Tighten local-tool-runtime comments

Author: danielhanchen

studio/backend/core/inference/external_agentic.py

@@ -176,14 +176,11 @@ async def stream_external_local_tool_loop(
     created = int(time.time())
     cancel_event = cancel_event or threading.Event()
 
-    # 9999 is the "no limit" sentinel; pass None so execute_tool never times out
-    # (mirrors the local GGUF loop).
+    # 9999 = "no limit" sentinel: pass None so execute_tool never times out (mirrors local GGUF loop).
     effective_tool_timeout = None if tool_call_timeout >= 9999 else tool_call_timeout
-    # Total tool calls executed across all rounds must not exceed the caller's
-    # per-message budget, even when a single completion returns parallel calls.
+    # Cap total tool calls across all rounds at the caller's per-message budget, even with parallel calls.
     calls_remaining = max_tool_iterations
-    # A forced tool_choice must only apply to the first round; after a tool
-    # runs, later rounds are freed to synthesize the answer.
+    # A forced tool_choice applies only to the first round; later rounds are freed to answer.
     active_tool_choice = tool_choice
 
     enabled_names = {
@@ -217,12 +214,10 @@ async def stream_external_local_tool_loop(
             stream = True,
         )
 
-        # Stop may fire while the remote is still in prefill and the read is
-        # blocked awaiting the next SSE line. Drive the generator one item at a
-        # time via a cancellable task raced against the cancel event, so Stop
-        # abandons the in-flight read immediately (calling aclose() on a
-        # generator that is mid-await raises RuntimeError, so it cannot be used
-        # to interrupt from another task).
+        # Stop may fire while the remote read is blocked awaiting the next SSE line. Drive
+        # the generator one item at a time via a cancellable task raced against the cancel
+        # event so Stop abandons the read immediately (aclose() on a mid-await generator
+        # raises RuntimeError, so it can't interrupt from another task).
         async def _wait_cancel() -> None:
             while not cancel_event.is_set():
                 await asyncio.sleep(0.1)
@@ -320,7 +315,7 @@ async def _wait_cancel() -> None:
         ordered_calls = [tc for tc in ordered_calls if (tc.get("function") or {}).get("name")]
 
         if not ordered_calls or finish_reason not in (None, "tool_calls", "stop"):
-            # No tool calls — emit a terminal finish if we streamed content.
+            # No tool calls: emit a terminal finish if we streamed content.
             if saw_content or finish_reason:
                 yield _openai_content_chunk_line(
                     completion_id = completion_id,
@@ -356,8 +351,7 @@ async def _wait_cancel() -> None:
         conversation.append(
             {
                 "role": "assistant",
-                # Retain any streamed explanation so the follow-up round keeps
-                # the model's own context for interpreting tool results.
+                # Retain streamed explanation so the follow-up round keeps the model's context.
                 "content": assistant_content or None,
                 "tool_calls": assistant_tool_calls,
             }
@@ -373,8 +367,7 @@ async def _wait_cancel() -> None:
             tool_call_id = tc["id"]
 
             _prov = tool_event_provenance()
-            # Later rounds must be free to answer; a forced choice only applies
-            # to the first round.
+            # Later rounds must be free to answer; a forced choice applies only to the first round.
             active_tool_choice = "auto"
 
             # Skip tools the caller did not enable (defense in depth).
@@ -465,7 +458,7 @@ async def _wait_cancel() -> None:
 
         continue
 
-    # Budget exhausted after tool rounds — one final synthesis pass without tools.
+    # Budget exhausted after tool rounds: one final synthesis pass without tools.
     final_finish_reason = None
     if not cancel_event.is_set():
         if max_tool_iterations > 0:

studio/backend/routes/inference.py

@@ -6789,9 +6789,8 @@ async def _proxy_to_external_provider(
     # `model_fields_set` tracks explicit-vs-default per request.
     _top_k_explicit = payload.top_k if "top_k" in payload.model_fields_set else None
 
-    # OAI-compat Connections (ollama / llama.cpp / vLLM / custom) can drive
-    # Unsloth's local tool runtime against the remote model (#7282). Hosted
-    # providers stay on the pure-proxy + supportsBuiltin* path.
+    # OAI-compat Connections (ollama / llama.cpp / vLLM / custom) drive Unsloth's local tool
+    # runtime against the remote model (#7282); hosted providers stay on the pure-proxy path.
     from core.inference.external_agentic import (
         provider_supports_local_tool_runtime,
         stream_external_local_tool_loop,
@@ -6806,8 +6805,7 @@ async def _proxy_to_external_provider(
         and payload.stream
         and (_tools_on_ext or _mcp_allowed_ext)
         and not payload.tools  # client-supplied tools stay on passthrough
-        # Callers opting out via tool_choice="none" skip the loop, mirroring
-        # the local GGUF routing gate.
+        # tool_choice="none" opts out, mirroring the local GGUF routing gate.
         and not (
             payload.tool_choice == "none" and not _explicit_studio_tool_loop_requested(payload)
         )
@@ -6820,13 +6818,12 @@ async def _stream_local_tools():
             payload, tools_on = bool(_tools_on_ext), mcp_allowed = _mcp_allowed_ext
         )
         if not tools_to_use:
-            # Nothing to run — reuse the plain proxy stream (incl. client.close).
+            # Nothing to run: reuse the plain proxy stream (incl. client.close).
             async for line in _stream():
                 yield line
             return
 
-        # Inject the same tool-action nudge local GGUF turns get so small
-        # remote models know when to call tools.
+        # Inject the same tool-action nudge local GGUF turns get so small remote models know when to call tools.
         _nudge = _build_tool_action_nudge(tools = tools_to_use, model_name = model)
         _nudge = _apply_rag_nudge(_nudge, tools_to_use, rag_scope = payload.rag_scope)
         loop_messages = list(chat_messages)
@@ -7208,9 +7205,8 @@ async def openai_chat_completions(
 
         untrack_current_request(request.scope)
 
-        # Resolve provider type early so OAI-compat Connections (ollama /
-        # llama.cpp / vLLM / custom) can use Unsloth's local tool runtime
-        # (#7282) instead of being rejected on confirm_tool_calls.
+        # Resolve provider type early so OAI-compat Connections (ollama / llama.cpp / vLLM /
+        # custom) can use the local tool runtime (#7282) instead of being rejected on confirm_tool_calls.
         _ext_provider_type = payload.provider_type
         if payload.provider_id and not _ext_provider_type:
             _cfg = providers_db.get_provider(payload.provider_id)

studio/frontend/src/features/chat/api/chat-adapter.ts

@@ -2232,8 +2232,7 @@ export function createOpenAIStreamAdapter(
             externalProvider.baseUrl,
           ),
       );
-      // OAI-compat Connections (ollama / llama.cpp / vLLM / custom) drive
-      // Unsloth's local tool runtime against the remote model (#7282).
+      // OAI-compat Connections drive Unsloth's local tool runtime against the remote model (#7282).
       const localToolRuntimeForThisTurn = Boolean(
         externalProvider &&
           providerSupportsLocalToolRuntime(externalProvider.providerType),
@@ -3056,10 +3055,8 @@ export function createOpenAIStreamAdapter(
               ...(externalCapabilities?.presencePenalty
                 ? { presence_penalty: params.presencePenalty }
                 : {}),
-              // Hosted providers: enabled_tools maps to server-side builtins.
-              // OAI-compat Connections: enable Unsloth's local tool runtime
-              // (web_search / python / terminal / MCP) against the remote
-              // model (#7282).
+              // Hosted providers: enabled_tools maps to server-side builtins. OAI-compat
+              // Connections: enable the local tool runtime (web_search / python / terminal / MCP) (#7282).
               ...(webSearchEnabledForThisTurn ||
               webFetchEnabledForThisTurn ||
               codeExecEnabledForThisTurn ||

studio/frontend/src/features/chat/chat-page.tsx

@@ -2079,10 +2079,8 @@ export function ChatPage({
           : true
         : state.reasoningEnabled,
       supportsPreserveThinking: false,
-      // Hosted providers have no local tool runtime, so supportsTools stays
-      // false and supportsBuiltin* carries server-side capability. OAI-compat
-      // Connections (ollama / llama.cpp / vLLM / custom) can drive Unsloth's
-      // local Search/Code/MCP tools against the remote model (#7282).
+      // Hosted providers have no local tool runtime (supportsTools false; supportsBuiltin*
+      // carries server-side capability). OAI-compat Connections drive local Search/Code/MCP (#7282).
       supportsTools: supportsLocalToolRuntime,
       supportsBuiltinWebSearch,
       supportsBuiltinCodeExecution,
@@ -2608,8 +2606,7 @@ export function ChatPage({
               : true
             : store.reasoningEnabled,
           supportsPreserveThinking: false,
-          // Hosted providers → supportsTools false + supportsBuiltin*.
-          // OAI-compat Connections can drive local Search/Code/MCP (#7282).
+          // Hosted providers: supportsTools false + supportsBuiltin*. OAI-compat Connections drive local Search/Code/MCP (#7282).
           supportsTools: supportsLocalToolRuntime,
           supportsBuiltinWebSearch,
           supportsBuiltinCodeExecution,

studio/frontend/src/features/chat/provider-capabilities.ts

@@ -160,11 +160,10 @@ function _inferProviderFromOpenrouterId(
 }
 
 /**
- * OAI-compat Connections (Ollama / llama.cpp / vLLM / Custom) on a remote
- * endpoint that can still drive Unsloth's *local* tool runtime (web_search /
- * python / terminal / MCP) on the Studio host. Unlike `providerSupportsBuiltin*`
- * (server-side hosted tools), these get `supportsTools=true` and the backend
- * runs the local tool loop against the model's function calls (#7282).
+ * OAI-compat Connections (Ollama / llama.cpp / vLLM / Custom) on a remote endpoint that
+ * can drive Unsloth's *local* tool runtime (web_search / python / terminal / MCP) on the
+ * Studio host. Unlike `providerSupportsBuiltin*` (server-side), these get `supportsTools=true`
+ * and the backend runs the local tool loop against the model (#7282).
  */
 const LOCAL_TOOL_RUNTIME_PROVIDER_TYPES = new Set([
   "ollama",