subagent: mimo-v2.5 stuck in infinite thinking loop (drains credits) #42923

Summary

When mimo-v2.5 (opencode-go/mimo-v2.5) is configured as a subagent model, spawned subagents enter an endless "thinking" loop that keeps emitting identical thinking output and never completes. The request runs indefinitely, never errors/times out, and drains Command Code credits ("Guthaben") — the model is unusable for subagent workloads.

Environment

• opencode version: opencode2 v0.0.0-next-17403
• OS: Darwin 25.6.0 (darwin arm64)
• Terminal: TERM_PROGRAM=Apple_Terminal, TERM=xterm-256color, COLORTERM=truecolor
• Shell: /bin/zsh
• Install/channel: next
• Active plugins: None found in config

Reproduction

1. Configure model opencode-go/mimo-v2.5 as default and for the subagent agents (e.g. experimental.subagent_depth: 2, all agents → mimo-v2.5).
2. In the TUI, trigger any flow that spawns a subagent.
3. The subagent starts "thinking" and never returns — it repeatedly emits the same thinking content.
4. No error/timeout occurs; the request runs forever and consumes credits.

Expected Behavior

The subagent should finish thinking and return a response, or fail with a clear error/timeout instead of looping.

Actual Behavior

The subagent's thinking indicator stays active forever, continuously producing the same thinking output. No completion, no error message. Credits are drained while it loops.

Additional Context

• All agents set to opencode-go/mimo-v2.5 in ~/.config/opencode/opencode.jsonc (plan, general, build, explore, title, summary; vision/document use gpt-5.6-luna).
• Observed in the TUI; not tied to a specific prompt — happens whenever a subagent uses mimo-v2.5.
• No stack trace / error text; the process just thinks perpetually.
• Suggested guard: a max-thinking/iteration timeout or reasoning-loop detector to prevent unbounded credit drain.