feat: add scoped model runtime configuration

Introduce provider, model, and invocation contracts with encrypted configuration persistence. Add web runtime fencing, route fallback, recovery middleware, workspace scoping, and comprehensive tests.

EvoScientist/llm/patches.py

@@ -11,6 +11,8 @@ Patches:
     - _patch_openai_capture_reasoning_content: capture provider
       reasoning_content into AIMessage.additional_kwargs (module-level,
       applied at import)
+    - _patch_openai_empty_sse_keepalive: ignore blank SSE keepalive events
+      emitted by some OpenAI-compatible Responses endpoints
     - _patch_deepseek_reasoning_passback: re-inject reasoning_content into
       outgoing DeepSeek assistant messages for thinking-mode multi-turn /
       tool_use scenarios
@@ -70,6 +72,65 @@ def _patch_anthropic_proxy_compat() -> None:
 _patch_anthropic_proxy_compat()
 
 
+# ---------------------------------------------------------------------------
+# Patch (module-level): some OpenAI-compatible endpoints emit SSE keepalive
+# frames with an empty ``data`` field. OpenAI SDK 2.x unconditionally passes
+# every event to ``json.loads``, so an otherwise harmless keepalive becomes a
+# JSONDecodeError and aborts the stream. Filter only blank data frames before
+# the SDK's parser; JSON events, error events, and [DONE] are unchanged.
+# ---------------------------------------------------------------------------
+_openai_empty_sse_keepalive_patched = False
+
+
+def _is_blank_sse_keepalive(event: Any) -> bool:
+    """Return whether an SSE event has no JSON payload to parse."""
+
+    data = getattr(event, "data", None)
+    return data is None or (isinstance(data, str) and not data.strip())
+
+
+def _patch_openai_empty_sse_keepalive() -> None:
+    global _openai_empty_sse_keepalive_patched
+    if _openai_empty_sse_keepalive_patched:
+        return
+    try:
+        import functools
+
+        from openai._streaming import AsyncStream as _AsyncStream
+        from openai._streaming import Stream as _Stream
+
+        original_async = _AsyncStream._iter_events
+        if not getattr(original_async, "_evoscientist_skips_blank_sse", False):
+
+            @functools.wraps(original_async)
+            async def _filtered_async_events(self: Any) -> Any:
+                async for event in original_async(self):
+                    if not _is_blank_sse_keepalive(event):
+                        yield event
+
+            _filtered_async_events._evoscientist_skips_blank_sse = True  # type: ignore[attr-defined]
+            _AsyncStream._iter_events = _filtered_async_events
+
+        original_sync = _Stream._iter_events
+        if not getattr(original_sync, "_evoscientist_skips_blank_sse", False):
+
+            @functools.wraps(original_sync)
+            def _filtered_sync_events(self: Any) -> Any:
+                for event in original_sync(self):
+                    if not _is_blank_sse_keepalive(event):
+                        yield event
+
+            _filtered_sync_events._evoscientist_skips_blank_sse = True  # type: ignore[attr-defined]
+            _Stream._iter_events = _filtered_sync_events
+        _openai_empty_sse_keepalive_patched = True
+    except Exception:
+        # The patch is only needed when the optional OpenAI SDK is available.
+        pass
+
+
+_patch_openai_empty_sse_keepalive()
+
+
 # ---------------------------------------------------------------------------
 # Patch: ccproxy-api 0.2.7 Codex compatibility.
 #
@@ -207,6 +268,9 @@ def _is_ccproxy_codex(
 # preserved, not flattened away.
 # ---------------------------------------------------------------------------
 _SKIP_CONTENT_TYPES = frozenset({"thinking", "reasoning", "reasoning_content"})
+_NONPORTABLE_REASONING_METADATA = frozenset(
+    {"reasoning_content", "reasoning_details"}
+)
 
 # Media block types preserved when flattening (positive allowlist;
 # thinking/reasoning still dropped).  Images + files (PDF/documents): both
@@ -407,6 +471,8 @@ def _copy_ai_message_with_tool_pairs(
                 # LangChain content blocks use id; the Responses converter later
                 # maps it to call_id.
                 block["id"] = call_id
+                if "call_id" in block:
+                    block["call_id"] = call_id
                 block["name"] = call_name
                 if isinstance(block.get("function"), dict):
                     block["function"] = {**block["function"], "name": call_name}
@@ -540,7 +606,7 @@ def _validate_openai_tool_history(messages: list[Any]) -> None:
                     }:
                         continue
                     block_id = str(
-                        block.get("id") or block.get("call_id") or ""
+                    block.get("call_id") or block.get("id") or ""
                     ).strip()
                     block_name = block.get("name") or block.get("tool_name")
                     function = block.get("function")
@@ -566,7 +632,12 @@ def _validate_openai_tool_history(messages: list[Any]) -> None:
         raise ValueError("assistant tool call is missing its tool result")
 
 
-def _sanitize_messages(messages: list[Any], hoist_tool_media: bool = True) -> list[Any]:
+def _sanitize_messages(
+    messages: list[Any],
+    hoist_tool_media: bool = True,
+    *,
+    drop_reasoning_metadata: bool = False,
+) -> list[Any]:
     """Flatten list content for OpenAI-compatible APIs, preserving media.
 
     Text/reasoning content is flattened to a string; image blocks are
@@ -593,6 +664,15 @@ def _sanitize_messages(messages: list[Any], hoist_tool_media: bool = True) -> li
             pending_media.clear()
 
     for msg in messages:
+        if drop_reasoning_metadata:
+            additional_kwargs = getattr(msg, "additional_kwargs", None) or {}
+            if set(additional_kwargs) & _NONPORTABLE_REASONING_METADATA:
+                msg = copy.copy(msg)
+                msg.additional_kwargs = {
+                    key: value
+                    for key, value in additional_kwargs.items()
+                    if key not in _NONPORTABLE_REASONING_METADATA
+                }
         is_tool = getattr(msg, "type", None) == "tool"
         if not is_tool:
             _flush()  # emit hoisted media before any non-tool message
@@ -714,7 +794,12 @@ def _strip_media_types(messages: list[Any], types: set[str]) -> list[Any]:
     return out
 
 
-def _patch_openai_compat_content(model: Any, hoist_tool_media: bool = True) -> None:
+def _patch_openai_compat_content(
+    model: Any,
+    hoist_tool_media: bool = True,
+    *,
+    drop_reasoning_metadata: bool = False,
+) -> None:
     """Flatten list content to strings before OpenAI-compatible API calls.
 
     Wraps ``_generate`` / ``_agenerate`` / ``_stream`` / ``_astream`` to prevent
@@ -749,11 +834,17 @@ def _patch_openai_compat_content(model: Any, hoist_tool_media: bool = True) -> N
 
     def _prepare(messages: list[BaseMessage]) -> list[BaseMessage]:
         msgs = _strip_media_types(messages, blocked) if blocked else messages
-        return _sanitize_messages(msgs, hoist_tool_media)
+        return _sanitize_messages(
+            msgs,
+            hoist_tool_media,
+            drop_reasoning_metadata=drop_reasoning_metadata,
+        )
 
     def _stripped(messages: list[BaseMessage], suspects: set[str]) -> list[BaseMessage]:
         return _sanitize_messages(
-            _strip_media_types(messages, blocked | suspects), hoist_tool_media
+            _strip_media_types(messages, blocked | suspects),
+            hoist_tool_media,
+            drop_reasoning_metadata=drop_reasoning_metadata,
         )
 
     orig_generate = getattr(model, "_generate", None)