OpenAI accounts with spendable credits are paused at 100% included weekly usage #6465

Client or integration

Other (T3 Code through Codex app-server)

Area

Authentication and account pool

Summary

An OpenAI account can reach 100% included weekly usage while still having spendable purchased credits. OpenCodex 2.63.0 ignores WHAM credits, marks the account exhausted and pauses it. Requests can then fall back to another account with no credits and return the weekly-limit 429. Reset tickets are a separate resource and were zero in the affected case.

Reproduction

1. Configure OpenAI Pool mode with a paid account and a second account with no included headroom or credits.
2. Return this redacted/synthetic WHAM fixture for the paid account:

{"plan_type":"pro","rate_limit":{"allowed":true,"primary_window":{"used_percent":100,"limit_window_seconds":604800}},"credits":{"has_credits":true,"unlimited":false,"overage_limit_reached":false,"balance":"42.5"},"rate_limit_reset_credits":{"available_count":0}}

3. Refresh quotas, then call PUT /api/codex-auth/accounts/pause-exhausted.
4. Observe the paid account being paused. A subsequent /v1/responses request may use the empty account and fail with 429.
5. Independently, computeCodexUsageScore reports 100 and drops an existing account affinity despite spendable credit capacity.

Expected: retain fresh spendable credits separately from reset tickets, keep the account eligible, and continue to honor upstream refusals, overage flags, real request cooldowns, and explicit local admission policies.

Version

@bitkyc08/opencodex 2.63.0; the same window-only exhaustion/selection checks remain on dev e0af52c.

Operating system

macOS on Apple Silicon; Bun 1.4.0.

Provider and model

openai / gpt-6.1-sol

Logs or error output

Before any reset-ticket redemption, a request through the funded account completed with HTTP 200 while included usage was 100%; shortly afterwards, fallback to the empty account returned HTTP 429: The usage limit has been reached. Account details, request identifiers and credentials are intentionally omitted. The synthetic regression fixture above reproduces the local classification error.

Screenshots and supporting files

Regression coverage is being submitted in a linked PR. A local hotfix also updates a private operator quota-recovery script to consult the same policy helpers; that script is not part of this upstream repository.

Redacted configuration

{"codexAccountMode":"pool","autoSwitchThreshold":99}

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.