Customer complaint: agent substituted inefficient workflow, consumed usage for 10m55s, and produced no clone artifacts

What version of the Codex App are you using (From “About Codex” dialog)?

26.908.4834.0 (installed Windows package OpenAI.Codex; verified locally)

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Windows x64 — Microsoft Windows NT 10.0.26200.0

What issue are you seeing?

This is a customer complaint about instruction adherence, resource stewardship, and execution controls, as well as an agent-behavior defect.

On September 15, 2026, in the desktop task “Quality Control Agent,” I requested an established, tightly specified conversation-cloning workflow. The model displayed was GPT-5.6 Terra High. The procedure normally takes approximately four minutes in my experience.

Instead, the agent repeatedly retrieved the parent conversation one turn at a time. I manually stopped the run after 10 minutes 55 seconds, as shown in the UI. It subsequently acknowledged deviating from the established method and said it had written no clone artifacts and created no successor. Its explanation was that it was trying to avoid transcript truncation.

The agent consumed limited usage without delivering the requested files, while requiring me to supervise and interrupt it. Exact token consumption and monetary cost for this incident have not been established.

I am concerned about the incentive conflict when inefficient agent execution consumes metered allowances or paid credits. These screenshots demonstrate the execution failure; they do not establish that OpenAI intentionally caused it for revenue. I want OpenAI to address the customer-protection and methodology issue, not simply recommend a longer prompt.

What steps can reproduce the bug?

Observed sequence (not independently re-run):

1. A long-running parent conversation reaches its conversation limit.
2. Give the work agent the established clone procedure: capture the full parent verbatim as Markdown, preserve the latest five relevant turns verbatim, prepare a concise current-state handoff and birth certificate, then create exactly one fresh successor and leave it awaiting human release. The instructions specify the exact parent and tightly constrain scope.
3. The agent chooses repeated single-turn retrieval instead of the established larger-page capture method.
4. Stop it manually after 10m55s.
5. Ask what happened. It acknowledges the deviation and states that no clone artifacts were written.

The complete original instructions and the agent acknowledgment are preserved in screenshots. Session identifiers and logs can be supplied privately to Support; they are not included in this public draft.

What is the expected behavior?

Follow the specified workflow efficiently and preserve verbatim content. If a capability or truncation problem prevents the established method, identify that obstacle before spending resources on a materially different approach. Persist useful output as required by the workflow and stop when an agreed resource/time boundary is reached.

Requested remedies:

• Investigate the instruction-versus-tool-call trace and explain the deviation.
• Quantify attributable usage and review whether a usage adjustment is appropriate.
• Provide enforceable per-task time, usage, and tool-call ceilings, independent of the model remembering to stop.
• Provide a way to lock an approved workflow against unapproved method substitution.
• Explain how repeated low-yield tool calls are detected and stopped.

Additional information

Please treat this as a formal customer complaint about product behavior and resource controls, not merely a request for prompting advice. If this issue tracker cannot handle account-specific usage review, please route that part to the appropriate Support team.

Evidence available: the original multi-part clone instructions; the UI showing “You stopped after 10m 55s”; and the agent acknowledgment that it repeatedly pulled one turn at a time and had not written any clone artifacts. The agent reported that it stopped after intervention.

The four-minute comparison is my prior operational benchmark, not a published OpenAI service guarantee. No exact financial loss is asserted without usage records.

No repository file, test, or code entry point is identified. Start by reviewing the instruction-versus-tool-call trace and the preserved screenshots and session logs; done means reproducing the single-turn retrieval behavior and addressing the requested resource-control outcome.