Unexpected 50% Weekly Usage Drop During Subagent Review/Wait Workflow

# Unexpected 50% Weekly Usage Drop During Subagent Review/Wait Workflow

## Summary

Codex weekly usage dropped from 86% remaining to 36% remaining overnight during a long-running coding task that
repeatedly dispatched one reviewer subagent and waited for it to return.

The root agent spent substantial periods suspended in wait_agent. Waiting itself should consume negligible
model tokens, but the reviewer continued performing model and tool work.

## Observed Usage

Activity Estimated tracked tokens Estimated share
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Root agent: context, fixes, tests, reports ~1,100,000 ~67%
─────────────────────── ──────────── ─────────
Reviewer: repeated adversarial reviews ~550,000 ~33%
──────────────────────── ──────────── ─────────
Root suspended in wait_agent Approximately 0 Approximately 0%
───────────────────── ────────────────── ────────────
Goal counter total 1,652,259 100%

Additional observations:

• Weekly capacity before the run: 86% remaining
• Weekly capacity after the run: 36% remaining
• Total decrease: 50 percentage points
• Goal duration: approximately 10.5 hours
• Only one reviewer was active at a time
• No parallel review fleet was intentionally running

The internal goal counter is not necessarily equivalent to the account billing meter. It does not provide a
per-model breakdown or separate input, cached-input, and output tokens.

## Expected Behavior

• Suspending the root agent in wait_agent should not consume model tokens.
• Usage generated by a subagent should be separately attributable.
• Long-running subagent loops should expose their current token or credit consumption.
• Users should receive a warning before a single workflow consumes a material percentage of the weekly

allowance.

## Actual Behavior

The workflow repeatedly followed this pattern:

1. Dispatch reviewer.
2. Suspend root agent and wait.
3. Reviewer performs an adversarial review.
4. Root processes the findings.
5. Root implements repairs and reruns tests.
6. Root dispatches the same reviewer again.
7. Repeat without a usage ceiling or checkpoint.

The same reviewer thread was reused, so its retained context grew after every review wave.

## Suspected Cause

The primary confirmed problem is an unbounded review-convergence loop, not the wait operation itself:

review → fix → full test → adversarial review → fix → full test → …

Each follow-up requested additional adversarial exploration instead of limiting verification to the named
repairs. The workflow had no:

• Token ceiling
• Credit ceiling
• Maximum review-wave count
• Weekly-usage checkpoint
• Warning threshold
• Automatic pause condition

A separate accounting problem remains possible if weekly capacity decreases while both root and subagents are
inactive, but this run does not prove that condition.

## Requested Investigation

• Confirm whether wait_agent itself incurs any metered usage.
• Provide per-thread and per-subagent token or credit attribution.
• Distinguish root activity, subagent activity, cached input, output, and tool-result context.
• Explain how the 1,652,259 tracked tokens contributed to the observed weekly allowance reduction.
• Verify that completed or idle subagents cannot continue consuming usage.
• Add an alert or automatic checkpoint before a workflow consumes a configurable percentage of weekly capacity.

## Acceptance Criteria

• [ ] An idle root waiting on an inactive agent consumes no metered model usage.
• [ ] Active subagent consumption is visible separately from root consumption.
• [ ] The usage dashboard identifies which thread or workflow consumed capacity.
• [ ] Long-running workflows can be assigned a token or credit ceiling.
• [ ] Codex warns before consuming a material portion of the weekly allowance.
• [ ] Review loops can enforce a maximum number of convergence waves.
• [ ] Raw tracked tokens can be reconciled with credits and weekly usage percentage.

I can provide account-specific details and screenshots privately to OpenAI Support, but I am not posting identifying or billing information publicly in this issue.