Sub-agent zoom (focus)
- Dominant language
- Shell
- Stars
- 11.2k
- Forks
- 1.9k
- Avg merge
- 14h 16m
- Merged PRs (30d)
- 6
Description
### Describe the feature or problem you'd like to solve
Ability to make a sub-agent directly observable or interactable.
### Proposed solution
"/focus [|session]" temporarily scopes the session to the context of the specified sub-agent allowing the user to observe it's activities and/or interact with the agent.
The nature of sub-agents means that interaction might be unadvisable, but an ability to checkpoint or rollback an agent would be incredibly helpful.
### Example prompts or workflows
- Sub-agent invokes a command incorrectly and is then consumed with addressing an artificial issue:
- - invokes "python" instead of using the virtualenv / uv, compounds by trying to "pip install", ...
- - invokes "gcc" instead of "clang",
- - tries to submit bash expressions to powershell,
- Sub-agent appears to run for 45+ minutes with no obvious indications of activity, but parent agent says it's "working",
- - allows user to see what it is that it's doing, diagnose a faulty loop
- Sub-agent invokes a unit test that spams output but is actually stuck,
- - parent agents typically see this as motion/progress,
- Sub-agent is performing commands that generate large volumes of text, eviscerating model's Attention,
- - hard to recognize/diagnose when you can't see the sub-agent progress,
- Two sub-agents with overlapping objectives (e.g concurrently run cmake) introducing a non-reproducible effect and one's Attention focuses on diagnosis,
- - user can diagnose/cancel the individual faulty agent,
- - if interaction is possible, user can correct or rewind the agent and avoid losing a long-running context,
### Additional context
_No response_
Contributor guide
Research direction
No files, tests, or entry points are identified. Start by tracing how sub-agent sessions and CLI commands are represented, then assess where a /focus workflow could observe or interact with one agent. Done should include a defined focus, cancellation or rollback behavior, and tests for concurrent or long-running sub-agents.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- shell
- Domain
- ai, cli
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100