version: "1.0.1" name: local-document-organizer description: Preview-first local file organizer. Scan a user-named folder, classify files with readable rules, show a plan, execute approved moves and undo them. Use for sorting Downloads, renaming into category folders, or reviewing organization history, including scoped metadata-only course persistence. Do not use for knowledge synthesis, source analysis or workflow orchestration.

Local Document Organizer

For a student-facing explanation of why this package exists and how the preview-then-apply workflow fits into the handbook, read README.md first. This file is the invocation contract for Codex.

Overview

Use this skill to propose and execute safe local file organization. The skill is preview-first and confirmation-gated: scan the user-named folder, write a proposed category structure as a Markdown report and a JSON plan, wait for explicit user approval, then run only the approved moves. Every applied move is logged so it can be reversed with undo.

The skill never deletes and never overwrites. See references/safety-rules.md for the full safety contract.

Safety Rules

• Never delete files. The script does not call unlink or rmtree.
• Never overwrite files. Destination collisions are skipped and logged as

conflict.

• Never move files before the user approves the preview and supplies

--confirm ORGANIZE.

• Refuse to scan /, /System, /Library, /Applications, ~/Library,

~/.ssh, ~/.gnupg, ~/.aws, or the home directory itself. The user must name a subfolder.

• Treat low-confidence and unmatched files as Unknown/ and skip them by

default; they are surfaced in the preview report so the user can opt in with --include-low-confidence.

• Permission errors and name collisions are recorded as skipped actions, not

silent failures; one skipped file does not abort the whole run.

• Runtime database, plans, action logs, and reports stay under

~/.codex/state/local-document-organizer/ and are not committed to git.

Default Workflow

1. Confirm with the user which folder to scan. Default to ~/Downloads only

if the user explicitly says so; otherwise ask.

1. Run scripts/local_document_organizer.py scan --folder <path>.
2. Read the generated Markdown report aloud to the user, including the

per-category counts and any skipped or low-confidence files.

1. Wait for explicit user approval. Do not invoke apply from inferred

consent.

1. On approval, run

scripts/local_document_organizer.py apply --plan <plan.json> --confirm ORGANIZE.

1. Report applied moves, conflicts, missing files, and the action log path so

the user knows where the undo log lives.

1. If the user asks to revert, run

scripts/local_document_organizer.py undo --log <log.json>.

Commands

Resolve scripts/local_document_organizer.py relative to this skill directory. When running from an installed Codex copy, that is usually ~/.codex/skills/local-document-organizer/scripts/local_document_organizer.py.

Preview a folder:

python3 scripts/local_document_organizer.py scan --folder "$HOME/Downloads"

Preview including low-confidence and hidden files:

python3 scripts/local_document_organizer.py scan \
  --folder "$HOME/Downloads" \
  --include-low-confidence \
  --include-hidden

Apply a generated plan after explicit user approval:

python3 scripts/local_document_organizer.py apply \
  --plan ~/.codex/state/local-document-organizer/plans/<run_id>-plan.json \
  --confirm ORGANIZE

Reverse the moves from a run:

python3 scripts/local_document_organizer.py undo \
  --log ~/.codex/state/local-document-organizer/logs/<run_id>-actions.json

List recent runs:

python3 scripts/local_document_organizer.py runs --limit 10

Classification Rules

Rules live in references/classification-rules.csv and are evaluated in CSV order; first match wins. Filename-keyword rules sit above extension rules so that, for example, invoice-uber-2026-04.pdf is classified as Invoices rather than the generic PDFs bucket.

Each row has columns: rule_id, category, match_type, pattern, confidence, enabled. match_type is extension or filename_keyword. pattern is a |-separated list of tokens (extensions without the dot, or whole-word keywords matched at word boundaries). confidence is high, medium, or low; only high and medium matches are eligible to move by default.

Add new rules by extending the CSV. Keep rule_id stable so action logs and report tables remain meaningful across edits.

Persistence

The SQLite database lives at:

~/.codex/state/local-document-organizer/organizer.sqlite

Schema:

organizer_runs(id, folder_path, created_at, status)
file_actions(id, run_id, old_path, new_path, action, confidence)

organizer_runs.status is one of: previewed, applying, ok, partial, no_moves. file_actions.action is one of: moved, conflict, missing, permission_error, failed.

Outputs

~/.codex/state/local-document-organizer/
  organizer.sqlite
  reports/YYYY-MM-DD-<folder-name>.md
  plans/<run_id>-plan.json
  logs/<run_id>-actions.json

Do not commit runtime artifacts unless the user explicitly asks for sample files.

Response Pattern

When reporting scan results to the user, include:

• folder scanned
• proposed category structure (per-category counts and total size)
• any low-confidence or skipped files that the user might want to include
• the report and plan paths
• a clear next step: "approve to apply, or ask me to refine the rules"

When reporting apply results, include:

• run status (ok, partial, or no_moves)
• moved, conflict, missing, permission, and failed counts
• the action log path so the user can run undo later if they change their

mind

When reporting undo results, include the restored count and any skipped files (e.g. because the original path now contains a different file).

Professional AI Agent Course adapter (Organize)

For classroom work, synthesis requests, or a storage-mode request, use scripts/course_organizer.py in this same package. The original helper remains the standalone baseline. Classify and safely relocate local files, with preview, approval and undo. Do not synthesize their knowledge.

Read the course section of README.md, then use the shared setup and synthetic fixtures. Put storage/context flags before the command. Keep source files and recovery journals local, honor all approval gates, and verify canonical reads after persistence. Remote mode requires the provisioned contract; a missing backend is a dependency failure, never a reason to invent credentials or silently use local state.

The course adapter reuses existing classification rules and preview generation. Moves stay within the selected directory and use atomic no-clobber hard links before removing the old directory entry; no file contents are deleted. It fails safely on filesystems without hard-link support. It refuses stale/tampered plans, changed sources, symlinks, collisions and replayed plans. Every intent and action is fsynced in a local journal. Undo checks content hashes and refuses collisions; use --storage local for recovery while the remote API is down. It does not upload file binaries.