Agent Immune System
Detect prompt bloat. Prevent rule autoimmunity. Regenerate healthier agent instructions.
github.com/inbusiness23/agent-immune-system
AGENTS.md
- Never run commands without asking first.
- Always check the branch before editing.
- Always verify git status before editing.
- Do not edit main.
- Stop if worktree is dirty.
- Never touch unrelated files.

+ Proceed with safe read-only discovery.
+ Ask before destructive commands, external writes, credential access, or irreversible actions.
+ Confirm worktree safety before editing: branch, metadata, dirty changes, writer collision.
Instruction Health Audit
74/100
My AGENTS.md is huge, contradictory, and every failed session turns into another rule.
Running audit. I found duplicate git safety rules, one autoimmune command rule, and a browser-proof lesson that should be watched before promotion.
Should the missed browser proof become a global rule?
Decision: watchlist. Track it for 3 visual-review sessions and promote only if it recurs.
What about all these git safety rules?
Regeneration candidate. Replace five overlapping rules with one durable worktree-safety principle.
A mistake is evidence, not automatic permanent memory.
Watchlist Prune Regenerate Autoimmune Finding Human Approval Required

Audit. Prune. Regenerate.

Proposal-first prompt bloat relief for self-healing AI agent workspaces.

Use agent immune system to audit my AGENTS.md.