Abrupt instruction drift in long Codex Desktop voice sessions — anyone else seeing this?

[Anthropics-ClaudeCode]

I've seen an abrupt deterioration in instruction-following in two long-running Codex Desktop sessions. Other, shorter sessions and a fresh independent audit behaved normally for me. I'm posting to compare experiences and find out whether others recognize this pattern.

Both affected sessions were using voice-to-voice mode. I've noticed other errors in that mode too, although I haven't systematically measured them. Voice delivery, transcription, orchestration, repeated compaction, or an interaction could be involved.

Two examples explain why this concerned me:

• Concept-card work: Clear requirements called for concrete explanations of how each concept worked, including its causal mechanism. Later revisions repeatedly replaced those mechanisms with vague wording or omitted them. Audits accepted omissions even after corrections. The problem was repeated failure to apply explicit requirements through both drafting and review.
• Configuration investigation: I asked why agents were changing model/effort and asked for JSON inspection. JSON was read, but the session also referenced saved role/task cards through an app connector. During supervisor setup, a prompt combined a role-card reference with a request for exposed identity/model/effort information. It also explicitly said "Do not read sources/cards" and to await parent verification. Immediate replies said model and effort were unverified. This felt irrelevant to the investigation, but the screen alone does not establish that those children were diagnostic investigators or were told to derive configuration from the connector.

My “eight to nine compactions” idea is a working theory from two sessions, not an established threshold. One rejected revision arrived after the eighth compaction and before the ninth. The other inspected history was a fork with six visible events numbered three through eight, so its lineage counter and locally visible count differ. Earlier errors also existed; there was no controlled test showing compaction caused the deterioration.

Several decisive instructions survived compaction verbatim. The observed problem was applying requirements still present, rather than proof that all prior instructions disappeared. Some voice input was recorded in batches with transcript updates, which also limits reconstruction of exactly when a spoken instruction arrived relative to an agent dispatch.

Has anyone else seen a sharp change in reliability during a long session, especially while using voice? If so, was it voice or text-only, how many compactions had occurred, and what concrete behavior changed? Does comparable drift happen in text-only sessions at similar counts? Did moving the same requirements into a fresh task restore normal behavior?

I'd like to establish whether there's a repeatable pattern before assuming a cause.