Looper — ideas

Grouped by size. small = an afternoon, medium = a day or two, large = a real project.

Small

  1. Handle merge conflictsWe already fetch mergeable. If a PR is CONFLICTING, send the agent a "rebase onto main and fix conflicts" round instead of leaving it stuck.
  2. Requeue from the TUIOne key to send a parked task back to the queue, so you don't have to touch the database or GitHub labels.
  3. Priority labelsPick up issues labelled p0/p1 first instead of oldest-first.
  4. Daily spend capThere's a per-issue cap already. Add a total-per-day cap so a bad day can't burn the budget.
  5. Notify when something needs youDesktop notification (notify-send) or a webhook when a PR goes ready or parks.

Medium

  1. Listen to human comments tooRight now only Greptile drives revisions. Feed comments from people on the PR into the next round.
  2. Switch to a stronger model when stuckStart on Sonnet; after two rounds with no score gain, retry on Opus.
  3. Ask before guessingIf an issue is too vague, the agent posts a question on it and parks as needs-info instead of writing the wrong thing.
  4. Plan first, code after 👍Optional mode: agent posts a short plan on the issue and waits for approval before opening a PR.
  5. Run tests before pushingHave the agent run the repo's test command locally before each push, saving slow CI round-trips.
  6. Clean up stale workClose parked PRs and delete their branches and worktrees after N days.
  7. Open a task's transcript in the TUIPress a key to read what the agent actually did on the selected task.

Large

  1. Stats pageSuccess rate, rounds to 5/5, cost per merged PR, over time. Tells you if Looper is worth it.
  2. Respect issue dependenciesDon't start issue B while it says "blocked by #A" and A isn't merged.
  3. Split big issuesAgent breaks a large issue into smaller sub-issues and works them one by one.
  4. Several repos at onceOne daemon watching a list of repos, with a repo column in the TUI.