feat(trace): record the skills that an agent used in a session #3540

Problem

Chainloop Trace records the skills that are available to an agent, but it does not record the skills that the agent used in a session.

The evidence holds three related items today:

• The ai-agent-config material lists the repository skill files (.claude/skills/*/SKILL.md, .cursor/skills/*/SKILL.md). This list shows the available skills only.
• tools_used.summary in the AI coding session evidence counts calls per tool name. For Claude Code, all skill calls go into one Skill entry. The count does not show which skills the agent called.
• raw_session holds the full transcript. Each Skill tool call holds the skill name in its input (for example {"skill": "superpowers:brainstorming"}). To find the used skills, a policy or the UI must scan the raw transcript.

Proposal

Add a structured list of the skills that the agent used in a session to the AI coding session evidence. For each skill, record the name and the number of calls.

Points for the spec to decide:

• The schema shape, for example skills_used: [{name, invocation_count}].
• How to detect a skill that the user starts with a slash command (for example /graphify). Such a skill can appear as a command marker in a user message, not as a Skill tool call.
• How to include skills that subagents use.
• Support for each provider: Claude Code, OpenCode, Cursor, and Pi. Each agent records skill use in a different transcript format.