@mzwing/pi-codex-downgrade-detector

Tells you when a Pi turn was served by a model other than the one you selected.

A substituted turn and a clean turn look identical in the transcript. This reads the model the server itself named — the openai-model response header, or the model field a completions-style relay echoes back — compares it against what Pi asked for, and puts the answer in the footer:

✓ codex gpt-6-astra                 the server confirmed your model
↓ codex gpt-6-astra→gpt-5.6-luna    something cheaper answered
↑ codex gpt-5.4→gpt-6-astra         something else answered, still not what you picked
⚠ codex gpt-6-astra≠claude-opus-5   different, with no ordering between them
⚠ codex gpt-6-astra · xhigh→medium  the model matched, the reasoning effort did not
? codex gpt-6-astra unverified      nothing named a model, so nothing is confirmed

The first time a session sees a substitution it also raises a notification, once per requested→served pair.

Nothing here judges answer quality, and nothing leaves your machine.

Install

pi install npm:@mzwing/pi-codex-downgrade-detector

Configuration

Optional. Both scopes are merged, project over global; tiers merges key by key.

Scope	Path
Global	~/.pi/agent/extensions/pi-codex-downgrade-detector/config.json
Project	.pi/extensions/pi-codex-downgrade-detector/config.json

Field	Default	Description
providers	["openai-codex"]	Provider ids to watch. [] watches every provider.
tiers	{}	slug to integer rank, higher meaning more capable. Extends the built-in table.
checkEffort	true	Also compare the reasoning effort Pi sent against the level you selected.
notify	true	Raise a notification the first time a pair is substituted.

Rank a slug under tiers whenever the footer reports a direction it could not work out — a brand-new model, or a relay's own house models.

/codex-downgrade

/codex-downgrade prints every turn this session observed, and which signal named the served model. /codex-downgrade show prints the resolved config and where it was read from.

Limits

The served model is only as good as what the server states. If the provider does not expose the openai-model header and does not echo a model back, the footer says unverified rather than reporting a clean turn — silence is reported as silence, never as a pass.

The effort comparison is client-side. It reports what Pi put on the wire against what your selected thinking level maps to for that model; nothing on the response side confirms the effort the server actually used.

Credits

https://t.me/c/2502727045/42320

Super thx for opening the detecting verdict out!

License

MIT