Codex Runtime stuck in a runaway inference loop burned credits with no project loaded
TWembley (OpenAI Codex Runtime user) · Jun 25, 2026 · Coding agent
What happened
A Codex Runtime user on Windows reported very high CPU during idle with no project loaded, plus high Defender CPU that normalized when Codex was closed. Two queries — a 15-line debug taking under five minutes and a follow-up on changed lines — wiped out a 5-hour usage limit. The app also created an empty .git directory in the project folder; deleting it collapsed the high CPU for both Codex and Defender. The issue was labeled bug, performance, and rate-limits, and remains open.
Causal vector
Runaway inference loop at idle / no project loaded; high CPU and token burn collapsed only when the app was closed
Source
Reported by GitHub issue, openai/codex #30061. Verified against the primary report.
A pre-spend gate that caps per-run spend and detects a no-progress loop would have terminated the session after the first query instead of silently consuming the full limit.
The six rule types that contain this class of failure
Per-transaction cap
Any single spend above your ceiling is BLOCKED before it moves.
Daily total
Cumulative spend across all agent calls, bounded per day.
Velocity limit
Stops runaway retry loops — the #1 cause of overnight losses.
Merchant allowlist
Only approved destinations can ever receive funds.
Category rules
Flag high-risk classes (crypto, infra, refunds) for review.
Approval threshold
Above a value, the action waits for a human.
Related incidents
Codex Windows app exhausted a weekly usage limit in about 62 hours across many parallel tasks
lty418905-blip (OpenAI Codex user) · Aug 5, 2026
Codex context auto-compaction loop rereads files, loses progress, and burns paid credits
edisonjoao1 (OpenAI Codex user) · Jul 24, 2026
OpenAI confirmed Codex background tasks were silently burning users' usage limits faster than normal
OpenAI (Codex) · Jun 30, 2026
Don't be the next entry
Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 80 documented failures, one control.