Closed Beta

See what your Codex agent actually understood.

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.

Get Started See How It Works
Built for
OpenAI Codex
Crux — Trace Analysis
12 Codex sessions analyzed — 2 misalignment patterns found
Model refactored instead of fixing the bug High confidence
What you asked
"Fix the failing test in auth_test.go"
What the model did
Rewrote the auth middleware and updated 4 files instead of patching the test assertion.
Why it happened
+ The prompt lacked scope constraint — no "only touch this file".
+ Model defaulted to a structural rewrite when the fix was a one-liner.
Misalignment Scope expanded without constraint
3 sessions
Aligned Clear file-scoped fix requests
9 sessions
How It Works

From raw traces to clear explanations

1

Collect traces

The CLI captures your Codex sessions — prompts, responses, reasoning, tool calls.

2

Analyze intent vs. outcome

Compares what you asked for against what the model actually did and why.

3

Find misalignment

Identifies where the model misunderstood scope, misread context, or chose the wrong approach.

4

Explain what happened

Produces a readable report: what you intended, what the model understood, what confused it.

5

You adjust

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 It Analyzes

Every trace tells a story

Your intent

What you actually wanted the agent to do

Model interpretation

How the agent framed and understood it

Confusion points

Where the model went off track and why

Actionable advice

Specific changes to prompt or scope better


Example Analysis

What a trace analysis looks like

Session: "Add retry logic to the API client" Misaligned
?

Your intent

Add exponential backoff retry to the existing HTTP client in pkg/api/client.go.

!

What the model did

Replaced the entire HTTP client with a new abstraction layer, added 3 new files, and changed the call signature used by 12 other files.

What confused the model

The prompt said "add retry logic" without specifying where to add it or what to leave unchanged. The model interpreted "add" as "redesign."

15 files
Changed (expected: 1)
Scope
Primary confusion source
High
Confidence

Stop guessing what your Codex agent did wrong

Sign in and connect your Codex CLI. Crux will read session traces and explain exactly where intent and execution diverged.

Closed Beta

Want access to the closed beta? Contact yechan@tooliense.com and mention your team, workflow, and how you want to use Crux.

Enter your beta credentials.