Survey, classify, and clean up `/mnt/local-analysis/` (or any sibling-to-workspace-hub directory holding orphan worktrees, codex-burn artifacts, agent log accumulations, and outer-clone duplicates) without losing useful code/work. Surfaces a tiered approval menu rather than baking decisions; defers all destructive ops until user confirms.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Survey, classify, and clean up `/mnt/local-analysis/` (or any sibling-to-workspace-hub directory holding orphan worktrees, codex-burn artifacts, agent log accumulations, and outer-clone duplicates) without losing useful code/work. Surfaces a tiered approval menu rather than baking decisions; defers all destructive ops until user confirms.
when_to_use
- User asks to "clean up local-analysis", "free up disk", "remove stale repos", "what's still useful in /mnt/local-analysis"
- User mentions "disk pressure", "orphan worktree", "codex-burn", "outer-clone duplicates", "agent-log accumulation"
- After a Hermes codex-burn run completes (post-run artifact sweep)
- Routinely (monthly or after disk pressure) to keep the workspace-hub sibling area lean
- Whenever `df -P /mnt/local-analysis` crosses a warning threshold (configurable)
A reusable routine for the dirty work that accumulates alongside workspace-hub/: orphan worktrees from completed Hermes codex-burn runs, outer-clone duplicates of nested repos, monitoring evidence from past audit passes, and agent log accumulations.
This skill was revised on 2026-05-12 after an adversarial cross-review by Hermes returned REJECT on its first draft. The 8 required patches are now applied; see references/hermes-adversarial-review-2026-05-12.md for the original findings and references/case-study-2026-05-12.md for the worked example.
Iron Law: NEVER skip the verification step
The single most-load-bearing rule. Before deleting anything that looks like a repo clone or worktree:
Confirm content is on origin — for git repos/worktrees:
Identify branch: git -C <bundle> branch --show-current (or read .git if it's a file: cat <bundle>/.git)
Compare tracked files: git -C <bundle> status --porcelain=v1 --untracked-files=all
Diff working tree vs git archive FETCH_HEAD-extracted branch tip with diff -rq
Explicitly list ignored files: git -C <bundle> status --ignored --short
Classify residue — the diff -rq residue goes into three buckets (see §3 below), not one blunt filter
Archive non-derived unique content — anything in the "evidence-bearing" or "inspect-first" bucket goes into the cleanup tarball before delete; with manifest, checksum, and atomic move
No rm -rf until a cleanup-verification.json artifact exists with delete_allowed: true
Reason for the law: it's how you honor "don't lose useful code/work" in a deletion-heavy routine.
Skill structure (numbered steps)
1. Survey
ls -la /mnt/local-analysis/
for d in /mnt/local-analysis/*/; dotimeout 20 du -sh "$d" 2>/dev/null; donedf -P /mnt/local-analysis # -P for machine-parseable, NOT -h
Capture: top-level entries, sizes, last-modified dates, mount usage %.
2. Classify each entry
Use this taxonomy (apply order matters — most-specific first):
a full .git/ clone of a canonical repo under an ad-hoc/staging/reconcile name; local branch is clean, upstream configured, ahead=0, and local HEAD is contained in origin history
reconcile-workspace-hub-YYYYMMDD-HHMMSS/
orphan worktree
.git is a file (gitdir: pointing at a path that no longer exists)
codex-burn-YYYYMMDD/<repo>-bundle/ after parent clone is gone
codex-burn run artifact
dated dir matching codex-burn-YYYYMMDD/ containing per-repo bundles, monitoring-evidence, logs, prompts
full Hermes codex-burn run output
agent log accumulation
dir holding provider-capacity-aware-YYYYMMDD-* or workspace-hub-{exit,closeout}-handoff-*.md
agent-logs/
top-level reconstructible Python dependency dir
dir contains pyvenv.cfg, no .git, mostly lib/ and bin/
marker-env/, capytaine-env/, raft-env/
nested open-issue Python dependency dir
dependency dir inside a workspace tied to an OPEN GitHub issue
ace2-gis-timelapse/.venv/
empty coordination meta-dir
dir created for an agent run that left no artifacts
agent-worktrees/, worktrees/
3. Derived-artifact classification (THREE buckets, not one)
When running diff -rq to identify residue, apply this three-bucket filter:
Bucket A — always disposable after diff (filter out, do not archive):
find <residue> -maxdepth 3 -type f -printf'%p|%s|%TY-%Tm-%Td %TH:%TM\n'
file <db> ; sqlite3 <db> '.tables'# if sqlite3 available
Build-output dirs (build/, dist/, target/, bin/, *.egg-info/, .eggs/) go in Bucket A only if their parent is unambiguously a build context; otherwise Bucket C.
4. Per-class verification
Class
Verification
outer-clone duplicate
Compare HEAD of outer vs nested. Check for-each-ref refs/heads/ --format='%(upstream:track)' for unpushed work. Check git stash list. Check git status --short for dirty.
clean duplicate clone
Fetch/prune first. Confirm remote URL matches the canonical repo; branch has upstream; git status --short is empty; git stash list is empty; git rev-list --left-right --count HEAD...@{upstream} returns 0 N (ahead=0); and git branch -r --contains HEAD includes the upstream/default branch. If clean but behind origin, deletion is acceptable after approval because origin already contains local HEAD.
orphan worktree
Read .git (file, contents = gitdir: <path>). If that path doesn't exist, orphan confirmed. Fetch the branch the orphan was tracking into a sibling clone, git archive it, diff -rq with bucket-A filter, build residue list, archive bucket-B/C, then delete.
codex-burn run artifact
Cross-check with Hermes via the full coordination protocol in §6 below. An old dated dir is a candidate, not vestigial-by-default.
agent log accumulation
Inspect timestamps; threshold for prune is the mtime policy below.
top-level reconstructible Python dependency dir
Confirm no active process has cwd/open files inside it; identify package intent from pyvenv.cfg and a small import probe; delete only after user confirms it is not a long-lived environment they still use. Archive normally not required because dependencies are reconstructible.
nested open-issue Python dependency dir
If parent workspace is tied to an OPEN GitHub issue, preserve outputs/, reports/, scripts/, logs, and other evidence-bearing artifacts; remove only the reconstructible dependency directory.
empty coordination meta-dir
find "$dir" -xdev -mindepth 1 -print -quit returns empty AND no permission errors. Fail closed if errors.
5. Risk-tier the proposal
Always present four tiers; let the user reject any.
Tier 0 — safe deletes: orphan worktrees with verified-clean origin-diff (residue is bucket-A only), clean duplicate clones with ahead=0/dirty=0/stash=0/HEAD-contained-on-origin, empty coordination meta-dirs (per find -quit check).
Tier 1 — archive then delete: codex-burn monitoring evidence, prompts, logs — anything in bucket B or C unique to the bundle. Archive workflow per §7.
Tier 2 — reduce in place: agent-logs with mtime policy (default: prune subdirs >14 days, keep standalone .md handoffs).
Tier 3 — leave alone: outer-clone duplicates with any unpushed/dirty/stashed state; anything Hermes coordination flagged; unclassified dotdirs; any orphan worktree whose branch is unmerged and parent issue is still open (call this evidence-preserving defer, not vestigial).
6. Hermes coordination protocol — old dirs are CANDIDATES, not vestigial
A dated codex-burn-YYYYMMDD/ dir can still matter even if Hermes will not poll it again. An old dated dir is deletable only after ALL of these are true:
# 1. No running process has cwd or open files under the dir
lsof +D <dir> 2>/dev/null || fuser -vm <dir> 2>/dev/null
pgrep -af 'hermes|codex|tui_gateway|slash_worker|tmux' | grep -F "<dir>"# 2. No Hermes cron/job/session/goal references the dir
grep -R --fixed-strings "<dir>" ~/.hermes/goals ~/.hermes/sessions ~/.hermes/logs 2>/dev/null
hermes cron list --all 2>&1 | grep -F "<dir>"# 3. No canonical repo still lists it in git worktree metadatafor repo in /mnt/local-analysis/workspace-hub/*/; do
git -C "$repo" worktree list --porcelain 2>/dev/null | grep -F "<dir>"done# 4. All branches named by the bundles exist on origin (verify per-branch, not just per-repo)# 5. Any GitHub issues referenced by prompts/logs are closed OR explicitly deferred with archive comment# 6. Archive verified before deletion (per §7)
Failure of ANY check → demote from Tier 0/1 to Tier 3 (defer).
Note on the orchestration skill location: the Hermes skill at ~/.hermes/skills/autonomous-ai-agents/agent-cli-delegation-operations/references/codex-background-burn-orchestration.md is a reference to an archived skill (.archive/umbrella-2026-04-29). The active orchestration source-of-truth may evolve. Treat this path as fragile; if it disappears, do not assume the protocol changed — re-derive via hermes skills list first.
Archive tracking in git: *.tar.gz is globally gitignored at .gitignore:330-332. Therefore:
The archive .tar.gz will NOT be committed by default
Commit the manifest + checksum (the .tsv and .sha256 are not gitignored) — these prove the cleanup happened
Only git add -f the archive itself after:
Secret scan: git secrets, gitleaks, or grep -RE '(sk-[A-Za-z0-9]{20,}|AKIA[0-9A-Z]{16}|ghp_[A-Za-z0-9]{36}|password|token|secret)' <extracted-archive>
Manual classification of any client names, prospect data, downloaded data licensing
Size review (committing >5 MB archives to git is generally wrong; prefer LFS or non-git storage)
8. Pre-delete final gate
Immediately before rm -rf (between approval and execution):
# Re-run race-detection checks
lsof +D <target> 2>/dev/null
pgrep -af 'hermes|codex|tui_gateway|slash_worker|tmux' | grep -F "<target>"
grep -R --fixed-strings "<target>" ~/.hermes/goals ~/.hermes/sessions 2>/dev/null
# Acquire cleanup lock
LOCK=/mnt/local-analysis/.cleanup-lock
if [ -e "$LOCK" ]; thenecho"Cleanup already in progress: $(cat $LOCK)"; exit 1
fiecho"pid=$$ host=$(hostname) date=$(date -Iseconds) targets=$*" > "$LOCK"# Two-phase delete: move to trash-stage first
TRASH=/mnt/local-analysis/.cleanup-trash/$(date +%Y%m%d-%H%M%S)
mkdir -p "$TRASH"mv <target> "$TRASH/"# Verify system behavior post-move (next codex-burn run, agent operations, etc.) before final rm# Final rm only after the trash-stage has lived for at least one cycle OR disk pressure justifiesrm -rf "$TRASH"rm"$LOCK"# Write verification artifactcat > "$ARCHIVE_BASE/${ARCHIVE_NAME}-cleanup-verification.json" <<EOF
{
"target_paths": [...],
"branch_compared": "...",
"diff_command": "...",
"residue_buckets": {"A": N, "B": N, "C": N},
"archive_path": "${ARCHIVE_BASE}/${ARCHIVE_NAME}.tar.gz",
"archive_sha256": "...",
"delete_allowed": true,
"approver": "user",
"timestamp": "$(date -Iseconds)"
}
EOF
Under severe disk pressure, the trash-stage MAY be skipped, but only after the archive is verified and the cleanup-verification.json is written.
9. User approval
Present the four-tier proposal to the user using the available user-interaction mechanism for the current agent harness:
Claude Code: AskUserQuestion tool
Hermes CLI: clarify tool
Codex CLI / non-interactive contexts: stop after writing the proposal to disk; do NOT delete
One question per tier (so user can reject independently). Always include "show me the diffs first" as an option for the tiers that delete content. Never auto-execute without per-tier approval.
If no user-interaction tool is available, halt and emit a proposal document to docs/sessions/YYYY-MM-DD-cleanup-proposal.md for a human to review later.
table of entries acted on, with class + verdict + residue-bucket breakdown
archive path, manifest path, checksum
cleanup-verification.json path
items deferred (and why — be specific: which check failed)
any follow-up GitHub issues filed
Commit the handoff + manifest + checksum. Do NOT commit the archive .tar.gz itself by default (it's globally gitignored; force-adding requires the security review above).
Disk-pressure trigger
If invoked autonomously (e.g. from a cron health check), gate execution on:
exceeding a threshold (default 75%). Below that, defer. Use -P not -h — human-readable parsing is brittle.
What NOT to clean
Anything under workspace-hub/ itself — different skill (per-repo housekeeping)
System-allowlisted caches: .pnpm-store/, .cache/, .cargo/, .npm/ — but NOT all dotdirs (unknown ones go Tier 3)
XDG .Trash-* (system-managed)
Any path referenced by an active Hermes goal, cron job, session, or process (per §6 protocol)
Any orphan worktree whose origin-diff bucket-C residue isn't classified
Any orphan worktree whose branch is unmerged AND parent issue is still open — that's evidence-preserving defer, not vestigial
Memory hooks
When this skill completes a successful run, optionally update auto-memory at ~/.claude/projects/-mnt-local-analysis-workspace-hub/memory/:
feedback_mnt_analysis_cleanup_gotcha.md — if a new gotcha surfaced (e.g., a check that should be in §6 but wasn't)
project_mnt_analysis_state.md — post-run state snapshot if useful for future sessions
The case-study reference (references/case-study-2026-05-12.md) captures the worked example that produced this skill. The Hermes adversarial review at references/hermes-adversarial-review-2026-05-12.md produced the patch set that hardened it. The clean duplicate clone deletion pattern is illustrated in references/clean-duplicate-clone-2026-05-18.md. Top-level reconstructible Python dependency cleanup and open-issue workspace reduction are covered in references/top-level-venv-and-open-issue-workspace-cleanup.md.
