Bug: opencode run exits before background tasks (explore/librarian) complete #16380

Description

opencode run terminates the process as soon as the LLM emits a text response, even when background tasks (explore agents, librarian agents) are still running. This means any one-shot invocation that triggers background agents returns incomplete results.

Steps to Reproduce

1. Run a one-shot query that triggers background agents:

   opencode run "did I leave any loose ends in my conversations this week?"

2. The LLM launches multiple background explore agents + direct tool calls
3. Direct tool calls complete, LLM emits text like "Let me wait for them to complete"
4. opencode run treats the text emission as session completion and exits
5. Background agents never get collected; output is incomplete

Expected Behavior

opencode run should wait until all pending background tasks have completed (or been cancelled) before considering the session done and exiting.

Actual Behavior

The process exits immediately after the LLM produces a text response, regardless of pending background tasks. The terminal shows the explore agents were launched and some completed, but the LLM never gets to collect their results or produce a final synthesis.

Workaround

Instructing the LLM (via AGENTS.md) to poll/collect all background task results before emitting any final text. This keeps the session alive because opencode run doesn't see a "complete" response until everything is collected. This is fragile — it depends on the LLM obeying the instruction.

Environment

• macOS (Apple Silicon)
• opencode latest (installed via bun)
• One-shot mode via opencode run
• TUI mode is unaffected (user can wait and interact)