mind_commit_guard v1.1: token-level clause parsing

bin/mind_commit_guard.py

@@ -67,6 +67,33 @@ def _mind_root(command: str, cwd: str) -> Path | None:
     return None
 
 
+def _clauses(command: str):
+    """Token-level clause split that respects quoting (v1.1).
+
+    v1.0 split clauses with a regex over the raw text, which cut inside
+    quoted commit messages and heredoc bodies — its first live firing was a
+    false positive on its own author's closeout command, minutes after
+    deployment. shlex with punctuation_chars keeps `&&`/`;`/`|` as tokens
+    while quoted strings (commit messages, gh comment bodies) stay single
+    tokens that cannot leak trigger words or swallow the `--`.
+    """
+    lex = shlex.shlex(command, posix=True, punctuation_chars=";&|")
+    lex.whitespace_split = True
+    clause: list[str] = []
+    try:
+        for tok in lex:
+            if tok and set(tok) <= set(";&|"):
+                if clause:
+                    yield clause
+                clause = []
+            else:
+                clause.append(tok)
+    except ValueError:
+        return  # unparseable (heredocs, unbalanced quotes) — fail open
+    if clause:
+        yield clause
+
+
 def check_command(command: str, cwd: str = "") -> str | None:
     """Return a denial reason, or None to allow."""
     if "PYAUTO_SKIP_MIND_GUARD=1" in command:
@@ -78,16 +105,14 @@ def check_command(command: str, cwd: str = "") -> str | None:
         return None
     root = _mind_root(command, cwd)
 
-    # Examine each `git ... commit ...` clause in the (possibly compound) command.
-    for clause in re.split(r"&&|\|\||;", command):
-        if not re.search(r"\bgit\b.*\bcommit\b", clause):
+    # Examine each clause of the (possibly compound) command at token level.
+    for tokens in _clauses(command):
+        if "git" not in tokens or "commit" not in tokens:
+            continue
+        if tokens.index("git") > tokens.index("commit"):
             continue
-        if "--amend" in clause or "--dry-run" in clause:
+        if "--amend" in tokens or "--dry-run" in tokens:
             continue
-        try:
-            tokens = shlex.split(clause)
-        except ValueError:
-            return None  # unparseable — fail open
         if "--" not in tokens:
             return (
                 "PyAutoMind is a SHARED checkout: concurrent sessions stage into "

tests/test_mind_commit_guard.py

@@ -77,3 +77,37 @@ def test_amend_and_dry_run_are_exempt(tmp_path):
     mind = tmp_path / "PyAutoMind"
     mind.mkdir()
     assert check_command(f'git -C {mind} commit --amend --no-edit') is None
+
+
+# --- v1.1: the two live false positives from the guard's first hour ------------
+
+def test_gh_comment_prose_does_not_trigger():
+    # First live firing: a `gh issue comment` whose BODY prose mentioned the
+    # trigger words. Quoted bodies are single tokens; no git-commit clause.
+    cmd = (
+        'gh issue comment 130 --repo X --body "the Mind commit guard denies '
+        'bare commits in the shared PyAutoMind checkout via git" '
+        "&& echo done"
+    )
+    assert check_command(cmd) is None
+
+
+def test_semicolon_inside_commit_message_keeps_pathspecs(tmp_path):
+    # Second live firing: a `;` INSIDE the quoted -m message made v1.0's raw
+    # regex split strand the `--` in the next pseudo-clause.
+    mind = tmp_path / "PyAutoMind"
+    mind.mkdir()
+    (mind / "active.md").write_text("x")
+    cmd = (
+        f'cd {mind} && git pull --ff-only -q && '
+        f'git commit -q -m "prompt: task complete; all phases complete" '
+        f"-- active.md && git push -q origin main"
+    )
+    assert check_command(cmd) is None
+
+
+def test_bare_commit_still_denied_in_compound_with_quotes(tmp_path):
+    mind = tmp_path / "PyAutoMind"
+    mind.mkdir()
+    cmd = f'cd {mind} && git commit -q -m "msg; with semicolon" && echo ok'
+    assert check_command(cmd) is not None