refactor: improve PDF extraction, error handling, and proxy configuration

app/services/claude_backend.py

@@ -41,7 +41,8 @@ async def call_claude(
         fix_errors: 上一轮验证错误列表（用于重试）
     """
     if session_id is None:
-        session_id = f"claude-summary-{uuid.uuid4().hex[:8]}"
+        # claude CLI 的 --session-id 要求合法 UUID；非 UUID 会被拒绝（Invalid session ID）
+        session_id = str(uuid.uuid4())
 
     cmd = [settings.CLAUDE_BIN, "-p", "--output-format", "text"]
 
@@ -51,22 +52,22 @@ async def call_claude(
     else:
         cmd += ["--session-id", session_id]
 
-    cmd.append(prompt)
-
     logger.info(
         "Calling claude (session=%s, fix=%s)",
         session_id,
         bool(fix_errors),
     )
 
+    # prompt 走 stdin，避免长文本（论文 PDF 动辄数十 KB）超出系统 ARG_MAX
     proc = await asyncio.create_subprocess_exec(
         *cmd,
+        stdin=asyncio.subprocess.PIPE,
         stdout=asyncio.subprocess.PIPE,
         stderr=asyncio.subprocess.PIPE,
     )
     try:
         stdout, stderr = await asyncio.wait_for(
-            proc.communicate(),
+            proc.communicate(input=prompt.encode("utf-8")),
             timeout=settings.SUMMARY_TIMEOUT_SECONDS,
         )
     except asyncio.TimeoutError:
@@ -77,8 +78,10 @@ async def call_claude(
         )
 
     if proc.returncode != 0:
-        raise ClaudeProcessError(
-            proc.returncode, stderr.decode("utf-8", errors="replace")
-        )
+        # claude CLI 把 API 错误输出到 stdout（stderr 常为空），优先用有内容的
+        detail = stderr.decode("utf-8", errors="replace")
+        if not detail.strip():
+            detail = stdout.decode("utf-8", errors="replace")
+        raise ClaudeProcessError(proc.returncode, detail)
 
     return stdout.decode("utf-8", errors="replace"), session_id