[BUG] SDK MCP server tool calls fail with "Stream closed" after ~70s - Python SDK v0.1.47

[waigitNV]

Environment

• Python SDK: claude-agent-sdk==0.1.47
• Python: 3.11
• OS: Linux (Docker, amd64)
• Bundled CLI version: 2.1.70 (from JSONL "version":"2.1.70")
• MCP setup: create_sdk_mcp_server with custom tools

Problem

SDK MCP server tool calls start failing with "Stream closed" after approximately 60-70 seconds of operation, even when tools are actively executing and returning results successfully. Built-in tools (Read, Grep) continue working fine during the same session.

Reproduction

1. Register custom MCP tools via create_sdk_mcp_server
2. Send a query that triggers many sequential MCP tool calls over 60+ seconds (e.g., log analysis workflow with find_errors → Read → Grep → custom tool calls)
3. After ~70s from session start, custom MCP tool calls return:

   {"type": "tool_result", "content": "Stream closed", "is_error": true}

4. Built-in tools (Read, Grep) called in the same session at the same time continue working normally

Evidence from JSONL session logs

Tool call and "Stream closed" timestamps differ by only 1-5ms — the tool never gets a chance to execute:

03:41:24.887  tool_call: mcp__my-server__git_get_commits_diff
03:41:24.892  tool_result: "Stream closed" (is_error=true)    ← 5ms, tool didn't run

03:41:27.619  tool_call: mcp__my-server__git_get_commits_diff
03:41:27.620  tool_result: "Stream closed" (is_error=true)    ← 1ms

03:41:30.747  tool_call: mcp__my-server__git_get_commits_diff
03:41:30.755  tool_result: "Stream closed" (is_error=true)    ← 8ms

Built-in Grep and Read tools called between these failures succeed normally. Only custom MCP tools are affected.

Root Cause

This is the same issue as anthropics/claude-agent-sdk-typescript#114:

The bundled CLI does not reset lastActivityTime when MCP server responses arrive. After the inactivity timer exceeds the stream close threshold (~15s default), subsequent MCP tool calls are immediately rejected with "Stream closed", while built-in tools (which don't go through MCP) are unaffected.

Request

Apply the resetLastActivityTime() fix from #114 to the bundled CLI binary shipped in the Python SDK

Related Issues

• anthropics/claude-agent-sdk-typescript#114 — Same root cause, verified fix, TypeScript SDK
• anthropics/claude-agent-sdk-typescript#41 — Concurrent MCP tool call failures
• anthropics/claude-agent-sdk-typescript#56 — Long messages causing Stream closed

[m13v]

we've hit similar timeout issues when running MCP tools that take longer than expected. in our desktop automation MCP server, some tool calls (like waiting for a page to load before taking a screenshot) can take 30-60 seconds, and the stream connection needs to stay alive during that time. our workaround was adding periodic heartbeat messages from the MCP server during long-running operations. not sure if the Python SDK supports that pattern directly, but keeping the stream active during long tool executions seems to be the core issue.

[m13v]

here's our MCP server where long-running tool calls are common (waiting for UI actions to complete): https://github.com/mediar-ai/mcp-server-macos-use/blob/main/Sources/MCPServer/main.swift - the screenshot + click + wait pattern can take 10-30s per tool call and we had to handle the connection lifecycle carefully.

[qing-ant]

This issue has the same root cause as #730 — wait_for_result_and_end_input() applies a 60-second timeout (CLAUDE_CODE_STREAM_CLOSE_TIMEOUT) that closes stdin prematurely when hooks or SDK MCP servers require the bidirectional control protocol to remain open. The "Stream closed" errors after ~70s are a direct consequence of this timeout.

PR #731 removes the timeout entirely when hooks or MCP servers are active, which should resolve this issue. Closing as duplicate of #730.