Problem
After the check-in report surfaces Do Today action items and the user approves executing them, the skill flow (Step 5c) says only:
Present actionable items to the user ... If the user approves, execute the actions
There are no instructions for how to execute. This causes ad-hoc behavior: no task list, no consistent agent structure, inconsistent handling of git operations, and no guidance on cleanup.
This gap was discovered on 2026-05-07 when the user responded to the check-in report with "Do all of them" and the skill had no framework for what that meant operationally.
Proposed additions to Step 5c
When the user approves executing "Do Today" items:
1. Task list first
Create a task entry per action item before spawning any agents.
2. Parallel subagents
Spawn one subagent per task simultaneously (not sequentially).
3. Temp-dir-only for repo access
Any subagent that needs to clone a repo MUST:
- Clone into a temp directory (
mktemp -d)
- Work exclusively in the temp directory
- Delete the temp dir when done (
rm -rf <tempdir>)
Never clone into the user's working directory or any persistent location.
4. Action type handling
| Action type |
How to handle |
| Rebase a PR branch |
Clone fork → add upstream remote → fetch → rebase → force push → clean up |
| Fix PR review blockers |
Clone fork branch → make changes → run tests if available → push → re-request review → clean up |
| Post a comment / tag a maintainer |
Use gh pr comment or gh issue comment directly — no clone needed |
| Watch / monitor items |
No action; note in report only |
5. Force-push authorization
If the user confirms "do all of them" in response to the check-in report's "Do Today" list, that is sufficient authorization to force-push PR branches (rebase workflow). Do not ask again per branch.
6. Subagent model
Follow the standard model cascade (one tier below orchestrator). For complex code-fixing tasks (unfamiliar repo + test infrastructure), override to Sonnet.
7. Report back
Each agent should report: what it did, any conflicts resolved, whether push/comment succeeded, and any blockers.
Related
- Session that triggered this: 2026-05-07 morning check-in
- Concrete examples of each action type executed this session:
Problem
After the check-in report surfaces Do Today action items and the user approves executing them, the skill flow (Step 5c) says only:
There are no instructions for how to execute. This causes ad-hoc behavior: no task list, no consistent agent structure, inconsistent handling of git operations, and no guidance on cleanup.
This gap was discovered on 2026-05-07 when the user responded to the check-in report with "Do all of them" and the skill had no framework for what that meant operationally.
Proposed additions to Step 5c
When the user approves executing "Do Today" items:
1. Task list first
Create a task entry per action item before spawning any agents.
2. Parallel subagents
Spawn one subagent per task simultaneously (not sequentially).
3. Temp-dir-only for repo access
Any subagent that needs to clone a repo MUST:
mktemp -d)rm -rf <tempdir>)Never clone into the user's working directory or any persistent location.
4. Action type handling
gh pr commentorgh issue commentdirectly — no clone needed5. Force-push authorization
If the user confirms "do all of them" in response to the check-in report's "Do Today" list, that is sufficient authorization to force-push PR branches (rebase workflow). Do not ask again per branch.
6. Subagent model
Follow the standard model cascade (one tier below orchestrator). For complex code-fixing tasks (unfamiliar repo + test infrastructure), override to Sonnet.
7. Report back
Each agent should report: what it did, any conflicts resolved, whether push/comment succeeded, and any blockers.
Related