Fix Ollama agent single-token responses (#1591)

Agent mode treated local /v1 endpoints, including Ollama on :11434, as native-tool-capable by host/model heuristics. On Ollama's OpenAI-compatible surface some models that advertise tool support stop after a single token when schemas are sent (issue #1567). Default local Ollama /v1 back to fenced tool blocks unless the endpoint explicitly has supports_tools=True.

Also compare both the runtime chat URL and the normalized endpoint base when reading ModelEndpoint.supports_tools. That keeps a saved base URL such as http://localhost:11434/v1 effective when the active session URL is /v1/chat/completions.

Tests: .venv/bin/python -m pytest tests/test_tool_support_heuristic.py

src/agent_loop.py

@@ -13,6 +13,7 @@ import re
 import time
 import logging
 from typing import AsyncGenerator, List, Dict, Optional, Set
+from urllib.parse import urlparse
 
 from src.llm_core import stream_llm, stream_llm_with_fallback, _is_ollama_native_url
 from src.model_context import estimate_tokens
@@ -475,6 +476,45 @@ _ADMIN_SCHEMA_NAMES = frozenset([
 ])
 _TOOL_SELECTION_TIMEOUT_SECONDS = 1.5
 
+
+def _is_ollama_openai_compat_url(endpoint_url: str) -> bool:
+    """Return True for local Ollama's OpenAI-compatible /v1 surface.
+
+    Ollama's /v1 endpoint accepts the OpenAI chat shape, but model-level tool
+    streaming is uneven. Some local models terminate after a token when schemas
+    are present. Keep native schemas opt-in via ModelEndpoint.supports_tools.
+    """
+    try:
+        parsed = urlparse(endpoint_url or "")
+    except Exception:
+        return False
+    path = (parsed.path or "").rstrip("/")
+    return parsed.port == 11434 and (path == "/v1" or path.startswith("/v1/"))
+
+
+def _endpoint_lookup_keys(endpoint_url: str) -> List[str]:
+    """Candidate ModelEndpoint.base_url keys for a runtime chat URL."""
+    raw = (endpoint_url or "").strip()
+    keys: List[str] = []
+
+    def add(value: str):
+        value = (value or "").strip()
+        if value and value not in keys:
+            keys.append(value)
+        trimmed = value.rstrip("/")
+        if trimmed and trimmed not in keys:
+            keys.append(trimmed)
+        if trimmed and f"{trimmed}/" not in keys:
+            keys.append(f"{trimmed}/")
+
+    add(raw)
+    try:
+        from src.endpoint_resolver import normalize_base
+        add(normalize_base(raw))
+    except Exception:
+        pass
+    return keys
+
 # Admin tool keywords — if the last user message contains any of these, include admin tools
 _ADMIN_KEYWORDS = [
     "session", "sessions", "chat", "chats", "conversation", "conversations",
@@ -1456,18 +1496,18 @@ async def stream_agent_loop(
     _model_lc = (model or "").lower()
     # Step 1: per-endpoint override (set at registration time from the
     # serve command — `--enable-auto-tool-choice` flips it on. UI can
-    # also toggle per endpoint). NULL = unknown, fall through to the
-    # keyword heuristic + host check.
+    # also toggle per endpoint). NULL = unknown; for local Ollama /v1 we
+    # default to fenced tools, otherwise fall through to keyword + host checks.
     _endpoint_supports: Optional[bool] = None
     try:
         from core.database import SessionLocal as _SL, ModelEndpoint as _ME
         _db = _SL()
         try:
-            _ep = _db.query(_ME).filter(_ME.base_url == endpoint_url).first()
-            if not _ep and endpoint_url:
-                _u = endpoint_url.rstrip("/")
-                _ep = _db.query(_ME).filter(_ME.base_url == _u).first() or \
-                      _db.query(_ME).filter(_ME.base_url == _u + "/").first()
+            _ep = None
+            for _key in _endpoint_lookup_keys(endpoint_url):
+                _ep = _db.query(_ME).filter(_ME.base_url == _key).first()
+                if _ep is not None:
+                    break
             if _ep is not None:
                 _endpoint_supports = _ep.supports_tools
         finally:
@@ -1503,9 +1543,15 @@ async def stream_agent_loop(
     # (via the endpoint settings toggle), treat Ollama-native as text-only so
     # the fenced-block path is used instead of native function calling.
     _is_ollama_native = _is_ollama_native_url(endpoint_url or "")
+    _ollama_openai_compat = _is_ollama_openai_compat_url(endpoint_url or "")
     if _endpoint_supports is True:
         _is_api_model = True
-    elif _endpoint_supports is False or _model_no_tools or _is_ollama_native:
+    elif (
+        _endpoint_supports is False
+        or _model_no_tools
+        or _is_ollama_native
+        or _ollama_openai_compat
+    ):
         _is_api_model = False
     else:
         _is_api_model = any(h in endpoint_url for h in _API_HOSTS) or _model_supports_tools