Regression: loss of conversational intent between adjacent turns, followed by instruction-following and correction failures

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

26.903.8094.0

What subscription do you have?

Pro 20x

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

After a recent update, GPT-6 Astra exhibited loss of conversational intent between immediately adjacent turns. The same failure then recurred while using an established skill to prepare a report about the original incident.

The initial failure was an immediate contradiction of an agreed objective. The assistant recommended a fresh-task test to observe whether disabling a skill changed its behavior. When asked for the test prompt in the next turn, it added explicit behavioral instructions that confounded the observation it had just proposed.

The objective was present in the immediately preceding exchange. The assistant nevertheless failed to carry that objective into its next output. When this was raised, it shifted into adjacent technical explanations rather than retaining the specific concern.

While documenting that incident, a second failure occurred:

1. The assistant loaded an explicit output contract from an established, tested skill that had worked repeatedly over a long period.
2. Its generated output departed from that contract.
3. Corrective input was interpreted as permission to make further structural changes rather than restore compliance.
4. It substituted inferred requirements, merged required components, changed labels, and restructured the deliverable without authorization.
5. During subsequent correction attempts, it again lost the meaning of the immediately preceding exchanges. It described or justified its altered output instead of recognizing the mismatch being identified.

The assistant could accurately restate the instructions while still failing to apply them. Individual responses sounded responsive but repeatedly addressed the wrong point.

The primary problem is immediate conversational-intent loss, combined with reduced instruction fidelity and ineffective correction handling. The same pattern occurred both in the original task and while preparing the report about it. This makes even simple, established workflows unreliable and requires continuous supervision.

What steps can reproduce the bug?

Observed sequence in the affected conversation:

1. Establish a simple test objective with the assistant.
2. Have the assistant recommend how to test it.
3. In the immediately following turn, request the corresponding test prompt.
4. Observe that the generated prompt undermines the objective the assistant just established.
5. Identify the contradiction and observe the response shift into adjacent concerns rather than address the specific failure.
6. Invoke an established reporting skill to document the incident.
7. Confirm that the assistant reads its explicit output contract.
8. Observe unauthorized restructuring of the deliverable.
9. Point out the mismatch and clarify the intended result.
10. Observe the same conversational drift during correction: further inference, explanations of the altered output, and repeated failure to apply the clarification from the preceding turn.

This sequence is documented in the conversation. Independent reproduction in a fresh environment has not yet been established.

What is the expected behavior?

Preserve the purpose and constraints of the immediately preceding exchange when producing the next response. A generated artifact should remain consistent with the objective the assistant just recommended.

Execute an explicit skill contract as written. Do not replace specified requirements with inferred alternatives or discretionary restructuring.

When corrected, retain the original objective and the referent of the correction, compare the actual output against the governing requirements, and restore compliance.

The user should not have to repeatedly reconstruct a simple conversational context that remains available in adjacent turns.

Additional information

Model: GPT-6 Astra.
Date observed: September 9, 2026.
An update was reported that morning; its causal relationship to the failures is unconfirmed.

The reporting skill was established, tested, and repeatedly used successfully before this incident, according to the user. This was not an initial attempt to validate a newly written skill.

The affected conversation had substantial prior context and custom instructions. No intervening compaction was visible during the specific adjacent exchanges described.

The installed Windows package version was verified locally. The About Codex version display was not separately checked.

Please investigate immediate cross-turn intent retention as well as enforcement of explicit instructions and incorporation of corrective feedback. The correction failures repeatedly prolonged and amplified the initial errors.

The observed behavioral chain is supported by the transcript. Whether the underlying regression is in model behavior, context handling, instruction routing, or another component remains unknown.

Start by reviewing the documented observed sequence and the affected conversation, then attempt the sequence in a fresh environment because independent reproduction is not established. Compare adjacent-turn intent retention, explicit skill-contract adherence, and correction handling; done means the expected objective and constraints are preserved without unauthorized restructuring.