If you're building agents that need to analyze Claude Code, Cursor, or Codex session files, this solves the context window problem. Session JSONLs can balloon to multi-megabyte sizes, so instead of reading the entire file, this extracts either a conversation skeleton (user messages, assistant responses, collapsed tool summaries) or just the error signals. You can optionally limit output with head or tail caps. It's an agent-facing primitive that knows the JSONL shape so your agent doesn't have to. The Socket audit flagged it, which is worth checking given it's processing session data, but it's from the Every engineering team and has solid adoption with 140 installs.
npx -y skills add everyinc/compound-engineering-plugin --skill ce-session-extract --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
shadcn/improve
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills