llm-trading-agent-security

Self Mutate

Overview

Use this when you need to change SOUL.md, USER.md, or files under skills/ in the live agent root. From the conversation workspace, those live root paths are typically ../../SOUL.md, ../../USER.md, and ../../skills/.... Protected changes must follow diff -> confirm -> snapshot -> write.

Rules

• Never modify AGENTS.md or anything under .history/.
• Do not use exec to write protected files. exec is only for read-only inspection or diff helpers.
• Protected writes are confirmed by the product, and the runtime snapshots the previous file into .history/ automatically.
• A rewritten skill is only loaded on the next top-level turn. The current turn must not rely on the freshly edited skill body mid-execution.
• Do not create conversation-local shadow files such as SOUL.md, USER.md, or skills/.... Target the reserved agent-root path directly.

Workflow

1. Read the current live root file directly. From the conversation workspace, use the root-relative path such as ../../SOUL.md, ../../USER.md, or ../../skills/<name>/SKILL.md.
2. Draft the new content or patch.
3. Generate a diff for review.
4. Ask the user to confirm the protected write.
5. After approval, use write, edit, or apply_patch on the live path.
6. Report the live path and the .history/ snapshot path when one is created.

Common Mistakes

• Writing a protected file with exec.
• Editing .history/ directly instead of letting the runtime create the snapshot.
• Assuming the new skill text will affect the same turn that wrote it.
• Creating SOUL.md, USER.md, or skills/... inside the conversation workspace instead of targeting the reserved root path.