CODEX CLI PLUGIN

You use Codex every day,
but can't see where you fall short.

Codex Radar replays the real sessions you've already run — not a quiz. Deterministic scripts extract the facts, compute the formula baselines, validate every evidence reference, and calculate the final scores across Communication, Engineering, and Outcome. Your active Codex session contributes only bounded evidence adjustments and the qualitative diagnosis, then gives you prompt rewrites to use in your next thread.

Every score traces to evidence Open, public rubric Runs in your own Codex
Codex Radar
sample report
FEATURE BUILD
Strong, steady delivery — but tests rarely back the changes you ship.
B
72 / 100
Comm
68B
Eng
76A
Out
73B
Strongest: Efficiency Bottleneck: Proof Check
Why Codex Radar

It grades what you actually did — not a quiz

It reads the sessions you've already run with Codex and turns them into a diagnosis you recognize and can act on: where you're strong, where you hold yourself back, and exactly what to do next.

Scored on how you really work, not how you describe it

It reads the JSONL Codex writes, so your score reflects how you actually work — not how you'd describe it. Every prompt you send, every shell command and exit code, every apply_patch edit, plan update, and MCP / search / sub-agent call — all of it counts.

Every score points back to your own words

A parser extracts facts and computes the formula baselines. Your active Codex session proposes bounded, evidence-cited adjustments and writes the diagnosis; a strict finalizer validates those references and calculates every final score. No hidden arithmetic and no uncited adjustment.

Paste-ready prompts that make the next run smoother

No "think harder" platitudes. Each weak spot comes with concrete wording you can paste straight into your next thread, tagged with the dimension it lifts and its expected impact — ready to use, so the next session with Codex really does go smoother.

9 Dimensions × 3 Categories

Measured, not guessed

Every score traces back to a formula and a countable signal from your session. No black box.

Communication
3 DIMS
How clearly you direct Codex through text.
Lock-On 瞄准力
Goal, expected behavior, file paths, constraints in your directives.
Scene Setting 铺场力
How much useful context your opening messages carry.
Steering 校准力
Quality of mid-session course-correction and verification asks.
Engineering
3 DIMS
How well you drive the Codex platform.
Toolcraft 工具调度
Shell, patch, plan, MCP, search, subagents + command success rate.
Architecture 工程脚手架
Durable setup — AGENTS.md, repo scaffolding, tests, manifests.
Tempo 推进节奏
Steady progress without aborts, compactions, or tool churn.
Outcome
3 DIMS
What actually gets done.
Efficiency 产出效率
Useful output per human message and per tool call.
Proof Check 验证意识
Tests, builds, lint, screenshots, renders, explicit checks.
Completion 闭环完成
Turns finish cleanly instead of trailing off mid-task.
How it works

Facts in, judgment out

Five controlled steps separate deterministic evidence and scoring from model-authored interpretation.

1

Parse & score facts

Matches the project, extracts redacted signals and evidence, then computes all formula baselines in code.

~/.codex/sessions
2

Prepare model input

Evaluates recipe triggers in code and emits a compact input with an explicit untrusted-data boundary.

model-input.json
3

Interpret

Your active Codex session writes evidence adjustments, observations, diagnosis, and typed suggestions.

analysis-1.json
4

Validate & finalize

Rejects invalid evidence or adjustments, then calculates final, category, and overall scores.

report-3.0.json
5

Render

Writes a validated, self-contained HTML dashboard with history and score deltas.

~/.codex-radar/reports
Fairness

A 3-message fix isn't a 50-session build

Every project is auto-classified into one of five profiles, each with its own category weights — and codex exec pipelines are judged as automation, not as conversation.

AOne-shot · few messages, light editing
BFeature build · real implementation work
CLong-running · many sessions or compactions
DLearning · conversation-heavy, few edits
EAutomation · codex exec batch pipelines
Honest N/A, not a faked 50. When a signal genuinely can't be evaluated — like Architecture when the project directory no longer exists on this machine — the dimension is marked N/A and dropped from the average.
Privacy

Local artifacts, controlled disclosure

Bundled scripts make no network calls and write private local artifacts. Standard mode redacts common credentials; strict mode omits message/command snippets, thread titles, web-search queries, and episode opening/closing text. Qualitative interpretation runs in your active Codex session and follows its configured data handling.

No plugin network calls
No separate API key
Credential redaction
Strict minimized-evidence mode
Private file permissions
Read-only sessions
Get started

Two commands, then ask

Requires the Codex CLI and Node.js 18+. No build step, no server.

codex
$ codex plugin marketplace add LeifDiao/codex-radar
$ codex plugin add codex-radar@codex-radar-marketplace

# start a new thread, then ask:
 Run Codex Radar on this project

Or develop locally: git clone the repo, then codex plugin marketplace add .

See exactly how you and Codex work together

Deterministic scoring runs through bundled scripts; qualitative interpretation uses the Codex session you're already in, with no separate API key.

Get Codex Radar