feat: one-call failing-test debugging (test-runner presets) #985

Activity

1. debugmcpdev

   added 2 commits that reference this issue on Oct 26, 2026

   fix(python): a module launch derives its cwd from the package, names …

   ...

   6f8ca12

   feat(python): a module launch goes out without program (#985 M0) (#992)

   ...

   c803ab3
2. debugmcpdev commented on Oct 26, 2026

   

   debugmcpdev

   on Oct 26, 2026

   CollaboratorAuthor

   More actions

   Status 2026-10-25: M0 merged, M1 in review

   • M0 — feat(python): a module launch goes out without program (#985 M0) #992 (merged, c803ab31): a Python module launch goes through. AdapterPolicy.launchTargetKeys (python: program/module/code), the connection manager fills program only when no target key is set, the Python transform drops the stand-in program, derives cwd as the directory above the top-most package containing the file, and says what it did; the dry run names module: <name>.
   • M1 — feat(python): one-call failing-test debugging — start_debugging { testRunner: "pytest", testTarget } (#985 M1) #994 (open): start_debugging { scriptPath: <test file>, testRunner: "pytest", testTarget }. The answer is paused at the failing assert with the Paused stop answers carry a crash snapshot (exception, top frames, bounded locals); attach answers gain landing parity #982 snapshot; a passing test runs to exit 0.

   Measured (docs/python/spike-notes-pytest.md in #994)

   • uncaught alone never pauses on a failing test (pytest catches it). debugpy's userUnhandled does: at the assert in the test frame with locals; for a raise in a helper, the raise-site frames sit above the live test frame (locals readable); never inside pytest.raises; once on pytest.skip(); in the fixture for a fixture error. Whole fixture file: 13 tests → 8 stops (6 failures + 1 fixture error + 1 skip), 0 on the 5 passing tests.
   • "all" (debugpy raised) is not a usable default: three false stops across two passing tests, one per user frame an expected exception crosses.
   • pydevd suffixes the exception name with (note: full exception trace is shown but execution is paused at: …) on text/exceptionId/typeName; the snapshot strips it. The frame name carries (Current frame).
   • Plugins: xdist dies at worker spawn (WinError 6: the debuggee runs under pythonw.exe, python (Windows): the debuggee runs under pythonw.exe, so a subprocess that inherits stderr fails with WinError 6 (pytest-xdist workers) #993) and would run the test undebugged anyway; cov coexists on 3.13 (unmeasured on ≤3.11). -p no:xdist -p no:cov is silent when absent, clean when present; the preset passes both unless the caller's args ask for the plugin.
   • pytest's capture holds a test's print until the report; -s is not forced (args: ["-s"] opts in).

   Shape M1 fixes for M2–M4

   1. testRunner/testTarget as optional start_debugging arguments, folded into adapterLaunchConfig at intake (no new tool; the facade and restart_debugging unchanged).
   2. AdapterModes.testRunners on the factory metadata (mirrored in the loader's list, shown by list_supported_languages, pinned in adapter-attach-contract.test.ts), and classifyTestRunner (src/session/launch/test-runner-gate.ts) refusing an undeclared runner by name before the session changes state.
   3. The adapter's transformLaunchConfig owns the expansion (packages/adapter-python/src/test-runners.ts is the template: node id, the runner's single-process flags, a rootdir walk for cwd, one launch-scoped note), with the keys in consumedLaunchKeys.
   4. AdapterPolicy.exceptionFiltersForLaunch(mode, launchConfig) for runner-specific filters; the launcher records what the launch armed beyond the mode, and the pending answer names it.
   5. A spike-notes file per milestone with the measured matrix before any claim.

   Open after M1

   • unittest as a second Python runner (same userUnhandled shape, unmeasured).
   • M2 jest/vitest: the single-worker flags plus js-debug's one-child adoption, measured; M3 cargo test --no-run / go test -c (and the dead cargo.test branch); M4 rspec / dotnet test / minitest.
   • python (Windows): the debuggee runs under pythonw.exe, so a subprocess that inherits stderr fails with WinError 6 (pytest-xdist workers) #993 for the Windows pythonw.exe debuggee.
3. debugmcpdev

   added 2 commits that reference this issue on Oct 26, 2026

   feat(python): one-call failing-test debugging — start_debugging { tes…

   ...

   66fcf7b

   feat(python): one-call failing-test debugging — start_debugging { tes…

   ...

   b080737
4. debugmcpdev commented on Nov 2, 2026

   

   debugmcpdev

   on Nov 2, 2026

   CollaboratorAuthor

   More actions

   Status note (2026-11-02): the "shape decision up front" in the body is made — M1 (#994) landed testRunner/testTarget as start_debugging arguments, AdapterModes.testRunners on the factory metadata, the policy-owned expansion, no new tool. Two things for the remaining milestones:

   • M2 (jest/vitest): forcing single-worker is the interim shape. Whether runner workers become child debug sessions instead is python: a debuggee that starts a Python child hangs — debugpy parks the child until a client attaches, which mcp-debugger never does (multi-process debugging design) #1005 Part B's question, now a pre-1.0 design task; M2 ships with the measured single-worker caveat either way and inherits the tree if it lands.
   • M3 (cargo test / go test): the dead cargo.test branch in packages/adapter-rust/src/rust-debug-adapter.ts:789 and the uncalled runCargoTest in packages/adapter-rust/src/utils/cargo-utils.ts:133 get revived or deleted in the same PR, not left as a third state.
5. debugmcpdev commented on Nov 5, 2026

   

   debugmcpdev

   on Nov 5, 2026

   CollaboratorAuthor

   More actions

   The session tree's last step (#1091) lifts the gate this issue names: pytest-xdist workers are child debug sessions by default now — python -m pytest -n 2 yields two, and the worker running the test stops on a breakpoint set before the run (tests/adapters/python/integration/python-pytest-xdist.test.ts; pytest-xdist is in requirements/pytest.txt). The preset keeps -p no:xdist for now: a worker's stop is reported by wait_for_stop, not by the launch answer the one-call autopsy promises, so turning xdist back on in the preset is this issue's call, not the tree's.