# Long Horizon Audit

> Exercise context compression by requiring many large tool observations before a final report.

- Skill: `cosmtrek/long-horizon-audit` (Agent Skill)
- Install (CLI): `npx skillmds@latest add cosmtrek/long-horizon-audit`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cosmtrek/long-horizon-audit/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: cosmtrek (https://skillmd.com/u/cosmtrek)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/cosmtrek/long-horizon-audit

---


# Long Horizon Audit

For this fixture, follow this exact workflow:

1. Call `chapter_probe` exactly once for each chapter number from 1 through 8, in ascending order.
2. After each tool result, retain the chapter id, the checkpoint code, the risk label, and the next action.
3. After chapter 8 has been observed, call `write` to save `reports/long-horizon-summary.md`.
4. The report must contain:
   - a title,
   - one bullet per chapter using the checkpoint code,
   - a final section named `Compression Probe Verdict`.
5. After the write succeeds, return a short completion note.

Do not return the final completion note before the report has been written.