| |
| """llm_request must surface server errors instead of hiding them. |
| |
| Before: an HTTP error status (e.g. llama-server's 400 "request exceeds the |
| available context size") was reported as "Cannot reach llama-server" and |
| retried twice; an error event inside the SSE stream, or a plain JSON reply |
| from a server that ignored "stream": true, became a silent empty answer. |
| """ |
|
|
| import json |
| import os |
| import sys |
| import threading |
| from http.server import BaseHTTPRequestHandler, HTTPServer |
|
|
| import pytest |
|
|
| sys.path.insert(0, os.path.dirname(os.path.dirname(__file__))) |
|
|
| import trashclaw |
|
|
|
|
| @pytest.fixture |
| def fake_server(monkeypatch): |
| state = {"status": 200, "ctype": "text/event-stream", "body": b"", "posts": 0} |
|
|
| class Handler(BaseHTTPRequestHandler): |
| def log_message(self, *args): |
| pass |
|
|
| def do_POST(self): |
| self.rfile.read(int(self.headers.get("Content-Length", 0))) |
| state["posts"] += 1 |
| body = state["body"] |
| self.send_response(state["status"]) |
| self.send_header("Content-Type", state["ctype"]) |
| self.send_header("Content-Length", str(len(body))) |
| self.end_headers() |
| self.wfile.write(body) |
|
|
| server = HTTPServer(("127.0.0.1", 0), Handler) |
| thread = threading.Thread(target=server.serve_forever, daemon=True) |
| thread.start() |
| monkeypatch.setattr(trashclaw, "LLAMA_URL", "http://127.0.0.1:%d" % server.server_address[1]) |
| monkeypatch.setattr(trashclaw, "LLM_RETRY_DELAY", 0) |
| monkeypatch.setattr(trashclaw, "_INTERRUPTED", False) |
| yield state |
| server.shutdown() |
| server.server_close() |
|
|
|
|
| MSGS = [{"role": "user", "content": "hi"}] |
|
|
|
|
| def test_http_400_reports_server_message_and_is_not_retried(fake_server): |
| fake_server.update(status=400, ctype="application/json", body=json.dumps( |
| {"error": {"code": 400, "message": "the request exceeds the available context size"}} |
| ).encode()) |
| result = trashclaw.llm_request_with_retry(MSGS) |
| assert "error" in result |
| assert "Cannot reach" not in result["error"] |
| assert "HTTP 400" in result["error"] |
| assert "exceeds the available context size" in result["error"] |
| assert fake_server["posts"] == 1 |
|
|
|
|
| def test_http_503_is_retried(fake_server): |
| fake_server.update(status=503, ctype="application/json", |
| body=b'{"error": {"message": "Loading model"}}') |
| result = trashclaw.llm_request_with_retry(MSGS) |
| assert "HTTP 503" in result["error"] |
| assert "Loading model" in result["error"] |
| assert fake_server["posts"] == trashclaw.LLM_RETRY_ATTEMPTS + 1 |
|
|
|
|
| def test_error_event_in_stream_is_reported(fake_server): |
| fake_server.update(body=b'data: {"error": {"message": "model crashed"}}\n\n') |
| result = trashclaw.llm_request(MSGS) |
| assert "error" in result |
| assert "model crashed" in result["error"] |
|
|
|
|
| def test_plain_json_reply_is_used(fake_server): |
| fake_server.update(ctype="application/json", body=json.dumps({ |
| "choices": [{"message": {"content": "hello there"}, "finish_reason": "stop"}] |
| }).encode()) |
| result = trashclaw.llm_request(MSGS) |
| assert result["choices"][0]["message"]["content"] == "hello there" |
|
|
|
|
| def test_plain_json_error_is_reported(fake_server): |
| fake_server.update(ctype="application/json", body=b'{"error": "bad model name"}') |
| result = trashclaw.llm_request(MSGS) |
| assert "bad model name" in result.get("error", "") |
|
|
|
|
| def test_normal_stream_still_works(fake_server): |
| chunks = [ |
| {"choices": [{"delta": {"content": "Hel"}}]}, |
| {"choices": [{"delta": {"content": "lo"}, "finish_reason": "stop"}]}, |
| ] |
| body = ": keep-alive\n\n" + "".join("data: %s\n\n" % json.dumps(c) for c in chunks) + "data: [DONE]\n\n" |
| fake_server.update(body=body.encode()) |
| result = trashclaw.llm_request(MSGS) |
| assert result["choices"][0]["message"]["content"] == "Hello" |
| assert result["choices"][0]["finish_reason"] == "stop" |
|
|
|
|
| def test_plain_json_reply_updates_stats(fake_server, monkeypatch): |
| monkeypatch.setattr(trashclaw, "LAST_GENERATION_STATS", {}) |
| monkeypatch.setattr(trashclaw, "SESSION_STATS", dict(trashclaw.SESSION_STATS, turns=0, |
| total_tokens=0, total_seconds=0.0)) |
| fake_server.update(ctype="application/json", body=json.dumps({ |
| "choices": [{"message": {"content": "x" * 40}, "finish_reason": "stop"}] |
| }).encode()) |
| trashclaw.llm_request(MSGS) |
| assert trashclaw.LAST_GENERATION_STATS["tokens"] == 10 |
| assert trashclaw.SESSION_STATS["turns"] == 1 |
|
|
