Community Plugin: OAuth bridge for Claude Max, ChatGPT Plus, and Google Gemini subscriptions #38246

Following up on my own question, since I found the answer and it is worth writing down.

The hook point is litellm_settings.callbacks. A CustomLogger instance loaded there gets
the proxy's FastAPI app during startup, which is enough to mount a sub-app and to reach the
Router. No sitecustomize.py, no PYTHONPATH games, no file in site-packages — the
whole installation is one line in config.yaml.

Rewrote the thing on that basis: litellm-mysubs.

What changed beyond the hook point, since the old version's problems were real:

• The credential API is gone. That endpoint served refresh tokens unauthenticated; it
  does not exist any more. The page requires proxy_admin, and the one route that has to
  be reachable without it — the pairing deposit — takes a 60-bit single-use code with a
  10-minute TTL and a per-origin rate limit.
• TLS verification is back on by default.
• Multi-worker actually works. The old one had no locking; with --num_workers 4 the
  workers raced on the rotating refresh token and eventually hit invalid_grant. There is
  now a flock and a single designated refresh owner, verified with four real worker
  processes and a token forced to expire.
• No fabricated numbers. The demo used to invent quota percentages to make the bars
  look good. Everything shown is now measured or labelled as unknown.

Three things I measured along the way that might save someone else the time:

• The official CustomLLM path loses the provider's usage — it gets replaced by a
  token_counter estimate and cached tokens disappear entirely. That is why the streaming
  generators are still a monkey-patch, and there is a script in tools/ that rechecks it
  against the installed LiteLLM so the decision can be reopened when it stops being true.
• Injecting into the Router with set_model_list() and persisting separately, rather than
  calling POST /model/new: the latter returns 200 and then the model is not in
  /v1/models and calls fail with no healthy deployments.
• The wire prefix has to follow the model family, not the provider, or LiteLLM finds no
  price table and every Antigravity model logs at zero cost.

Ten decisions like those, each with the measurement behind it, are in docs/DECISIONS.md.

On provenance, since the old version was vague about it: the wire-level work is a port of
oh-my-pi (@oh-my-pi/pi-ai and siblings, MIT), not
reverse engineering and not my design. It is about half the source — everything under
wire/, plus the provider auth rules — and OMP stays the source of truth: when a provider
changes something, the fix lands there first and this follows. Every borrowed constant carries an # omp: anchor naming the source file and
symbol — 168 of them — and CI checks each one against the published tarballs on every
push, values included. When upstream renames a symbol or changes an endpoint path, the
build goes red with the file and the new value in the message, instead of the plugin
failing quietly against a provider months later.

Same caveat as before, stated more plainly this time: this is a personal tool for your own
subscription on your own machine, and the providers' terms treat subscription OAuth
differently from API keys. That is at the top of the README now, where it belongs.