Codex Inspector

Codex Inspector uses Codex to review your sessions and generate helpful insights
Codex Inspector Dashboard provides insights about token usage
Session Inspector gives you a birds eye view of a session, broken down by turns
You can dive into each turn to see what Codex saw under the hood
Reviews provide both positive and critical feedback that you can use to be a better AI Engineer
The dashboard contains a drilldown into your most token intensive sessions

Inspiration

We are Codex power users and we wanted to be able to have more insight into how our tokens are spent and how we could use Codex more effectively. The app hides a lot of technical detail from the user. We want to be able to see what context Codex had received, or how work was distributed across agents. Answering those questions meant manually searching large session logs and reconstructing events by hand. We wanted the observability tools we expect from other developer systems, but designed specifically for private, local Codex workflows.

What it does

Codex Inspector turns the Codex session logs already on your computer into a local dashboard. It shows where tokens and capacity went, lets you trace a root session through its turns, spawned agents, tool calls, compactions, and recorded evidence, and can launch an opt-in Effectiveness Review that offers evidence-backed strengths and improvements. Every insight remains connected to the session events behind it.

How we built it

We built Inspector as three connected pieces: a Codex plugin with lifecycle hooks and skills, a Go CLI that creates a versioned local SQLite index, and a React dashboard for exploring usage, sessions, and reviews. Codex and GPT-5.6 helped us move from brainstorming and product-design grilling through architecture, implementation, browser-driven design reviews, and end-to-end verification. The team then combined the strongest ideas from independent MVP explorations into one coherent workflow.

Challenges we ran into

Codex history is large, evolving, and spread across several formats that do not always have identical coverage or identifiers. We had to correctly connect root sessions with spawned agents, distinguish exact facts from derived, estimated, or unavailable information, and keep searches responsive across large histories. Preserving a strict privacy boundary while supporting exact evidence and model-powered reviews also required careful scope previews, explicit consent, compatibility checks, and failure-safe indexing.

Accomplishments that we're proud of

We shipped a complete local-first journey from aggregate usage to a causal session map, chronological event evidence, and an explicitly started GPT-5.6 review whose citations return to the original session. Inspector handles large histories incrementally, keeps model judgment separate from deterministic indexing, and never rewrites the source sessions. We are especially proud that the full installation and hero workflow were exercised repeatedly through clean-install, browser, crash-recovery, security, and synthetic end-to-end tests.

What we learned

We learned that Codex’s local history contains enough information to build a surprisingly rich observability experience, but trustworthy interpretation depends on careful provenance and honest fidelity labels. Token usage alone rarely tells the whole story: context growth, retries, tool choice, task framing, and delegation matter just as much. Most importantly, improvement advice becomes far more useful when users can inspect its evidence and remain in control of whether Codex acts on it.

What's next for Codex Inspector

Next, we want to support more platforms and Codex history versions, improve live inspection of active sessions, and add stronger privacy controls for sharing evidence. We also see opportunities for recurring reviews, longitudinal workflow trends, and detectors that identify repeated processes worth turning into skills, hooks, or automations. The long-term goal is to help every Codex user build a workflow that continuously becomes more understandable, reliable, and effective.