Skills, subagents, profiles

Skills as short playbooks that attach automatically, plus sub-agents and profiles.

Skills are short playbooks (skills/*.md) with triggers. Matching ones attach automatically (keyword match, no model call); the model can pull any other with use_skill(name) or one section with use_skill(name, section=...). Every skill is listed with a one-line "use when" so the model can choose. A playbook already in the conversation is named, not pasted again.
- Always on: ponytail (do the least that fully works), grounded, house-rules.
- On match: debugging, code-review, security, refactor, testing, performance, database, api-design, git-workflow, review-pr, review-changes, review-commits, bug-hunt, scalability, typescript, feature-planning.
- Skills are read from ~/.zerd/skills first, then Claude Code's, agents and pi's folders. Turn one off with disabled_skills.
Subagents are declared in a profile ([[subagents]]: name, description, goal_prompt, allowed_tools, role). role picks a routing chain, so exploration runs cheap and debugging runs strong. Sub-agents start no MCP servers, can't spawn more, and cap answers at 6000 chars.
Repo profiles (profiles/<name>.toml): per-project instructions, forbidden_paths, forbidden_commands, verify_rules, routing, enforced in the tool layer. Auto-selected by path, or profile = "name". profiles/tunyt.toml is a worked example.