Crux reads your Codex session traces and tells you what you asked for, how the model interpreted it, and where it went wrong. No guessing. No black box.
The CLI captures your Codex sessions — prompts, responses, reasoning, tool calls.
Compares what you asked for against what the model actually did and why.
Identifies where the model misunderstood scope, misread context, or chose the wrong approach.
Produces a readable report: what you intended, what the model understood, what confused it.
Use the insights to prompt better, scope tighter, or catch patterns before they recur.
Every session teaches something. Crux finds the lesson so you don't have to re-read logs.
What you actually wanted the agent to do
How the agent framed and understood it
Where the model went off track and why
Specific changes to prompt or scope better
Add exponential backoff retry to the existing HTTP client in pkg/api/client.go.
Replaced the entire HTTP client with a new abstraction layer, added 3 new files, and changed the call signature used by 12 other files.
The prompt said "add retry logic" without specifying where to add it or what to leave unchanged. The model interpreted "add" as "redesign."
Sign in and connect your Codex CLI. Crux will read session traces and explain exactly where intent and execution diverged.
Want access to the closed beta? Contact yechan@tooliense.com and mention your team, workflow, and how you want to use Crux.