The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
fix(llm): retry Ollama capability lookups that failed
FenjuFu7556360412b1
diff --git a/src/llm/providers/ollama.py b/src/llm/providers/ollama.pyindex 65256b179..47f5f1cc6 100644--- a/src/llm/providers/ollama.py+++ b/src/llm/providers/ollama.py@@ -92,26 +92,29 @@ class OllamaProvider(LLMProvider):Catalogued models answer from :attr:`ModelInfo.supports_tools`. Any othermodel is looked up once through ``/api/show``, whose ``capabilities``- list is what the installed model itself declares. A failed lookup counts- as no tool support.+ list is what the installed model itself declares. Only an answered lookup+ is cached: a failed one counts as no tool support for that request and+ is retried on the next."""for info in self._models:if info.name == model:return info.supports_tools- if model not in self._tool_support:- try:- req = urllib.request.Request(- f"{self.base_url}/api/show",- data=json.dumps({"model": model}).encode("utf-8"),- headers={"Content-Type": "application/json"},- )- with urllib.request.urlopen(req, timeout=10) as response:- capabilities = json.loads(response.read().decode("utf-8")).get("capabilities") or []- self._tool_support[model] = "tools" in capabilities- except Exception as e:- logger.warning("Could not read capabilities for Ollama model '%s': %s", model, type(e).__name__)- self._tool_support[model] = False- return self._tool_support[model]+ if model in self._tool_support:+ return self._tool_support[model]+ try:+ req = urllib.request.Request(+ f"{self.base_url}/api/show",+ data=json.dumps({"model": model}).encode("utf-8"),+ headers={"Content-Type": "application/json"},+ )+ with urllib.request.urlopen(req, timeout=10) as response:+ capabilities = json.loads(response.read().decode("utf-8")).get("capabilities") or []+ except Exception as e:+ logger.warning("Could not read capabilities for Ollama model '%s': %s", model, type(e).__name__)+ return False+ supports_tools = "tools" in capabilities+ self._tool_support = {**self._tool_support, model: supports_tools}+ return supports_toolsdef generate(self, input: LLMInput) -> LLMOutput:try:diff --git a/tests/test_provider_tools.py b/tests/test_provider_tools.pyindex caa0290c4..accdef762 100644--- a/tests/test_provider_tools.py+++ b/tests/test_provider_tools.py@@ -3,6 +3,7 @@ import jsonimport urllib.requestfrom io import BytesIOfrom types import SimpleNamespace+from typing import Anyimport pytest@@ -173,14 +174,19 @@ def test_ollama_provider_serializes_generation_options(class _FakeOllama:"""Serves /api/show and replays queued /api/chat replies, recording requests."""- def __init__(self, monkeypatch, replies, capabilities=None):+ def __init__(+ self,+ monkeypatch: pytest.MonkeyPatch,+ replies: list[dict[str, Any]],+ capabilities: list[str] | None = None,+ ) -> None:self.replies = list(replies)self.capabilities = capabilities- self.chats = []- self.shows = []+ self.chats: list[dict[str, Any]] = []+ self.shows: list[str] = []monkeypatch.setattr(urllib.request, "urlopen", self.urlopen)- def urlopen(self, request, timeout):+ def urlopen(self, request: urllib.request.Request, timeout: float) -> BytesIO:body = json.loads(request.data)if request.full_url.endswith("/api/show"):self.shows.append(body["model"])@@ -191,7 +197,7 @@ class _FakeOllama:return BytesIO(json.dumps(self.replies.pop(0)).encode())-def _ollama_tool_calls(*calls):+def _ollama_tool_calls(*calls: tuple[str, dict[str, Any]]) -> dict[str, Any]:return {"message": {"content": "",@@ -201,7 +207,7 @@ def _ollama_tool_calls(*calls):}-def _ollama_answer(text):+def _ollama_answer(text: str) -> dict[str, Any]:return {"message": {"content": text}, "done_reason": "stop"}@@ -213,7 +219,7 @@ def _lookup_tool() -> ToolDefinition:)-def test_ollama_provider_serializes_tools_and_parses_tool_calls(monkeypatch):+def test_ollama_provider_serializes_tools_and_parses_tool_calls(monkeypatch: pytest.MonkeyPatch) -> None:server = _FakeOllama(monkeypatch, [_ollama_tool_calls(("search", {"query": "ollama"}))], capabilities=["completion", "tools"])@@ -246,12 +252,16 @@ def test_ollama_provider_serializes_tools_and_parses_tool_calls(monkeypatch):("codellama", None, False, []),("qwen3", ["completion", "tools"], True, ["qwen3"]),("gemma3", ["completion", "vision"], False, ["gemma3"]),- ("unknown", None, False, ["unknown"]),+ ("unknown", None, False, ["unknown", "unknown"]),],)def test_ollama_provider_sends_tools_only_to_models_that_declare_them(- monkeypatch, model, capabilities, expect_tools, expect_shows-):+ monkeypatch: pytest.MonkeyPatch,+ model: str,+ capabilities: list[str] | None,+ expect_tools: bool,+ expect_shows: list[str],+) -> None:server = _FakeOllama(monkeypatch, [_ollama_answer("ok"), _ollama_answer("ok")], capabilities=capabilities)provider = OllamaProvider(default_model=model)request = LLMInput(messages=[Message(role=Role.USER, content="hi")], tools=[_tool()])@@ -260,11 +270,25 @@ def test_ollama_provider_sends_tools_only_to_models_that_declare_them(provider.generate(request)assert all(("tools" in chat) is expect_tools for chat in server.chats)- # The capability lookup is cached per model.+ # An answered capability lookup is cached per model; a failed one is retried.assert server.shows == expect_shows-def test_ollama_react_agent_correlates_multiple_calls_without_ids(monkeypatch):+def test_ollama_provider_retries_capability_lookup_after_a_failure(monkeypatch: pytest.MonkeyPatch) -> None:+ server = _FakeOllama(monkeypatch, [_ollama_answer("ok"), _ollama_answer("ok"), _ollama_answer("ok")])+ provider = OllamaProvider(default_model="qwen3")+ request = LLMInput(messages=[Message(role=Role.USER, content="hi")], tools=[_tool()])++ provider.generate(request)+ server.capabilities = ["completion", "tools"]+ provider.generate(request)+ provider.generate(request)++ assert ["tools" in chat for chat in server.chats] == [False, True, True]+ assert server.shows == ["qwen3", "qwen3"]+++def test_ollama_react_agent_correlates_multiple_calls_without_ids(monkeypatch: pytest.MonkeyPatch) -> None:server = _FakeOllama(monkeypatch,[@@ -298,7 +322,7 @@ def test_ollama_react_agent_correlates_multiple_calls_without_ids(monkeypatch):assert second == {"role": "tool", "content": "query=weather", "tool_call_id": ids[1], "tool_name": "search"}-def test_ollama_react_agent_reports_unknown_tool_under_its_own_name(monkeypatch):+def test_ollama_react_agent_reports_unknown_tool_under_its_own_name(monkeypatch: pytest.MonkeyPatch) -> None:server = _FakeOllama(monkeypatch,[_ollama_tool_calls(("missing", {}), ("search", {"query": "x"})), _ollama_answer("done")],@@ -317,7 +341,7 @@ def test_ollama_react_agent_reports_unknown_tool_under_its_own_name(monkeypatch)]-def test_ollama_react_agent_without_tool_support_answers_directly(monkeypatch):+def test_ollama_react_agent_without_tool_support_answers_directly(monkeypatch: pytest.MonkeyPatch) -> None:server = _FakeOllama(monkeypatch, [_ollama_answer("I cannot call tools.")])agent = ReActAgent(OllamaProvider(default_model="codellama"), ToolExecutor())