| name | clean-my-loop |
| description | Audit and clean the persistent context that feeds a recurring autonomous agent loop — memory files and indexes, scheduled-task / automation prompts, and CLAUDE.md / AGENTS.md — so the loop stops degrading into a self-reinforcing echo chamber. Works for Claude Code /loop crons and Codex automations. |
clean-my-loop
A recurring autonomous loop (Claude Code /loop cron, Codex automation) re-reads the same
context every iteration: its memory/state files, the scheduled-task prompt, and CLAUDE.md /
AGENTS.md. Those artifacts decay: memory grows into a changelog, task rules acquire
no-exit "skip on error" clauses, instruction files give the loop cover to feel productive
while defeating its own purpose. The decay compounds — each iteration re-reads the bloated
state and writes more of it back. The result is a degenerate loop: lots of motion, no
progress. This skill diagnoses that decay with evidence and fixes it.
This is a destructive-edit + behavior-change task on the user's automation. Move carefully,
show evidence before editing, and put the irreversible / judgment calls behind
AskUserQuestion (see Step 4).
When this applies
The user points at a loop and says it feels stuck, repetitive, bloated, expensive, or "off" —
or just asks to clean / health-check a /loop cron or Codex automation. If no loop is named,
ask which one (Step 1).
Workflow
1. Scope the target
Identify exactly one loop to clean. If the user named a repo/automation, use it. Otherwise
enumerate candidates and ask via AskUserQuestion:
- Claude Code crons: call the
CronList tool; also cat <repo>/.claude/scheduled_tasks.json
and ~/.claude/scheduled_tasks.json if present.
- Codex automations:
ls ~/.codex/ and look for an automations/scheduled config; check the
repo for AGENTS.md.
Confirm the platform (Claude Code vs Codex) — it decides where the context lives. See
references/platform-layout.md for exact file locations and
the discovery commands for each.
2. Map the loop's context surface
Gather every artifact the loop re-reads or re-writes each iteration. Read them, and measure
sizes (wc -l -w -c). The surface is typically:
| Artifact | Claude Code | Codex |
|---|
| The task prompt | scheduled_tasks.json prompt, or CronList | the automation's prompt/config |
| Persistent memory | ~/.claude/projects/<enc>/memory/MEMORY.md + *.md | whatever STATUS/state file the prompt maintains |
| Instruction files | ~/.claude/CLAUDE.md + <repo>/CLAUDE.md | ~/.codex/AGENTS.md + <repo>/AGENTS.md |
| Run history (for forensics) | ~/.claude/projects/<enc>/*.jsonl | ~/.codex/sessions/ or history logs |
Anchor each artifact to its own last-touch date, not a shared one. The surface decays on
independent cadences — a memory file rewritten yesterday and a CLAUDE.md untouched for months
need different staleness windows, and a single global date gives wrong reads for both. Per
artifact, record its last-modified time: git log -1 --format=%ai -- <path> for tracked files,
stat -f %Sm <path> (mtime) for untracked ones like ~/.claude/.../memory/*.md. That per-file
date is the window you measure churn and staleness against in Step 3.
3. Diagnose against the decay heuristics
Run the checks in references/detection-heuristics.md.
Each smell has a quantitative trigger and a fix. The high-value ones:
- Memory-as-changelog — a "one fact" memory file that's grown to a per-iteration log
(large size, append-not-replace, dense
DONE/deferred/done markers).
- Bloated index — Claude
MEMORY.md with a single line far over a one-line hook.
- No-exit task rule — a "skip / defer on error" clause with no "stop after N failures /
escalate after M hours" exit. A transient guard becomes a permanent stall.
- Dead-end human handoff — the prompt tells the unattended loop to "ask the user to do X";
nobody reads it in real time, so the loop churns forever.
- Unbounded queue growth — each iteration piles up unvalidated/unadopted work. Confirm
with
git: commits since the last deploy/adopt marker growing monotonically while reported
status stays upbeat.
- Doc-as-cover — CLAUDE.md / AGENTS.md framing that lets "keep building locally" feel like
discipline while the loop's actual purpose goes unmet.
Use the run history as evidence, not vibes. Extract each iteration's final report from the
*.jsonl (or Codex session) logs and look for: the same action repeated K times, a
"queue/count ahead" number rising monotonically, real progress (deploys/merges) stalled. Cite
concrete numbers (file sizes, repeat counts, queue depth) in your findings — the
detection-heuristics ref has the extraction one-liners.
Scale how much history you read to how much the loop has churned. A loop that fired 200
times since its last clean hides its decay differently than one that fired 8 times — and
reading only the most recent few sessions samples the tail, exactly where a degenerate loop
looks calmest. Count the iterations in the window (sessions since the last-touch date from Step
2), then sample accordingly:
| iterations since last clean | how to read run history |
|---|
| < 10 | deep-read every session |
| 10 – 49 | deep-read the latest ~10, spot-check the oldest 2–3 for when the smell began |
| ≥ 50 | sample across the whole window (early / middle / recent), not just the tail; the early sessions date the onset |
When you sample rather than read exhaustively, say what you skipped in your findings — a
silent tail-only read reads as "the loop is fine lately" when the bloat is upstream.
4. Ask the key judgments (AskUserQuestion)
Before any destructive or behavior-changing edit, surface findings and put the calls only
the user can make behind AskUserQuestion. Don't ask what you can verify yourself; do ask:
- Which fixes to apply (multiSelect) — list each detected smell as an option so the user
opts in per-fix rather than all-or-nothing.
- History: archive or delete — when slimming a bloated memory/state file, archive the old
body to a cold sibling file vs. drop it (git already has the history).
- Edit the task prompt itself? — changing the cron/automation prompt changes the loop's
future behavior (e.g. adding an exit condition). This is higher-stakes than cleaning files;
confirm before touching it.
- The exit / "blocked" policy — when adding a no-exit fix, ask what should happen on
persistent failure: stop generating new work and only health-check + alert, keep going, or a
threshold (after N fails / M hours). This is a real ops-judgment call.
Recommend a default in each question (first option, "(Recommended)") but let the user steer.
5. Apply the approved fixes
Common, proven moves:
- Snapshot + cold archive a bloated memory/state file. Copy the full file to a sibling
archive (e.g.
<name>-archive.md, type: reference, with a header pointing back to the live
file). Rewrite the live file as a short current-state snapshot — identity, a single
## CURRENT STATE block, what's blocking, what's next. End it with an explicit rule:
"This is a SNAPSHOT — overwrite the state block in place each iteration; do NOT append a
per-commit log here." That last line is what stops re-bloat.
- Shrink the index (Claude
MEMORY.md) back to one hook line per memory; add a dim pointer
for the archive so it's findable but not loaded every run.
- Add an exit condition to the task prompt: after N consecutive failures / M hours blocked,
stop producing new deferred work and switch to health-check + a single loud BLOCKED line +
(if available) a real notification.
- Add a counter-rule where an instruction file gives unproductive cover.
Keep edits minimal and reversible; the archive preserves everything.
6. Report and prevent recurrence
Summarize before/after with numbers (sizes, queue depth, repeat count). Flag the recurrence
risk explicitly: if you cleaned files but the loop's prompt still says "append progress each
run", the next fire re-bloats them — so the prompt fix in Step 5 is what makes the cleanup
durable. If the user declined the prompt edit, say so plainly.
External references (prior art, security-clean — read for ideas, don't run)
This skill's niche — auditing the loop (its prompt, exit conditions, and run-history), not
just a static memory file — appears unoccupied. But three adjacent, MIT/local-only projects are
worth reading when you need depth on the memory half:
wan-huiyan/memory-hygiene (GitHub) — the closest analogue: tiered memory architecture,
index truncation-risk thresholds, inline-content-extraction, and an approval gate before
destructive edits. Best single reference for the memory-file cleanup mechanics.
/slim self-audit (ai-muninn blog) — per-turn token-cost measurement and the
index→one-line-pointer + subfile shrink mechanic (mirrors Step 5's "shrink the index").
- Pensyve (
major7apps/pensyve) — FSRS decay scoring for "stale memory." Borrow the
recency-decay concept for smell #7; don't pull in its runtime.
Avoid hosted "memory hygiene" SaaS that ask you to paste memory contents off-machine, and don't
copy any rm -rf/service-restart commands from third-party SKILL.md files verbatim.
Boundaries
- Don't run the loop's actual work, deploy, or rotate secrets — this skill cleans the
loop's context, it doesn't operate the system the loop manages.
- Don't delete history without an archive unless the user explicitly chose delete in Step 4.
- Treat recalled memory as possibly stale: if it names a file/flag/commit, verify it still
exists before trusting or repeating it.