AgentSkillsCN

Merge Winners

合并优胜者

SKILL.md
--- frontmatter
name: merge-winners
description: Merge winning experiment branches into main via PRs and clean up all branches
disable-model-invocation: true
allowed-tools: Bash(git *, gh *, python *), Read, Edit, Write, Glob, Grep
argument-hint: [branch1,branch2,...] or "auto" to use check-results report

Merge Winners

Merge winning branches into main via GitHub PRs, then clean up all experiment branches.

Input

Winner branches: $ARGUMENTS

If "auto" is passed, first run the report to determine winners:

bash
python .claude/skills/check-results/check_results.py report

Workflow

Phase 1: Validate

  1. •

    Confirm winners

    • •List the winning branches and their key metrics
    • •List the losing branches that will be discarded
    • •Ask user for confirmation before proceeding
  2. •

    Pre-merge checks

    • •Ensure main is up to date: git checkout main && git pull
    • •For each winner branch, verify it has a finished WandB run
    • •Check for merge conflicts between winners: attempt git merge --no-commit --no-ff <branch> then git merge --abort for each

Phase 2: Code Review Gate

  1. •Review each winner branch
    • •Spawn code-reviewer agent per winner branch (parallel via Task tool with run_in_background: true)
    • •Each agent diffs the branch against main, reads modified files, and checks for bugs, security issues, breaking changes, and missing docs
    • •Collect all review results
    • •If any review has blocker severity issues: stop and ask the user whether to proceed, fix, or discard the branch
    • •Only branches that pass review proceed to merge

Phase 3: Merge via PRs

  1. •

    Post final results to all PRs

    • •For each branch (winner and loser):
      bash
      python .claude/skills/check-results/check_results.py post-pr --branch <branch-name>
  2. •

    Mark winning PRs as ready

    • •For each winner:
      bash
      gh pr ready <pr-number>
gh pr edit <pr-number> --add-label "experiment:winner"
  3. •

    Merge winning PRs one by one

    • •For each winning branch (in order):
      bash
      gh pr merge <pr-number> --merge --subject "Merge <branch-name>: <short description>"
    • •If conflicts occur, Claude must manually resolve them:
      • •git checkout main && git pull
      • •git merge <branch-name>
      • •Read both versions of the conflicting file
      • •Understand the intent of each branch's change
      • •Combine both changes logically (both features should work together)
      • •If two branches modify the same function/block, integrate both modifications
      • •If changes are fundamentally incompatible, stop and ask the user which to prioritize
      • •After resolving: git add <files> and git commit and git push
    • •Pull main after each merge: git pull origin main
  4. •

    Verify merged baseline

    • •git checkout main && git pull
    • •Run python -c "from src.train import train; from src.data import load_dataset; print('OK')" to check imports
    • •Check that main.py entry point is intact
    • •If verification fails, fix the issue and commit before continuing

Phase 4: Close losers

  1. •Close losing PRs
    • •For each losing branch:
      bash
      gh pr close <pr-number> --comment "Experiment did not outperform baseline or other branches. Closing."
gh pr edit <pr-number> --add-label "experiment:loser"

Phase 5: Cleanup

  1. •

    Remove all worktrees and branches

    • •Remove all experiment worktrees:
      bash
      git worktree remove .worktrees/<branch-name> --force
      Repeat for every experiment branch (winners and losers).
    • •Winners' remote branches are already deleted by PR merge
    • •Delete remaining local branches: git branch -d <branch> for each
    • •Delete remaining remote branches: git push origin --delete <branch> for each
    • •Prune stale worktree references: git worktree prune
  2. •

    Archive results

  • •Update docs/baseline-history.md (create if needed) with:
    • •Date of merge
    • •Which branches were merged (winners) with PR links
    • •Which branches were discarded (losers) with PR links
    • •Key metrics of the new baseline
  • •Commit and push this doc

Phase 6: Report

  1. •Summary
    • •Print the new baseline state
    • •List merged branches and their contributions
    • •List discarded branches
    • •Confirm ready for next experiment cycle

Rules

  • •Always ask for user confirmation before merging
  • •Never force-push to main
  • •Merge via gh pr merge — keeps PR history clean
  • •Close losing PRs with explanation comment
  • •Delete ALL experiment branches after merge (clean slate)
  • •Document everything in baseline history with PR links
  • •If merge conflicts are complex or fundamentally incompatible, stop and ask the user
  • •Claude must resolve conflicts manually by reading and integrating both sides — never blindly accept one side
  • •After all merges, run /run-baseline to establish new baseline metrics