loopguard: a here-string is not a heredoc

tests/test_envfacts.py

@@ -71,6 +71,9 @@ def test_classify_forms_and_intents():
     # heredoc bodies never become commands
     assert observe.classify_command("python - <<'PY'\npytest_is_a_word = 1\nPY")[0].intent == \
         "run_interpreter"
+    # a here-string (`<<<`) has no body: the commands after it are still read
+    assert [o.intent for o in observe.classify_command("read -r x <<< x\npytest -q\nx=1")] == \
+        ["run_tests"]
 
 
 def test_error_classes_are_environmental_only():

tests/test_live_core.py

@@ -679,11 +679,22 @@ def test_the_opencode_side_call_adds_only_the_flags_its_cli_lists(monkeypatch):
     # a quoted `|` in a sed script must not split the command (seen live: sympy-13091)
     ("sed -i -E 's/x (<|<=|>|>=) y/z/' numbers.py expr.py", ["numbers.py", "expr.py"]),
     ("sed -i 's/a/b/' x.py 2>&1 | tail -2", ["x.py"]),
+    # a here-string (`<<<`) is not a heredoc: the lines after it are still commands
+    ("grep -q ok <<< \"ok\"\nsed -i 's/a/b/' pkg/mod.py", ["mod.py"]),
+    ("read -r v <<< done\ncat > fix.py <<'EOF'\nx = 1\nEOF", ["fix.py"]),
 ])
 def test_edits_made_through_the_shell_are_recognised(cmd, want):
     assert guard.shell_edit_targets(cmd) == want
 
 
+def test_an_edit_after_a_here_string_is_counted():
+    """`<<<` feeds one word to a command; reading it as a heredoc tag made every later line of
+    the command a heredoc body, so a `sed -i` after it was not counted as an edit."""
+    g = guard.TurnGuard(turn_key="t")
+    g.fold("Bash", {"command": "grep -q ok <<< ok\nsed -i 's/a/b/' pkg/mod.py"}, "", 0)
+    assert g.edits == 1 and g.whole.edit_counts == {"mod.py": 1}
+
+
 def test_shell_edits_reach_the_state_the_judge_reads_and_the_ended_green_proxy():
     g = guard.TurnGuard(turn_key="t")
     edit = "python - <<'E'\np='pkg/mod.py'; s=open(p).read()\nopen(p,'w').write(s[::-1])\nE"

tests/test_loopguard_productive_turns.py

@@ -26,6 +26,8 @@
     ("SP=/tmp/x; cd $SP && python3 - <<'EOF'\nimport json\nprint('error outputs: 3')\nEOF",
      True),
     ("python3 - <<'EOF'\nprint(1)\nEOF\npython -m pytest -q | tail -3", False),
+    ("python repro.py <<< y", True),
+    ("python repro.py <<< \"y\"\npython -m pytest -q | tail -3", False),
     ("", False),
 ])
 def test_an_exit_status_covers_a_command_unless_its_work_could_fail_unseen(cmd, covers):

tokeneconomizer/loopguard/shell_edits.py

@@ -13,7 +13,8 @@
 
 from .state import basename
 
-_HEREDOC_START = re.compile(r"<<-?\s*(['\"]?)([A-Za-z_]\w*)\1")
+# `<<<` is a here-string: its word is the input itself, so no body follows on the next lines.
+_HEREDOC_START = re.compile(r"(?<!<)<<(?!<)-?\s*(['\"]?)([A-Za-z_]\w*)\1")
 _OPS = {"&&", "||", ";", ";;", "|", "|&", "&", "\n", "(", ")"}
 _OUT = {">", ">>", ">|", "&>", "&>>"}                  # redirects that write their operand
 _SKIP_OPERAND = {"<", "<<", "<<<", "<&", ">&", "<>"}   # input, heredoc tags, fd duplication

tokeneconomizer/microskills/observe.py

@@ -116,7 +116,7 @@ class Vocabulary:
 _ENV_ASSIGN_RE = re.compile(r"^(?:[A-Za-z_][A-Za-z0-9_]*=(?:\"[^\"]*\"|'[^']*'|\S+)\s+)+")
 _ASSIGN_TOKEN_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*=")
 _WRAP_RE = re.compile(r"^/bin/(?:ba)?sh\s+-l?c\s+(.*)$", re.S)
-_HEREDOC_RE = re.compile(r"<<-?\s*['\"]?(\w+)['\"]?[^\n]*\n.*?\n\s*\1\b", re.S)
+_HEREDOC_RE = re.compile(r"(?<!<)<<(?!<)-?\s*['\"]?(\w+)['\"]?[^\n]*\n.*?\n\s*\1\b", re.S)
 _URL_HOST_RE = re.compile(r"https?://([^/\s'\"`]+)")
 
 # Error classes an ENVIRONMENT produces (as opposed to the code under test). Order matters: