JesseRaber Multi-Agent Folder Cleanup
Create a workspace where an agent arriving cold can quickly tell what is true now, what is proposed, and what is historical without choosing between plausible copies.
Report as: Loaded Multi-Agent Folder Cleanup v1.1.0.
Choose the operating mode
State the mode in the first line of the response. Default to Audit when execution is not explicitly authorized.
| Mode | Typical request | Allowed work |
|---|---|---|
| Audit | “What is here?” “Why are agents confused?” | Read-only inspection and an evidence report. |
| Plan | “How should this be organized?” | Read-only inspection, a proposed tree, and a literal source-to-target map labeled PROPOSED. |
| Execute | Explicit approval of a specific map or record-correction list | Only the approved moves or exact factual patches. |
Read references/workflow.md before any Audit, Plan, or Execute run. Read references/audit-tools.md when a filesystem is mounted and the bundled helpers can run. Read references/navigation-templates.md only when creating or reviewing navigation files.
Keep authorization narrow
- The approved map is the mutation boundary. Discovery never expands it.
- Git operations, deployments, databases, credentials, sync settings, sharing permissions, and external service writes remain out of scope unless separately authorized.
- Deletion is never implied. Moving a named path is not approval to delete documents, archives, duplicates, or nearby folders.
- Remove only verified staging copies and confirmed-empty source folders when the approved protocol names that cleanup and the platform offers recoverable trash semantics.
- Protect instruction files by default. Propose a demonstrably stale factual correction separately and apply it only after explicit approval; never rewrite behavioral rules as housekeeping.
- A failed safety check overrides prior approval. Stop on a collision, changed source, missing file, placeholder, concurrent write, path-length failure, or unapproved target.
Audit evidence, not filenames
Discover root and nested instruction files such as AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, README.md, and README_FIRST.md. Include instructions configured outside the folder when the host exposes them. Record conflicts rather than silently choosing a winner. Follow the folder’s own required logging policy, but do not replace a concurrently written journal when the access route cannot append safely.
Confirm one exact root, the storage substrate, the access route, and any companion or predecessor roots. A mounted filesystem can support hashes and Execute mode. A connector or web listing can support a real but degraded Audit only; identify unverified hash, hydration, OS path-length, and byte-comparison claims. Never Execute through a connector-only view.
Inventory substantive documents, archives, handoffs, transcripts, data, scripts, outputs, mirrors, generated state, and duplicate groups. Verify every required entrypoint exists. Check indexes in both directions: broken references and live folders omitted from authoritative-looking indexes.
Treat a filename such as FINAL_v3.md, STATUS.md, or AUTHORITY.md as a reading cue, not proof. Open the supporting artifacts and label material claims verified, contradicted, or unverifiable. Never upgrade “unverifiable” to current.
Prioritize these confusion sources:
- Divergent control documents, especially across separate roots.
- Missing required entrypoints or no visible agent instructions.
- One filename resolving to different content.
- Stale indexes, manifests, status files, or authority maps.
- Identical copies with no canonical marker.
- Generated machine state overwhelming search and inventory.
Flag credential-bearing filenames but never open, stage, copy, quote, or index their contents. Keep archives closed unless central-directory inspection is explicitly useful; never bulk-extract them to make them searchable.
On large trees, keep structural checks exhaustive but read selectively: all instructions, authority or state claims, handoffs, and representative files from duplicate groups. Mark other judgments classified-by-metadata and report exclusions. An exclusion limits this review; it does not make files invisible to another agent or cloud index.
For OneDrive or SharePoint, distinguish the local and provider views. A junction skipped locally may be materialized as searchable cloud files. Before Execute, verify hydration and quiet sync on Windows. A successful local move does not verify cloud retention, version history, sharing links, conflict copies, or provider-side synchronization.
Classify each substantive file once
Use exactly one bucket:
- Owner direction and verified current state
- Documentary authority
- Current analysis and provenance
- Active backlog
- Historical or superseded evidence
- Raw or unreviewed incoming material
- Read-only mirrors and external checkouts
- Generated machine state
Roadmaps remain backlog until implementation is independently verified. Source snapshots remain history unless tied to a reproducible commit. Generated caches, browser profiles, build output, and runtime dependencies are bucket 8, not documentary history and not automatic deletion targets.
Design for fast navigation
Prefer a short root with an entrypoint, instructions, status, authority map, index, and one current handoff. Adapt names to the project’s vocabulary; do not impose a universal numbered tree. Keep one authority per topic and clearly separate current material, backlog, history, incoming material, and read-only mirrors.
Execute with measured verification
Follow the staged sequence in references/workflow.md exactly:
- Freeze the approved literal move map and write a SHA-256 baseline outside the source and target trees.
- Preflight missing sources, collisions, target roots, hydration, and a conservative 240-character Windows path limit.
- Copy to a clearly labeled staging tree that mirrors final relative paths.
- Verify every staged hash.
- Move the exact approved map in one controlled pass. Do not pause for a second approval already covered by the map.
- Verify final hashes, intended source absence, and instruction/index integrity.
- Remove only verified staging and approved confirmed-empty source folders using recoverable semantics.
Use one filesystem environment end to end and literal paths throughout. Do not construct move or deletion commands from globs, unresolved environment variables, or a second shell. If the map empties a parent, name that parent in advance or report the resulting empty folder as follow-up; never silently broaden the cleanup.
For approved navigation-only or factual record corrections, use the record-only Execute protocol in references/workflow.md: pre-write hashes, immediate re-read on mismatch, exact-context patches, and post-write checks for both the new facts and stale claims that should be gone.
Finish with proof
Cleanup is incomplete while staging remains, an index describes proposed paths, a current record retains a superseded claim, or verification has unresolved gaps. Report:
- exact scope and mode;
- inspected, inferred, excluded, and inaccessible material;
- verified findings separated from owner decisions;
- every moved or intentionally changed path;
- hash, collision, hydration, index, and instruction-file checks;
- remaining limitations and separately optional follow-ups.
Do not call the work complete merely because the files look tidy.