Agent Routing Waste Audit

Paste an agent job, cron, routing, or run summary and get an immediate read-only audit of possible routing, retry, fallback, or model-assignment waste. Progressive evidence handling — thin input produces a preliminary finding, richer input produces a stronger audit.

Overview

A cross-agent routing waste audit skill. It inspects pasted run summaries — from Hermes cron lists, OpenClaw job lists, LiteLLM/OpenRouter usage exports, PilotDeck routing sessions, or generic JSON/CSV/log snippets — and produces an immediate read-only audit of possible routing, retry, fallback, or model-assignment waste.

One principle: paste what you have, get a ranked audit immediately. Do not fill a template first.

This skill is a sibling to waste-audit. They are not interchangeable:

• waste-audit → recurring OpenClaw job waste (schedule, delivery, token burn)
• agent-routing-waste-audit → routing decisions within runs (model tier, retry, fallback, sub-agent, local/cloud split)

Use waste-audit for OpenClaw recurring job waste. Use this skill when the main question is routing, retry, fallback, sub-agent model choice, or local/cloud split.

This skill is read-only. It does not edit configs, switch providers, disable jobs, or auto-apply routing changes.

Activation

Activate when a user types:

audit agent routing waste

Also activate for questions about:

• whether a task used an overpowered model
• whether retry loops burned tokens silently
• whether fallback chains escalated cost without recovery
• whether sub-agents were assigned to unnecessarily strong models
• whether local/cloud model split was unclear or wasteful
• whether a scheduled agent job still justifies its model tier

Do Not Use For

• Recurring OpenClaw job waste with no routing question → use waste-audit instead
• Full provider or API key configuration → not this skill's scope
• Auto-fixing or auto-disabling anything → read-only by design
• Financial, legal, or security advice → not this skill's scope

Start Here

Paste any one of these:

hermes cron list

or

• an OpenClaw job list
• an agent run summary
• a LiteLLM / OpenRouter usage summary
• a PilotDeck routing / session summary
• a redacted JSON / CSV / log snippet

The skill will first produce a preliminary audit from whatever is available, then ask for only the next most useful evidence. Do not prepare a structured form before pasting.

Evidence Depth Levels

The skill operates at three evidence levels. Each level adds weight to findings but none are required upfront.

Level	What you have	What you get
Level 1	Schedule / job list only	Preliminary ranking, highest-frequency candidates flagged
Level 2	Job metadata + prompt summary	Stronger signal on whether the task needs an LLM every run
Level 3	Token usage + retry + fallback + output usefulness	High-confidence audit with policy recommendation

Do not block on Level 3. Paste what you have now.

Hermes Cron List Parser

When the input is raw hermes cron list output, the skill parses this block structure:

<job_id> [active|paused|completed]
    Name:      <name>
    Schedule:  <cron expression>
    Repeat:    <∞|number>
    Next run:  <ISO timestamp>
    Deliver:   <target>
    Last run:  <ISO timestamp>  <ok|error|...>

Extract these fields:

• job_id
• state (active / paused / completed)
• name
• schedule (cron expression)
• repeat (∞ or a number)
• next_run (ISO timestamp)
• delivery_target
• last_run_time
• last_status (ok / error / other)

Frequency mapping (approximate runs per day):

Schedule	Runs/day	Priority
*/5 * * * *	~288	High
*/10 * * * *	~144	High
*/15 * * * *	~96	Medium
*/30 * * * *	~48	Medium
0 * * * *	~24	Medium
0 9 * * *	~1	Low
Weekly / monthly	<1	Low
Unknown / natural language	—	Mark: schedule parse uncertain

Output Structure

---

Initial Finding

One concise sentence. Example: "One high-priority audit candidate found."

---

Top Routing / Waste Candidate

Field	Value
Name	job / run / session name
ID	job_id or run identifier
Why flagged	Specific signal: schedule frequency, retry pattern, model tier, fallback chain, etc.
Estimated frequency	Approximate runs/day if schedule is available
Confidence	High / Medium / Low
Evidence depth	Level 1 / 2 / 3

---

Lower-Priority Candidates

Only include if there are meaningful candidates below the top one. For each:

• Name + ID
• Why it is lower priority
• Estimated frequency

---

Missing Evidence

Short list only. Do not overwhelm the user. Example:

• model / provider used
• whether LLM is invoked every run
• token usage for normal runs
• whether normal "ok" outputs are useful

---

Suggested Manual Check

One conservative next check. No edits. Example:

Inspect whether hermes-health-watchdog can become script-first — LLM-only-on-anomaly — instead of running an LLM on every scheduled tick when last run was "ok".

---

Copy-Paste Prompt for Your Agent

A bounded prompt for your agent to manually investigate the top candidate only.

Do not include instructions to edit, disable, delete, or mutate anything.

Please inspect this agent run for possible routing waste.

Run: <name / identifier>
Flagged because: <what triggered the flag — schedule frequency, model tier, retry pattern, etc.>
Evidence depth so far: <Level 1 / 2 / 3>

Do not edit, disable, delete, or mutate anything yet.

Inspect and report:
1. whether this is real routing or model-assignment waste
2. whether a normal "ok" run actually needed an LLM
3. whether script-first / anomaly-only / silent-on-ok is a safer policy
4. the safest manual next step, with no changes applied yet

Redact secrets. Do not expose raw private payloads.

---

NEEDS INFO

Only output NEEDS INFO when:

• input cannot be parsed at all (no identifiable job / run / session)
• no schedule, model, retry, fallback, task, or recurrence signal exists
• the user asks for a final policy decision but provides only a free-text description with no identifiable signals

Do not use NEEDS INFO for missing token counts, model names, or retry counts alone. A parseable cron list with a schedule is audit-eligible at Level 1.

---

Safety Boundaries

This skill will not:

• Edit, disable, delete, or mutate any config, job, or routing policy
• Switch providers or models automatically
• Apply hidden optimization or auto-fix behavior
• Suggest downgrading code review, security, production, wallet, payment, or irreversible-action workflows without explicit human approval
• Expose secrets or raw private logs
• Fabricate missing fields or assume unknown values

---

Relationship to waste-audit

	waste-audit	agent-routing-waste-audit
Focus	Recurring OpenClaw job waste	Routing decisions within runs
Input	OpenClaw cron jobs + runs	Any agent runtime: Hermes, OpenClaw, LiteLLM, PilotDeck, etc.
Signals	Token burn, delivery silence, error rate	Model tier, retry, fallback, sub-agent, local/cloud
Blocking requirement	None — schedule list enough for Level 1	No token data required for Level 1; at least one parseable job/run/routing signal is required

Do not run both on the same input without a clear reason. If the input is an OpenClaw cron job and the question is "should this keep running?", use waste-audit. If the question is "was this routed correctly?", use this skill.

---

Common Pitfalls

1. Fabricating missing fields — use "missing" in evidence tables; do not assume model names, token counts, or retry counts
2. Treating Level 1 as Level 3 — a preliminary finding from a schedule list has lower confidence; be honest about evidence depth
3. Flagging all candidates as high-priority — rank by frequency and compounding cost; a daily job at 1 run/day is not the same as 288 runs/day
4. Suggesting changes to production routing without a human review gate — every suggested policy change must include an explicit review step
5. Running this instead of waste-audit for simple recurring job hygiene — they are siblings; use the right one
6. Claiming broad native support for platforms not confirmed — parse what you can, label what you cannot

---

Verification Checklist

• Input parsed at correct evidence level — Level 1 / 2 / 3 stated explicitly
• Highest-frequency candidate ranked first
• Frequency approximation correct for the cron expression (288 for */5, 1 for 0 9 * * *, etc.)
• NEEDS INFO only when input is truly unparseable or signal-free
• Missing evidence listed but does not block the preliminary finding
• Copy-paste prompt instructs do-not-edit clearly
• Safety boundaries present in output
• If OpenClaw recurring job input, relationship to waste-audit noted