pr_runtime: run vitest with its verbose reporter and parse it

src/repo2rlenv/log_parsers/jest_parser.py

@@ -30,8 +30,19 @@
   ○ / "skipped"   → SKIPPED
 
 Mocha (the other common JS runner) uses the same ✓/✕ glyphs, so this parser
-covers most mocha output too. Vitest's default reporter is jest-compatible
-by design — also covered.
+covers most mocha output too.
+
+Vitest (3+) is detected by its ` RUN  vX.Y.Z` / ` Test Files` markers and read
+from `--reporter=verbose`, which prints one fully qualified line per test:
+
+     ✓ src/foo.test.ts > Foo > returns 200 1ms
+     × |unit| src/foo.test.ts > Foo > returns 500 3ms
+
+The name is kept as printed, minus the duration and retry/heap/note suffixes.
+Vitest's default reporter collapses fully passing files to one summary line,
+so its per-test lines are ignored: a test would vanish from the log as soon as
+a patch fixed the rest of its file. Vitest also colors output whenever TERM is
+unset, as in a non-TTY `docker exec`, so escape codes are stripped first.
 
 Released under Apache-2.0.
 """
@@ -44,8 +55,9 @@
 
 # File header: `PASS src/foo.test.ts (123 ms)` or `FAIL src/foo.test.ts`.
 # Captures the file path so we can prefix it onto test names.
+# With color, the label is padded (` FAIL `), leaving one leading space.
 _JEST_FILE_RE = re.compile(
-    r"^(?:PASS|FAIL)\s+(?P<path>\S+\.(?:ts|tsx|js|jsx|mjs|cjs))\b",
+    r"^ ?(?:PASS|FAIL)\s+(?P<path>\S+\.(?:ts|tsx|js|jsx|mjs|cjs))\b",
 )
 
 # Per-test glyph line. Indented arbitrarily; the glyph is the discriminator.
@@ -65,6 +77,44 @@
     "◯": "SKIPPED",
 }
 
+# Terminal escape codes (colors), stripped before any matching.
+_ANSI_RE = re.compile(r"\x1b\[[0-9;?]*[A-Za-z]")
+
+# ` RUN  v5.0.1 /repo` starts a vitest run; ` Test Files  1 passed (1)` ends it.
+_VITEST_MARKER_RE = re.compile(r"^\s*(?:RUN\s+v\d+\.\d+|Test Files\s+\d)", re.MULTILINE)
+
+# A `--reporter=verbose` test line. The optional project label is `|unit| `
+# without color and ` unit  ` with it. Suffixes follow vitest's
+# getTestCaseSuffix: duration, retries, repeats, heap usage, skip note.
+_VITEST_TEST_RE = re.compile(
+    r"^\s*(?P<glyph>[✓×↓□]) (?:\|(?P<project>[^|]+)\| | (?P<label>\S+)  )?"
+    r"(?P<name>\S+ > .+?)(?P<duration> \d{1,9}ms)?"
+    r"(?: \(retry x\d{1,9}\))?(?: \(repeat x\d{1,9}\))?(?: \d{1,9} MB heap used)?(?: \[[^\[\]]*\])?$"
+)
+
+_VITEST_STATUS: dict[str, TestStatus] = {
+    "✓": "PASSED",
+    "×": "FAILED",
+    "↓": "SKIPPED",
+    "□": "SKIPPED",
+}
+
+
+def _parse_vitest(log: str) -> dict[str, TestStatus]:
+    out: dict[str, TestStatus] = {}
+    for raw in log.split("\n"):
+        m = _VITEST_TEST_RE.match(raw.rstrip())
+        if not m:
+            continue
+        glyph = m.group("glyph")
+        name = m.group("name")
+        # Only finished tests print a duration; keep a skipped title's own "5ms".
+        if glyph in "↓□" and m.group("duration"):
+            name += m.group("duration")
+        project = m.group("project") or m.group("label")
+        out[f"|{project}| {name}" if project else name] = _VITEST_STATUS[glyph]
+    return out
+
 
 def parse_jest(log: str) -> dict[str, TestStatus]:
     """Return {test_name -> status} parsed from Jest / Mocha / Vitest output.
@@ -78,6 +128,9 @@ def parse_jest(log: str) -> dict[str, TestStatus]:
     out: dict[str, TestStatus] = {}
     if not log:
         return out
+    log = _ANSI_RE.sub("", log)
+    if _VITEST_MARKER_RE.search(log):
+        return _parse_vitest(log)
 
     current_file: str | None = None
     # describe stack indexed by indent depth (in characters). On a new test

src/repo2rlenv/pipelines/_pr_runtime_verifier.py

@@ -173,7 +173,7 @@ def parse_cargo_test(log: str) -> dict[str, str]:
     return out
 
 
-_JEST_FILE_RE = re.compile(r"^(?:PASS|FAIL)\s+(?P<path>\S+\.(?:ts|tsx|js|jsx|mjs|cjs))\b")
+_JEST_FILE_RE = re.compile(r"^ ?(?:PASS|FAIL)\s+(?P<path>\S+\.(?:ts|tsx|js|jsx|mjs|cjs))\b")
 _JEST_TEST_RE = re.compile(
     r"^(?P<indent>\s*)(?P<glyph>✓|√|✕|×|✗|○|◯)\s+(?P<name>.+?)(?:\s+\(\d+(?:\.\d+)?\s*m?s\))?$"
 )
@@ -186,13 +186,40 @@ def parse_cargo_test(log: str) -> dict[str, str]:
     "○": SKIPPED,
     "◯": SKIPPED,
 }
+# Keep these patterns in sync with log_parsers/jest_parser.py.
+_ANSI_RE = re.compile(r"\x1b\[[0-9;?]*[A-Za-z]")
+_VITEST_MARKER_RE = re.compile(r"^\s*(?:RUN\s+v\d+\.\d+|Test Files\s+\d)", re.MULTILINE)
+_VITEST_TEST_RE = re.compile(
+    r"^\s*(?P<glyph>[✓×↓□]) (?:\|(?P<project>[^|]+)\| | (?P<label>\S+)  )?"
+    r"(?P<name>\S+ > .+?)(?P<duration> \d{1,9}ms)?"
+    r"(?: \(retry x\d{1,9}\))?(?: \(repeat x\d{1,9}\))?(?: \d{1,9} MB heap used)?(?: \[[^\[\]]*\])?$"
+)
+_VITEST_STATUS = {"✓": PASSED, "×": FAILED, "↓": SKIPPED, "□": SKIPPED}
+
+
+def _parse_vitest(log: str) -> dict[str, str]:
+    """{test_name -> status} from vitest `--reporter=verbose` lines only."""
+    out: dict[str, str] = {}
+    for raw in log.split("\n"):
+        m = _VITEST_TEST_RE.match(raw.rstrip())
+        if not m:
+            continue
+        glyph, name = m.group("glyph"), m.group("name")
+        if glyph in "↓□" and m.group("duration"):
+            name += m.group("duration")
+        project = m.group("project") or m.group("label")
+        out[f"|{project}| {name}" if project else name] = _VITEST_STATUS[glyph]
+    return out
 
 
 def parse_jest(log: str) -> dict[str, str]:
     """{test_name -> status} from Jest / Mocha / Vitest output."""
     out: dict[str, str] = {}
     if not log:
         return out
+    log = _ANSI_RE.sub("", log)
+    if _VITEST_MARKER_RE.search(log):
+        return _parse_vitest(log)
     current_file: str | None = None
     describe_stack: list[tuple[int, str]] = []
     last_test_indent: int | None = None

src/repo2rlenv/pipelines/pr_runtime.py

@@ -629,9 +629,12 @@ def normalize_test_cmds_for_runtime(test_cmds: list[str]) -> list[str]:
         - Add `-v` if missing (default `go test` doesn't print --- PASS lines)
       cargo test:
         - Default output is already parseable; no transform needed
-      jest / npm test:
+      jest / mocha / npm test:
         - Add `--verbose` if not present, so per-test ✓/✕ lines are emitted
         - Some configs swallow stdout via `--silent`; we strip that
+      vitest:
+        - Use `--reporter=verbose` instead: vitest rejects `--verbose`, and its
+          default reporter collapses fully passing files to one summary line
     """
     out: list[str] = []
     for cmd in test_cmds:
@@ -674,9 +677,14 @@ def normalize_test_cmds_for_runtime(test_cmds: list[str]) -> list[str]:
         # --- jest / npm test / yarn test / pnpm test ---
         elif re.search(r"\b(?:jest|mocha|vitest|npm\s+test|yarn\s+test|pnpm\s+test)\b", cleaned):
             cleaned = re.sub(r"\s+--silent\b", "", cleaned)
-            # Add --verbose if the cmd is the runner itself (skip wrappers
-            # where flags need to go after `--`)
-            if re.search(r"\b(?:jest|mocha|vitest)\b", cleaned) and not re.search(
+            # Add a per-test reporter if the cmd is the runner itself (skip
+            # wrappers where flags need to go after `--`)
+            if re.search(r"\bvitest\b", cleaned):
+                # `vitest --verbose` exits with `CACError: Unknown option`.
+                cleaned = re.sub(r"\s+--verbose\b", "", cleaned)
+                if not re.search(r"\s--reporter\b", cleaned):
+                    cleaned = cleaned.rstrip() + " --reporter=verbose"
+            elif re.search(r"\b(?:jest|mocha)\b", cleaned) and not re.search(
                 r"\s--verbose\b|\s--reporter\b", cleaned
             ):
                 cleaned = cleaned.rstrip() + " --verbose"

tests/test_pipeline_pr_runtime.py

@@ -529,6 +529,21 @@ def test_normalize_jest_adds_verbose_and_strips_silent():
     assert normalize_test_cmds_for_runtime(["jest --verbose"]) == ["jest --verbose"]
 
 
+def test_normalize_vitest_uses_verbose_reporter():
+    # `vitest --verbose` exits with `CACError: Unknown option`.
+    assert normalize_test_cmds_for_runtime(["npx vitest run"]) == [
+        "npx vitest run --reporter=verbose"
+    ]
+    assert normalize_test_cmds_for_runtime(["npx vitest run --verbose --silent"]) == [
+        "npx vitest run --reporter=verbose"
+    ]
+    # An explicit reporter is the repo's choice ⇒ keep
+    assert normalize_test_cmds_for_runtime(["vitest run --reporter=junit"]) == [
+        "vitest run --reporter=junit"
+    ]
+    assert normalize_test_cmds_for_runtime(["npx mocha"]) == ["npx mocha --verbose"]
+
+
 def test_normalize_strips_trailing_pipe_head_or_tail():
     """Bootstrap agents sometimes save commands with a `| head -N` tail truncator.