nodecode-codex-truly
Codex Truly: the real context window of a Codex subscription, measured and written. A Nodecode rewrite of kyssta-exe/hermes-codex-truly by kyssta-exe.
The Codex catalog a subscription is served understates what the backend takes: it advertises 272K or 500K per model and publishes an 872K ceiling, while a request carrying ~920K input tokens completes. Nodecode reads providers.<id>.models[].context_window before anything it learned, so one scoped row per model hands you the window you pay for.
| What | Does |
|---|---|
codex_truly |
The model's verb: status, apply, revert, probe. |
/codex-truly |
The same from the prompt: status, full, apply [dry-run], revert, probe [MODEL]. |
| a manual topic | (help :codex-truly), the window and the trade it makes, in words. |
Get it
Tell the agent "add the codex-truly cell", or run nodecode add nc://codex-truly.
Set it up
Sign in to the subscription once, which is what puts the token in Nodecode's shared auth.json under oauth_tokens.openai-codex:
/openai-codex → Sign in
Then, in /settings (or in the codex-truly section of config.jsonc):
"codex-truly": {
"provider": "openai-codex",
"target_context_window": 900000,
"models": [],
"timeout": 60
}
| Setting | What it does |
|---|---|
provider |
The provider id whose sign-in and model rows are read. openai-codex unless a custom provider speaks the Codex wire. |
token_env |
A variable holding an access token, used before the saved sign-in. Empty when unset. |
auth_file |
The auth store to read, when it is not this home's. Empty means the organism's own. |
target_context_window |
The window apply grants each model, 272000 to 918000. 918000 is the highest value backed by measurement. |
models |
Model ids to manage. Empty manages every model the account's catalog exposes. |
timeout |
How long one request to the backend may take, in seconds. |
Use it
/codex-truly status # the one-line picture, free /codex-truly full # the table: advertised, ceiling, configured, measured /codex-truly apply dry-run # what a write would change /codex-truly apply # write it; a new session reads the wider window /codex-truly revert # take back exactly those values /codex-truly probe gpt-6.1-sol # re-measure the true ceiling live
From eval, in the codex-truly package:
(codex-truly:status) ; the table, as text (codex-truly:apply-windows) ; the tool's own body (codex-truly:probe "gpt-6.1-sol")
probe bisects the live backend: a refused request is free, an accepted one bills its input to the subscription, so it costs about one accepted request and is never run by itself. There is no full 1M of input on this route — about 1.05M total context shares a 128K output reserve, leaving ~920K — and a window past ~925K fails late, mid-stream, as context_length_exceeded.
What it does not do
The original also selected a Codex-backed image provider and generated images through the subscription. That is Nodecode's own image_gen choice and its own image cells, so this cell leaves it alone and only reports the window. The interactive settings wizard and the /codex-truly desktop panel are gone with the harness that had them: the section is the settings, and the command is the panel.
