Description
When a subagent's LLM stops responding (gets stuck — happens with various models/providers, especially on long-running tasks), there is no way to nudge or retry within the existing subagent session.
The main-agent workaround doesn't work for subagents:
For the main agent, when the LLM stops responding, you can press Esc Esc to cancel and then type cont — the agent resumes with full context. This is a well-known and effective workaround.
For subagents, this doesn't work:
- Esc Esc cancels the stuck subagent
- Typing
cont causes the main agent to interpret the cancellation as "subagent is done"
- The main agent spawns a brand new subagent with zero context of the previous one's work
All accumulated tool calls, file reads, and reasoning from the stuck subagent are lost. The parent agent has no access to the partial results.
This is especially painful when the subagent has been running for a long time (10+ minutes of exploration) and gets stuck near the end.
Plugins
None
OpenCode version
1.17.7
Steps to reproduce
- Launch a subagent that does substantial work (e.g., multi-file exploration with many tool calls)
- Wait for the subagent to get stuck (LLM stops responding — reproduces with models that have connection issues or rate limiting)
- Press Esc Esc to cancel the stuck subagent
- Type
cont to try to resume
- Main agent creates a new subagent from scratch — previous subagent's session and work are inaccessible to both the user and the parent agent
Expected behavior
At minimum one of:
- Nudge/retry: ability to retry the last LLM call within the same subagent session, preserving full context
- Resume after cancel: cancelled subagent sessions should be viewable (related: cancelled subagents can't be opened in TUI) and their partial results should be available to the parent
- Explicit restart: ability to restart a subagent within its existing session/context rather than spawning a fresh one
Screenshot and/or share link
N/A
Operating System
macOS
Terminal
TUI
Related issues
Description
When a subagent's LLM stops responding (gets stuck — happens with various models/providers, especially on long-running tasks), there is no way to nudge or retry within the existing subagent session.
The main-agent workaround doesn't work for subagents:
For the main agent, when the LLM stops responding, you can press Esc Esc to cancel and then type
cont— the agent resumes with full context. This is a well-known and effective workaround.For subagents, this doesn't work:
contcauses the main agent to interpret the cancellation as "subagent is done"All accumulated tool calls, file reads, and reasoning from the stuck subagent are lost. The parent agent has no access to the partial results.
This is especially painful when the subagent has been running for a long time (10+ minutes of exploration) and gets stuck near the end.
Plugins
None
OpenCode version
1.17.7
Steps to reproduce
contto try to resumeExpected behavior
At minimum one of:
Screenshot and/or share link
N/A
Operating System
macOS
Terminal
TUI
Related issues