Write/Edit tool hangs 2–4 min (ep_poll, no network/CPU) when a stdio MCP server is connected

Summary: Write/Edit tool calls intermittently hang for 2–4 minutes when an MCP (stdio) server is connected. The file is written to disk instantly, but the tool result returns [Tool result missing due to internal error] and the turn stalls until the user interrupts. A no-MCP session does not hang.

Version: Claude Code 2.1.177 (also seen on 2.1.170). Linux (Bazzite / Fedora atomic), Node-based CLI, running inside tmux.

Symptom:

• A Write or Edit tool call hangs 2–4 min. The target file lands on disk immediately (verified via mtime).
• The tool result comes back as [Tool result missing due to internal error].
• The status line shows Recalling/writing memory… / a long-running spinner during the stall.
• Force-stopping the turn recovers; the next turn proceeds. Intermittent.

Process state captured during a live ~200s hang (sampled every 15s on the claude PID):

stat=Ssl  cpu=4.4%  wchan=ep_poll  :443 sockets=NONE  hot-thread=ep_poll

i.e. the process is parked idle in the Node event loop (ep_poll) for the entire hang — no API/network socket open, ~0 CPU, not in uninterruptible IO, not a futex deadlock. It is awaiting something internal that never arrives, then the tool result fails internally.

Ruled out by that capture: slow API/network (no socket), CPU/compute (flat ~4%), disk/fsync (stat=Ssl, vmstat IO-wait 0%), context-size/cache-miss (would show API socket + CPU), local futex deadlock, low memory/disk (23 GiB free, disk 28%).

Suspected cause: the only child of the claude process is an MCP server (stdio). ep_poll is consistent with the event loop blocked on a pipe fd, i.e. the harness awaiting an MCP-server response/ack on tool completion. The mcp logs show the harness attempting channel notifications (Channel notifications skipped: server did not declare claude/channel capability).

Control test: a session launched with --strict-mcp-config --mcp-config '{}' (no MCP servers; confirmed no MCP child) ran 15 sequential Write tool calls with zero hangs, completing in <60s. The MCP-connected session hangs on single writes. (Bug is intermittent, so this is directional, not absolute.)

Workaround: writing files via a Bash heredoc (cat > file) does not hang; only the Write/Edit tools do — consistent with the stall being in the tool-completion → MCP path, not file IO.

Repro (approx): in a session with a stdio MCP server connected, issue repeated Write/Edit tool calls; intermittently one hangs ~2–4 min with the process idle in ep_poll and the tool result returning an internal error.