You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[FEATURE]: Silent result collection for background tasks to avoid consuming LLM turns #20460
I have verified this feature I'm about to request hasn't been suggested before.
Describe the enhancement you want to request
Problem
Background task notifications (<system-reminder>) wake up the main agent's LLM every time a subagent finishes. On per-request pricing like GitHub Copilot (1 prompt = 1 premium request), this burns through quota fast.
If I spawn 3 background agents, that's roughly 8 requests total: my original prompt, the agent spawning them, then 3 separate notifications each triggering a full LLM turn to collect results. That's 5x the cost of just running things sequentially, with no real benefit if the main agent has nothing else to do while waiting.
The current model assumes cheap per-token pricing (Anthropic, OpenAI direct). It doesn't work well when each LLM turn costs a premium request.
What I'd like to see
A way to collect background task results without forcing an LLM turn on completion. A few options:
Silent mode — Config option so task completion doesn't inject a system reminder. Results sit in session state until the next natural turn or until the user asks for them.
Batched notifications — Instead of one reminder per completed task, combine all finished results into a single message.
Attach to next turn — Don't wake the LLM at all. When the user sends their next message, attach any pending results as context to that existing turn.
Why this matters
Parallel agent patterns are a big selling point of OpenCode, but they're economically unviable on Copilot/Cursor pricing right now. Users on those platforms end up avoiding run_in_background=true entirely, which defeats the purpose.
The issue title should start with [FEATURE]: as required by the Feature Request template (e.g. [FEATURE]: Silent result collection for background tasks...)
The description appears to be AI-generated (highly structured prose, formatted comparison tables, multiple labeled options A/B/C, tool-by-tool comparisons). AI-generated walls of text are not acceptable per our contributing guidelines. Please rewrite this in your own words, focusing on the core problem and what you'd like to see.
Please edit this issue to address the above within 2 hours, or it will be automatically closed.
This issue might be a duplicate of existing issues. Please check:
changed the title [-]Silent result collection for background tasks to avoid consuming LLM turns on per-request pricing[/-][+][FEATURE]: Silent result collection for background tasks to avoid consuming LLM turns[/+]on Apr 1, 2026
Thanks for updating the issue! It now meets our contributing guidelines. We appreciate the clear description of the problem and the concrete suggestions for how to address it.
Activity
github-actions commentedon Apr 1, 2026
This issue doesn't fully meet our contributing guidelines.
What needs to be fixed:
[FEATURE]:as required by the Feature Request template (e.g.[FEATURE]: Silent result collection for background tasks...)Please edit this issue to address the above within 2 hours, or it will be automatically closed.
This issue might be a duplicate of existing issues. Please check:
If you believe this was flagged incorrectly, please let a maintainer know.
[-]Silent result collection for background tasks to avoid consuming LLM turns on per-request pricing[/-][+][FEATURE]: Silent result collection for background tasks to avoid consuming LLM turns[/+]github-actions commentedon Apr 1, 2026
Thanks for updating the issue! It now meets our contributing guidelines. We appreciate the clear description of the problem and the concrete suggestions for how to address it.
github-actions commentedon Jul 4, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.