Inspect the input.
Find truncated descriptions, overlapping skills, and instructions written for older models.
Astra Xray
Inspect truncated descriptions, duplicates, and outdated instructions. Review what Astra receives, then back up and clean up only approved changes.
Find truncated descriptions, overlapping skills, and instructions written for older models.
See what will change and why. Approved files are backed up to a ZIP before editing.
Run the checks again and compare. Built-in skills, plugins, model settings, and approval settings stay unchanged.
Why skill descriptions get cut, and what changes after cleanup, with real results.
Codex gives the skill list only 2% of the context. With many skills installed, descriptions reach the model cut short, so it never reads the part that says when to use a skill. Overlapping skills and instructions written for older models make it worse.
The list math follows the Codex 0.154.0 source line by line and matched a real session log on 84 of 84 lines.
A month of those savings is about four dollars at Astra API prices. The point is not money but making the model read each description to the end. Where tokens really go (tool results and skill bodies) is what Codex Token Xray measures.
Scan results and backups live only under ~/.astra-xray. Every file is zipped before it changes, so you can undo at any time.
Download the repository and copy skills/astra-xray into ~/.agents/skills/.
OpenAI’s article on Astra skills and prompts explains the background.
After installing, type $astra-xray in Codex. Requires Python 3.9 or newer.