Session hangs indefinitely after MCP tool-call — no timeout recovery (deadlock) #35207

Description

Occasionally an opencode session becomes permanently stuck after an MCP tool-call (most often a playwright_browser_* call) does not return. The process stays alive but is unresponsive.

Observed on a session that had been hung for ~36 minutes:

• Process: alive, 35 threads, ~970 MB working set, CPU drifting near-idle (~0.1 CPU-sec/min).
• TCP connections owned by the opencode PID: 5 Bound, 0 ESTABLISHED — not waiting on any LLM API response.
• No bash/pwsh/cmd child process running a user command — only the MCP server children (playwright, z_ai, context7). So it is not waiting on a shell command either.

It looks like the tool-call state machine is waiting forever for an MCP response that will never arrive. /interrupt and Esc have no visible effect. Only killing the opencode.exe process from another terminal recovers it, and the session is lost.

Workaround I'm now using: disabled the Playwright MCP by default ("enabled": false) and added "timeout": 10000 to each MCP entry — but per the docs that timeout only covers the initial tool-list fetch, not runtime tool-calls, so this is only partial mitigation.

Plugins

None

OpenCode version

1.17.13

Steps to reproduce

Non-deterministic — depends on MCP server flakiness, but has happened multiple times with Playwright MCP enabled.

1. Start opencode with Playwright MCP enabled: true.
2. In a session, ask the agent to use a playwright_browser_* tool against a slow-loading or crashing page.
3. Occasionally the tool-call never returns (browser hang or stdio pipe stall).
4. The session UI shows a spinner forever; /interrupt and Esc do nothing; the conversation cannot continue.

Suggested fixes (any one would unblock users):

• Add a configurable per-tool-call timeout (default e.g. 60s) in the MCP tool executor; on expiry, return a synthetic error to the LLM (tool_call timed out).
• Make /interrupt actually cancel the in-flight tool-call (send cancellation to the MCP server, drop the pending response, restore the session loop).
• Detect MCP child-process health (stdio EOF / heartbeat); if the child crashes or the pipe stalls, fail any pending tool-call for that server with a clear error instead of waiting forever.

Screenshot and/or share link

No response

Operating System

Windows 10 Pro 19045

Terminal

PowerShell 7