编排自定义AI智能体框架

The one write-capable skill in the plugin. Every other skill is pure-read. This one calls metaharness new, which writes a new directory tree.

Safety (load-bearing)

1. Dry-run by default. Without --confirm, the script prints what it would do and exits 0 without touching disk.
2. Refuses project root. If --target resolves to the current working directory OR any path inside it, the script errors out with exit 2. Target must be an absolute path OUTSIDE the calling repo (default is a fresh /tmp/ruflo-mint-<ts>-<name>/ dir).
3. Refuses existing target. Won't overwrite — must scaffold into a non-existent dir.
4. Subprocess + 60s timeout. No library import, no in-process execution. The mint stays sandboxed from ruflo's runtime.

Algorithm

Implementation: scripts/mint.mjs.

1. Validate --name, --template. Default --host to claude-code.
2. Resolve --target (default: temp dir).
3. Run safety checks (no project-root writes; target must not exist).
4. Without --confirm: emit dry-run plan, exit 0.
5. With --confirm: shell npx metaharness new <name> --template <id> --host <id> --target <abs> --yes.

Templates

minimal, vertical:coding, vertical:devops, vertical:support, vertical:legal, vertical:research, vertical:trading, vertical:health, vertical:education, vertical:sales, vertical:business, vertical:crm, vertical:marketing, vertical:advertising, vertical:ai, vertical:agentics, vertical:ruview, vertical:gaming, vertical:repo-maintainer, vertical:exotic.

Hosts

claude-code, codex, pi-dev, hermes, openclaw, rvm, copilot, opencode, github-actions.

Example dry-run

$ node scripts/mint.mjs --name my-harness --template vertical:coding --host claude-code
# harness-mint (dry-run)

- action: metaharness new
- name: my-harness
- template: vertical:coding
- host: claude-code
- target: /tmp/ruflo-mint-1718560000-my-harness
- confirm: false
- willWrite: false

Re-run with `--confirm` to actually scaffold.

Why dry-run by default

Ruflo's behavioral rules say "executing actions with care" — destructive or repo-touching actions need confirmation. The dry-run output makes the WHAT visible before the WHEN. A human sees target, decides, then adds --confirm if happy.