diff --git a/pyproject.toml b/pyproject.toml
index 58af9ce..088021d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "filthy-tool-fixer"
-version = "0.3.0"
+version = "0.3.1"
 requires-python = ">=3.11"
 dependencies = [
     "fastapi>=0.109.0",
diff --git a/src/filthy_tool_fixer/backends/ollama.py b/src/filthy_tool_fixer/backends/ollama.py
index 727757d..4c7b1bd 100644
--- a/src/filthy_tool_fixer/backends/ollama.py
+++ b/src/filthy_tool_fixer/backends/ollama.py
@@ -90,11 +90,17 @@ async def chat_completion_stream(
             json=payload,
             timeout=httpx.Timeout(effective_timeout, connect=10.0),
         ) as resp:
-            resp.raise_for_status()
+            try:
+                resp.raise_for_status()
+            except httpx.HTTPStatusError as e:
+                error = {"error": {"message": f"Backend returned {e.response.status_code}", "type": "backend_error"}}
+                yield f"data: {json.dumps(error)}\n\n".encode()
+                yield b"data: [DONE]\n\n"
+                return
             try:
                 async for line in resp.aiter_lines():
                     if line:
-                        yield (line + "\n").encode()
+                        yield (line + "\n\n").encode()
             except asyncio.CancelledError:
                 log.info("stream_cancelled_by_client")
                 raise
diff --git a/src/filthy_tool_fixer/backends/vllm.py b/src/filthy_tool_fixer/backends/vllm.py
index e1121a2..e280923 100644
--- a/src/filthy_tool_fixer/backends/vllm.py
+++ b/src/filthy_tool_fixer/backends/vllm.py
@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import asyncio
 import json
 from typing import Any, AsyncIterator
 
@@ -28,12 +29,12 @@ def build_tool_call_schema(tools: list[ToolDefinition]) -> dict[str, Any]:
     tool_schemas = []
     for tool in tools:
         func = tool.function
-        # Each tool call has name + arguments matching the tool's parameter schema
+        # Each tool call has name + arguments as a JSON string
         tool_schema: dict[str, Any] = {
             "type": "object",
             "properties": {
                 "name": {"type": "string", "const": func.name},
-                "arguments": func.parameters if func.parameters else {"type": "object"},
+                "arguments": {"type": "string"},
             },
             "required": ["name", "arguments"],
             "additionalProperties": False,
@@ -87,7 +88,8 @@ async def chat_completion(
         timeout: float | None = None,
         keep_alive: str | None = None,
     ) -> ChatCompletionResponse:
-        assert self._client is not None
+        if self._client is None:
+            raise RuntimeError("VLLMAdapter not started — call startup() first")
         payload = self._build_payload(request, keep_alive)
         payload["stream"] = False
 
@@ -106,7 +108,8 @@ async def chat_completion_stream(
         timeout: float | None = None,
         keep_alive: str | None = None,
     ) -> AsyncIterator[bytes]:
-        assert self._client is not None
+        if self._client is None:
+            raise RuntimeError("VLLMAdapter not started — call startup() first")
         payload = self._build_payload(request, keep_alive)
         payload["stream"] = True
 
@@ -117,14 +120,25 @@ async def chat_completion_stream(
             json=payload,
             timeout=httpx.Timeout(effective_timeout, connect=10.0),
         ) as resp:
-            resp.raise_for_status()
-            async for line in resp.aiter_lines():
-                if line:
-                    yield (line + "\n").encode()
+            try:
+                resp.raise_for_status()
+            except httpx.HTTPStatusError as e:
+                error = {"error": {"message": f"Backend returned {e.response.status_code}", "type": "backend_error"}}
+                yield f"data: {json.dumps(error)}\n\n".encode()
+                yield b"data: [DONE]\n\n"
+                return
+            try:
+                async for line in resp.aiter_lines():
+                    if line:
+                        yield (line + "\n\n").encode()
+            except asyncio.CancelledError:
+                log.info("stream_cancelled_by_client")
+                raise
 
     async def health_check(self) -> bool:
         try:
-            assert self._client is not None
+            if self._client is None:
+                return False
             resp = await self._client.get("/health", timeout=5.0)
             return resp.status_code == 200
         except Exception:
diff --git a/src/filthy_tool_fixer/main.py b/src/filthy_tool_fixer/main.py
index 46af30d..9e423b1 100644
--- a/src/filthy_tool_fixer/main.py
+++ b/src/filthy_tool_fixer/main.py
@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import asyncio
+import json
 import time
 from contextlib import asynccontextmanager
 from typing import AsyncIterator
@@ -24,20 +25,43 @@
 
 log = get_logger(__name__)
 
+
+class _RequestCounter:
+    """In-flight request counter with drain support for graceful shutdown.
+
+    All methods are synchronous — safe under asyncio's cooperative model
+    since no await can interleave between the read-modify-write steps.
+    """
+
+    def __init__(self) -> None:
+        self.count = 0
+        self.drain_event = asyncio.Event()
+        self.drain_event.set()  # Initially drained
+
+    def enter(self) -> None:
+        self.count += 1
+        self.drain_event.clear()
+
+    def exit(self) -> None:
+        self.count -= 1
+        if self.count == 0:
+            self.drain_event.set()
+
+
 # Module-level references set during lifespan
 _orchestrator: ProxyOrchestrator | None = None
 _profile_loader: ProfileLoader | None = None
 _primary_backend: OllamaAdapter | None = None
 _escalation_backend: OllamaAdapter | None = None
-_in_flight: int = 0
-_drain_event: asyncio.Event | None = None
+_requests = _RequestCounter()
 
 _SHUTDOWN_DRAIN_TIMEOUT = 15.0  # Max seconds to wait for in-flight requests
+_STREAM_CHUNK_TIMEOUT = 60.0  # Max seconds to wait between stream chunks
 
 
 @asynccontextmanager
 async def lifespan(app: FastAPI) -> AsyncIterator[None]:
-    global _orchestrator, _profile_loader, _primary_backend, _escalation_backend, _drain_event
+    global _orchestrator, _profile_loader, _primary_backend, _escalation_backend
 
     setup_logging()
     log.info(
@@ -46,24 +70,27 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
         escalation_url=settings.escalation_backend_url,
     )
 
-    _drain_event = asyncio.Event()
-    _drain_event.set()  # Initially "drained" (no requests)
-
     # Load profiles
     _profile_loader = ProfileLoader(settings.profiles_dir)
 
-    # Initialize backends
+    # Initialize backends (tolerate startup failures for graceful degradation)
     _primary_backend = OllamaAdapter(
         base_url=settings.backend_url,
         default_timeout=settings.backend_timeout,
     )
-    await _primary_backend.startup()
+    try:
+        await _primary_backend.startup()
+    except Exception:
+        log.exception("primary_backend_startup_failed")
 
     _escalation_backend = OllamaAdapter(
         base_url=settings.escalation_backend_url,
         default_timeout=settings.escalation_timeout,
     )
-    await _escalation_backend.startup()
+    try:
+        await _escalation_backend.startup()
+    except Exception:
+        log.exception("escalation_backend_startup_failed")
 
     _orchestrator = ProxyOrchestrator(
         primary_backend=_primary_backend,
@@ -76,15 +103,24 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     yield
 
     # Wait for in-flight requests to drain before closing backends
-    if _in_flight > 0:
-        log.info("shutdown_draining", in_flight=_in_flight)
+    if _requests.count > 0:
+        log.info("shutdown_draining", in_flight=_requests.count)
+        try:
+            await asyncio.wait_for(
+                _requests.drain_event.wait(), timeout=_SHUTDOWN_DRAIN_TIMEOUT
+            )
+        except asyncio.TimeoutError:
+            log.warning("shutdown_drain_timeout", in_flight=_requests.count)
+
+    # Shutdown backends with timeout to prevent hanging
+    for backend, name in [(_primary_backend, "primary"), (_escalation_backend, "escalation")]:
         try:
-            await asyncio.wait_for(_drain_event.wait(), timeout=_SHUTDOWN_DRAIN_TIMEOUT)
+            await asyncio.wait_for(backend.shutdown(), timeout=5.0)
         except asyncio.TimeoutError:
-            log.warning("shutdown_drain_timeout", in_flight=_in_flight)
+            log.warning(f"{name}_backend_shutdown_timeout")
+        except Exception:
+            log.exception(f"{name}_backend_shutdown_error")
 
-    await _primary_backend.shutdown()
-    await _escalation_backend.shutdown()
     log.info("stopped")
 
 
@@ -135,8 +171,6 @@ async def list_models() -> JSONResponse:
 
 @app.post("/v1/chat/completions", response_model=None)
 async def chat_completions(request: Request):
-    global _in_flight
-
     if _orchestrator is None or _profile_loader is None:
         return JSONResponse(
             status_code=503,
@@ -155,8 +189,7 @@ async def chat_completions(request: Request):
                 status_code=413,
                 content={"error": {"message": "Request body too large", "type": "invalid_request_error"}},
             )
-        import json as _json
-        body = _json.loads(body_bytes)
+        body = json.loads(body_bytes)
         chat_request = ChatCompletionRequest.model_validate(body)
     except Exception:
         log.exception("invalid_request")
@@ -183,9 +216,7 @@ async def chat_completions(request: Request):
         last_role=chat_request.messages[-1].role if chat_request.messages else "",
     )
 
-    _in_flight += 1
-    if _drain_event is not None:
-        _drain_event.clear()
+    _requests.enter()
     try:
         result = await asyncio.wait_for(
             _orchestrator.handle_request(chat_request, profile),
@@ -207,9 +238,7 @@ async def chat_completions(request: Request):
             content={"error": {"message": "Backend request failed", "type": "proxy_error"}},
         )
     finally:
-        _in_flight -= 1
-        if _in_flight == 0 and _drain_event is not None:
-            _drain_event.set()
+        _requests.exit()
 
     elapsed = time.monotonic() - start
 
@@ -217,7 +246,7 @@ async def chat_completions(request: Request):
     if hasattr(result, "__aiter__"):
         log.info("response_streaming", elapsed_ms=round(elapsed * 1000))
         return StreamingResponse(
-            result,
+            _timeout_stream(result),
             media_type="text/event-stream",
             headers={"X-FilthyToolFixer-Request-ID": rid},
         )
@@ -254,19 +283,42 @@ async def chat_completions(request: Request):
     )
 
 
+async def _timeout_stream(
+    stream: AsyncIterator[bytes],
+    chunk_timeout: float = _STREAM_CHUNK_TIMEOUT,
+) -> AsyncIterator[bytes]:
+    """Wrap a stream with per-chunk timeout to prevent hanging on stalled backends."""
+    aiter = stream.__aiter__()
+    while True:
+        try:
+            chunk = await asyncio.wait_for(aiter.__anext__(), timeout=chunk_timeout)
+            yield chunk
+        except StopAsyncIteration:
+            break
+        except asyncio.TimeoutError:
+            log.warning("stream_chunk_timeout", timeout=chunk_timeout)
+            error = {"error": {"message": "Backend stream stalled", "type": "proxy_timeout"}}
+            yield f"data: {json.dumps(error)}\n\n".encode()
+            yield b"data: [DONE]\n\n"
+            break
+        except asyncio.CancelledError:
+            raise
+
+
 async def _synthesize_sse(response) -> AsyncIterator[bytes]:
     """Synthesize SSE events from a buffered ChatCompletionResponse.
 
     Emits the response as a single chunk followed by [DONE].
     """
-    import json
-
     data = response.model_dump(exclude_none=True)
     # Convert to streaming format
     data["object"] = "chat.completion.chunk"
     for choice in data.get("choices", []):
         if "message" in choice:
             delta = choice.pop("message")
+            # Ensure content field is present (some clients expect it even as null)
+            if "content" not in delta:
+                delta["content"] = None
             # Streaming tool calls require an index field on each tool call
             for i, tc in enumerate(delta.get("tool_calls", [])):
                 tc["index"] = i
diff --git a/src/filthy_tool_fixer/profiles/loader.py b/src/filthy_tool_fixer/profiles/loader.py
index ee92912..f2f1026 100644
--- a/src/filthy_tool_fixer/profiles/loader.py
+++ b/src/filthy_tool_fixer/profiles/loader.py
@@ -50,7 +50,7 @@ def _parse(self, data: dict) -> ModelProfile:
             backend_url=model_data.get("backend_url", ""),
             tool_calling=ToolCallingConfig(
                 system_suffix=tc_data.get("system_suffix", ""),
-                temperature_override=tc_data.get("temperature_override", 0.0),
+                temperature_override=tc_data.get("temperature_override"),
                 tool_choice_override=tc_data.get("tool_choice_override", ""),
                 strip_thinking=tc_data.get("strip_thinking", False),
                 think_tag_pattern=tc_data.get("think_tag_pattern", "<think>.*?</think>"),
diff --git a/src/filthy_tool_fixer/profiles/types.py b/src/filthy_tool_fixer/profiles/types.py
index 9c57178..d12337b 100644
--- a/src/filthy_tool_fixer/profiles/types.py
+++ b/src/filthy_tool_fixer/profiles/types.py
@@ -16,7 +16,7 @@ class EscalationConfig:
 @dataclass
 class ToolCallingConfig:
     system_suffix: str = ""
-    temperature_override: float | None = 0.0
+    temperature_override: float | None = None  # None = passthrough client temperature
     strip_thinking: bool = False
     think_tag_pattern: str = "<think>.*?</think>"  # Regex pattern for thinking tags
     tool_choice_override: str = ""  # Override tool_choice sent to backend (e.g. "required")
diff --git a/src/filthy_tool_fixer/proxy.py b/src/filthy_tool_fixer/proxy.py
index 548f8f5..f140984 100644
--- a/src/filthy_tool_fixer/proxy.py
+++ b/src/filthy_tool_fixer/proxy.py
@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import asyncio
+import functools
 import re
 import time
 from typing import AsyncIterator
@@ -22,8 +23,10 @@
 
 log = get_logger(__name__)
 
-# Cache compiled think-tag patterns per profile
-_think_re_cache: dict[str, re.Pattern] = {}
+# Cache compiled think-tag patterns per profile (bounded)
+@functools.lru_cache(maxsize=32)
+def _compile_think_re(pattern: str) -> re.Pattern:
+    return re.compile(pattern, re.DOTALL)
 
 # Regex patterns for condensing tool descriptions
 _EXAMPLE_BLOCK_RE = re.compile(r"<example>.*?</example>", re.DOTALL)
@@ -232,9 +235,7 @@ def _strip_thinking(self, response: ChatCompletionResponse, pattern: str) -> Cha
         """Remove thinking blocks from response content using the profile's pattern."""
         if not pattern:
             return response
-        if pattern not in _think_re_cache:
-            _think_re_cache[pattern] = re.compile(pattern, re.DOTALL)
-        compiled = _think_re_cache[pattern]
+        compiled = _compile_think_re(pattern)
         for choice in response.choices:
             if choice.message and choice.message.content:
                 choice.message.content = compiled.sub("", choice.message.content).strip()
diff --git a/src/filthy_tool_fixer/retry/loop.py b/src/filthy_tool_fixer/retry/loop.py
index a6a1225..09105c9 100644
--- a/src/filthy_tool_fixer/retry/loop.py
+++ b/src/filthy_tool_fixer/retry/loop.py
@@ -7,6 +7,9 @@
 import time
 import uuid
 from difflib import get_close_matches
+
+import httpx
+
 from filthy_tool_fixer.backends.base import BackendAdapter
 from filthy_tool_fixer.logging import get_logger
 from filthy_tool_fixer.models import (
@@ -68,7 +71,7 @@ async def execute(
             remaining = budget_remaining - elapsed
 
             # Don't start a retry if insufficient time remains
-            if attempt > 0 and remaining < 10.0:
+            if attempt > 0 and remaining < 20.0:
                 log.info("retry_budget_exhausted", attempt=attempt, remaining=remaining)
                 break
 
@@ -81,6 +84,23 @@ async def execute(
                     timeout=effective_timeout,
                     keep_alive=self._profile.tool_calling.keep_alive,
                 )
+            except httpx.HTTPStatusError as e:
+                status = e.response.status_code
+                if status < 500:
+                    # 4xx = permanent error, don't retry
+                    log.warning("backend_client_error", attempt=attempt, status=status)
+                    break
+                # 5xx = transient, retry
+                log.warning("backend_server_error", attempt=attempt, status=status)
+                if attempt < max_retries:
+                    continue
+                break
+            except (httpx.TimeoutException, httpx.ConnectError) as e:
+                # Transient network/timeout errors — retry
+                log.warning("backend_transient_error", attempt=attempt, error=str(e))
+                if attempt < max_retries:
+                    continue
+                break
             except Exception:
                 log.exception("backend_request_failed", attempt=attempt)
                 break
@@ -286,7 +306,8 @@ async def _try_escalation(
         if failed_response:
             # Add a brief summary of what went wrong, not the full failed conversation
             summary = self._build_escalation_summary(failed_response, tools)
-            messages.append(ChatMessage(role="user", content=summary))
+            if summary:
+                messages.append(ChatMessage(role="user", content=summary))
 
         escalation_request = original_request.model_copy(
             update={
@@ -379,7 +400,10 @@ def _extract_tool_calls(self, response: ChatCompletionResponse):
     # Llama 4 pythonic: [func_name(param="val", param2=val2)]
     # May contain multiple calls: [func1(a=1), func2(b=2), func3(c=3)]
     _PYTHONIC_BRACKET_RE = re.compile(r"\[([^\]]+)\]", re.DOTALL)
-    _PYTHONIC_CALL_RE = re.compile(r"(\w+)\(([^)]*)\)", re.DOTALL)
+    # Match function calls, allowing quoted strings to contain parens
+    _PYTHONIC_CALL_RE = re.compile(
+        r"""(\w+)\(((?:[^()"']*|"[^"]*"|'[^']*')*)\)""", re.DOTALL
+    )
 
     def _rescue_tool_calls_from_text(
         self,
@@ -411,8 +435,9 @@ def _rescue_tool_calls_from_text(
                 return rescued
 
         # Try Llama 4 pythonic format: [func_name(param="val")]
-        bracket_match = self._PYTHONIC_BRACKET_RE.search(content)
-        if bracket_match:
+        # Try all bracket groups, not just the first (avoids false match on
+        # non-tool brackets like [1, 2, 3] appearing earlier in the text)
+        for bracket_match in self._PYTHONIC_BRACKET_RE.finditer(content):
             inner = bracket_match.group(1)
             pythonic_matches = self._PYTHONIC_CALL_RE.findall(inner)
             rescued = self._parse_pythonic_calls(pythonic_matches, tool_names)
diff --git a/src/filthy_tool_fixer/validation/schema.py b/src/filthy_tool_fixer/validation/schema.py
index cd574cd..6678870 100644
--- a/src/filthy_tool_fixer/validation/schema.py
+++ b/src/filthy_tool_fixer/validation/schema.py
@@ -157,7 +157,7 @@ def _validate_single_call(
         validator = jsonschema.Draft7Validator(schema)
         for error in validator.iter_errors(args):
             # Skip errors we already reported (required, additionalProperties)
-            if error.validator in ("required",):
+            if error.validator in ("required", "additionalProperties"):
                 continue
             errors.append(
                 ValidationError(