audit(token_audit): one submission carrying two slash commands is one message

engine/skills/reflect/scripts/tests/test_token_audit.py

@@ -670,6 +670,40 @@ def test_product_blame_and_plain_questions_do_not_flag_intervention(self):
         finally:
             os.unlink(path)
 
+    def test_two_slash_commands_in_one_submission_are_not_a_repeat(self):
+        """`/reflect /cat-mode` on one line injects the same text twice, tens
+        of milliseconds apart. That is one message, not a re-send."""
+        text = "please prevent this shit from happening again"
+        lines = [
+            codex_response_item("user", text, ts="2026-09-11T05:40:00.000Z"),
+            codex_response_item("user", text, ts="2026-09-11T05:40:00.041Z"),
+            codex_response_item("assistant", "Working on it.", ts="2026-09-11T05:40:05.000Z"),
+        ]
+        path = write_jsonl(lines)
+        try:
+            with redirect_stdout(io.StringIO()):
+                result = token_audit.audit_codex(path)
+            kinds = {k for f in result["frustration"]["flagged"] for k in f["kinds"]}
+            self.assertNotIn("verbatim-repeat", kinds)
+        finally:
+            os.unlink(path)
+
+    def test_a_real_resend_minutes_later_still_counts_as_a_repeat(self):
+        text = "please prevent this shit from happening again"
+        lines = [
+            codex_response_item("user", text, ts="2026-09-11T05:40:00.000Z"),
+            codex_response_item("assistant", "Working on it.", ts="2026-09-11T05:40:20.000Z"),
+            codex_response_item("user", text, ts="2026-09-11T05:44:00.000Z"),
+        ]
+        path = write_jsonl(lines)
+        try:
+            with redirect_stdout(io.StringIO()):
+                result = token_audit.audit_codex(path)
+            kinds = {k for f in result["frustration"]["flagged"] for k in f["kinds"]}
+            self.assertIn("verbatim-repeat", kinds)
+        finally:
+            os.unlink(path)
+
     def test_handed_over_something_broken_flags_intervention(self):
         """The wording that scored zero: no profanity, no told-you. The user
         says the artifact was bogus, then that it did not run."""

engine/skills/reflect/scripts/token_audit.py

@@ -150,6 +150,8 @@ def _direct_run_targets(command):
 # class (told-you / accusation / agent-blame twice, two of those kinds in
 # one session, or a verbatim re-send) is the automate-me trigger. Product
 # blame ("the ui is messed up") does not match agent-blame.
+MIN_RESEND_GAP_SECS = 5
+
 INTERVENTION_KINDS = frozenset({
     "told-you", "accusation", "agent-blame", "restated-ask", "proof-challenge",
     "cheap-way-out", "explicit-invocation",
@@ -299,7 +301,8 @@ def frustration_signals(user_msgs, interruptions=0, failed_turn_indices=None):
         norm = re.sub(r"\s+", " ", t).casefold()
         if len(norm) >= 12:
             for prev_secs, prev_norm, prev_idx in seen:
-                if prev_norm == norm and (secs is None or prev_secs is None or 0 <= secs - prev_secs <= 600):
+                gap = None if (secs is None or prev_secs is None) else secs - prev_secs
+                if prev_norm == norm and (gap is None or MIN_RESEND_GAP_SECS <= gap <= 600):
                     if _failed is not None and _has_index_between(_failed, prev_idx, idx):
                         continue
                     kinds.append("verbatim-repeat")