[MelbourneDeveloper]

TLDR

The autopilot's leak offer no longer reads "Suspected memory leak at :0" for a leak with a type and no source line. The tracking toast no longer tells a direct (.NET) runtime that each pause is captured. Two flaky tests are made deterministic: the server's mid-scan cancel, and Windows hot-line navigation.

Details

Both bugs showed up watching SharpLsp's memory autopilot e2e in a real VS Code window on Windows.

• The leak offer ([PROF-UI-AUTOPILOT]).
  • The UI's decoration diff (toDiffResult) dropped the runtime's type_name, and the offer named every leak by basename(file):line.
  • A .NET byte[] grown by user code has a type and no line, so every local run of SharpLsp's money-flow e2e offered :0 (HIGH): 8 of 8 offers.
  • New memory-leak-offer.ts:
    • describeLeak names the type and the source line as far as they were reported, for example LoopLeak.Widget at Program.cs:4 or System.Byte[];
    • leakToOffer offers the most confident leak, and among equals one with a source line.
  • LeakActionOffer (the e2e seam) now records typeName and the message shown.
  • worstLeak is gone: the capture ledger's max confidence comes from leakToOffer.
• The tracking toast ([PROF-MEMORY-MODE], [PROF-UI-AUTOPILOT]).
  • Both start toasts said "press Continue … each pause is captured". A direct runtime is never captured at a pause, so a .NET user waited for captures that never came.
  • New memory-guidance.ts, trackingGuidance(direct, intervalSecs):
    • a courier runtime keeps its words;
    • a direct runtime is told it is captured every N s while it runs, or to take a snapshot while it runs when interval capture is off. Either way, never while paused, and its last snapshot opens when it ends.
  • captureIntervalSeconds() is now the one reader of the interval settings, shared by the autopilot timer and the toast.
• Spec. [PROF-UI-AUTOPILOT] now states both rules.
• The stop toast ([PROF-MEMORY-MODE]).
  • Stopping before any snapshot said "Take a snapshot while paused", directly under the new start toast saying "never while paused".
  • stoppedWithoutSnapshot(direct) now tells a direct runtime to take one while the program runs.
• Language names ([PROF-UI-HOST]).
  • "Run & profile" with no matching file open said "Open a csharp/fsharp file".
  • ProfilerHost now carries the required languageNames, and the message reads "Open a file in F# or C# to run and profile". The spec lists the new field.
• Windows: the UI suite deleted a tracked fixture.
  • editor-path-key.test.ts junctions dotnet/fixtures/HotLeakApp into a temp dir. Its teardown's recursive rmSync followed the junction on Windows and deleted the fixture's Program.cs and HotLeakApp.csproj on every run.
  • The link is now unlinked first.
• Flake: cancel_during_a_scan_stops_the_stream_with_the_standard_error.
  • It failed on this PR's first CI run.
  • It required a second partial after the barrier, but a cancel racing the first partial sometimes left none.
  • The scan is now held on the native read of its last document (HeldSource; waiting.json sorts after every scan-NNNN.json), so the cancel lands mid-scan by construction.
  • It now asserts that all 64 documents before the hold streamed exactly once, the held one never did, and the reply is -32800.
• Windows: a measured hot line navigates to the actual runtime source.
  • The contract compared activeTextEditor.document.uri.fsPath to the runtime's file as raw strings. VS Code spells the drive c:, and a fixture built from C:\... reported C:.
  • The right file had opened at the right line, but the test failed in SharpLsp's Windows profiler chunk.
  • It now compares through editorPathKey, like every other overlay, and asserts the hot line's own file opened.

How Do The Automated Tests Prove It Works?

src/test/suite/memory-leak-offer.test.ts has 9 tests:

• a diff keeps the type the runtime named… pushes a wire diff of three leaks through toMemoryIngestResult and toDiffResult: type-only, located, and neither. It asserts each leak's file, line, type and confidence. With the mapping line reverted it fails on typeName: '' vs 'System.Byte[]' and 'LoopLeak.Widget'.

• Naming:

  • a type-only leak is named System.Byte[] and contains no :;
  • a located one is LoopLeak.Widget at Program.cs:4, or Program.cs:4 without a type;
  • a file without a line, or a line without a file, is not a location.

• Choice:

  • equal confidence prefers the located leak in either order, and keeps the first of true equals;
  • higher confidence beats a source line, including DEFINITE over HIGH;
  • an empty list offers nothing.

• Toast:

  • a direct runtime is never told "Continue" at intervals off, 1 and 30;
  • it is told its interval when on, and to snapshot when off;
  • a courier runtime keeps "Press Continue to auto-capture each pause".

• stopped before any snapshot…: a direct runtime is told to snapshot while it runs, never while paused, and a courier one keeps its words.

• host-language-names.test.ts: two names are joined in words in the host's order, one name stands alone, and no / or language ID leaks through.

• removing the link leaves the project it linked to in place fails without the unlink: the fixture is gone. It passes with it.

The full UI suite passes (189) and npm run lint is clean.

• Cancel test. dispatcher:: passed 20 of 20 runs on Windows after the change.
• Hot-line navigation.
  • SharpLsp's profiler-cpu-e2e on Windows failed before the change: the editor was at c:\...\ProfileTarget\Program.cs line 251, and the test timed out on the raw compare.
  • It passes 9 of 9 with the change. SharpLsp's money-flow e2e will assert the offer it shows once this lands.