Yukti

Yukti lets you ask Codex about everything visible above the current notebook cell. The notebook stays the conversation: markdown, code, outputs, earlier questions, and earlier answers all become context for the next %%ask question.

Install for development

You need JupyterLab 4, Node.js, and an authenticated Codex CLI.

codex login
cd /Users/yeshwanth/Code/Personal/yukti
npm install
npm run build:prod
uv pip install -e .

Restart JupyterLab after installation.

Use Yukti

Load the kernel extension once near the top of the notebook:

%load_ext yukti

Then ask a question in a cell:

%%ask
Why did the previous query return these rows?

Yukti chooses how to apply the response. Questions remain rendered answers in the %%ask cell. Requests for new notebook content add code and markdown cells below it. Requests to change earlier cells update those cells.

Yukti sends cells above that question through a Jupyter Comm channel. Outputs are limited to 8 KB each, and the complete request is limited to 512 KB. Yukti starts a disposable Codex App Server thread with its own visible base instruction, an empty working directory, a read-only sandbox, and no project instruction files.

Yukti accepts only ChatGPT subscription authentication. It stops before the model call when Codex reports API-key authentication.

Change what %%ask may do

By default Codex only reads the transcript. A %%yukti cell raises that limit for every later %%ask cell in the same kernel session:

%%yukti
permissions: elevated

The cell prints the settings it applied. An empty %%yukti cell, or %yukti, prints the help with the settings in force. Three profiles set them together:

Any line after permissions overrides one value:

%%yukti
permissions: elevated
cwd: ~/Code/report
writable_roots: /tmp/scratch
network: on
approvals: auto_review

• sandbox: read-only, workspace-write, or danger-full-access
• cwd: the directory Codex works in, or empty for a disposable one
• network, tools: on or off
• run: on or off, whether Yukti runs the code cells it inserts
• writable_roots: extra writable directories, separated by commas
• approvals: never, or auto_review to let a Codex subagent decide

With elevated or full, Codex may edit files and run commands. Yukti prints one line per command and per file it changes. A notebook cannot show an approval prompt, so Yukti accepts any approval request that still reaches it: the sandbox and the working directory remain the real limits. Project instruction files stay off in every profile, so an AGENTS.md never competes with Yukti's instruction.

Let Yukti run the cells it writes

Yukti runs a code cell it has just inserted and reads what that cell printed, so one question can end in a cell that works instead of a cell that looks right. The run appears where the code is: the prompt shows [*] while the kernel is inside the cell, and the outputs land under it. If the run fails, Yukti reads the traceback, rewrites the cell, and runs it again.

The code runs in your kernel, in the namespace your earlier cells built, so it sees your variables and can change them. Yukti runs only the code cells it inserted in the same turn, never a cell of your own, and never a cell that starts with %%ask or %%yukti. One turn runs at most 20 cells.

That cell holds no execute request of its own, so input() and live widgets do not work in it, and it shows the same execution count as the %%ask cell that ran it. Running is on by default, and one line turns it off:

%%yukti
run: off

See the exact Codex request

Add --debug to show the authentication type, base instruction, thread settings, and complete notebook transcript in the cell output:

%%ask --debug
Why did the previous query return these rows?

Debug mode starts the local App Server protocol and stops before the model turn.

Measure a slow turn

Add --trace to write every message of the turn to ~/.cache/yukti/traces/<time>.jsonl, and to print the timing table under the answer:

%%ask --trace
Find the number check() is hiding.

The table names each wait of half a second or more, and the header splits the turn three ways: model is Codex thinking and streaming, Yukti is the notebook work itself, and unread is a tool call that had arrived and was waiting to be read, which should stay near zero. Read any earlier trace the same way:

python -m yukti.trace ~/.cache/yukti/traces/*.jsonl

Pass --gap-ms 2000 to name only the waits over two seconds.

Copy a cell

Click a cell. Three buttons join the cell toolbar above it. Each button shows the cell as two bars, and the solid bar marks the part it copies:

• top bar solid: the source
• bottom bar solid: the output text
• both bars solid: the source, then the output

Every button wraps what it copies in a code fence, and the source fence carries the kernel's language. A cell without output copies nothing for the middle button.

The three buttons also appear in the command palette as Copy Input, Copy Output, and Copy Input and Output, so you can bind a keyboard shortcut to each one. The buttons follow the active cell, because JupyterLab keeps one cell toolbar and moves it to the cell you click.

Copied output holds the same text that %%ask sends to Codex, so a rich output such as a DataFrame arrives as its plain-text form.

Verify the installation

jupyter labextension list
python -m compileall -q yukti

The extension list should show yukti as enabled.