feat: Agent skills

[lvndry]

Summary

This PR introduces an Agent Skills subsystem (progressive disclosure) and integrates it end-to-end: skill discovery + loading for agents, a user-facing /skills chat command, and safer interactive CLI prompts with Escape cancellation.

What changed

• Skill runtime + DI

  • Added SkillService implementation at src/core/skills/skill-service.ts.
  • Wired SkillsLive.layer into src/app-layer.ts so skills are available wherever ToolRequirements/chat/agent runner are used.
  • Skill discovery:
    • Scans global skills under ~/.jazz with a cache file at ~/.jazz/global-skills-index.json.
    • Scans local skills in the current project (process.cwd()).
    • Merges results with local overriding global by name.
  • Parses SKILL.md frontmatter (name, description) using gray-matter.
  • Uses fast-glob to locate **/SKILL.md with node_modules/.git ignored.
  • Provides loadSkillSection(...) with path normalization and directory-escape protection.

• Agent prompt + execution context

  • Extended AgentPromptOptions and AgentRunContext to include knownSkills metadata.
  • Appends skills guidance + <available_skills> XML to the system prompt when skills exist (src/core/agent/agent-prompt.ts, src/core/agent/prompts/shared.ts).
  • AgentRunner now:
    • Lists skills at start of run (metadata only) and passes them into both prompt builder and tool execution context.
    • Always includes load_skill and load_skill_section in the tool allowlist (even if the agent config omits them).

• Tools: Skill system tools

  • Added src/core/agent/tools/skill-tools.ts implementing:
    • load_skill(skill_name)
    • load_skill_section(skill_name, section_name)
  • Registered tools under a new Skills tool category (SKILLS_CATEGORY) in src/core/agent/tools/register-tools.ts.

• Chat command: /skills

  • Added command parsing and handling for /skills (src/services/chat/commands/parser.ts, src/services/chat/commands/handler.ts).
  • Interactive flow:
    • Lists discovered skills with name + description.
    • Loads selected skill content and renders it via formatMarkdown(...) for terminal-friendly display.

• CLI prompt cancellation / UX hardening

  • Added optional reject to PromptState and Escape handling via useInput in the Ink prompt component (src/cli/ui/types.ts, src/cli/ui/Prompt.tsx).
  • Updated TerminalService.select and TerminalService.search to return T | undefined on cancellation (src/core/interfaces/terminal.ts) and implemented it in src/services/terminal.ts.
  • Updated create/edit agent flows to treat cancelled selections as cancellation instead of proceeding with undefined (src/cli/commands/create-agent.ts, src/cli/commands/edit-agent.ts).

Dependencies

• Updated @openrouter/ai-sdk-provider to ^2.0.0.
• Added fast-glob, gray-matter, fs-extra (and @types/fs-extra) to support skill discovery/parsing.

User-facing behavior

• In chat:
  • /skills lets users browse and view available skills (global + local).
• For agents:
  • Agents are informed about available skills and can load skill instructions on demand via tools.
• In interactive prompts:
  • Escape cancels select/search prompts cleanly; create/edit flows now exit safely when cancelled.

Test plan

• Manual
  • Create a local skill (e.g. skills/example/SKILL.md with name/description frontmatter).
  • Run jazz agent chat <name> and execute /skills; verify skills are listed and selecting one prints formatted content.
  • In jazz agent create / jazz agent edit, press Escape during provider/model/type selections and confirm the flow cancels cleanly.
• Automated
  • Run bun test.