diff --git a/core.py b/core.py
index a4cfdbe..b5cfeb5 100644
--- a/core.py
+++ b/core.py
@@ -421,6 +421,14 @@ def _expand_tools_for_call(current_tools: list[dict], called_names: list[str]) -
 # Telegram bot token: digits:alphanumeric (e.g. 1234567890:ABCdefGHI...)
 _TG_TOKEN_RE = _re.compile(r'\b(\d{8,12}:[A-Za-z0-9_\-]{30,50})\b')
 
+# Re-prompt when a provider reports finish_reason="tool_calls" but streams no
+# tool call at all — a narration-only stall, or an interrupt mid-retry. Taking
+# the narration as the answer would end the turn with the task unstarted.
+_DROPPED_TOOLCALL_NUDGE = (
+    "Your previous turn indicated a tool call but none was included. Do not narrate a plan or "
+    "restate intent — issue the actual tool call now to continue the task."
+)
+
 
 def _select_tools(user_input: str, messages: list[dict] = None, router_groups: set[str] | None = None) -> list[dict]:
     """
@@ -889,6 +897,7 @@ def _run_conversation_loop(self, user_input: str, image_path: str | None = None)
             empty_retries = 0
             length_continuation_count = 0
             invalid_json_retries = 0
+            _dropped_toolcall_retries = 0
             for _round in range(MAX_ROUNDS):
                 if self._cancel.is_set():
                     yield {"type": "interrupted"}
@@ -1109,6 +1118,37 @@ def _run_conversation_loop(self, user_input: str, image_path: str | None = None)
 
                 # ── No tool calls → pure text response, done ─────────────────────
                 if not _tool_buf:
+                    # finish_reason="tool_calls" with nothing buffered means the
+                    # provider promised a call and streamed none: the model
+                    # narrated a plan instead of issuing it, or an interrupt cut
+                    # the retry. Delivering `full` as the answer ends the turn
+                    # with the task unstarted, so re-prompt for the actual call —
+                    # bounded to 3 consecutive stalls, after which the narration
+                    # is delivered rather than looping forever.
+                    if _finish_reason == "tool_calls" and _dropped_toolcall_retries < 3:
+                        _dropped_toolcall_retries += 1
+                        import logging
+                        logging.getLogger(__name__).warning(
+                            "finish_reason=tool_calls with no buffered call (narration only) — "
+                            f"re-prompting to emit the call ({_dropped_toolcall_retries}/3)."
+                        )
+                        # Both halves are retry scaffolding, stripped from the
+                        # final history by the loop's cleanup.
+                        self.messages.append({
+                            "role": "assistant",
+                            "content": full,
+                            "_dropped_toolcall_nudge": True,
+                        })
+                        self.messages.append({
+                            "role": "user",
+                            "content": _DROPPED_TOOLCALL_NUDGE,
+                            "_dropped_toolcall_nudge": True,
+                        })
+                        if self.provider.supports_thinking:
+                            yield {"type": "thinking"}
+                        continue
+
+                    _dropped_toolcall_retries = 0
                     self.messages.append({"role": "assistant", "content": full})
                     return
 
@@ -1206,6 +1246,9 @@ def _run_conversation_loop(self, user_input: str, image_path: str | None = None)
                     continue
 
                 invalid_json_retries = 0
+                # A batch that reached execution is real progress: the stall
+                # budget counts *consecutive* narration-only rounds, so reset it.
+                _dropped_toolcall_retries = 0
 
                 # ── Execute each tool call ────────────────────────────────────────
                 permanent_failure = False
@@ -1439,7 +1482,10 @@ def run_one(call):
                 except RuntimeError:
                     pass  # already released
             # Clean up empty recovery synthetic messages from self.messages
-            self.messages = [m for m in self.messages if not m.get("_empty_recovery_synthetic")]
+            self.messages = [
+                m for m in self.messages
+                if not m.get("_empty_recovery_synthetic") and not m.get("_dropped_toolcall_nudge")
+            ]
 
     def _refresh_memory_context(self, user_input: str, prompt_sections: set[str] | None = None) -> None:
         """
diff --git a/tests/test_dropped_tool_calls.py b/tests/test_dropped_tool_calls.py
new file mode 100644
index 0000000..238710a
--- /dev/null
+++ b/tests/test_dropped_tool_calls.py
@@ -0,0 +1,170 @@
+"""finish_reason="tool_calls" with no buffered call must not deliver the plan text.
+
+A provider can report ``finish_reason="tool_calls"`` and stream no tool-call
+chunk at all: the model narrated a plan ("I'll read the file now…") instead of
+issuing the call, or an interrupt cut a retry short. Koza used to append that
+narration as the final assistant message and end the turn, so the task silently
+stopped half-done. It now re-prompts for the real call, bounded to 3
+*consecutive* stalls; a round that actually executes a tool clears the budget.
+"""
+import os
+import sys
+import threading
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from core import _DROPPED_TOOLCALL_NUDGE, Agent, ToolLoopGuardrail
+
+
+class _FakeProvider:
+    name = "fake"
+    supports_vision = False
+    supports_thinking = False
+    _model = "fake-model"
+
+    def __init__(self, rounds):
+        self._rounds = list(rounds)
+        self.rounds_started = 0
+
+    def stream_chat(self, messages, tools=None, cancel_event=None):
+        self.rounds_started += 1
+        if not self._rounds:
+            return
+        yield from self._rounds.pop(0)
+
+
+def _tool_chunk(index, name, args_chunk, call_id=None):
+    return {
+        "__tool_chunk__": True,
+        "index": index,
+        "id": call_id or f"{name}-{index}",
+        "name": name,
+        "args_chunk": args_chunk,
+    }
+
+
+def _finish(reason="tool_calls"):
+    return {"__finish_reason__": reason}
+
+
+def _make_agent(rounds):
+    """Agent with only the attributes the conversation loop actually reads."""
+    agent = Agent.__new__(Agent)
+    agent.provider = _FakeProvider(rounds)
+    agent.messages = [{"role": "system", "content": "test"}]
+    agent._cancel = threading.Event()
+    agent._busy = False
+    agent._router = None
+    agent._guardrail = ToolLoopGuardrail()
+    agent.permission_callback = None
+    agent.tool_progress_callback = None
+    agent._pre_fetch_links = lambda user_input: iter(())
+    agent._refresh_memory_context = lambda *a, **k: None
+    agent._resolve_available_tools = lambda *a, **k: []
+    # Echo the path back so each result is attributable to its own call.
+    agent._execute_tool = lambda name, args: f"read {args.get('path')}"
+    return agent
+
+
+def _events(agent, text="do something"):
+    return list(agent._run_conversation_loop(text))
+
+
+def _tool_results(agent):
+    return [m for m in agent.messages if m.get("role") == "tool"]
+
+
+def _last_assistant(agent):
+    return [m for m in agent.messages if m.get("role") == "assistant"][-1]
+
+
+# ── the stall is re-prompted, not delivered ──────────────────────────────────
+
+
+def test_narration_only_stall_is_reprompted_until_the_call_arrives():
+    agent = _make_agent(
+        [
+            ["I will read a.py now.", _finish("tool_calls")],  # narration only
+            [_tool_chunk(0, "read_file", '{"path": "a.py"}'), _finish("tool_calls")],
+            ["Read it.", _finish("stop")],
+        ]
+    )
+
+    _events(agent)
+
+    assert agent.provider.rounds_started == 3, (
+        "a narration-only finish_reason=tool_calls must re-prompt, not end the turn"
+    )
+    assert [m["content"] for m in _tool_results(agent)] == ["read a.py"], (
+        "the re-prompted round must actually execute the call"
+    )
+    assert _last_assistant(agent)["content"] == "Read it."
+
+
+def test_a_model_that_only_ever_narrates_still_terminates():
+    agent = _make_agent([["thinking out loud", _finish("tool_calls")]] * 6)
+
+    _events(agent)
+
+    assert agent.provider.rounds_started == 4, (
+        "1 round + at most 3 re-prompts, then the narration is delivered"
+    )
+    assert _last_assistant(agent)["content"] == "thinking out loud"
+
+
+def test_a_real_tool_batch_resets_the_stall_budget():
+    agent = _make_agent(
+        [
+            ["narrating 1", _finish("tool_calls")],  # stall 1
+            ["narrating 2", _finish("tool_calls")],  # stall 2
+            [_tool_chunk(0, "read_file", '{"path": "a.py"}'), _finish("tool_calls")],  # progress
+            ["narrating 3", _finish("tool_calls")],  # stall 1 again
+            ["narrating 4", _finish("tool_calls")],  # stall 2
+            ["narrating 5", _finish("tool_calls")],  # stall 3
+            ["narrating 6", _finish("tool_calls")],  # budget spent → deliver
+        ]
+    )
+
+    _events(agent)
+
+    assert agent.provider.rounds_started == 7, (
+        "a round that really executed a tool must clear the consecutive-stall budget"
+    )
+
+
+# ── regressions: normal turns are untouched ─────────────────────────────────
+
+
+def test_plain_stop_response_is_returned_without_a_re_prompt():
+    agent = _make_agent([["the answer", _finish("stop")]])
+
+    _events(agent)
+
+    assert agent.provider.rounds_started == 1
+    assert _last_assistant(agent)["content"] == "the answer"
+
+
+def test_text_response_without_a_finish_reason_is_returned_immediately():
+    # gemini/anthropic only emit __finish_reason__ on truncation, so a normal
+    # completion leaves it None — that must never be mistaken for a stall.
+    agent = _make_agent([["just text"]])
+
+    _events(agent)
+
+    assert agent.provider.rounds_started == 1
+    assert _last_assistant(agent)["content"] == "just text"
+
+
+def test_retry_scaffolding_is_not_left_in_history():
+    agent = _make_agent(
+        [
+            ["narrating", _finish("tool_calls")],
+            ["done", _finish("stop")],
+        ]
+    )
+
+    _events(agent)
+
+    assert all(m.get("content") != _DROPPED_TOOLCALL_NUDGE for m in agent.messages)
+    assert all(not m.get("_dropped_toolcall_nudge") for m in agent.messages)
+    assert [m["content"] for m in agent.messages] == ["test", "do something", "done"]