Undo map (undo)

Turn "the agent broke something" into a concrete list: which files changed, where each one can be recovered from, and which actions are already past recovery.

Language rule: write every user-facing output in the language the user is speaking with you. Safety rule: this skill produces a plan. It does not restore anything until the user says to, one file or one group at a time. Never restore in bulk on your own initiative.

Step 0 — Pick the session

Use the id the user gave, or latest. When they describe it instead ("az önceki", "dsg-platform oturumu"), run --list --days 7 --md first and match on directory and time. If two sessions touched the same files, map both — a file restored to the wrong version is worse than one left alone.

Step 1 — Build the map

node "${CLAUDE_PLUGIN_ROOT}/scripts/blackbox.mjs" --undo --session <id|latest> --md

For an agent that records diffs rather than file backups (Codex), also write a draft reverse patch (expand the home directory yourself; never pass a literal ~):

node "${CLAUDE_PLUGIN_ROOT}/scripts/blackbox.mjs" --undo --session <id> \
  --patch "<home>/.agentlens/blackbox/<slug>-undo.patch"

Requires Node.js 18+. Read the output fully before saying anything.

Step 2 — Sort files by what you can actually do

Each row carries a backup version, a git state and a suggested recovery. Group them:

Before recommending anything, check whether the file changed after the session ended (modifiedAt versus the session's end time). If it did, someone edited it since — restoring would silently discard that work. Call this out.

Step 3 — Read the irreversible list

The report's second half lists destructive, data, git and publish actions that actually ran. These are not in the file map because no file map can fix them. Report them in the order that matters: deleted data first, then anything published or pushed outward, then history rewrites. For each, say what the realistic remedy is — a remote copy, a database backup, a new release — or say that there is none.

Step 4 — Propose, do not perform

Present the plan as a short numbered list, most valuable first, with the exact command for each step. Then stop and ask which of them to run. When the user picks:

• Restore one file or one named group at a time, and show the diff you are about to overwrite when it is short enough to read.
• Never chain a restore with anything else in the same command.
• If the working directory is a git repository and it is dirty, propose git stash -u (or a WIP commit) before the first restore, so this operation is itself reversible.
• If it is not a git repository, say so — the agent's backups are then the only safety net, and they are per-session.

Step 5 — Reply

Short: how many files changed, how many are recoverable and by which route, how many are not, and the irreversible actions if any. End with the one question the user has to answer to move forward. If nothing changed on disk, say that in a single line.