LLM Usage Burn Guard — 額度秒燒／失敗仍扣費

When to use

• Credits vanish in minutes across Codex / Claude / ChatGPT / other agents.
• Failed or empty responses still seem to burn quota.
• Multiple subscriptions run in parallel and nobody can answer “where did the money go?”

Problem framing

Public signal: users burn all credits and get no useful response; burn rate feels invisible until zero. Need metering + kill-switches, not another model.

Metering map (decide per vendor)

Fill this table for your stack before trusting any dashboard.

Guardrails (install before the long run)

1. Hard budget — max $ or max turns per job; stop when hit.
2. Soft warning — alert at 50% / 80%.
3. Confirm before expensive tools — browse, huge context, image, deploy.
4. Loop breaker — same error 3× → pause and ask human.
5. No unbounded auto-retry on empty/failed responses.

Run ledger template

Keep one ledger file per day. After a burn event, you must answer: which hour, which task, which failed paid attempts.

Playbook — credits hit zero

1. Freeze new agent jobs.
2. Classify root cause: infinite loop / retry storm / wrong model tier / surprise tool fan-out.
3. Write one rule for tomorrow (e.g. “max 20 rounds”, “no retry on empty”).
4. Only then reopen the cheapest model for triage.

Anti-patterns

• Agent loops with no turn cap
• Auto-retry forever on 5xx / empty
• Running four “smart” agents on the same bug overnight
• Inventing spend numbers when the UI is blank — log “unknown” instead

Acceptance test

You can point to the ledger line for the burn and name the failed-but-charged calls — or explicitly mark metering as opaque (blocked path).

Output template

Stack + metering table:
Hard budget / soft alerts:
Loop breaker rule:
Today's ledger path:
Burn postmortem (if any):

Out of scope

Refund disputes; hacking billing systems; recommending unpaid usage; publishing without operator approval.