sync-agenticat

# Sync Agenticat Harvest agents and skills from an external source, keep only what clears this repo's quality bar, rewrite to its conventions, route each keeper to its home, and **propose before writing**. ## Destinations | Pick kind | Home | Rule | |---|---|---| | Agent worth shipping | `agents/<name>.md` | rewritten as a thin implementer, never copied verbatim | | Knowledge worth a first-party skill | `skills/<name>/SKILL.md` | only as an original distillation authored here; declares `metadata.version` | | Good third-party skill as-is | README link | `npx skills add <owner>/<repo> --skill <name>` (drop `--skill` for single-skill repos); **never vendored** | | Curation-process rule or gotcha | `CLAUDE.md` learnings / `docs/publish-backlog.md` | | | Everything else | drop | listed in the drop list, not silently skipped | Routing test: *does it act (method + gates + output contract)* → agent. *Is it knowledge an agent or session loads* → skill or README link, split by authorship. *Is it about how this repo curates* → learnings. ## Pipeline 1. **Resolve the source** (`$1`): local path in place; `owner/repo` → github URL; git URL → shallow clone to a temp dir (remove at the end); raw URL → fetch. A trailing free-text arg (`agents only`, `just shell stuff`) narrows the harvest — honor it. 2. **Harvest.** Agent defs in any harness format (`agents/*.md`, `.claude/agents/`, opencode / codex / gemini dirs — same content is usually duplicated, read one canonical copy),...