Codex CLI Text Generator

Resumable bulk text generation through the official Codex CLI and your existing ChatGPT login.

No OpenAI API key is required. codex-text uses the ChatGPT login already configured by codex login and runs the documented codex exec command.

What it handles

• TXT, JSON, JSONL/NDJSON, and CSV input
• {{field}} prompt templates for structured records
• concurrent workers, retries, timeouts, and configurable batches
• one atomic output file per item
• append-only JSONL checkpoints with automatic resume
• prompt/model/profile/schema fingerprints to detect changed jobs
• macOS and Linux available-memory guards
• Codex 5-hour and weekly usage-limit detection with deferred resume
• isolated temporary directories, read-only sandboxes, and ephemeral sessions

Requirements

• Node.js 20+
• Codex CLI installed
• Codex CLI logged in with ChatGPT, not an API key

codex login status
# Logged in using ChatGPT

Install

npm install -g github:nathankim0/codex-cli-text-generator

For local development:

git clone https://github.com/nathankim0/codex-cli-text-generator.git
cd codex-cli-text-generator
npm link

Single prompt

codex-text --prompt "Write a concise release note for a calendar app."
codex-text --prompt "Return a JSON object with title and summary." \
  --schema ./examples/article.schema.json \
  --output ./article.json

Bulk prompts

Plain text uses one non-empty line per job:

codex-text --input prompts.txt --concurrency 3

JSONL can provide stable IDs and prompts:

{"id":"welcome-ko","prompt":"Write a Korean welcome email."}
{"id":"welcome-en","prompt":"Write an English welcome email."}

codex-text --input jobs.jsonl --output-dir output/emails

Structured records can be combined with a template. Given products.csv:

id,name,audience,tone
starter,Starter Plan,freelancers,friendly
team,Team Plan,small teams,professional

And product-prompt.md:

Write a 100-word product description for {{name}}.
Audience: {{audience}}
Tone: {{tone}}
Return only the finished description.

Run:

codex-text --input products.csv \
  --template-file product-prompt.md \
  --output-dir output/descriptions \
  --results output/descriptions.jsonl \
  --concurrency 2 \
  --retries 2 \
  --batch-size 20 \
  --batch-delay 5

Resume behavior

The results JSONL file is an append-only checkpoint. A later invocation skips IDs that already have a successful record with the same prompt fingerprint:

codex-text --input jobs.jsonl
# interrupted after 40 jobs

codex-text --input jobs.jsonl
# resumes with job 41

Use --no-resume to run every item again. If an existing ID's prompt, model, profile, or schema path changes, it is regenerated automatically. Keep IDs unique and stable across runs. Prompt text itself is not stored in checkpoints.

Options

--model <model>           Codex model override
--profile <name>          Codex config profile
--schema <file>           JSON Schema for structured output
--concurrency <n>         parallel Codex processes (default: 2)
--retries <n>             retries after failure (default: 2)
--batch-size <n>          jobs per batch (default: 20)
--batch-delay <seconds>   pause between batches (default: 5)
--min-free-memory <MiB>   pause below available memory (default: 1024)
--memory-per-worker <MiB> reserve per active job (default: 512)
--memory-poll <seconds>   memory recheck interval (default: 15)
--timeout <seconds>       per-job timeout (default: 300)
--no-resume               rerun successful IDs
--extension <ext>         output file extension (default: txt)
--quiet                   hide per-job progress

Run codex-text --help for the complete CLI reference.

Cost, memory, and usage limits

codex-text does not set or read OPENAI_API_KEY. Authentication and usage follow the account shown by codex login status. Your ChatGPT plan limits still apply, so begin with the default two workers.

When Codex reports that the 5-hour or weekly limit is exhausted, the affected item is checkpointed as deferred, new work stops, and the process exits with status 2. Run the same command after reset; completed items are skipped and deferred or unstarted items continue automatically.

Before launching each process, the memory guard checks OS-available memory. On macOS it includes free, inactive, speculative, and purgeable VM pages; on Linux it uses MemAvailable. It waits below the configured floor plus the active worker reservation. Use --min-free-memory 0 --memory-per-worker 0 to disable it.

Security

• Prompts go to codex exec over stdin, not command-line arguments.
• Jobs run under --sandbox read-only in a new empty temporary directory.
• Sessions use --ephemeral and do not fill Codex session history.
• The CLI never reads ~/.codex/auth.json directly.
• Protect output and checkpoint files if generated content is sensitive.

Library API

import { checkCodex, generateText, loadJobs, runBulk } from 'codex-cli-text-generator';

await checkCodex();
const jobs = await loadJobs('./jobs.jsonl');
const summary = await runBulk({ jobs, concurrency: 2 });
console.log(summary);

License

MIT