ACP: opencode acp stays alive after its private server dies and every request fails with a generic -32603 #51716

Description

In 2.0.3, opencode acp runs its own private opencode serve --stdio --port 0 child. If that child dies, the acp process keeps running and answers every later request with the same generic error:

{"code":-32603,"message":"Internal error: Internal service failure","data":{"errorName":"ClientError"}}

session/new, session/prompt, and session/list all fail this way until the client restarts opencode acp. The ACP process never exits, so a client that restarts the agent when its process exits never notices.

The error also can't be told apart from an ordinary request failure. On a healthy service, session/new with a cwd that does not exist returns exactly the same code, message, and data. That leaves clients two options: parse the message text, or send a second request to check whether the service is still alive.

Either of these would let clients recover cleanly:

1. Exit opencode acp (non-zero) when its private service exits, or restart the service in place.
2. Return a distinct error for "the service is unreachable", for example a separate data.errorName or data.service value.

Plugins

None

OpenCode version

2.0.3, driven over ACP (opencode acp) on stdio.

Steps to reproduce

1. Start opencode acp, then send initialize and session/new. Both succeed.
2. Find the child with pgrep -P <acp pid> (opencode serve --stdio --port 0) and kill -9 it.
3. The acp process stays alive. Send session/new, session/prompt, or session/list: each returns the error above.
4. For comparison, on a fresh healthy process, send session/new with cwd: "/does/not/exist". The response is the same error. A session/list sent right afterwards succeeds.

Screenshot and/or share link

N/A (JSON-RPC responses above)

Operating System

macOS 26 (Darwin 25.6.0), arm64

Terminal

N/A (ACP over stdio)