Tiny agentic loop with Docker sandbox

Changes since 1.1.0

Skill corrections

• delegate: dropped unenforceable "no nesting" clause; replaced broken wait <pid> with kill -0 polling (detached children can't be waited on from a later bash invocation).
• debug: moved tdd hand-off into step 4 so verify no longer runs on unwritten code.
• verify: removed hardcoded per-stack commands. Now directs the agent to repo docs (AGENTS.md / CLAUDE.md / README) → declared build metadata → CI config, and forbids invented defaults.
• plan / tasks: state moved from fixed /tmp/mi-{plan,tasks}.md to per-task /tmp/mi-<slug>/ so concurrent sessions and parallel delegated subprocesses don't collide.
• tasks: mandates full-file rewrites via quoted heredoc; explicit ban on sed/awk so special chars in task titles can't corrupt state.

System prompt

• Added a single-line evidence clause: "When you declare a task done, your final message must include the actual command output that proves it — not a summary of what you did. Unreproduced work is unfinished work." Framing borrowed from Droid/Factory.ai's "evidence" convention.

Published artifacts

• index.mjs and skills/ — unchanged structure, updated content.