メニュー
Use when auditing stale worktrees in Freshell to determine which contain novel unmerged work, which should be landed, which need finishing, and which can be safely deleted.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when preparing a Freshell release — before bumping versions, writing release notes, tagging, etc on GitHub.
Use when reviewing, triaging, or landing pull requests — especially batches of open PRs that need inspection, fixes, and merging. Also triages open issues after the PR queue is clear.
Use when interacting with Freshell panes, panels, or tabs from the CLI for tmux-style automation and multi-pane workflows, outside external-browser automation tasks.
Use when installing, creating, or setting up Freshell extensions — from GitHub repos, local directories, or from scratch as custom panes.
Use when producing screen-recorded demos that need scenario-specific pane layouts, live interaction walkthroughs, and machine-readable timecodes for automated video editing.
Use when producing screen-recorded demos that need scenario-specific pane layouts, live interaction walkthroughs, and machine-readable timecodes for automated video editing.
| name | triage-old-worktrees |
| description | Use when auditing stale worktrees in Freshell to determine which contain novel unmerged work, which should be landed, which need finishing, and which can be safely deleted. |
Use this skill when the repo has accumulated worktrees in .worktrees/ whose branches may or may not have been merged into main. This skill provides a systematic, agent-driven process to classify each worktree and produce a clear action plan.
origin/main is the integration branch (Freshell policy)origin/main when mergedorigin/main and target PRs to main01-baseline → Establish safe-to-ignore criteria & filter worktrees
02-first-pass → One subagent scans all candidate worktrees for novel work
03-second-pass → Parallel deep-dive subagents for meaningful worktrees
04-aggregate → Produce final report (md + csv + html) with verdicts
Each step produces files in the analysis worktree. The final output drives deletion decisions and work-to-land prioritization.
Deploy a subagent to review the codebase and worktree inventory (.worktrees/branch-inventory.json if it exists) and establish criteria for "safe to ignore."
Category A — Auto-Skip / Safe to Delete (no deep dive needed):
origin/main AND working tree is clean AND branch has 0 ahead commitsplan/*, docs/*, proof-*)branch-inventory.jsonCategory B — First-Pass Only (quick inspection):
git diff for lost work)origin/main)plan/*, docs/*, proof-*, port/*, debug/*, test/*, chore/*) AND ahead <= 20Category C — Deep-Dive Required:
feat/*, fix/*, codex/*, freshagent-*, freshcodex-*, freshopencode-*, opencode-*, rollback/* branchesFor each worktree in scope:
git -C .worktrees/<name> rev-parse HEAD
git merge-base --is-ancestor <HEAD> origin/main # YES/NO
git -C .worktrees/<name> symbolic-ref --short HEAD # branch name
git -C .worktrees/<name> log -1 --format=%ci # last commit date
git rev-list --count origin/main..<branch> # ahead count
git rev-list --count <branch>..origin/main # behind count
git -C .worktrees/<name> status --porcelain # working tree check
Filter to the desired time window (e.g., past 4 weeks = date > "2026-05-23").
Write baseline criteria to .worktrees/<analysis-wt>/baseline-criteria.md.
Deploy one fresh (no-context) subagent to run the process over all worktrees in scope. For each worktree, it determines whether the worktree contains novel work not landed on origin/main.
git -C .worktrees/<name> merge-base --is-ancestor HEAD origin/main
git -C .worktrees/<name> status --porcelain
git -C .worktrees/<name> log --oneline origin/main..HEAD # unmerged commits
git -C .worktrees/<name> diff --stat origin/main...HEAD # change scope
first-pass-table.md — one-line-per-worktree table with columns: worktree, branch, date, ancestor?, status, commits, files Δ, meaningful?, summaryworktrees-to-deep-dive.txt — list of worktrees needing second passDeploy multiple fresh subagents in parallel, grouped by topic area (3-4 worktrees per subagent). Each subagent dives deep on its assigned worktrees and produces a verdict.
Ready for landing — work is done, never landed, seems useful. After thorough static analysis, code review, and optional test runs, the work is complete and worthwhile.
Finish work — significant progress towards something useful, but still has bugs, open questions, or integration gaps.
Throw away — useless — superseded by different work, the user/session history made clear this was a mistake or dead end.
Throw away — in main already — the same functionality already landed on origin/main, possibly via a different implementation.
For each worktree, the subagent should:
log origin/main..HEAD, diff origin/main...HEAD, diff origin/main...HEAD --statgit merge-base --is-ancestor, grep for key identifiers in origin/main, compare blob hashesorigin/main for related keywords and implementer commitsnpm run test:vitest -- run <test-pattern> to verify the work still works~/.claude/projects/freshell/sessions/, ~/.codex/, ~/.config/opencode/ for session historyGroup worktrees by topic to minimize context switching within a subagent:
Each deep-dive report goes to .worktrees/<analysis-wt>/deep-dive/<NN>-<topic>.md.
Deploy one subagent to read all deep-dive reports, baseline criteria, and first-pass table, then produce three files:
final-report.mdfinal-report.csvnum,worktree,branch,date,verdict,category,analysisready-landing, finish-work, in-main, skipped-plan, skipped-trivialfinal-report.htmlAfter completing all steps, the analysis worktree contains:
.worktrees/<analysis-wt>/
├── baseline-criteria.md
├── first-pass-table.md
├── worktrees-to-deep-dive.txt
├── deep-dive/
│ ├── 01-<topic>.md
│ ├── 02-<topic>.md
│ └── ...
├── final-report.md
├── final-report.csv
└── final-report.html
The final report tells you: