Support Agent Skills

The specification is here: https://agentskills.io/specification

Basically we need:

* Read all the skills files and inject frontmatters as system prompt
* pass map of tools name? (users pass `allowed-tools` but the tool names should not directly exist in our agents' tools.)

When agent runs, agent selects an appropriate tool and then read corresponding SKILL.md using bash tool.

Start by reading the linked Agent Skills specification and tracing the agent's tool-selection and execution entry points. Determine how skills files, frontmatter, allowed-tools, and SKILL.md reads should interact, then define tests that verify the selected skill is loaded through the bash tool and that tool names are mapped correctly.

Written by the indexing model from the issue text.