Imported from organvm-iv-taxis/a-i--skills (distributions/claude/skills/shell-history-hygiene/SKILL.md). Install upstream with npx skills add organvm-iv-taxis/a-i--skills --skill shell-history-hygiene. Copyright stays with the author (MIT).

shell-history-hygiene — Shell-History Audit + Targeted Cleanup

When to invoke

• User explicitly asks to audit / clean / dedup / prune their shell history (atuin or otherwise)
• User pasted an ad-hoc atuin stats / prune / dedup one-liner and asked for review or improvement
• Before any history-store migration (atuin → atuin v2, zsh-only → atuin, etc.) where you want a clean baseline
• During machine-setup audits where the history corpus is being inspected for size or quality

Backends

• atuin (current): atuin stats today, atuin history prune, atuin history dedup
• zsh native (future): .zsh_history line dedup, size limits
• bash native (future): .bash_history line dedup
• mcfly (future): mcfly's own pruning

This skill currently ships the atuin backend at scripts/atuin-audit.sh. Add sibling scripts under scripts/ as backends are added.

How it works (preview-then-apply)

The audit is dry-run only. It produces three things:

1. A headline summary on the terminal — atuin stats today + counts.
2. Two dated preview artifacts on disk — one per operation (prune, dedup) — for human inspection.
3. Echoed apply commands at the end — copy/paste shapes that the human runs manually.

The skill never runs the apply commands itself. Separation of audit from apply is deliberate — it's the kill-switch against accidental history destruction.

Running the audit

${CLAUDE_SKILL_ROOT}/scripts/atuin-audit.sh           # dedup --before defaults to today
${CLAUDE_SKILL_ROOT}/scripts/atuin-audit.sh 2026-05-01  # dedup --before <date>

If ${CLAUDE_SKILL_ROOT} isn't resolved by the harness, use the absolute path:

/Users/4jp/Code/organvm/a-i--skills/skills/tools/shell-history-hygiene/scripts/atuin-audit.sh

Preview artifacts land alongside the script:

• atuin-prune-preview-YYYY-MM-DD.txt
• atuin-dedup-preview-YYYY-MM-DD.txt

Dated filenames preserve a trail — re-runs don't clobber prior previews.

Interpreting output

Applying the cleanup

After the human eyeballs the preview and approves:

atuin history prune                                          # apply prune
atuin history dedup --before "$BEFORE_DATE" --dupkeep 1     # apply dedup (keep 1 copy per duplicate string)

Run in that order. Prune first removes the structural no-ops; dedup then collapses the remaining exact-string repeats. Reverse order works but does redundant scanning.

--dupkeep 0 is destructive (removes ALL copies). Never recommend it — --dupkeep 1 is the right default.

Safety / failure modes

• Script preserves dry-run guarantee — set -euo pipefail, never invokes apply commands.
• atuin not on PATH — script exits 1 with clear error. No partial action.
• Multi-line commands in atuin — the script's headline parses atuin's own "Found N" lines (not wc -l) so the count is accurate even when multi-line commands inflate the line total.
• --before date in the future — atuin accepts it; dedup just considers all history. Inspect preview before applying.

Heritage / why this skill exists

Triggered by a 2026-05-21 review of an ad-hoc one-liner that had a tee cmd > file.txt bug — wrote two files, swallowed dedup output. The fix evolved into a dated-preview script, then into this skill so the workflow lives somewhere discoverable rather than as scratch material under ~/_dot-config/scratch/. Per Rule #34 ("merge into ideal form") the function clarified beyond scratch/'s purpose, so it migrated here.

Original artifact birthplace: ~/_dot-config/scratch/atuin-audit.sh (worktree commit 6d6ef39 on _dot-config main). This skill's scripts/atuin-audit.sh is the canonical-from-here copy.