fix(standing_rules): a mention is not a command — stop hijacking the turn

skills/.manifest.json

@@ -79,7 +79,7 @@
     "self_improve.py": "74c8b9cb0ce81f8600d35e72b73e543bdeccb3c8e5a4cdc79ddb4df5282c1846",
     "shift_report.py": "2c4bc669dec391f18329e0fe7b955c6b911322404e7e627e102776c22b448da2",
     "skill_forge.py": "07ea8fb2d026eca050154e5e60ea1968e77bc2d7786958d60f6e1a6d50f07182",
-    "standing_rules.py": "9f7f4a2af7cf8fbe11aaa11440c61eeab299e28da0e3bb0d7a52703787e22ee1",
+    "standing_rules.py": "1a3faa0da2213d5dba91625a34c8485190363eeef330b1ead23a0ca9bc0e45ea",
     "stuck.py": "dfddfe4dfa1a9d5c017f53e53033a7eb33114a517cd6fa937d9d44e65083f1c9",
     "system_info.py": "4cc32e0ad9cb81f34309734ef3388f7cde63407de9b23bb0cd589571a4c43ecb",
     "terminal.py": "1fe6c1cd1241ea6d328401f1c0fb4c362482e3ec2a61a4f7592b2d4822cd51c1",

skills/standing_rules.py

@@ -19,24 +19,36 @@
     "reply, in every session. Persisted to disk, not just remembered in chat."
 )
 SKILL_MCP_EXPOSE = True
+# Commands only — never bare mentions. "standing rule(s)" and "my rules" used to
+# be triggers, so merely DISCUSSING standing rules hijacked the turn: a message
+# quoting "…was there ever a rule you had to delete?" fired this skill, which
+# then tried to remove a rule named "you" and the reply never reached the LLM.
+# Same class as the file_ops chat-hang (#282): a mention is not an instruction.
 SKILL_TRIGGERS = [
     "add a standing rule",
     "add standing rule",
-    "standing rule",
-    "standing rules",
-    "my rules",
-    "always remember to",
-    "from now on always",
+    "remove standing rule",
+    "delete standing rule",
+    "list standing rules",
+    "show standing rules",
+    "show my standing rules",
+    "what are my standing rules",
+    "clear standing rules",
 ]
 
 _ADD = re.compile(
     r"^(?:add\s+(?:a\s+)?standing\s+rule|standing\s+rule|from\s+now\s+on\s+always|"
     r"always\s+remember\s+to)\b[:\s]*(.+)$", re.I | re.S)
+# All anchored to the START of the message. Unanchored, any sentence CONTAINING
+# these words matched — "was there ever a rule you had to delete?" read as a
+# delete command. A command is what the message opens with, not what it mentions.
 _REMOVE = re.compile(
-    r"\b(?:remove|delete|drop|forget)\b.*\b(?:standing\s+)?rule\b\s*#?\s*(\w+)?", re.I)
+    r"^\s*(?:remove|delete|drop|forget)\s+(?:the\s+)?(?:standing\s+)?rule\s*#?\s*(\w+)",
+    re.I)
 _LIST = re.compile(
-    r"\b(?:list|show|what(?:'s| are)|see)\b.*\b(?:standing\s+)?rules?\b", re.I)
-_CLEAR = re.compile(r"\b(?:clear|wipe|reset)\b.*\b(?:standing\s+)?rules\b", re.I)
+    r"^\s*(?:list|show|what(?:'s| are)|see)\b[^.?!]{0,20}\b(?:standing\s+)?rules?\b", re.I)
+_CLEAR = re.compile(
+    r"^\s*(?:clear|wipe|reset)\s+(?:all\s+)?(?:my\s+)?(?:standing\s+)?rules\b", re.I)
 
 
 def _render(rules):

tests/test_standing_rules.py

@@ -187,3 +187,48 @@ def test_listing_labels_injection_honestly(monkeypatch, tmp_path):
     assert "sent to the model" in out
     assert "counts INJECTION" in out
     assert "fired" not in out.lower(), "must not claim the rule 'fired'"
+
+
+# ── mentions must not hijack the turn ─────────────────────────────────────────
+# "standing rule(s)" and "my rules" were bare triggers, so merely DISCUSSING
+# standing rules fired the skill. Quoting "…was there ever a rule you had to
+# delete?" matched the unanchored remove-regex, the skill tried to remove a rule
+# named "you", and the turn never reached the LLM — so the reply was a skill
+# error and the claim/premise checks never ran. Same class as the file_ops
+# chat-hang (#282): a mention is not an instruction.
+
+@pytest.mark.parametrize("text", [
+    "was there ever a rule you had to delete?",
+    "In that cleanup process, was there ever a standing rule you thought was important?",
+    "I like the idea of standing rules for agents.",
+    "You mentioned you delete the rules that never fire.",
+    "Rules you never invoke are just noise you pay for on every message.",
+    "he ends each session by listing which standing rules actually fired",
+])
+def test_mentions_do_not_dispatch(text):
+    import codec_dispatch as cd
+    cd.load_skills()
+    hit = cd.check_skill(text)
+    assert (hit or {}).get("name") != "standing_rules", f"hijacked by: {text!r}"
+
+
+@pytest.mark.parametrize("text", [
+    "add a standing rule: be brief",
+    "show my standing rules",
+    "list standing rules",
+    "remove standing rule 2",
+])
+def test_real_commands_still_dispatch(text):
+    import codec_dispatch as cd
+    cd.load_skills()
+    hit = cd.check_skill(text)
+    assert (hit or {}).get("name") == "standing_rules", f"command missed: {text!r}"
+
+
+def test_remove_regex_requires_a_leading_command(monkeypatch, tmp_path):
+    s = _skill(monkeypatch, tmp_path)
+    s.run("add a standing rule: keep me")
+    # a question that merely mentions deleting must not remove anything
+    s.run("was there ever a rule you had to delete?")
+    import codec_standing_rules as mod
+    assert len(mod.list_rules()) == 1, "a mention must never delete a rule"