Tiny agentic loop with Docker sandbox

Skills overhaul

All 9 bundled skills improved across 12 iterations:

• Sharper descriptions — every skill now leads with concrete trigger phrases so the model selects the right one
• Cross-references — debug→tdd, tdd→refactor, plan↔tasks wired up at decision points
• Red Flags sections — added to debug, tdd, verify, refactor, review so agents know when to stop and ask
• Concrete examples — review FINDINGS format, debug observation template, verify per-tool output patterns
• File-writing guidance — debug and tdd now warn against echo+\n and sed for multi-line files
• Portability — harness exports MI_PATH; delegate skill uses it instead of hardcoded paths
• plan — explicit /tmp/mi-/plan.md path with copy-pasteable command
• tasks — explicit work loop (pick → mark in-progress → do → mark done); slug collision guidance
• delegate — log-reading strategies (tail, grep, wc); blocked-subagent handling in review

Fixes

• Fixed test suite deleting bundled skills/ directory after test runs
• Harness now sets MI_PATH env var so subagents can spawn further subagents
• .gitignore added; .npmignore tightened