AgentDust

View saved version →

---
name: cleanup
description: Explain AgentDust process reports and guide safe local cleanup. Use when the user asks about processes left by Claude Code sessions, AgentDust doctor output, or cleanup.
---

# AgentDust cleanup

This skill package contains guidance only. It does not connect to a Mac or provide live process access. Live inspection requires a separately configured local AgentDust MCP connection. AgentDust 0.3.0 supports Claude Code on Apple silicon; setting it up there does not connect it to ChatGPT or Codex.

## Explain a report the user provides

Ask the user to paste `agentdust doctor --json` for process classifications or `agentdust disk --json` for disk totals. Do not ask for the terminal `doctor` output unless needed, because it can include command arguments and directory paths. The JSON doctor report still contains process IDs, executable basenames, ages, classes, and evidence, so the user should remove anything they do not want to share.

Explain classes and evidence as reported. A process name, parent PID, age, or CPU use alone does not show that a process is safe to stop. A pasted report is a snapshot and cannot authorize or support a later signal.

## If a local AgentDust MCP connection is available

1. For read-only inspection, call `agentdust_doctor` and explain the classes and evidence.
2. For cleanup, call `agentdust_plan` to obtain a fresh plan. Show the eligible processes and evidence, then use only the plan ID and item IDs from that response for items the user requested.
3. Call `agentdust_apply` for those items. AgentDust presents the approval form and generates the code. The user must review the items and type the code. Never invent, copy, or submit an approval code.
4. If the user declines or cancels, stop. If the client cannot show the form, tell the user to run `agentdust apply` locally. Do not run that command yourself.
5. Report the result returned by `agentdust_apply`. Do not repeat an apply call for an item that already has a result.

If live tools are unavailable, say this package does not create that connection. Explain the report the user provides, or direct them to AgentDust's local Claude Code setup. Do not change client settings or install AgentDust as part of this skill.

View saved version →

---
name: disk
description: Explain AgentDust disk reports and guide local setup. Use when the user asks about Claude Code storage, history, caches, worktrees, or AgentDust disk output.
---

# Claude Code disk usage

This skill package contains guidance only. It does not connect to a Mac or provide live disk access. Live inspection requires a separately configured local AgentDust MCP connection. AgentDust 0.3.0 supports Claude Code on Apple silicon; setting it up there does not connect it to ChatGPT or Codex.

If the user provides `agentdust disk --json`, explain the logical and allocated byte totals by root and category. Allocated bytes are not reclaimable space. History, memory, plugins and worktrees can contain work the user still needs. The report contains sizes and scan status, not file paths or contents.

Report missing, unavailable and partial roots and skipped counts. A zero in a partial or unavailable root does not mean it is empty. The worktree root is scoped to the MCP server's working directory; other projects and custom locations are excluded.

Do not read transcripts, credentials or other file contents to complete the report. Do not delete files or infer permission to delete from a category. If the user has not provided a report and live `agentdust_disk` is unavailable, say this package does not create that connection. Explain how to obtain the report through an explicitly configured local AgentDust MCP connection or the local CLI. Do not change client settings or install AgentDust as part of this skill.