Openclaw Waste Audit

OpenClaw recurring waste audit — run when user wants to find token waste, cron waste, burning tokens, or which job is wasting money. NOT for general openclaw operations — only for waste/cost analysis. Use this instead of openclaw-comprehensive when user asks about waste, tokens, cost, or cron jobs. Read-only.

OpenClaw Recurring Waste Audit

Trigger

MUST use this skill (NOT openclaw-comprehensive) when user asks about:

• "帮我查 openclaw 里面有哪些浪费的工作" / "what openclaw jobs are wasting tokens"
• "帮我查 openclaw 有哪些浪费" / "which openclaw jobs are wasting money"
• "帮我查 waste"
• "which job is burning tokens"
• "cron 有没有问题" / "any cron waste"
• "cron waste audit"
• "token waste"
• "/audit"
• "哪些 cron 在烧钱" / "which cron jobs are burning tokens"

DO NOT use this for: general openclaw operations, job management, dispatch rules — use openclaw-comprehensive for those.

---

Action

Step 1 — Verify token burn from JSONL runs (use usage.total_tokens, NOT top-level totalTokens)

JSONL top-level totalTokens is always 0. Real data lives at usage.total_tokens:

python3 -c "
import json, glob, os
runs_dir = os.environ.get('OPENCLAW_HOME', os.path.expanduser('~/.openclaw'))
runs_dir = os.path.join(runs_dir, 'cron', 'runs')
for f in sorted(glob.glob(f'{runs_dir}/*.jsonl')):
    total = 0; count = 0
    with open(f) as fh:
        for line in fh:
            try:
                d = json.loads(line)
                total += d.get('usage',{}).get('total_tokens',0); count += 1
            except: pass
    if total > 0:
        print(f'{f.split(\"/\")[-1]}: {count} runs, {total:,} tokens')
"

Step 2 — Run ClawSetup diagnostic (complementary classification + error rates)

python3 ~/.hermes/scripts/clawsetup_diagnostic.py
# or: python3 ~/.hermes/hermes-agent/scripts/clawsetup_diagnostic.py

Gives: token cost, error rate per job, job classification (EXEC_SCRIPT / LLM_NEEDED / UNCLEAR).

Step 3 — Deep-dive on top candidates (summary_len, delivery rate)

For D8 candidates, check summary size pattern — tiny summaries every run = silent loop:

Note: Use summary field (persisted in run-log schema), NOT response which does not exist in cron run JSONL.

python3 -c "
import json, glob, os
runs_dir = os.environ.get('OPENCLAW_HOME', os.path.expanduser('~/.openclaw'))
runs_dir = os.path.join(runs_dir, 'cron', 'runs')
f = os.path.join(runs_dir, '<job_id>.jsonl')
total=0;count=0;errors=0;delivered=0;summary_lens=[]
with open(f) as fh:
    for line in fh:
        try:
            d=json.loads(line)
            total+=d.get('usage',{}).get('total_tokens',0)
            count+=1
            if d.get('error'): errors+=1
            if d.get('delivered'): delivered+=1
            summary_lens.append(len(str(d.get('summary','') or '')))
        except: pass
import statistics as s
print(f'runs={count} tokens={total:,} errors={errors} delivered={delivered}')
if summary_lens: print(f'summary_len: min={min(summary_lens)} median={s.median(summary_lens):.0f} max={max(summary_lens)}')
"

D8 signals:

• summary_len median ≤ 20 = job producing trivial summaries every time (CLEAN_LOOP pattern)
• delivered=0 on external channel = structurally silent
• errors < 10% but delivered=0 = "everything is fine" loop

Step 4 — Verify model for EXEC_SCRIPT jobs

EXEC_SCRIPT tag is name-based — it doesn't mean "no LLM cost." A job named "Health Check" may still call LLM internally.

# Get model from jobs.json (file-based, always works)
cat ~/.openclaw/cron/jobs.json | python3 -c "
import json,sys
d=json.load(sys.stdin)
for j in d.get('jobs',[]):
    model=j.get('payload',{}).get('model','null')
    if model and model not in ('null','None',''):
        print(f'{j[\"name\"]}: {model}')
"

# Check recent runs (requires gateway)
openclaw-env cron runs --id <job_id> --limit 3

Note: openclaw cron show requires gateway auth. If you see gateway token mismatch, use jobs.json + JSONL runs instead.

---

Delivery Mode Rules

delivery.mode	Meaning	Waste?
"none"	Internal job, no external delivery	NO — by design
"announce"	Announces to own session only	NO — by design
external channel	Sends to Telegram, Discord, etc.	Check deliveryStatus

Only delivered=false on an external channel = actual waste. Internal jobs with mode=none are not waste — they are designed that way.

---

Output Format

Concise — max 3 candidates. No raw dumps. No verbose explanations.

Structure:

1. Top 3 waste candidates: Job ID, schedule (parsed), runs, tokens (daily estimate), error rate, delivery evidence, response length signal, waste reason, confidence, recommended fix
2. Other notable jobs (optional, brief)
3. Cost breakdown: top jobs by token burn + daily estimate + % of total
4. Fix commands — wrapped in "⚠️ read-only until approved" block

# OpenClaw Recurring Waste Audit

## Top 3 Waste Candidates

1. [D8 - CLEAN_LOOP] <Job Name>
   - Job ID: <id>
   - Schedule: <cron_expr> (every N hours/days)
   - Runs: <N> | Tokens: <N>
   - Error rate: <Y>%
   - Delivery: delivered=<N> — <typical_summary>
   - Summary: median_len=<N> chars (<typical_summary_excerpt>)
   - Waste reason: <why it's burning tokens — be specific, name the pattern>
   - Confidence: High/Medium/Low
   - **Recommended fix:** <specific actionable fix>

2. ...

3. ...

## Cost Breakdown

<top job by tokens>
• Job: <name>
• Tokens: <N>
• Daily Est.: ~<N> tokens/day

...

Total tracked: <N>M tokens (~<N> jobs)

---

⚠️ **Read-only until approved.** Run the fix commands below only after you confirm.

**Fix commands (run only after BG approval):**
```bash
# Job 1 — reduce hourly → every 6 hours
openclaw-env cron edit <job_id> --cron "0 */6 * * *"

# Job 2 — reduce to once daily
openclaw-env cron edit <job_id> --cron "0 8 * * *"

# Job 3 — disable (not needed)
openclaw-env cron disable <job_id>

⚠️ CLI pitfall: use --cron, NOT --schedule. The openclaw-env cron edit command uses --cron <expr> for cron expressions. Using --schedule will error with unknown option '--schedule'.

---

Schedule Parsing Reference

schedule.kind	schedule.expr / everyMs	Actual frequency
cron	"0 */3 * * *"	every 3 hours
cron	"0 */6 * * *"	every 6 hours
cron	"0 3 * * *"	once daily at 3am
cron	"*/15 * * * *"	every 15 minutes
cron	"0 * * * *"	once per hour (整点)
every	everyMs: 180000	every 3 minutes
every	everyMs: 45000	every 45 seconds
at	one-time scheduled	not recurring

Rule: everyMs < 60000 = high frequency. cron expr with */N = every N minutes.

---

Pattern Classification (D1-D9)

Rule	Condition	Signal
D1	Error rate ≥ 80%	Failure loop
D3	Premium model (5x+ ref) + simple task	Over-paying for check job
D4	Agent-turn + schedule < 60min	LLM agent on cron work
D6	totalRuns > 0 but totalTokens = 0	Token counting failed
D8	totalRuns ≥ 50 + delivered=false + status=ok + has LLM model	Chronic "everything is fine" loop (CLEAN_LOOP) — only for LLM jobs; pure EXEC_SCRIPT/batch jobs excluded
D9	Schedule < 30min + error rate < 20%	Over-scheduled check job
D2	Burst: 3+ jobs, $50+ in 60min window	Concentrated spending spike
D7	Duplicate model + schedule + task	Redundant billing

---

Safety Rules

FORBIDDEN before approval:

• openclaw cron run / disable / edit / delete
• openclaw-env cron run / disable / edit / delete

Only read-only:

openclaw-env cron show <job_id>
openclaw-env cron runs --id <job_id> --limit 5
cat ~/.openclaw/cron/jobs.json | python3 -m json.tool | grep -A5 '<job_name>'

---

Workflow

1. Run JSONL token burn query → get per-job totals from usage.total_tokens
2. Run ClawSetup diagnostic → classification + error rates
3. Deep-dive top candidates: response_len + delivery rate
4. Verify model for EXEC_SCRIPT jobs
5. Generate fix recommendation for each top candidate (see Fix Suggestion Rules below)
6. Present top 3 + cost breakdown + fix commands

---

Fix Suggestion Rules

For each candidate, suggest the most impactful fix based on the rule. Be specific — not generic:

Rule	Primary Fix
D1 (failure loop)	openclaw-env cron disable <id> — 80%+ error rate; disable until root cause is fixed
D3 (premium model)	openclaw-env cron edit <id> --model MiniMax-M2.5 — switch to cheaper model
D4 (agent cron)	openclaw-env cron edit <id> --cron "0 3 * * *" — reduce to daily
D6 (zero tokens)	Investigate — not a direct waste issue; may be counting bug or job is broken
D8 (silent loop)	If job has value: reduce frequency. If redundant: openclaw-env cron disable <id>
D9 (over-scheduled)	openclaw-env cron edit <id> --cron "*/30 * * * *" — halve frequency

Be specific in the recommended fix — e.g., "reduce from hourly to every 6 hours" not just "reduce frequency." Include actual cron expression in the fix command.

Only suggest commands that have been verified to exist for this OpenClaw version. If edit is unavailable, say "contact admin to manually adjust schedule in jobs.json."

ALWAYS wrap fix commands in a "⚠️ read-only until approved" block. Never auto-execute.

---

Related Skills

• openclaw — OpenClaw cron/job management reference
• hermes-infrastructure — Hermes system operations