Cleaning atuin history

Skill narumiruna/skills/deprecated/cleaning-atuin-history

Install

npx -y skills add narumiruna/skills --skill cleaning-atuin-history

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

8 stars8 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Audit Atuin history for duplicate pressure and high-confidence typo/retry pairs, then prepare fixed-scope deduplication or user-run inspector steps without direct SQLite deletion. This deprecated internal workflow keeps cleanup-typos disabled.

SKILL.md

3.7 KB, as published. Nobody here has run it

Atuin History Cleanup (Deprecated Reference)

Audit first. Do not treat a preview, uniqueness estimate, or skill invocation as deletion approval.

Guardrails

Resolve the active database with atuin info and use the bundled audit before proposing mutation.
Never delete rows directly from SQLite or edit Atuin config as part of this workflow.
atuin history dedup mutates the selected history window globally.
atuin search --delete deletes every match under active search semantics; do not use it manually for a single row.
The transactional typo command snapshots the database but also runs atuin store push, recomputes and deletes candidates, may drive a TUI, and finishes with atuin sync. Its current interface cannot bind execution to approved candidate IDs or guarantee non-interactive operation.
Do not invoke cleanup-typos through this skill. Keep it disabled until it gains a plan-only output plus approved-ID and non-interactive execution checks; its source and recovery artifacts remain historical reference material.
Re-run the audit and matching dry run immediately before every mutation, whether user- or agent-executed. Use a fixed cutoff instead of a relative value such as now; if candidates, scope, or counts differ from the approved set, stop for renewed approval.
Do not auto-restore a full live database after failure; preserve live state, snapshot, and verification artifacts for recovery.
The interactive inspector is user-run. An agent must not open or drive its TUI.

Read references/atuin-cli.md when exact dedup, search, or prune semantics matter.

Audit

Resolve this skill directory and run the script by absolute path:

uv run python "$SKILL_DIR/scripts/atuin_history_cleanup.py" audit

Use --format json for structured review. Optional scope flags include --db-path, --dupkeep, --before, --typo-window-seconds, and --max-typos.

The audit reads selected history columns, groups duplicates by command/cwd/host, and proposes typo pairs only when session, time, arguments, exit status, frequency, and edit-distance evidence align. Treat its output as candidates, not proof that deletion is wanted.

Prepare Cleanup

Duplicates

Resolve the audit's --before value to a fixed cutoff, then run the reported atuin history dedup --dry-run ... and present the exact window, keep count, groups, and potential deletion count. Immediately before applying, repeat the audit and matching dry run with that same fixed cutoff. Run the identical non-dry command only when the results still match the approved scope and loss; otherwise obtain renewed approval.

Typos

Present each candidate ID, timestamp, cwd, original, correction, and reason. The supported path is the user-run inspector: give the user the prefix/cwd preview, candidate details, and inspector keys so they can confirm and delete the exact row themselves.

Do not offer or execute the bundled transactional command as an approved automation path. It can run atuin store push, delete a recomputed set, open a TUI, and run atuin sync in one invocation before an agent can verify that the executed plan matches the reviewed IDs.

Verify

Stop after the approved cleanup. Re-run the same audit scope and report remaining counts. If diagnosing a historical transactional run, preserve pre_audit.json, plan.json, post_verify.json, and the snapshot path; report remote-sync and recovery status without exposing unrelated history.