yctimlin Agent Usage Analyzer
Goal
Produce an accurate, local, read-only report of structured skill, tool, command, and MCP usage across agent sessions.
Inputs to infer from the user's request
- agent or source filters
- since and until date filters
- repo or cwd filters
- raw versus canonical mode
- kind filter
- output format
- whether the user supplied an export directory
- whether samples, partial sessions, or heuristic parsing are allowed
Resources
- Analyzer script:
scripts/analyze.mjs - Install helper:
scripts/install.mjs - Output contract:
references/OUTPUT.md - Alias example:
references/ALIASES.example.yaml
Workflow
- Do not count usage manually from transcript prose.
- Prefer running
scripts/analyze.mjs. - Use
--provider autounless the user supplied an explicit export directory. - If the user asks for "skill usage", "agent skills", or named skill counts, prefer
--kind skill. - For Codex and Claude Code event-log exports, count skill usage from structured
.../skills/<name>/SKILL.mdload signals rather than generic transport tools. - If the source only exposes tool or MCP events and does not expose any structured skill-load signal, say so explicitly instead of presenting tool counts as skill counts.
- If the analyzer reports missing prerequisites, explain what is missing instead of guessing.
- Present the result as a markdown table unless the user asked for JSON or CSV.
- Always state coverage and caveats.
Output rules
- Treat
skill,tool,command, andmcpas separate categories unless the user asks for a broader combined report. - Default to canonical mode.
- Use raw mode only when the user asks for raw names.
- Keep samples off unless the user asks for them.
- Sort rows by invocation count descending.
- Never imply skipped sessions were counted.
- When reporting skill usage from Codex or Claude Code archives, note that counts come from structured skill-load records, typically reads or prompts that reference
SKILL.md.