Bug: session ends after MCP tool calls when using Anthropic-compatible gateways #11417

Description

When using the Anthropic provider behind an Anthropic-compatible gateway/proxy, MCP tools can run successfully but the chat stops without producing the final assistant reply.

What happens

• The model triggers one or more MCP tools
• The tools complete successfully
• The session loop exits right after the tool parts, so the model never gets a follow-up turn to read tool results and produce the final text response

What I expected

• If the latest assistant message contains tool parts, the session should continue until the model outputs a final text reply.

Why this seems to happen
Some gateways/proxies may return finish_reason=stop even though the assistant message still contains tool parts (sometimes even text + tool in the same message). OpenCode treats this as “done” and exits early.

How to reproduce (high level)

• Configure Anthropic provider via an Anthropic-compatible gateway/proxy
• Enable any MCP server(s)
• Ask the model to call two MCP tools (serial or parallel) and then output a single final line (e.g. “DONE”)
• Observe: tools run, but the final line is never produced

Notes

• OpenAI provider does not show this behavior in the same setup.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response