Problem:
Visible UI State ≠ Complete Context State
Users see the chat, memory summary and some sources, but not always which context actually influenced a specific response. OpenAI docs confirm ChatGPT may use past chats, memories, custom instructions, Library files and connected apps, and that Memory Sources may not show every factor.
This makes long-running work hard to debug — e.g. after a persistent maximum-length warning (2h54m case) or after Codex outage Sep 26 2026 when OpenAI did a banked reset for paid users across Codex and Work.
Proposal — optional per-response manifest (collapsed by default):
Current chat used: YES/NO Past-chat context used: YES/NO Memory used: YES/NO Custom instructions used: YES/NO Connected source used: YES/NO Context compressed: YES/NO Conversation limit reached: YES/NO
Goal:
Make Documented Behavior ?= Observed Behavior testable.
This is NOT about alleging manipulation — Partial Observability ⇏ Malicious Manipulation. Just about making context testable so ChatGPT is easier to debug and trust for professional work.
Code in Justice and Respect