diff --git a/.cursor/rules/run-tests.mdc b/.cursor/rules/run-tests.mdc
new file mode 100644
index 0000000..fbc5b0e
--- /dev/null
+++ b/.cursor/rules/run-tests.mdc
@@ -0,0 +1,16 @@
+---
+description:
+globs:
+alwaysApply: true
+---
+# Running tests
+
+After finishing any big code changes, and before yielding to the user, run with your terminal tool the following comand:
+
+```
+source .venv/bin/activate && pytest -k ""
+```
+
+You don't need to preface it with /bin/bash -c, nor add | cat at the end. Just run it as is.
+
+If you want to run linting and tests, perform 2 parallel function calls
\ No newline at end of file
diff --git a/.env.example b/.env.example
index bfcca79..924f570 100644
--- a/.env.example
+++ b/.env.example
@@ -4,6 +4,8 @@ FLASK_DEBUG=1
FLASK_ENV=development
GUNICORN_WORKERS=1
LOG_LEVEL=debug
+LOG_CONTEXT=on
+LOG_COMPLETION=on
RECORD_TRAFFIC=off
# Arbitrary API key to protect your service.
diff --git a/README.md b/README.md
index 1e8b6b2..bf1f4a4 100644
--- a/README.md
+++ b/README.md
@@ -5,7 +5,6 @@
[](#)
[](#)
-
@@ -26,8 +25,6 @@ This project originates from Cursor's lack of support for Azure models that are
Upcoming features:
- Multimodal: Will be implemented as soon as better testing is in place and there is demand (PRs welcome).
-- Multiple models simultaneously: Even Cursor's Azure configuration only supports a single deployment at a time. It would be fairly easy to implement support for multiple models in this service, covering even more needs.
-- Full test coverage: See the [Testing](#testing) section for an explanation of the current low coverage.
Feel free to create or vote on any [project issues](https://github.com/gabrii/Cursor-Azure-GPT-5/issues), and star the project to show your support.
@@ -59,6 +56,8 @@ Alternatively, you can pass them through the environment where you run the appli
| `AZURE_TRUNCATION` | Truncation strategy for long inputs. | `auto` |
| `FLASK_ENV` | Flask environment. Use `development` for dev or `production` for prod. | `production` |
| `RECORD_TRAFFIC` | Toggle writing request/response traffic to `recordings/` | `off` |
+| `LOG_CONTEXT` | Enable rich pretty-printing of request context to console. | `on` |
+| `LOG_COMPLETION` | Enable logging of completion responses (not yet implemented). | `on` |
diff --git a/app/azure/response_adapter.py b/app/azure/response_adapter.py
index b063c36..38f10c5 100644
--- a/app/azure/response_adapter.py
+++ b/app/azure/response_adapter.py
@@ -11,9 +11,12 @@ import time
from string import ascii_letters, digits
from typing import Any, Dict, Iterable, Optional
-from flask import Response, stream_with_context
+from flask import Response, current_app, stream_with_context
+from rich.live import Live
+from ..common.logging import console, create_message_panel
from ..common.sse import chunks_to_sse, sse_to_events
+from ..exceptions import ClientClosedConnection
# Centralized events that should end a block before handling
THINKING_STOP_EVENTS = {"response.output_text.delta", "response.output_item.added"}
@@ -158,10 +161,27 @@ class ResponseAdapter:
self._tool_calls = 0
def gen_dicts() -> Iterable[Dict[str, Any]]:
- try:
+ # Initialize message object for completion logging
+ completion_msg: Dict[str, Any] = {
+ "role": "assistant",
+ "content": "",
+ "tool_calls": [],
+ }
+
+ events = 0
+ with Live(
+ None,
+ console=console,
+ refresh_per_second=2,
+ ) as live: # update 4 times a second to feel fluid
for ev in sse_to_events(
upstream_resp.iter_content(chunk_size=8192)
):
+ if current_app.config["LOG_COMPLETION"]:
+ if events > 1:
+ live.update(create_message_panel(completion_msg, 1, 1))
+ events += 1
+
handler_name = "_" + (ev.event or "").replace(
"response.", ""
).replace(".", "__")
@@ -176,19 +196,61 @@ class ResponseAdapter:
)
self._thinking = False
+ if current_app.config["LOG_COMPLETION"]:
+ completion_msg["content"] += "\n\n"
+
res = handler(ev.json)
if res is not None:
yield res
- finally:
- # Emit finish reason at the end of stream
+
+ if current_app.config["LOG_COMPLETION"]:
+ delta = res.get("choices", [{}])[0].get("delta", {})
+ content = delta.get("content")
+
+ if content is not None:
+ # Append content to the message
+ completion_msg["content"] += content
+ else:
+ # Handle tool calls
+ tool_calls_delta = delta.get("tool_calls", [])
+ for tool_call_delta in tool_calls_delta:
+ function = tool_call_delta.get("function", {})
+ name = function.get("name")
+ arguments = function.get("arguments", "")
+
+ if name:
+ # New tool call - add to the list
+ completion_msg["tool_calls"].append(
+ {
+ "id": tool_call_delta.get("id", ""),
+ "type": "function",
+ "function": {
+ "name": name,
+ "arguments": arguments,
+ },
+ }
+ )
+ else:
+ # Append arguments to the last tool call
+ completion_msg["tool_calls"][-1][
+ "function"
+ ]["arguments"] += arguments
+
if self._tool_calls > 0:
yield self._build_completion_chunk(finish_reason="tool_calls")
else:
yield self._build_completion_chunk(finish_reason="stop")
+ if current_app.config["LOG_COMPLETION"]:
+ live.update(create_message_panel(completion_msg, 1, 1))
- # Wrap as SSE with [DONE]
try:
yield from chunks_to_sse(gen_dicts())
+ except GeneratorExit:
+ # Downstream client closed the connection mid-stream
+ # Translate to a clearer exception; upstream will be closed in finally
+ raise ClientClosedConnection(
+ "Client closed connection during streaming response"
+ ) from None
finally:
upstream_resp.close()
diff --git a/app/blueprint.py b/app/blueprint.py
index 461b009..16b7950 100644
--- a/app/blueprint.py
+++ b/app/blueprint.py
@@ -4,7 +4,7 @@ This module defines the application blueprint, configures logging, and
forwards incoming HTTP requests to the configured backend implementation.
"""
-from flask import Blueprint, jsonify, request
+from flask import Blueprint, current_app, jsonify, request
from .auth import require_auth
from .azure.adapter import AzureAdapter
@@ -46,7 +46,8 @@ def catch_all(path: str):
implementation, returning the backend's response. If forwarding fails,
returns a 502 JSON error payload.
"""
- log_request(request)
+ if current_app.config.get("LOG_CONTEXT"):
+ log_request(request)
init_last_recording()
increment_last_recording()
record_payload(request.json, "downstream_request")
diff --git a/app/common/logging.py b/app/common/logging.py
index 4d9f793..dbb6aaf 100644
--- a/app/common/logging.py
+++ b/app/common/logging.py
@@ -105,6 +105,82 @@ def escape_tags(text: str) -> str:
).replace(">`\n\n\n`<", ">`\n\n`<")
+def create_message_panel(msg: Dict[str, Any], idx: int, total: int) -> Panel:
+ """Create a Rich Panel for displaying a message.
+
+ Args:
+ msg: Message object with 'role', 'content', optional 'name', 'tool_call_id', 'tool_calls'
+ idx: Current message index (1-based)
+ total: Total number of messages
+
+ Returns:
+ A Rich Panel object ready to be printed
+ """
+ role = str(msg.get("role", ""))
+ content_val = msg.get("content", "")
+ name = msg.get("name")
+ tool_call_id = msg.get("tool_call_id")
+ message_title = (
+ f"[italic]{idx}/{total}[/italic] [bold]<{role}>[/bold]"
+ if not name
+ else f"[italic]{idx}/{total}[/italic] [bold]<{role} name={name} id={tool_call_id}>[/bold]"
+ )
+ message_elements = []
+ message_elements.append(
+ Padding(
+ Markdown(
+ escape_tags(content_val),
+ ),
+ (1, 0),
+ )
+ )
+ tool_calls = msg.get("tool_calls", [])
+ for tool_call in tool_calls:
+ function = tool_call.get("function", {})
+ arguments = function.get("arguments")
+ tool_elements = []
+ tool_title = f"[bold][/bold]"
+ tool_elements.append(
+ f"[bold][magenta]{function.get('name')}[/magenta] [blue]([/blue][/bold]",
+ )
+ try:
+ tool_elements.append(Padding(JSON(arguments), (0, 4)))
+ except json.JSONDecodeError:
+ tool_elements.append(
+ Padding("[red]Invalid JSON generated by the model:[/red]", (0, 4))
+ )
+ tool_elements.append(arguments)
+ tool_elements.append("[bold][blue])[/blue][/bold]")
+ message_elements.append(
+ Panel(
+ Group(*tool_elements),
+ title=tool_title,
+ border_style=ROLE_COLORS["tool"],
+ )
+ )
+ message_style = ROLE_COLORS.get(role, "red")
+ message_subtitle = message_title[message_title.find("<") :].replace("<", "")
+ )
+ ]
+ + ">[/bold]"
+ )
+ return Panel(
+ Group(*message_elements),
+ title=message_title,
+ title_align="left",
+ subtitle=message_subtitle,
+ subtitle_align="right",
+ border_style=message_style,
+ )
+
+
def log_request(req: Request) -> str:
"""Pretty-print a Flask request using Rich and return the request id."""
request_id = uuid.uuid4().hex[:8]
@@ -131,130 +207,65 @@ def log_request(req: Request) -> str:
messages = json_payload.get("messages", [])
tools = json_payload.get("tools", [])
- for idx, tool in enumerate(tools, start=1):
- table = Table(
- caption="[italic]Required fields are marked with *[/italic]",
- pad_edge=False,
- box=box.SIMPLE,
- leading=2,
+ # Render tools section once (no duplicate panels)
+ table = Table(
+ caption="[italic]Required fields are marked with *[/italic]",
+ pad_edge=False,
+ box=box.SIMPLE,
+ leading=2,
+ )
+ table.add_column("Name", justify="right", style="cyan", no_wrap=True)
+ table.add_column("Description", style="white")
+
+ for tool in tools:
+ function = tool.get("function")
+ parameters = function.get("parameters", {}) or {}
+ required = parameters.get("required", []) or []
+ props = parameters.get("properties", {}) or {}
+
+ params_table = Table(
+ show_header=False,
+ show_edge=True,
+ show_lines=True,
+ title="Parameters",
+ title_justify="left",
+ expand=True,
+ leading=3,
+ )
+ params_table.add_column("Name", justify="right", no_wrap=True, style="cyan")
+ params_table.add_column("Description", style="white")
+ for param_name, param_value in props.items():
+ param_type = param_value.get("type")
+ if param_type == "array":
+ param_type += f"({param_value.get('items').get('type')})"
+ if param_name in required:
+ param_type = f"[bold]*{param_type}[/bold]"
+ param_type = f"[magenta]{param_type}[/magenta]"
+ params_table.add_row(
+ Group(param_name, param_type),
+ f"{param_value.get('description')}",
+ )
+ table.add_row(
+ f"{function.get('name')}",
+ Group(
+ Markdown(escape_tags(function.get("description"))),
+ params_table,
+ ),
)
- table.add_column("Name", justify="right", style="cyan", no_wrap=True)
- table.add_column("Description", style="white")
-
- for tool in tools:
- function = tool.get("function")
- required = function.get("parameters").get("required")
-
- params_table = Table(
- show_header=False,
- show_edge=True,
- show_lines=True,
- title="Parameters",
- title_justify="left",
- expand=True,
- leading=3,
- )
- params_table.add_column("Name", justify="right", no_wrap=True, style="cyan")
- params_table.add_column("Description", style="white")
- for param_name, param_value in (
- function.get("parameters").get("properties").items()
- ):
- param_type = param_value.get("type")
- if param_type == "array":
- param_type += f"({param_value.get('items').get('type')})"
- if param_name in required:
- param_type = f"[bold]*{param_type}[/bold]"
- param_type = f"[magenta]{param_type}[/magenta]"
- param_name = f"{param_name}"
- params_table.add_row(
- Group(param_name, param_type),
- f"{param_value.get("description")}",
- )
- table.add_row(
- f"{function.get('name')}",
- Group(
- Markdown(escape_tags(function.get("description"))),
- params_table,
- ),
- )
-
- console.print(
- Panel(
- table,
- title=f"[italic]{0}/{len(messages)}[/italic] [bold][/bold]",
- title_align="left",
- subtitle="[bold][/bold]",
- subtitle_align="right",
- border_style=ROLE_COLORS["tool"],
- )
+ console.print(
+ Panel(
+ table,
+ title=f"[italic]{0}/{len(messages)}[/italic] [bold][/bold]",
+ title_align="left",
+ subtitle="[bold][/bold]",
+ subtitle_align="right",
+ border_style=ROLE_COLORS["tool"],
)
+ )
for idx, msg in enumerate(messages, start=1):
- role = str(msg.get("role", ""))
- content_val = msg.get("content", "")
- name = msg.get("name")
- tool_call_id = msg.get("tool_call_id")
- message_title = (
- f"[italic]{idx}/{len(messages)}[/italic] [bold]<{role}>[/bold]"
- if not name
- else f"[italic]{idx}/{len(messages)}[/italic] [bold]<{role} name={name} id={tool_call_id}>[/bold]"
- )
- message_elements = []
- message_elements.append(
- Padding(
- Markdown(
- escape_tags(content_val),
- ),
- (1, 0),
- )
- )
- tool_calls = msg.get("tool_calls", [])
- for tool_call in tool_calls:
- function = tool_call.get("function", {})
- arguments = function.get("arguments")
- tool_elements = []
- tool_title = f"[bold][/bold]"
- tool_elements.append(
- f"[bold][magenta]{function.get('name')}[/magenta] [blue]([/blue][/bold]",
- )
- try:
- tool_elements.append(Padding(JSON(arguments), (0, 4)))
- except json.JSONDecodeError:
- tool_elements.append(
- Padding("[red]Invalid JSON generated by the model:[/red]", (0, 4))
- )
- tool_elements.append(arguments)
- tool_elements.append("[bold][blue])[/blue][/bold]")
- message_elements.append(
- Panel(
- Group(*tool_elements),
- title=tool_title,
- border_style=ROLE_COLORS["tool"],
- )
- )
- message_style = ROLE_COLORS.get(role, "red")
- message_subtitle = message_title[message_title.find("<") :].replace("<", "")
- )
- ]
- + ">[/bold]"
- )
- console.print(
- Panel(
- Group(*message_elements),
- title=message_title,
- title_align="left",
- subtitle=message_subtitle,
- subtitle_align="right",
- border_style=message_style,
- )
- )
+ panel = create_message_panel(msg, idx, len(messages))
+ console.print(panel)
return request_id
diff --git a/app/common/sse.py b/app/common/sse.py
index 08b5b1d..4b82557 100644
--- a/app/common/sse.py
+++ b/app/common/sse.py
@@ -155,10 +155,10 @@ def chunks_to_sse(chunks: Iterable[Dict[str, Any]]) -> Iterator[bytes]:
sse = encode_sse_json(obj)
buffer += sse
yield sse
- finally:
sse = done_event_bytes()
buffer += sse
yield sse
+ finally:
record_sse(buffer, "downstream_response")
diff --git a/app/exceptions.py b/app/exceptions.py
index e0397f4..129f3c7 100644
--- a/app/exceptions.py
+++ b/app/exceptions.py
@@ -23,3 +23,10 @@ class CursorConfigurationError(ConfigurationError):
"""Exception raised for configuration errors in Cursor configuration."""
preamble = "Cursor configuration error, check your Cursor settings."
+
+
+class ClientClosedConnection(Exception): # noqa: N818
+ """Raised when the downstream client closes the HTTP connection mid-stream.
+
+ This helps distinguish client disconnects from other server-side errors.
+ """
diff --git a/app/settings.py b/app/settings.py
index 6a31680..965e4cf 100644
--- a/app/settings.py
+++ b/app/settings.py
@@ -14,6 +14,8 @@ env.read_env()
ENV = env.str("FLASK_ENV", default="production")
DEBUG = ENV == "development"
RECORD_TRAFFIC = env.bool("RECORD_TRAFFIC", False)
+LOG_CONTEXT = env.bool("LOG_CONTEXT", True)
+LOG_COMPLETION = env.bool("LOG_COMPLETION", True)
SERVICE_API_KEY = env.str("SERVICE_API_KEY", "change-me")
diff --git a/tests/settings.py b/tests/settings.py
index 46d6910..b06daf5 100644
--- a/tests/settings.py
+++ b/tests/settings.py
@@ -14,6 +14,8 @@ AZURE_VERBOSITY_LEVEL = "medium"
AZURE_TRUNCATION = "auto"
RECORD_TRAFFIC = False
+LOG_CONTEXT = True
+LOG_COMPLETION = True
AZURE_RESPONSES_API_URL = (
diff --git a/tests/test_client_closed_connection.py b/tests/test_client_closed_connection.py
new file mode 100644
index 0000000..882b4f3
--- /dev/null
+++ b/tests/test_client_closed_connection.py
@@ -0,0 +1,51 @@
+"""Tests for client-closed connection during streaming.
+
+This test simulates a downstream client closing the connection mid-stream by
+forcing a GeneratorExit inside the streaming pipeline. The outer boundary in
+ResponseAdapter should translate that into ClientClosedConnection.
+"""
+
+import pytest
+
+from app.exceptions import ClientClosedConnection
+from tests.replay_base import ReplyBase
+
+
+def _raise_generator_exit(
+ _chunks,
+): # pragma: no cover - behavior is validated via exception
+ """Stub encoder that simulates generator shutdown.
+
+ Raising GeneratorExit here mimics the server-side observation that the
+ response iterable was closed. We don't yield any bytes.
+ """
+ raise GeneratorExit
+
+
+class TestClientClosedConnection(ReplyBase):
+ """Ensure a client disconnect is surfaced as ClientClosedConnection."""
+
+ # Use any existing recording that yields a normal streaming response
+ recording: str = "one_ping_pong"
+
+ def test(self, testapp, requests_mock, monkeypatch):
+ """Test chunks_to_see raising GeneratorExit."""
+ # Patch the exact symbol used inside ResponseAdapter to encode SSE
+ monkeypatch.setattr(
+ "app.azure.response_adapter.chunks_to_sse", _raise_generator_exit
+ )
+
+ # Mock upstream with recorded SSE so the adapter enters streaming code
+ mock = self.mock_upstream(requests_mock)
+
+ # When the streaming pipeline is interrupted, the adapter should raise
+ # our domain-specific exception. No response bytes should be produced.
+ with pytest.raises(ClientClosedConnection):
+ testapp.post(
+ "/chat/completions",
+ params=self.downstream_request_body,
+ headers=self.downstream_request_headers,
+ )
+
+ # Sanity check: upstream was called with the expected payload
+ self.assert_upstream_request(mock)
diff --git a/tests/test_logging.py b/tests/test_logging.py
index e4b7d0f..fcac3d5 100644
--- a/tests/test_logging.py
+++ b/tests/test_logging.py
@@ -50,3 +50,45 @@ def test_should_not_refact(mocker):
}
redacted_headers = redact_headers(headers)
assert redacted_headers == headers
+
+
+class TestLogContextEnabled(ReplyBase):
+ """Test that log_request is called when LOG_CONTEXT=True."""
+
+ def modify_settings(self, app) -> None:
+ """Ensure LOG_CONTEXT is enabled."""
+ app.config["LOG_CONTEXT"] = True
+
+ def test(self, testapp, requests_mock, mocker):
+ """Test that log_request is called when LOG_CONTEXT is True."""
+ log_request_mock = mocker.patch("app.blueprint.log_request")
+ super().test(testapp, requests_mock)
+ log_request_mock.assert_called_once()
+
+
+class TestLogContextDisabled(ReplyBase):
+ """Test that log_request is NOT called when LOG_CONTEXT=False."""
+
+ def modify_settings(self, app) -> None:
+ """Disable LOG_CONTEXT."""
+ app.config["LOG_CONTEXT"] = False
+
+ def test(self, testapp, requests_mock, mocker):
+ """Test that log_request is NOT called when LOG_CONTEXT is False."""
+ log_request_mock = mocker.patch("app.blueprint.log_request")
+ super().test(testapp, requests_mock)
+ log_request_mock.assert_not_called()
+
+
+class TestLogCompletionDisabled(ReplyBase):
+ """Test that logging of create_message_panel is NOT called when LOG_COMPLETION=False."""
+
+ def modify_settings(self, app) -> None:
+ """Disable LOG_COMPLETION."""
+ app.config["LOG_COMPLETION"] = False
+
+ def test(self, testapp, requests_mock, mocker):
+ """Test that Rich.Live.update is NOT called when LOG_COMPLETION is False."""
+ live_update_mock = mocker.patch("rich.live.Live.update")
+ super().test(testapp, requests_mock)
+ live_update_mock.assert_not_called()