# cleanvibe — skill / template update index This page is the canonical, hand-maintained index of every skill, section, and convention that cleanvibe currently ships. Every generated `CLAUDE.md` carries a `## Skills` pointer naming this page; the `cleanvibe-update-check` skill reads it weekly. **How to read this page (v1.14.0+).** As of **v1.14.0** the workflow behaviors ship as standalone **skills** under `.claude/skills/`, not as inlined `CLAUDE.md` prose. Entries are keyed by the cleanvibe version that introduced or revised a skill. If a `.claude/skills//SKILL.md` is older than what's described here, update that file to match the wording here; don't paraphrase. Then bump the *Last cleanvibe update check* date in the repo's `## Skills` section and commit with a message naming which skills were refreshed. (Historical: before v1.14.0 these same behaviors were inlined `CLAUDE.md` sections and the check folded new *sections* into `CLAUDE.md`. The pre-v1.14.0 entries below are kept as a record; their content now lives in the skills.) **Scope.** Skills are vendored into `cleanvibe new`, `convert`, `clone`, and `research` projects. `cleanvibe replicate` projects are a bounded paper-replication workflow with their own definition of done and are not auto-vendored the skill set. The `autonomous-loop` (three-cron) skill is itself self-exempting for replication-style bounded work. --- ## v1.15.0 (2026-06-05) — Copyright fix: replication papers are NEVER committed `cleanvibe replicate` was committing the paper into the generated repo (the extracted arXiv LaTeX `source/`, the clawRxiv `paper.md`, the downloaded URL source), which redistributes a copyrighted work. Fixed: the **entire** `replication_target/` tree is now gitignored (`replication_target/*` + `!replication_target/.gitkeep`) and the paper is only ever populated **locally** by `download_paper.py`. clawRxiv and URL modes gained their own `download_paper.py` (re-fetch from the API / recorded URL); the arXiv scaffolder still downloads before launch but commits nothing. Authors' code added as a submodule now needs `git submodule add -f` (gitignored path; only the gitlink is committed). **If you have an older replication repo:** check whether the paper text was committed (`git ls-files replication_target`). If so, replace your `.gitignore`'s `replication_target/*.pdf`-style lines with `replication_target/*` + `!replication_target/.gitkeep`, then `git rm -r --cached replication_target` (keeping `.gitkeep`) and commit — the paper should never have been in history. (Replication projects are not auto-vendored the skill set, so this is a template/behavior change rather than a skill update.) --- ## v1.14.0 (2026-05-30) — Workflow behaviors become standalone skills The reusable workflow prose that used to be inlined into every generated `CLAUDE.md` is now **six standalone skills**, vendored into each repo's `.claude/skills/` and installable globally at `~/.claude/skills/`. `CLAUDE.md` keeps only a short `## Skills` pointer plus project-specific content. Single source of truth: `cleanvibe/skills.py` (`write_skills()`). The six skills (slug — trigger): 1. **`emergency-stop`** — repeated "stop" / explicit halt demand → kill all running processes, background jobs, and this repo's GitHub Actions runs, then take no further actions until "emergency stop ended". 2. **`cron-is-local`** — user says "cron"/"schedule" → the in-session `CronCreate` tool, local, standing consent; just set it up. 3. **`autonomous-loop`** — starting extensive/large-scale autonomous work → the three-cron playbook (work-loop `3 * * * *`, auto-flush `15 * * * *`, status-report `42 * * * *`) with its full lifecycle. 4. **`queue-driven-workflow`** — any multi-step/planning work → plan into `queue.md` first; the `todo.md` → `queue.md` → `devlog.md` flow; delete-don't-check completion; task-tool mirroring; tests/CI discipline. 5. **`writing-style`** — writing any prose → avoid the self-congratulatory "honest"/"frank"/"candid"/"transparent" move; name failures flatly. 6. **`cleanvibe-update-check`** — session start, weekly → fetch this page and refresh `.claude/skills/` to the latest shipped versions. **Migrating from v1.13.x and earlier:** delete the inlined Workflow Rules, Writing, Cron, three-cron, weekly-update-check, and Emergency Stop sections from your `CLAUDE.md`; add the six skill files under `.claude/skills/` (copy from a freshly-scaffolded project or from cleanvibe's `skills.py`); replace the removed prose with the `## Skills` pointer. The `migrate_repos_to_skills.py` script in the cleanvibe repo automates exactly this. --- ## v1.11.0 (2026-05-26) — Autonomous productivity loop (three-cron playbook) + this update mechanism ### New section: "Autonomous productivity loop — the three-cron playbook" Replaces the prior "Hourly status-report cron for extensive work" section (single cron at the top of the hour) with a generalized **three-cron** playbook that has shown the strongest empirical productivity in the maintainer's own large-scale autonomous sessions. The three crons stagger across the hour and play different roles: 1. **Work-loop cron — `3 * * * *` (hourly at :03)** — the engine. Sync remote (never force-push or `reset --hard`), take the top actionable `queue.md` item or promote one from `todo.md`, hold the hard rails (never fake, never weaken / skip / delete a test, never claim "works" without measuring, name unbuilt things plainly, verify CI not just local), commit + push, one-line report. 2. **Auto-flush cron — `15 * * * *` (hourly at :15)** — the backstop. Commit + push pending work; no empty commits. 3. **Status-report cron — `42 * * * *` (hourly at :42)** — the heartbeat, reporting only, no code changes. **Lifecycle:** - A fresh session **starts all three crons** as the opening queue item. - A mid-session large-scale queue **re-fill kills the running crons** as its first item. - Entering planning mode **disables** the crons. - The last two queue items, always pinned at the tail, **ensure the three crons are running** and then **run the status-report independently** as an end-of-session summary. Migrating from v1.10.x: delete the old single-cron section and the single-cron bootstrap-queue references; add the three-cron section using the wording in `cleanvibe/templates.py` `claude_md()` and `queue_md()`. ### New section: "Check cleanvibe for skill updates (weekly)" The section you're reading from. Every generated `CLAUDE.md` carries a small self-update pointer with three fields: - *Generated by cleanvibe version:* the version that wrote the file. - *Last cleanvibe update check:* the date the check last ran. **Weekly** — if the last-check date is more than 7 days ago, fetch this page and fold in any newer entries. - *Updates source:* `https://cleanvibe.emmaleonhart.com/updates.md`. The check is opportunistic: if `WebFetch` fails (offline, DNS, page down), leave the date alone and try next session. --- ## v1.10.0 (earlier) — Bootstrap queue actually starts the cron The bootstrap queue's opening step actually creates the local cron rather than only describing the kill/restart lifecycle. (Now generalized to the three-cron playbook in v1.11.0.) ## v1.9.0 (earlier) — Hourly status-report cron pinned in tail `## Always last` section pinned at the tail of every generated `queue.md`, restarting the hourly status cron and running a final status report. (Generalized to ensure all three crons in v1.11.0.) ## v1.8.0 (earlier) — Cron jobs and scheduled work LOCAL by default "Cron requests are local and immediate" section. A generic mention of "cron" means the in-session `CronCreate` tool, run locally while the user is away. ## Earlier — Emergency stop mode + writing rules Pinned in `claude_md()` since the first templated CLAUDE.md.