Cleanup

Run the script, show its listing as it is, propose a set, and wait for the user's word. Speak in short sentences when proposing, when clarifying, after deleting, and on every refusal or error. Read the JSON yourself for the numbers; never read it, paths or exit codes to the user, and never retell the listing in your own words.

Every sentence quoted below is a model of what to say, not text to copy out. Say it in the user's own language, keeping the names, counts and reasons the command gave and inventing none; a listing block is the one thing shown exactly as the command printed it.

The cycle

1. Run the listing with the Bash description "List files left by entrust."

   F="$(mktemp "${TMPDIR:-/tmp}/entrust-cleanup.XXXXXXXX")"
   CLAUDE_PLUGIN_DATA="${CLAUDE_PLUGIN_DATA}" node "${CLAUDE_SKILL_DIR}/../codex/scripts/cleanup.mjs" --list --json >"$F" && cat "$F" && echo "snapshot: $F"
   

   mktemp gives each listing its own file. A name built from the shell's $$ does not: two listings in one shell would share it, and a number from the first would then be read against the second. Keep the snapshot path from the last line; step 3 needs that exact path. Show the text field to the user in one code block, unchanged. It is the listing: numbered items, their sizes, when they last changed, whether each is suggested, selectable by its number or kept, and why. Say nothing about the items yourself. When it says the agent scan used Node's fallback temporary directory, keep that warning in the block so the user sees that agent scratch elsewhere may not have been found. An empty inventory is "I found no items covered by this cleanup."

2. Propose in one sentence exactly what proposed holds, by those rows' names and their total size: "I suggest deleting the temporary files for agent u1-astra and 172 temporary directories from the lock tests, about 11 MB; shall I?" When selectable holds numbers that are not in proposed, add one sentence naming them: "Item 1, the 9 September 2026 cleanup, item 7, the standalone report from run 42, and item 18, 43 saved conversations from the tests, can go too if you say their numbers." Then wait. With nothing suggested and nothing else selectable, say "I have no cleanup to suggest; the listed items are being kept for the reasons shown."

3. Map the answer to numbers yourself. "Yes", "yes please", "go", "go ahead", "apply it" or "да" is exactly the numbers in proposed. Digits are those numbers; an affirmative with digits adds them to the suggestion; "only" or "instead" restricts to the digits alone. "All" or "everything": ask "Do you mean the items I suggested, or the runs, standalone reports and saved conversations as well?" and wait. "No": "I'll leave the listed items in place." A question: answer it, delete nothing. Silence: wait. A number that is not in selectable is not yours to send — say " is being kept; the listing says why." and leave it out. Then run, with the description "Delete the cleanup items the user selected.":

   CLAUDE_PLUGIN_DATA="${CLAUDE_PLUGIN_DATA}" node "${CLAUDE_SKILL_DIR}/../codex/scripts/cleanup.mjs" --delete --from "<SNAPSHOT>" <numbers>
   

   Do not run the listing again between the user's word and this call: the snapshot is what binds each number to what was shown, and an item that changed since then is left in place and reported.

4. The command prints one paragraph per outcome, then the fresh listing. Report every outcome it printed in your own message, keeping its names and reasons and adding none — "I deleted the temporary files for agent u1-astra and 172 temporary directories from the lock tests, and left the 9 September 2026 cleanup in place because it changed since it was listed." Then show the fresh listing in a code block when anything remains, and say "I have no further cleanup to suggest." only when its last lines say nothing is suggested and nothing else is selectable.

Reading the result

Read the whole output even on exit 10: deletions and refusals occur together. Use the listing's names; omit outcomes that did not occur.

What it never touches

Answers, managed worktrees and their ledger, write locks and the shared Codex home are listed and never removed: the driver prunes answers and reconciles the next two itself, and the last is shared by every agent. The data directory of another copy of this plugin is the user's own to remove — when they ask how, say "This command removes it." and show that row's command from manual in its own block. The one directory this does not cover is the plugin's own under its previous name, codex-delegate-<marketplace>: listed as such, proposed, removed by number like scratch, unless a plugin of that name is still installed. The notCovered commands apply only to other entries in the coordinator's temporary directory. Say "To list those entries without removing them, run this command." and show notCovered.listCommand; for removal, notCovered.removeCommand. An agent started under another temporary root is outside the agent scan; its report is kept while report.json is absent. The driver's private <state>/tmp directories are guarded and not listed: the driver owns them, and this cleanup never makes them removable.

Forward CLAUDE_PLUGIN_DATA as shown. The script uses ENTRUST_STATE_DIR first, then CLAUDE_PLUGIN_DATA; setup follows the sibling's One call. On the clone route both commands need the codex skill linked beside this one; the installation recipe links the two together.