M2: Skills system — SKILL.md loading and shell execution

[bug-ops]

Goal

Agent loads SKILL.md instructions, injects them into system prompt, and can execute shell commands generated by the LLM.

Scope

• SKILL.md parser (YAML frontmatter + markdown)
• Skill registry (directory scanning)
• System prompt formatter
• 3 bundled skills
• Shell command detection and execution in agent loop

Acceptance Criteria

• Skills loaded from skills/ at startup
• Skill instructions visible in LLM context
• LLM generates bash commands based on skill instructions
• Commands executed, output fed back to LLM
• Working end-to-end: "list files" -> skill -> command -> result

Child Issues

• Implement SKILL.md parser with YAML frontmatter #12 Implement SKILL.md parser with YAML frontmatter
• Implement skill registry with directory scanning #13 Implement skill registry with directory scanning
• Format skills as XML for system prompt injection #14 Format skills as XML for system prompt injection
• Create 3 bundled SKILL.md files #15 Create 3 bundled SKILL.md files
• Add shell command detection and execution to agent loop #16 Add shell command detection and execution to agent loop

Depends On

#2 (M1)