When Codex runs opencode run "task" via bash, can OpenCode ask clarifying questions and wait for answers? The answer is no.
How opencode run Works
opencode run is explicitly a non-interactive mode. It takes a prompt, processes it, writes the result to stdout, and exits. There is no stdin conversation.
This is by design. From the docs: "Run opencode in non-interactive mode… useful for scripting, automation, or when you want a quick answer without launching the full TUI."
When OpenCode encounters ambiguity:
- Guesses and proceeds, or returns an error
- Cannot ask clarifying questions
- Does not wait for follow-up input
- Exits immediately after producing output
So What Is It Good For?
The value is not conversation but division of labor:
- Codex (expensive, powerful model) handles architecture, planning, review
- OpenCode Go Flash (cheap, fast model) handles localized execution
Codex unilaterally tells Flash "do this" via opencode run and gets back a result. No back-and-forth needed.
The Alternative: opencode serve + HTTP API
You can build a conversational bridge by running OpenCode as a persistent server:
- Start opencode serve in the background
- The server exposes a REST API at localhost:4096
- Send messages via POST /session/:id/message to create multi-turn conversations
- The OpenCode JS SDK (@opencode-ai/sdk) provides a type-safe client for this
With this setup, Codex can use webfetch to communicate with OpenCode over HTTP instead of bash. Each API call maintains session state, enabling real back-and-forth.
However, Codex does not do this automatically. You would need to give Codex explicit instructions or write a wrapper script.
OpenCode's Internal Subagents
OpenCode (in TUI mode) has built-in subagents like @general and @explore. The primary agent delegates tasks to subagents via the task tool. But this is internal to OpenCode and unrelated to Codex.
Bottom Line
- opencode run from Codex via bash: one-shot, no conversation
- opencode serve + HTTP API: multi-turn possible but requires manual setup
- opencode (TUI): full conversation, for direct human use
The current Codex-to-OpenCode pattern works because conversation is unnecessary for the task at hand: cheap models execute, expensive models supervise.
