Fix native tool-calling follow-up round on Gemini and Ollama (#867)

The agent's multi-round (tool-result) follow-up request was rejected with
HTTP 400 on two providers, so tools ran but the agent never produced an answer:

- OpenAI-compatible streaming (Gemini 3) dropped the per-call thought_signature
  and collided parallel tool calls, which arrive with index=None: they all
  landed in slot 0, overwriting the first call's name and corrupting its
  arguments by concatenation, so the follow-up request 400'd. Capture and replay
  each call's extra_content (thought_signature), and give every parallel call
  its own accumulator slot (allocated above the max key, so sparse or mixed
  indices can't collide).
- Native Ollama /api/chat expects object tool-call arguments, but Odysseus
  carries them as a JSON string, which Ollama rejected ("Value looks like
  object, but can't find closing '}' symbol"). Convert them to objects in the
  Ollama payload builder.

Both compose with the no-prose null-content sanitize fix from #862.

Tested: python -m pytest tests/test_llm_core_streaming.py
tests/test_llm_core_ollama.py tests/test_agent_loop.py (53 pass), and
python -m py_compile src/llm_core.py src/agent_loop.py.

tests/test_llm_core_ollama.py

@@ -41,3 +41,75 @@ def test_llm_call_posts_native_ollama_payload(monkeypatch):
     assert seen["headers"]["Authorization"] == "Bearer ollama-key"
     assert seen["json"]["stream"] is False
     assert seen["json"]["options"] == {"temperature": 0.2, "num_predict": 7}
+
+
+# ---------------------------------------------------------------------------
+# Tool-call argument serialization for native Ollama
+#
+# Odysseus carries assistant tool calls in the OpenAI shape, where
+# `function.arguments` is a JSON *string*. Native Ollama /api/chat expects a
+# JSON *object* and rejects the string form with HTTP 400 ("Value looks like
+# object, but can't find closing '}' symbol"), aborting every follow-up
+# (tool-result) round. _build_ollama_payload must parse it back to an object.
+# ---------------------------------------------------------------------------
+
+def _assistant_tool_call_msgs():
+    """A canonical OpenAI-style assistant tool call + tool result, as produced by
+    agent_loop._append_tool_results (arguments are a JSON string)."""
+    return [
+        {"role": "user", "content": "what do you know about me?"},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call_0",
+                    "type": "function",
+                    "function": {"name": "app_api", "arguments": '{"action": "get_memory"}'},
+                }
+            ],
+        },
+        {"role": "tool", "tool_call_id": "call_0", "content": "Memory: user is James."},
+    ]
+
+
+def test_ollama_payload_parses_string_arguments_to_object():
+    payload = llm_core._build_ollama_payload(
+        "gpt-oss:120b", _assistant_tool_call_msgs(), temperature=0.0, max_tokens=0,
+    )
+    asst = payload["messages"][1]
+    args = asst["tool_calls"][0]["function"]["arguments"]
+    # The whole point: arguments must be a dict, not the JSON string.
+    assert args == {"action": "get_memory"}
+    assert not isinstance(args, str)
+    assert asst["tool_calls"][0]["function"]["name"] == "app_api"
+    assert asst["tool_calls"][0]["id"] == "call_0"
+
+
+def test_ollama_payload_drops_gemini_thought_signature():
+    """A cross-provider fallback can hand Ollama a tool call that still carries
+    Gemini's opaque extra_content; it is meaningless to Ollama and must not leak."""
+    msgs = _assistant_tool_call_msgs()
+    msgs[1]["tool_calls"][0]["extra_content"] = {"google": {"thought_signature": "AAAA"}}
+    payload = llm_core._build_ollama_payload(
+        "gpt-oss:120b", msgs, temperature=0.0, max_tokens=0,
+    )
+    tc = payload["messages"][1]["tool_calls"][0]
+    assert "extra_content" not in tc
+    assert tc["function"]["arguments"] == {"action": "get_memory"}
+
+
+def test_ollama_payload_leaves_plain_messages_untouched():
+    msgs = [{"role": "user", "content": "hello"}]
+    payload = llm_core._build_ollama_payload("m", msgs, temperature=0.0, max_tokens=0)
+    assert payload["messages"][0] == {"role": "user", "content": "hello"}
+
+
+def test_ollama_payload_tolerates_malformed_arguments():
+    msgs = [{
+        "role": "assistant",
+        "tool_calls": [{"function": {"name": "x", "arguments": "{not json"}}],
+    }]
+    payload = llm_core._build_ollama_payload("m", msgs, temperature=0.0, max_tokens=0)
+    # Falls back to an empty object rather than raising.
+    assert payload["messages"][0]["tool_calls"][0]["function"]["arguments"] == {}