Skill Call Extract

Leading word: the trace index — one line per session entry (role, tool calls with arg sizes, result sizes, per-turn usage). The compact map of a multi-MB session log; build it once, navigate the raw file only where it points.

Two retrieval modes:

1. By skill name — two signals, either qualifies a session:
   • the skill-injection marker (automatic trigger), or
   • a read-load: a read tool call whose path argument ends in /<name>/SKILL.md (the agent manually loaded the contract — the marker cannot see this; matches both the installed layout and source checkouts)
2. By session id — resolve an id the user already has (reports even without either signal; the user already knows it ran the skill)

Process

1. Resolve inputs

Skill name (as written in its SKILL.md frontmatter) for marker mode; session id or filename prefix for direct mode. Both given → id mode, but record the marker count anyway. Neither → ask; do not guess a skill name.

2. Search