Fix MiniMax MCP startup handling

app/ai/image_understand.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 
 import asyncio
+from collections import deque
+from contextlib import suppress
 import json
 import os
 from dataclasses import dataclass
@@ -107,6 +109,9 @@ class MinimaxImageUnderstandClient:
         self._request_id = 0
         self._stdout: asyncio.StreamReader | None = None
         self._stdin: asyncio.StreamWriter | None = None
+        self._stderr: asyncio.StreamReader | None = None
+        self._stderr_task: asyncio.Task[None] | None = None
+        self._stderr_tail: deque[str] = deque(maxlen=20)
         self._lock = asyncio.Lock()
 
     async def __aenter__(self) -> "MinimaxImageUnderstandClient":
@@ -129,7 +134,7 @@ class MinimaxImageUnderstandClient:
                 *self.config.command,
                 stdin=asyncio.subprocess.PIPE,
                 stdout=asyncio.subprocess.PIPE,
-                stderr=asyncio.subprocess.DEVNULL,
+                stderr=asyncio.subprocess.PIPE,
                 env=env,
             )
         except FileNotFoundError as exc:
@@ -138,25 +143,32 @@ class MinimaxImageUnderstandClient:
             ) from exc
 
         if self._process.stdin is None or self._process.stdout is None:
+            await self._terminate_process()
             raise MCPError("MiniMax MCP process did not expose stdio pipes")
 
         self._stdin = self._process.stdin
         self._stdout = self._process.stdout
+        self._stderr = self._process.stderr
+        self._stderr_task = asyncio.create_task(self._drain_stderr())
         request_timeout = self.config.timeout if self.config.timeout is not None else 120.0
 
-        await self._request(
-            "initialize",
-            {
-                "protocolVersion": "2024-11-05",
-                "capabilities": {},
-                "clientInfo": {
-                    "name": "multi-bot-tg",
-                    "version": "1.0.0",
+        try:
+            await self._request(
+                "initialize",
+                {
+                    "protocolVersion": "2024-11-05",
+                    "capabilities": {},
+                    "clientInfo": {
+                        "name": "multi-bot-tg",
+                        "version": "1.0.0",
+                    },
                 },
-            },
-            timeout=request_timeout,
-        )
-        await self._notify("notifications/initialized", {})
+                timeout=request_timeout,
+            )
+            await self._notify("notifications/initialized", {})
+        except Exception:
+            await self._terminate_process()
+            raise
 
     async def close(self) -> None:
         process = self._process
@@ -186,6 +198,8 @@ class MinimaxImageUnderstandClient:
 
         self._stdin = None
         self._stdout = None
+        self._stderr = None
+        await self._stop_stderr_task()
 
     async def list_tools(self) -> list[dict]:
         if self._process is None:
@@ -274,7 +288,13 @@ class MinimaxImageUnderstandClient:
 
             if timeout is None:
                 return await _wait_for_response()
-            return await asyncio.wait_for(_wait_for_response(), timeout=timeout)
+            try:
+                return await asyncio.wait_for(_wait_for_response(), timeout=timeout)
+            except TimeoutError as exc:
+                raise MCPError(
+                    f"Timed out waiting for MiniMax MCP '{method}' response after {timeout:.0f}s"
+                    f"{self._format_stderr_tail()}"
+                ) from exc
 
     async def _write_message(self, payload: dict[str, Any]) -> None:
         if self._stdin is None:
@@ -289,10 +309,11 @@ class MinimaxImageUnderstandClient:
         if self._stdout is None:
             raise MCPError("MiniMax MCP stdout is not available")
 
-        header_block = await self._stdout.readuntil(b"\r\n\r\n")
+        header_block = await self._read_header_block()
         headers = header_block.decode("ascii", errors="ignore").split("\r\n")
         content_length = None
         for header in headers:
+            header = header.strip()
             if header.lower().startswith("content-length:"):
                 content_length = int(header.split(":", 1)[1].strip())
                 break
@@ -302,6 +323,70 @@ class MinimaxImageUnderstandClient:
         payload = await self._stdout.readexactly(content_length)
         return json.loads(payload.decode("utf-8"))
 
+    async def _read_header_block(self) -> bytes:
+        if self._stdout is None:
+            raise MCPError("MiniMax MCP stdout is not available")
+
+        buffer = bytearray()
+        while True:
+            chunk = await self._stdout.read(1)
+            if not chunk:
+                raise MCPError(f"MiniMax MCP stdout closed before a response{self._format_stderr_tail()}")
+            buffer.extend(chunk)
+            if buffer.endswith(b"\r\n\r\n"):
+                return bytes(buffer)
+            if buffer.endswith(b"\n\n"):
+                return bytes(buffer).replace(b"\n", b"\r\n")
+            if len(buffer) > 64 * 1024:
+                raise MCPError("MiniMax MCP response headers exceeded 64KB")
+
+    async def _drain_stderr(self) -> None:
+        if self._stderr is None:
+            return
+        while True:
+            line = await self._stderr.readline()
+            if not line:
+                return
+            text = line.decode("utf-8", errors="replace").strip()
+            if text:
+                self._stderr_tail.append(text)
+
+    def _format_stderr_tail(self) -> str:
+        if not self._stderr_tail:
+            return ""
+        return "\nMiniMax MCP stderr tail:\n" + "\n".join(self._stderr_tail)
+
+    async def _stop_stderr_task(self) -> None:
+        task = self._stderr_task
+        self._stderr_task = None
+        if task is None:
+            return
+        task.cancel()
+        with suppress(asyncio.CancelledError):
+            await task
+
+    async def _terminate_process(self) -> None:
+        process = self._process
+        self._process = None
+        if process is None:
+            return
+
+        try:
+            process.terminate()
+        except ProcessLookupError:
+            pass
+        try:
+            await asyncio.wait_for(process.wait(), timeout=5)
+        except Exception:
+            try:
+                process.kill()
+            except ProcessLookupError:
+                pass
+        await self._stop_stderr_task()
+        self._stdin = None
+        self._stdout = None
+        self._stderr = None
+
     def _extract_text(self, result: Any) -> str:
         if isinstance(result, dict):
             content = result.get("content")

app/ai/service.py

@@ -336,16 +336,17 @@ class AIService:
     async def _fetch_mcp_tool_schemas(self) -> list[dict]:
         if self._mcp_tools_fetched:
             return self._mcp_tool_schemas or []
-        self._mcp_tools_fetched = True
         if self.config.image_understand is None:
             log.info("MCP tools: image_understand not configured, skipping")
             self._mcp_tool_schemas = []
+            self._mcp_tools_fetched = True
             return []
         log.info("MCP tools: fetching tool list from %s", self.config.image_understand.command)
         try:
             async with MinimaxImageUnderstandClient(self.config.image_understand) as mcp_client:
                 raw = await mcp_client.list_tools()
             self._mcp_tool_schemas = mcp_tools_to_openai_schemas(raw)
+            self._mcp_tools_fetched = True
             names = [s["function"]["name"] for s in self._mcp_tool_schemas]
             log.info("MCP tools: loaded %d tool(s): %s", len(names), ", ".join(names) or "(none)")
         except Exception: