Test Triage

You narrow down failing or relevant tests using Roslyn MCP validation tools — prefer tools over raw shell unless the user insists.

Input

$ARGUMENTS may include a test project name, a filter expression, or changed file paths. If missing, start from full discovery then narrow.

Server discovery

Use `discover_capabilities` (testing / all) or MCP prompts `debug_test_failure` (after a test run) and `review_test_coverage` for gap analysis.

Workflow

1. `workspace_load` if needed; `workspace_status` for health.
2. `test_discover` — summarize frameworks, projects, and test counts.
3. `test_run` — full suite, or scoped with project/filter from arguments.
4. On failures: use output stack traces; `symbol_search` / `go_to_definition` to jump to code; MCP prompt `debug_test_failure` to structure the next steps.
5. After local edits: `test_related` for symbols or `test_related_files` for changed file paths, then `test_run` on the smallest useful scope.
6. Optionally `build_workspace` if compilation state is uncertain before tests.

Flake detection mode

Invoke with --detect-flakes[=N] (default N=5) or ask "is this test flaky?" to bucket failing tests by determinism.

1. After Step 3 produces a list of failures, take the failing test filter and call test_run with that filter N times in sequence. Pass runSettings that disable test parallelism within a run if the framework supports it (xUnit: parallelizeTestCollections=false; NUnit: --workers=1) to reduce cross-run contamination.
2. Bucket each failing test across the N runs:

• deterministic-fail: failed in N/N runs — real bug, fix the assertion or the system under test
• flake: failed in 1 < k < N runs — flaky, likely timing/ordering/state-leak
• transient-pass: failed once and passed in all retries — likely environmental (clock skew, tmp dir race)

1. Produce a per-test row with bucket, pass-rate (k passes / N), and suggested next action:

• deterministic-fail → inspect the assertion + surrounding code via symbol_info / get_source_text
• flake → look for DateTime.Now, Task.Delay, new Random() without seed, shared statics in the SUT — scan via symbol_search on likely suspects
• transient-pass → annotate as suspected flake; do not act unless it recurs

Honor a cap on N (default max 20) to avoid runaway CI time. If the user requests a huge N, warn and ask.

Output

Give a short summary: failing tests, likely root cause (deterministic vs flake when detection ran), next command or tool calls, and whether `compile_check` is clean.