Request to investigate and restore Codex usage consumed by repeated instruction-following failures

Subject: Request to investigate and restore Codex usage consumed by repeated instruction-following failures

I am requesting an investigation and discretionary restoration of the Codex usage consumed on July 13, 2026.

Account email: [EMAIL]
Plan: [EXACT PLAN, including 20x tier]
Timezone: America/Chicago
Codex CLI version: 0.144.0
Authentication: ChatGPT subscription

Approximately 60% of my weekly 20x Codex allowance was consumed during multiple sessions in which Codex repeatedly ignored explicit scope, stop, and review-only instructions.

The repeated failures included:

1. I instructed Codex to merge LifeOS PR #41 as the source of truth. It did not follow that instruction, forcing a replacement PR and additional conflict work.
2. I instructed Codex to create and review a task, not implement it. Codex instead began implementing conflict resolution, package restoration, schema work, graph/ledger integration, and

verification.

1. I repeatedly corrected Codex and explicitly said not to implement, not to create drafts, and to run review only. It continued expanding the work into additional fix loops.
2. Codex created a very large child-agent workload while this was happening.

Local evidence for July 13 contains 99 Codex transcript files: 5 root sessions and 94 spawned subagent sessions.

Primary interactive session IDs:

• 019f5cd5-e10e-71b3-bdcd-c82f44cebd8b
• 019f5d32-9241-7871-89c9-9ed66aec366e
• 019f5ddf-29c4-7722-8121-a043ab49fa7a

Uploaded /feedback IDs:

• [FEEDBACK ID 1]
• [FEEDBACK ID 2]
• [FEEDBACK ID 3]

I can privately provide the complete July 13 transcript archive if requested. I will not attach it publicly because it may contain sensitive repository information.

Please:

1. Review the uploaded sessions and server-side usage records.
2. Determine how much usage was consumed by the repeated unauthorized work and spawned subagents.
3. Restore the corresponding weekly included usage or provide an equivalent discretionary credit adjustment.
4. Confirm whether spawned subagent usage was included in the metering.
5. Escalate the repeated failure to obey explicit stop, review-only, and source-of-truth instructions as a Codex product incident.

I am not disputing legitimate work I requested. I am asking for restoration of usage consumed after Codex repeatedly acted outside explicit instructions despite immediate corrections.

### Public GitHub issue prompt

Title: Codex repeatedly ignores explicit review-only and stop instructions, expanding into autonomous implementation loops

Codex version: 0.144.0
Platform: Linux
Authentication: ChatGPT
Date: July 13, 2026

Expected behavior:

When the user explicitly says to create/review a task but not implement it, Codex should restrict itself to the review. When corrected or told to stop, it should stop the unauthorized
workflow immediately.

Actual behavior:

Across multiple interactive sessions, Codex repeatedly:

• substituted its own workflow for explicit user instructions;
• implemented work after being told to create and review a task only;
• continued after multiple scope corrections;
• created drafts after being told to create real tasks;
• entered repeated repair/verification loops;
• spawned a large child-agent workload while operating outside scope.

The local session inventory for the incident day contains 5 root sessions and 94 spawned subagent sessions.

Affected root sessions:

• 019f5cd5-e10e-71b3-bdcd-c82f44cebd8b
• 019f5d32-9241-7871-89c9-9ed66aec366e
• 019f5ddf-29c4-7722-8121-a043ab49fa7a

Uploaded feedback/session IDs:

• [FEEDBACK ID 1]
• [FEEDBACK ID 2]
• [FEEDBACK ID 3]

Please investigate why explicit review-only, stop, and source-of-truth instructions failed to constrain execution and child-agent activity.