fix(titles): reject trivial-echo new session titles (gate-rebase of #6529)

Co-authored-by: Hermes Agent <hermes@nousresearch.com>

api/streaming.py

@@ -2839,13 +2839,19 @@ def _sanitize_generated_title(text: str) -> str:
     s = re.sub(r'^\s*title\s*:\s*', '', s, flags=re.IGNORECASE)
     s = s.strip(" \t\r\n\"'`*_~")
     s = re.sub(r'\s+', ' ', s).strip()
-    # Guard against chain-of-thought leakage and meta-reasoning patterns.
+    # Guard against chain-of-thought leakage, meta-reasoning, and trivial echo.
-    if _looks_invalid_generated_title(s):
+    if _is_bad_new_title(s):
         return ''
     return s[:80]
 
 
 def _looks_invalid_generated_title(text: str) -> bool:
+    """True when an existing/persisted title is structurally invalid (CoT leak).
+
+    Intentionally does NOT reject short conversational words like "Done" or
+    "Cool" — those can be legitimate stored titles. Use ``_is_bad_new_title``
+    when validating a freshly generated candidate.
+    """
     s = str(text or '')
     if not s.strip():
         return True
@@ -2857,7 +2863,31 @@ def _looks_invalid_generated_title(text: str) -> bool:
         or re.search(r'^\s*(i|we)\s+(should|need to|will|can)\b', s, flags=re.IGNORECASE)
         or re.search(r'^\s*let me\b', s, flags=re.IGNORECASE)
         or re.search(r"^\s*here(?:'s| is) (?:a |my )?(?:thinking|thought)", s, flags=re.IGNORECASE)
-        or re.search(r'^\s*(ok|okay|done|all set|complete|completed|finished)\b[\s.!?]*$', s, flags=re.IGNORECASE)
+    )
+
+
+def _is_bad_new_title(text: str) -> bool:
+    """True when a freshly generated title candidate should be discarded.
+
+    Includes structural CoT rejection plus trivial single-token echo replies
+    (pong, yes, done, cool, …) that are useless as first-generation titles.
+    """
+    if _looks_invalid_generated_title(text):
+        return True
+    s = str(text or '').strip()
+    _token = re.sub(r'[\s.!?]+$', '', s, flags=re.IGNORECASE)
+    if re.fullmatch(
+        r'(?:pong|ping|yes|no|yep|nope|hi|hello|hey|thanks|thank you|sure|k|kk|cool|nice|lol|ok|okay|done)',
+        _token,
+        flags=re.IGNORECASE,
+    ):
+        return True
+    return bool(
+        re.search(
+            r'^\s*(ok|okay|done|all set|complete|completed|finished)\b[\s.!?]*$',
+            s,
+            flags=re.IGNORECASE,
+        )
     )
 
 
@@ -2992,7 +3022,7 @@ def _first_exchange_snippets(messages):
             # is asking...", etc.). Assistant rows that carry tool_calls but
             # also contain a substantive answer text are kept — those are
             # agentic first-turn plans that are legitimate title candidates.
-            if m.get('tool_calls') and (not candidate or _looks_invalid_generated_title(candidate)):
+            if m.get('tool_calls') and (not candidate or _is_bad_new_title(candidate)):
                 continue
             if candidate:
                 asst_text = candidate
@@ -3016,7 +3046,7 @@ def _latest_exchange_snippets(messages):
         if role == 'assistant' and not asst_text:
             candidate = _message_text(m.get('content'))
             # Skip tool-call-only preambles
-            if m.get('tool_calls') and (not candidate or _looks_invalid_generated_title(candidate)):
+            if m.get('tool_calls') and (not candidate or _is_bad_new_title(candidate)):
                 continue
             if candidate:
                 asst_text = candidate

tests/test_issues_853_857.py

@@ -98,3 +98,54 @@ class TestThinkingPreambleStripping:
         assert not _looks_invalid_generated_title("Python import debugging"), (
             "Real titles must still pass the invalid-title guard"
         )
+
+    def test_looks_invalid_generated_title_rejects_trivial_echo(self):
+        """Trivial echo rejection applies to NEW titles only, not existing titles."""
+        from api.streaming import _is_bad_new_title, _looks_invalid_generated_title
+        assert _is_bad_new_title("pong")
+        assert _is_bad_new_title("Yes!")
+        assert _is_bad_new_title("Done")
+        assert _is_bad_new_title("Cool")
+        # Existing-title validity must accept short conversational titles so
+        # every subsequent turn does not re-fire the title LLM.
+        assert not _looks_invalid_generated_title("pong")
+        assert not _looks_invalid_generated_title("Done")
+        assert not _looks_invalid_generated_title("Cool")
+        assert not _looks_invalid_generated_title("Repo Metrics Overview")
+
+    def test_snippet_extraction_skips_tool_preamble_echo(self):
+        """First/latest exchange helpers skip tool rows whose content is echo/CoT."""
+        from api.streaming import _first_exchange_snippets, _latest_exchange_snippets
+
+        messages = [
+            {"role": "user", "content": "ping"},
+            {
+                "role": "assistant",
+                "content": "Let me check my memory first.",
+                "tool_calls": [{"id": "1", "function": {"name": "memory"}}],
+            },
+            {
+                "role": "assistant",
+                "content": "pong",
+                "tool_calls": [{"id": "2", "function": {"name": "echo"}}],
+            },
+            {"role": "assistant", "content": "Repo metrics look healthy overall."},
+        ]
+        u, a = _first_exchange_snippets(messages)
+        assert u == "ping"
+        assert "healthy" in a.lower()
+        assert "pong" not in a.lower()
+        assert "let me" not in a.lower()
+
+        latest = [
+            {"role": "user", "content": "status?"},
+            {
+                "role": "assistant",
+                "content": "Done",
+                "tool_calls": [{"id": "9", "function": {"name": "status"}}],
+            },
+            {"role": "assistant", "content": "All services green."},
+        ]
+        u2, a2 = _latest_exchange_snippets(latest)
+        assert u2 == "status?"
+        assert "green" in a2.lower()