Troubleshooting

Agent Deleted Files by Mistake?

A fast containment + recovery workflow for accidental deletes, with safer defaults to prevent it happening again.

If an agent ran a delete command you didn’t expect, move fast but stay calm. Most damage gets worse from panic retries. This guide gives you a deterministic recovery sequence.

First rule: stop issuing new write/delete commands until you verify what changed. Extra commands can overwrite recovery evidence.

1) Contain immediately (60 seconds)

  1. Pause the workflow and stop retries.
  2. Ask the agent for the exact command it executed.
  3. Capture path(s), timestamp, and whether it used trash or permanent delete.

2) Ask for proof of what changed

Before doing anything else, return: 1) exact delete command executed 2) absolute path(s) affected 3) whether files were moved to Trash or permanently removed 4) current directory listing for parent folder 5) git status --short output (if repo) Do not run more write/delete commands.

3) Recover based on delete type

Do not run cleanup commands (like broad resets or cache cleaners) before you verify recoverability. They can make recovery harder.

4) Known-good recovery prompt

Recovery mode ON. Goal: Restore accidentally deleted files safely. Rules: - No new delete commands. - No broad cleanup/reset. - Show command + path proof before each recovery step. Steps: 1) Identify deleted file list with absolute paths. 2) Attempt Trash restore where possible. 3) If git-tracked, restore from git and show diff/status. 4) After each restore, verify file exists and show first 3 lines. 5) Stop and summarize what was restored vs still missing.

5) Prevent repeat incidents

Copy/paste prevention block

Destructive Action Safety Policy - Never run rm -rf or permanent delete without explicit approval. - Prefer recoverable delete (trash) for normal cleanup. - Before any delete: show absolute path(s) + why each file is safe to remove. - If uncertain, stop and ask. - After delete: return proof list of removed files and current folder state.

Bottom line: contain first, recover with proof, then add guardrails so delete mistakes become rare and recoverable.