Description

When a provider ends a turn with a content filter (e.g. Anthropic stop_reason: refusal, which the protocol layer maps to the content-filter finish reason), the assistant message records the finish reason but no error is attached. If the turn produced no visible output (only empty reasoning + finish), the session goes idle with nothing rendered — it looks like opencode just stopped responding.

Hit this 3 times in one session on an ordinary code-review task with an Anthropic model. Related but distinct from #30411 (that one is empty turns under context pressure; this is specifically the content-filter finish path).

Plugins

none

OpenCode version

1.17.3

Steps to reproduce

1. Run a prompt where the provider's safety classifier fires (Anthropic returns stop_reason: refusal)
2. The turn ends with finish reason content-filter and no text/tool parts
3. Session goes idle with no error box, no notification, nothing visible

Operating System

macOS

Terminal

Ghostty