kotlin-debug-adapter (fwcd) is unmaintained: its evaluate only resolves plain local names and it reports breakpoints verified without checking. Kotlin debugging now uses the debugger inside JetBrains kotlin-lsp once its project import has finished, attached to the JVM the IDE starts through the existing JDWP link. kotlin-debug-adapter stays the fallback, with a notice saying why (not installed, still importing, import failed, cannot start).
kotlin-lsp's DAP quirks (step stop reasons, thread ids from earlier stops, frame names, file: URIs) are handled in one place, and its invalidated events reload the Variables pane and watches.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
Capture a heap dump of the debugged JVM over the attach listener (GC.heap_dump, after checking free disk space and the JDK's rules for suspended VMs) or open an .hprof file, and browse it in an editor tab: classes with shallow and retained sizes, biggest objects, dominators, the path from an object to a GC root, and a comparison of two snapshots. A snapshot captured at a stop adds a Retained Size column to that stop's Memory tab.
New crate ide-heap: HotSpot hprof parser (bounds-checked; unfinished or crafted dumps are refused), object graph, dominator tree (semi-NCA), retained sizes and GC-root paths. Snapshots live in private directories removed with their window; an opened dump is analyzed from a private copy so changes to the user's file can't affect the mapped data.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
Adds a Memory tab to the Debug tool window for JVM sessions: loaded classes with instance counts and shallow sizes (with deltas between refreshes), read over the HotSpot attach listener (`GC.class_histogram`), falling back to counting reachable objects through the debugger on JDKs that cannot answer while suspended. While paused, a class's instances, an instance's fields and its referrers are listed through a JDWP link (the Kotlin adapter's path, and a jdtls bundle `java/jvm-inspect` for java-debug). Debug launches started by the IDE get `-XX:+StartAttachListener` unless the user's VM options decide otherwise (not on Windows).
New crates: ide-jdwp, ide-jvm-attach, ide-jvm-inspect. The bundled jar rebuilds byte-for-byte from build.py against the pinned JDT LS plugins.
Co-authored-by: ntbang0901 <121003585+ntbang0901@users.noreply.github.com>
* feat(dap): debug C/C++ with GDB's native DAP interpreter
Fixes#2338
Native debugging was LLDB-only (lldb-dap, CodeLLDB), so toolchains and
embedded setups that ship GDB (pretty printers, .gdbinit, gdbserver
targets) had no debugger in the IDE. Adds a built-in "gdb" adapter:
`gdb -i dap` over stdio, with an install hint. GDB's DAP interpreter
shipped in GDB 14, so a launch with an older GDB is refused with the
found and the required version in the message (prepare_gdb, a probe
like debugpy's). C/C++ run configurations pick the debugger in the
Debug adapter control - GDB is an explicit choice there, the default
stays lldb-dap. The c-app fixture (C and C++ entry points) is debugged
with GDB in the language smoke; real_adapters gains a gdb case that
skips itself without a GDB 14+ on PATH.
* fix(dap): start GDB 15 sessions with the launch after the configuration
GDB 15 starts the program inside its launch request, so the launch-first
order every other adapter uses races the breakpoint upload: the run can
be past the breakpoints before setBreakpoints lands, and configurationDone
is refused with notStopped while the program is off. GDB 14 runs the
program at configurationDone and 16+ defers the whole launch there, so
both are fine with launch-first; 15.x is the one version that needs the
launch to go out after configurationDone. The order now follows a
gdb --version probe (only for the gdb adapter); a fake adapter speaking
GDB 15's wire behavior pins the ordering and the notStopped refusal.
* fix(dap): harden GDB startup and expression evaluation
---------
Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>
PHP had no regression coverage. Adds a php-app fixture (Composer + Pest) with language, run and debug smoke cases, and a php entry in the Language smoke workflow: PHP 8.5 with Xdebug via setup-php, Intelephense pinned like the other servers. Other languages' selectors are unchanged.
Co-authored-by: lleviathe <95321986+lleviathe@users.noreply.github.com>
run.test.ts's afterEach kills openidea-server and removes its home
directory, but a language server's own process (jdtls's JVM) outlives
the server and keeps writing .metadata underneath it, so the removal
races its writes (ENOTEMPTY). Retries only narrow the window: three
occurrences after the retry bump prove the JVM can linger past them.
The server now spawns detached (its own process group, which its
children inherit), and close() reaps the whole group after the server
exits (SIGKILL + gone-poll on POSIX, taskkill /T /F on Windows), so
the directory removal can't race anyone. debug.test.ts's teardown now
awaits close() too.
Language servers and debug adapters still write into the fixture's
temp dir as the test's afterEach tears it down, so a bare rmSync raced
them and failed the suite on ENOTEMPTY (Language smoke kotlin, java).
Give the removal maxRetries + retryDelay so a few late writes can't
flake the job.
Debugs each language's fixture the way the editor does, against openidea-server and the debugger the IDE installs: a breakpoint on the entry point's marker line, Debug from the file's context, then the paused frame, its variables, Evaluate Expression, Step Over and Resume. Runs per language in the Language smoke workflow. The build model is loaded first, as the Build Tools feature does, so a Gradle Application is debugged through Gradle; Step Over allows a repeated stop on a line (Python 3.12 reports a conditional expression's line twice).
ntbang0901
Vadim Lazukandl0ng-ai
lleviathe
arookieofc
arookieofc