的 TCC 權限還沒被 gate 驗過"). If it reports different grants, re-measure with--measureand updateEXPECTEDanddocs/CAPABILITY_MATRIX.mdrather than loosening the gate.Changes
MCP
tools/callruns as a stoppable executor run (mcp-…).notifications/cancelledstops anac_execute_actionslist before its next action and wakesac_wait_for_image/ac_wait_for_pixelmid-poll;ac_stop_executionends MCP calls too.tools/callpath moved out ofserver.py(745 → 627 lines) into_tool_calls.py; MCP adapters were split into strict domain modules.GUI
Remote desktop / WebRTC
Security
read_data.Executor and mobile
AC_parallelbranches, so the Script Builder offers a branch's token as its own value.DeviceError. Both were found by reading the uiautomator2 3.7.1 and facebook-wda 1.5.4 sources.allow_app_switch=True, because facebook-wda switches to WebDriverAgent through Siri to read it.Platform
LongPathsEnabled.Docs and records
Progress.mdreconciled with the tree: finished, stale and superseded items removed and recorded indocs/updates/2026-11-b.md(U-20261111-*, U-20261112-01 to -14).Compatibility
See
CHANGELOG.md› Unreleased. The ones a caller may notice:ExecutionStopped), a cancelled one with-32800;ac_list_executionslists MCP calls in progress.AutoControlGUIWidget()gets them on first access.swipeignoresduration_s;AC_ios_get_clipboardneedsallow_app_switch=True.read_data.Testing
AutoControl Dev CIondevhead65116dbc: all fourpytest-headlesssquares (Windows 3.11 / 3.14, Ubuntu 3.14, macOS 3.11) passed.test_agent_transfer_payloads.py, which fail the same way on the commit before these changes and pass in CI.On merge
pyproject.tomlonmaindeclares 1.2.1, which is tagged, soscripts/stable_release.pyreleases 1.2.2.