[debugmcpdev]

Python launches now debug child processes by default. Children become sessions with parentSessionId, inherit breakpoints, and report stops through the parent's wait_for_stop.

Behavior

• Send subProcess: true unless dapLaunchArgs.subProcess explicitly overrides it. An explicit false keeps children undebugged and produces a diagnostic; explicit true is silent.
• Preserve the per-tree limit, DEBUG_MCP_MAX_CHILD_SESSIONS (default 32; 0 releases all reported children).
• A by-port attach target decides subprocess debugging at listen time. The attach request forwards subProcess, but an explicit false now explains that disabling child debugging requires the target's --configure-subProcess false or debugpy.configure(subProcess=False).
• Update tool descriptions, snapshots, docs, the debugging skill, and release fragments to describe the default.

Lifecycle coverage and fixes

• Exercise child and parent termination, grandchildren, subtree close, siblings, restart, attach-root close/detach, multiprocessing, and pytest-xdist workers against real debugpy. Add hash-pinned pytest-xdist to CI requirements.
• Synchronize attach tests with a parent breakpoint before spawning. Verify the wait identifies that parent before continuing it; this also keeps opt-out targets alive through attach verification.
• Allow Python's POSIX resource-tracker sibling while verifying the intended multiprocessing worker's breakpoint, locals, and completion.
• Track attach targets before readiness, handle spawn errors, close sessions on setup failure, and await drained output before asserting target completion.
• Fix the existing Windows Dart/.NET bridge cleanup race (bridge-tree-kill.test.ts flakes on Windows: one survivor when taskkill loses the 300 ms race to the plain kill (#937 follow-up) #1089) exposed during validation: wait for the tree-termination command to finish before a direct-child fallback can orphan descendants. Bound a stuck command and preserve exited-child and duplicate-cleanup protections.

Scope

Closes #1081 and #1089. Ancestor/descendant idle protection remains in #1025 step 2; session-tree M2 is #1082. The pytest preset retains its in-process default; xdist coverage uses a module launch.

Validation

• Full pre-push gate passed: lint, source/test type checks, clean build, 8,091 unit tests passed (5 skipped), 138 integration tests passed (1 skipped).
• Python suites use the repository-pinned debugpy 1.8.22 and pytest dependencies.
• Both real Windows bridge tree-termination tests and all 37 bridge unit tests passed.
• Documentation, changelog, personal-path checks, and git diff --check passed.
• GitHub CI passed on 3d0a17fc917c9c57a32de285f0e0a3c765117b64: full Linux/Windows suites with coverage, Windows Python focus, MCP conformance, container tests, and all language host lanes. CodeQL and patch coverage also passed.

[codecov]

Codecov Report

❌ Patch coverage is 98.59155% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...adapter-dotnet/src/utils/netcoredbg-bridge-core.ts	96.66%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[debugmcpdev]

Reviewed the runtime defaults, attach diagnostics, child-session lifecycle, fixtures, and platform behavior. Fixed the attach synchronization/cleanup defects and the POSIX multiprocessing assertion in 6d564ec. Full validation also reproduced #1089; 3d0a17f fixes the Windows Dart/.NET tree-termination race and adds six regression cases.

No outstanding review findings on this revision. The full local gate passed (8,091 unit tests, 138 integration tests), and GitHub CI, CodeQL, and patch coverage are green for this exact head.