diff --git a/routes/chat.py b/routes/chat.py
index c8252c9..df6939f 100644
--- a/routes/chat.py
+++ b/routes/chat.py
@@ -717,7 +717,12 @@ def _maybe_escalate_suggestion(user_text: str, session_id: str):
     (>= 60 chars + an action verb) pays for classification. Returns the
     suggestion dict for the UI chip, or None."""
     try:
-        if len(user_text or "") < 60 or not _ESCALATE_HINT_RE.search(user_text):
+        # 2026-07 fix: the floor was 60, which silently dropped legitimate complex
+        # asks like "plan and build me a 5-page competitor report with charts"
+        # (56 chars) — the offer never fired, the prompt just degenerated in chat.
+        # 24 still filters trivial single-verb messages; the action-verb regex +
+        # the Qwen classifier are the real gate.
+        if len(user_text or "") < 24 or not _ESCALATE_HINT_RE.search(user_text):
             return None
         verdict = _should_escalate_to_project(user_text, session_id)
         if not verdict.get("escalate"):
@@ -1087,9 +1092,9 @@ def _resolve_skill_tag(raw_tag):
                             )
                         else:
                             yield _frame(
-                                "⚠️ The model returned an empty reply — it may be "
+                                "*The model returned an empty reply — it may be "
                                 "busy, restarting, or out of context. Please try "
-                                "again in a moment."
+                                "again in a moment.*"
                             )
                     # Step 10 Q11 (2026-07): post-reply Project suggestion.
                     _sugg = _maybe_escalate_suggestion(last_user_text, _session_id)
diff --git a/tests/test_chat_escalation.py b/tests/test_chat_escalation.py
index 384689c..ed72279 100644
--- a/tests/test_chat_escalation.py
+++ b/tests/test_chat_escalation.py
@@ -11,6 +11,8 @@
 """
 from __future__ import annotations
 
+import pytest
+
 import json
 import sys
 from pathlib import Path
@@ -131,3 +133,35 @@ def test_kill_switch_disables_all_escalation(monkeypatch):
 
     decision = cd._should_escalate_to_project(user_text="x", session_id="s5")
     assert decision["escalate"] is False
+
+
+# ── 2026-07: the length floor dropped legit 56-char asks; lowered 60 -> 24 ─────
+
+def test_maybe_escalate_fires_for_short_but_complex_ask(monkeypatch):
+    """'plan and build me a 5-page competitor report with charts' is 56 chars —
+    it used to be silently dropped by the len<60 gate and just degenerated in
+    chat. It must now reach the classifier and produce an offer."""
+    import routes.chat as rc
+    monkeypatch.setattr(rc, "_should_escalate_to_project",
+                        lambda text, sid: {"escalate": True, "estimated_checkpoints": 5,
+                                           "reason": "multi-step report", "is_project": True})
+    out = rc._maybe_escalate_suggestion(
+        "plan and build me a 5-page competitor report with charts", "sess-1")
+    assert out is not None, "offer must fire for a complex 56-char ask"
+    assert out["estimated_checkpoints"] == 5
+
+
+def test_maybe_escalate_skips_trivial_messages(monkeypatch):
+    import routes.chat as rc
+    monkeypatch.setattr(rc, "_should_escalate_to_project",
+                        lambda text, sid: pytest.fail("classifier must not run on trivial text"))
+    assert rc._maybe_escalate_suggestion("build", "s") is None            # too short
+    assert rc._maybe_escalate_suggestion("what's the weather today?", "s") is None  # no action verb
+
+
+def test_maybe_escalate_respects_classifier_no(monkeypatch):
+    import routes.chat as rc
+    monkeypatch.setattr(rc, "_should_escalate_to_project",
+                        lambda text, sid: {"escalate": False, "estimated_checkpoints": 1})
+    assert rc._maybe_escalate_suggestion(
+        "research the best pizza place near me right now", "s") is None