Summary
Active subagent cards show the current tool activity, but not the size of the working-tree change being produced. For implementation tasks, a compact live summary such as 2 files · +82 −16 would make the card as informative as the change indicators in Claude Code/Codex, without requiring the subagent to report it in its prompt or final response.
Proposed behavior
- While an agent is running, automatically refresh a compact working-tree diff summary after tool activity settles, rendered alongside the existing activity line.
- Example:
⎿ editing 2 files · +82 −16
- Calculate from the agent's
cwd; debounce refreshes so rapid edits do not spawn a Git command per tool event.
- Degrade quietly when the cwd is not a Git worktree or Git cannot provide a summary.
Scope / caveat
The indicator is a working-tree summary, not a perfect attribution ledger: pre-existing changes in a shared checkout can be included. Worktree-isolated agents naturally avoid that ambiguity. The feature should not parse model text, require a builder prompt convention, or change Git state.
Implementation and verification
I plan to implement this with a focused regression test for the activity/card update, then run:
npm run lint
npm run typecheck
npm run test
npm run build
I noticed open PR #181 also changes the activity-tracker path; I will rebase/adjust against the current master rather than modify or depend on that PR.
Summary
Active subagent cards show the current tool activity, but not the size of the working-tree change being produced. For implementation tasks, a compact live summary such as
2 files · +82 −16would make the card as informative as the change indicators in Claude Code/Codex, without requiring the subagent to report it in its prompt or final response.Proposed behavior
⎿ editing 2 files · +82 −16cwd; debounce refreshes so rapid edits do not spawn a Git command per tool event.Scope / caveat
The indicator is a working-tree summary, not a perfect attribution ledger: pre-existing changes in a shared checkout can be included. Worktree-isolated agents naturally avoid that ambiguity. The feature should not parse model text, require a builder prompt convention, or change Git state.
Implementation and verification
I plan to implement this with a focused regression test for the activity/card update, then run:
npm run lint npm run typecheck npm run test npm run buildI noticed open PR #181 also changes the activity-tracker path; I will rebase/adjust against the current
masterrather than modify or depend on that PR.