Please add a developer-mode Context Inspector that shows what information is being supplied to the model for the current Codex session.
This would help users troubleshoot unexpected behaviour, refine AGENTS.md files, identify conflicting instructions, detect unnecessary context and discover when installed plugins or skills are adding significant overhead.
The inspector should show, in effective precedence order:
-
AGENTS.md files, including their paths and applicable scope
-
Project and user instructions
-
Compaction summaries
-
Conversation history currently retained in active context
-
Loaded files and relevant tool output
-
Enabled plugins, apps, skills and tool definitions
-
Memory or other retrieved context
-
Approximate token usage for each source
-
Any excluded, truncated or superseded context
Security-sensitive platform instructions could remain redacted, but their presence, precedence and token contribution should still be visible.
Ideally, users could inspect the context used for a specific response and export a sanitized diagnostic snapshot.
At present, troubleshooting context problems requires guesswork. A Context Inspector would make Codex considerably easier to configure, optimize and trust—especially when multiple AGENTS.md files, plugins, skills and automatic compaction are involved.