race-worktrees
Race independent droid exec candidates in separate worktrees and combine their strongest parts. Use for competing designs; use Missions for multiple features or Spec mode for one plan.
Cynnamoroll · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add Cynnamoroll/factory-stack --skill race-worktreesAbout this skill
name: race-worktrees description: "Race independent droid exec candidates in separate worktrees and combine their strongest parts. Use for competing designs; use Missions for multiple features or Spec mode for one plan." license: MIT compatibility: droid disable-model-invocation: true
Race worktrees
Race isolated candidates, pick a base, and graft their best parts. Before launching, open a todo list with one item per phase.
1. Frame
Give every candidate the same prompt.
- State the artifact each candidate produces.
- Write a rubric of 3 to 6 gradeable criteria for this task. Candidates do not see it.
- Pick N, usually 2 to 4. Use more when the design directions differ widely.
- Commit or stash your own work first. Candidates branch from
HEAD.
2. Fan out
Write the task to a file, then start every candidate at once from the repo root:
for i in $(seq 1 "$N"); do
droid exec -m auto --auto medium --worktree "race-$i" -o json \
-f /tmp/race-task.md > "/tmp/race-$i.json" 2>&1 &
done
wait
The task file tells each candidate to produce the artifact, leave its changes uncommitted in its worktree, and end with a short rationale naming the alternatives it rejected. droid exec removes a worktree that is clean on exit, so uncommitted changes keep it. Find each worktree with git worktree list.
If a run exits non-zero or produces nothing, continue with N-1 and note the dropout. If every run fails on authentication, stop and ask the user to export FACTORY_API_KEY in their shell. Never write a key to a file.
3. Judge
Issue one explorer Task as the judge. Give it the rubric and each worktree path by label. It scores each criterion per candidate and recommends a base. Read every candidate end to end yourself while it runs.
4. Pick a base
Score every candidate on every criterion. Agreement with the judge confirms the pick; otherwise read both rationales for bias or an unclear rubric. Prefer a base maintainers can extend without breaking invariants. Break ties with the smaller interface.
5. Graft
Walk each losing candidate once more for the one or two things worth porting. Fold each in by hand so the result reads as one design. If the candidates converge, ship the shared shape. If they diverge wildly, the frame was underspecified, so reframe and rerun.
6. Verify and clean up
Run prove-it-works on the merged result in the main checkout. If it finds a problem the race missed, go back to Frame or Graft. Then remove the candidate worktrees with git worktree remove <path> and git worktree prune.
Done when
The reply names the base and why, the judge's verdict, each graft with its source candidate, what was rejected, any dropouts, and the prove-it-works result. No race-* worktree remains.