| name | sync |
| description | Self-update — read the latest releases across the engine's whole surface (Anthropic, stack, services, practice), diff against the engine, adopt or propose |
| when_to_use | Use when self-updating the engine against upstream releases — reads the Anthropic/stack/services/practice release feeds per engine.json sync tiers, diffs findings against the engine, adopts mechanical wins, proposes the rest, and stamps sync dates. Triggers on: sync, what's new in claude code, refresh the engine, release sweep. |
| argument-hint | [anthropic|stack|services|practice|all] [--force] |
Sync — the engine updates itself
Kun configures the best of what exists rather than rebuilding it — which only works if it always knows what exists. /sync is that loop, across four tiers. Anchor: .claude/engine.json → sync.
The keyword sync in natural conversation ("are we current?", "sync the engine", "anything new in Next.js?") routes here — no slash needed.
Tiers
| Tier | Surface | Cadence | Stamp |
|---|
| anthropic | Claude Code CHANGELOG · platform release notes · anthropic.com/news | 7 days | sync.anthropic |
| stack | Next.js · React · TypeScript · Tailwind · Prisma · shadcn/ui release surfaces | 14 days | sync.stack |
| services | Vercel · Neon · GitHub · Figma changelogs | 14 days | sync.services |
| practice | Agent-config craft: Anthropic engineering blog, AGENTS.md ecosystem, BMAD-method, high-signal community Claude Code configs | 30 days | sync.practice |
Exact URLs live in engine.json → sync.sources — when a source moves, fix it there, not here.
Steps
1. Decide scope
- No argument → every tier whose stamp is older than its cadence (or
null = never synced). Nothing due → report per-tier freshness and stop.
anthropic / stack / services / practice → that tier only.all → every tier. --force ignores staleness.
2. Fetch (parallel WebFetch; gh api for GitHub releases)
For each in-scope tier, extract only what is NEW since the tier's stamp and RELEVANT to the engine:
- anthropic — new tools, config surfaces (settings keys, frontmatter fields, hook events,
.claude/* dirs), models, permissions/sandbox, autonomy/long-horizon features, plan/billing changes
- stack — new versions, breaking changes, deprecations, new idioms for the versions kun's products pin (Next.js 16, React 19, TypeScript 5, Tailwind 4, Prisma 6, shadcn/ui)
- services — platform changes that touch deploys (Vercel), databases (Neon), repo workflows (GitHub), design handoff (Figma)
- practice — agent-config patterns worth stealing: multi-agent methods (e.g. BMAD), CLAUDE.md/AGENTS.md conventions, skill/hook/workflow/subagent techniques
Sources of the form gh:owner/repo mean gh api repos/owner/repo/releases --jq '.[0:5] | .[] | {tag_name, published_at, body}' — cheaper and more structured than fetching release pages.
3. Classify each finding
| Class | Meaning | Action |
|---|
| ADOPT-SAFE | Pure config/docs addition — no behavior risk, no spend, no team-workflow change | Apply now |
| ADOPT-DECIDE | Changes behavior, cost, dependencies, or how the team works — including any product version bump | /decide entry + GitHub issue (type:chore, label engine-upgrade); PushNotification Abdout if Type-1 |
| TRACK | Real but not useful to kun yet | Deferred list in docs/CONFIG-BENCHMARK.md |
Standing constraints veto adoption regardless of class:
- Subscription-only — nothing that requires usage credits or per-token spend (
engine.json → billing)
- Main-only git — nothing that reintroduces branches/worktrees/PRs (
.claude/rules/github-workflow.md)
- Anthropic-native — no third-party substitutes for anything Anthropic ships (
docs/CONFIG-BENCHMARK.md)
Stack findings diff against reality, not memory: product package.json versions + the rule corpus (.claude/rules/<domain>/, since: frontmatter). A breaking change with no matching rule → write the rule (ADOPT-SAFE). A version bump for products → ADOPT-DECIDE, executed via /package.
3.5 Harness audit (on model releases only)
When the anthropic tier surfaces a new model (or a major Claude Code version), run the reverse pass: every rule, hook, gate, and fix-loop in the engine encodes an assumption about what the model couldn't do when it was written. Sample the highest-friction ones (fix loops with retry caps, mandatory verification passes, deny-list breadth) and ask: does the new model still need this scaffold? Propose removals as ADOPT-DECIDE — guardrails get retired, not just added. (Anthropic harness-engineering guidance, adopted 2026-07-10.)
4. Apply + record
- Apply ADOPT-SAFE to canonical sources; regenerate plugins if any plugin source changed (
bash .claude/scripts/build-plugin.sh)
- Stamp each synced tier:
engine.json → sync.<tier> = today; tier anthropic also updates sync.claude_code_version
- Material adoptions → dated line in
docs/CONFIG-BENCHMARK.md → Adoption log
- Commit:
chore(engine): sync <tiers> — <headline> + git pull --rebase origin main && git push origin main
5. Report
One table per synced tier — finding → class → action taken — then per-tier freshness ("anthropic 0d · stack 0d · services 12d · practice 21d").
Cadence
- Weekly:
/weekly Monday plan runs /sync first (due tiers only — usually just anthropic).
- Session start: the captain flags any overdue tier and syncs before allocating.
- Event-driven: a release mentioned in conversation ("did you see the Next.js release?") → sync that tier now.
Exit gate
Every in-scope tier stamped today, every finding classified, ADOPT-SAFE applied and committed, ADOPT-DECIDE escalated with a durable record.
