BUG: opencode run --agent plan hangs at init, before any session is created #42779

Description

opencode run --agent plan hangs indefinitely. It never creates a session and never sends a request to the model, so it has to be killed.

This looks like a regression of #3828 ("Plan Agent gets stuck in run mode", closed as fixed in 0.15.31), but the failure is earlier and broader than the original report. #3828 needed a prompt that triggered a bash ask permission. This reproduces with a prompt that uses no tools at all, and it hangs before a session exists, so no permission could have been requested yet.

OpenCode version

1.18.18

Steps to reproduce

opencode run --agent plan -m <provider>/<model> "di solo OK"
# hangs forever, no output

Reproduced with a minimal config (provider + model only, no custom agents, no MCP servers, no instructions, no permission overrides), so it is not caused by user customisation.

What the logs show

With --print-logs --log-level DEBUG, it reaches init and stops there:

level=INFO message=bootstrapping directory=/home/macuartin/www/knowledge-base
level=INFO message="all LSPs are disabled"
level=INFO message="all formatters are disabled"
level=INFO message=init
level=INFO message=cleanup prune=7.days      <- 60s later, this is my timeout killing it

A working run with the same config and the default agent gets past init immediately:

level=INFO message=init
level=INFO message=created id=ses_... agent=undefined model=undefined
     permission="[{\"permission\":\"question\",...,\"action\":\"deny\"},
                  {\"permission\":\"plan_enter\",...,\"action\":\"deny\"},
                  {\"permission\":\"plan_exit\",...,\"action\":\"deny\"}]"
level=INFO message="event connected"
level=INFO message=loop session.id=ses_... step=0

So with --agent plan no created line is ever emitted. Note the plan_enter / plan_exit permissions present in the working session; I suspect the plan agent is waiting on one of those in a context where nothing can answer, but I have not read the source to confirm.

What I ruled out

Environment

• Provider: OpenAI-compatible local endpoint (llama.cpp router)
• Model: a local 30B; token counters on the server confirm zero prompt tokens ever arrive, so this is client-side
• OS: Linux

I could not test against a hosted provider, so I cannot rule out that this is specific to a custom OpenAI-compatible provider.