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, '\x3Cjob_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 \x3C 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 \x3Cjob_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

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] \x3CJob Name>
   - Job ID: \x3Cid>
   - Schedule: \x3Ccron_expr> (every N hours/days)
   - Runs: \x3CN> | Tokens: \x3CN>
   - Error rate: \x3CY>%
   - Delivery: delivered=\x3CN> — \x3Ctypical_summary>
   - Summary: median_len=\x3CN> chars (\x3Ctypical_summary_excerpt>)
   - Waste reason: \x3Cwhy it's burning tokens — be specific, name the pattern>
   - Confidence: High/Medium/Low
   - **Recommended fix:** \x3Cspecific actionable fix>

2. ...

3. ...

## Cost Breakdown

\x3Ctop job by tokens>
• Job: \x3Cname>
• Tokens: \x3CN>
• Daily Est.: ~\x3CN> tokens/day

...

Total tracked: \x3CN>M tokens (~\x3CN> 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 \x3Cjob_id> --cron "0 */6 * * *"

# Job 2 — reduce to once daily
openclaw-env cron edit \x3Cjob_id> --cron "0 8 * * *"

# Job 3 — disable (not needed)
openclaw-env cron disable \x3Cjob_id>

⚠️ CLI pitfall: use --cron, NOT --schedule. The openclaw-env cron edit command uses --cron \x3Cexpr> for cron expressions. Using --schedule will error with unknown option '--schedule'.

Schedule Parsing Reference

Rule: everyMs \x3C 60000 = high frequency. cron expr with */N = every N minutes.

Pattern Classification (D1-D9)

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 \x3Cjob_id>
openclaw-env cron runs --id \x3Cjob_id> --limit 5
cat ~/.openclaw/cron/jobs.json | python3 -m json.tool | grep -A5 '\x3Cjob_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:

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