Tiny agentic loop with Docker sandbox

Skills

Three new subagent-oriented skills for common agentic workflows:

• explore — question-driven codebase comprehension. Fans clusters to mi -p subagents writing citation-grade summaries at /tmp/mi-explore-<cluster>.md. Enforces verbatim citation preservation in aggregation and forbids freeform file:line references in the prose answer field.
• review — pragmatic diff review with a size gate (≤3 files and ≤200 lines → in-context, otherwise partition by concern). Each partition subagent reports against four axes: correctness, scope, tests, convention.
• refactor — structural changes only. Green-gate via explicit skill("verify") load at step 1; callsite sweep delegated with an AMBIGUOUS: field for false-positive matches (strings, comments, docstrings). One transformation per pass; revert on red rather than patch forward.

Cross-reference cleanup

Skill bodies referencing other skills have been adjusted across debug/plan/tasks/tdd. Small transfers (bg=truthy detach mechanics, kill -0 poll semantics) are inlined directly where they're needed. Larger bodies are invoked via explicit call skill("<name>") and follow its body directives — empirically the model does not chain-load a referenced skill from a naked mention in another skill's body.