Bring in what matters.
Read a function or a focused slice of a file. Keep full logs on disk and retrieve the parts you need.
find → read → inspectA LAUNCHER + COMPANION CODEX PLUGIN
Spend fewer tokens.
Keep the work moving.
Use Codex with a more organized workbench. Start tasks in the LeanTask launcher to give the model focused code, short task notes, and compact logs—with the potential to spend fewer tokens on repeated context.
Designed to reduce repeated context. Actual token savings vary and are not yet proven.
macOS & Linux · Python 3.11+ · Your Codex login
▧One task.Shared context.
01 Find the function ✓
02 Make the change ✓
03 Keep building ↗
notes · code · logs · usage
BUILT TO BE INSPECTED
98 automated tests
Python 3.11–3.13 CI
Offline demo included
See beta test evidence ↗A LITTLE LESS REPEATING YOURSELF
Useful context, across the whole task.
Think of tokens as the pieces of text an AI reads and writes. LeanTask tries to give Codex the useful pages from a project, instead of repeatedly handing it the whole folder. It still uses your Codex account and model.
Carry decisions forward.
Save requirements and observations with the task. Notes backed by source files are marked stale when that evidence changes.
same task, next promptKnow what actually ran.
See reported input, cached input and output tokens. Inspect command receipts and the underlying logs.
evidence over estimatesTWO PARTS. ONE SETUP GUIDE.
The launcher does the heavy lifting.
The setup below installs the launcher and the companion plugin. They have different jobs.
Where you start your coding tasks.
Open tasklean ui, choose your project, and describe the work. The launcher gives Codex tools to read selected code, keep short notes, and retrieve saved logs. Continue with follow-up prompts in the same task.
Use this for the main context-efficiency features. You also get task usage and account-limit visibility.
Help inside your regular Codex workspace.
The plugin adds an efficiency-audit skill and a hook that records prompt-size estimates locally. When invoked, its skill guides Codex in using the tools.
Installing it does not automatically optimize every Codex chat. It does not silently rewrite prompts or change models.
A SIMPLE EXAMPLE
“Fix the cart total. Now add a test.”
LeanTask can return just the total-calculation function, save a short note about the quantity rule, and return a test summary with the full log available on demand. That can mean less text for Codex to process across follow-up prompts.
Ordinary Codex already supports continued conversations and efficient reads. The added value here is a consistent set of tools and saved evidence—not a special smaller model or a guarantee of fewer tokens.
THE WORKSPACE
A clearer view of the work.
Tasks, conversations and evidence in one local browser window.
PERSISTENT TASK
Fix cart quantity behavior.
Honor quantity while preserving the default and empty-cart behavior.
The total now multiplies each price by its quantity. Default quantity remains 1. The existing tests pass.
Saved answerRetrievable test logContinue this task… ↗
Preserve the public API.
Quantity defaults to 1. An empty cart returns 0.
The calculation multiplies unit price by quantity. If the source changes, this note is marked stale.
cart.pyTest run completed.
Inspect the exit status, a compact summary, and the saved command output.
The real dashboard displays reported tokens after a model turn. Cached input is included within input, not counted twice.
No savings claim inferredThis preview uses example content. The installed dashboard runs real tasks through your own Codex account.
FROM CLONE TO FIRST TASK
Start locally.
Try it for yourself.
Follow all the commands to install both the launcher and the Codex plugin, then open the dashboard and choose Try the offline demo. It creates a disposable project and checks the workflow without making a model call.
For real coding, create a task using an existing project folder. Prompts use your Codex login and account allowance. Project access starts read-only.
Full setup & troubleshooting guide ↗TERMINAL / macOS & Linux
git clone https://github.com/Thinkelution/codex-toptimizer.git
cd codex-toptimizer
python3 -m venv .venv
source .venv/bin/activate
# 1. Install the launcher + browser UI
python -m pip install ./plugins/tasklean
# 2. Register this plugin source and install the companion
codex plugin marketplace add .
codex plugin add tasklean@codex-leantask
# 3. Check setup and open the launcher
tasklean doctor
tasklean uiRequires Python 3.11+, Git and a recent Codex CLI on your PATH. Real model turns require Codex login. Start a new Codex chat after installing the plugin; review and trust its hook before it runs.
↳ Already installed? Run
tasklean ui in your environment. Keep the launcher terminal open while working.AN HONEST BETA
Measure first.
Optimize with evidence.
For the same task, LeanTask may use fewer tokens than a direct Codex run when its focused tools avoid repeated large reads and logs. It can also use the same or more tokens: tool instructions, notes, and summaries add overhead. We have not yet proven net savings in a controlled comparison.
The beta has been exercised with real multi-prompt coding and SSH/dataset reuse. A fair comparison must use the same task, starting code, model, and quality checks, and include follow-up prompts, retries, and all overhead. Smaller prompts alone do not prove a smaller bill or less weekly allowance.
Read the benchmark approach ↗GOOD TO KNOW
A few practical details.
Does this website run my coding tasks?
No. This is the public setup site. The dashboard runs on your own machine, binds to a local address and uses your existing Codex configuration. Your task files are not uploaded to this website. Codex receives the information used in model calls you launch.
Is it only a command-line tool?
No. Start with tasklean ui to use the browser dashboard. The CLI and chat loop remain available for terminal workflows, and existing task folders can be imported into the UI.
What does the Codex plugin add?
The setup guide installs both components. The launcher supplies the main source and memory tools for tasks you start through it. The companion plugin adds an audit skill and passive prompt-size metrics in ordinary Codex. The launcher can also run on its own; pip alone does not install or enable the Codex plugin.
Can it reuse remote connections and datasets?
The CLI supports task-scoped SSH connection reuse and a separate remote dataset worker. This keeps parsed data available for repeated queries. It does not put entire datasets into model context or preserve arbitrary shell variables. Read the session guide.
What is still missing?
This beta targets macOS and Linux. Exact symbol reads currently support Python; other languages use line reads. The dashboard shows running status and completed results, with no live event stream or cancel button yet. Hosted team accounts, automatic model routing, billing and automatic test skipping are not implemented.