integrations: fix four things that make openzoo fail out of the box

[staccDOTsol]

Audited integrations/ against the live gateway and the shipped npx openzoo CLI. Four things are broken; all four are verified with commands, not inferred.

Ran the whole thing end to end first: open-webui + this pipe + npx openzoo → 982 zoo models in the dropdown, paid completion succeeds, receipt renders. Screenshot of the receipt line at the bottom.

---

1. The OpenWebUI receipt is dead code — and it's the plugin's headline feature

_format_receipt read billedUsd / savesVsDirect out of usage. They're in a top-level x402 block. Live response:

usage keys: completion_tokens, cost, cost_details, is_byok, prompt_tokens, total_tokens
  billedUsd in usage?     False
  savesVsDirect in usage? False
x402: {"billedUsd":0.00700472, "savesVsDirect":0.333…, "cogsUsd":…, "directUsd":…,
       "subscription":{"tier":"pro","wouldHaveBilled":0.02101416,"invoiced":"stripe"}}

So "users see the per-call cost and the savings from the leCore spill" never fired — it could only print N tokens read.

Second, subtler bug in the same function. savesVsDirect is the ratio paid, not the fraction saved:

0.00700472 / 0.02101416 = 0.3333  ==  savesVsDirect

f"saved {savesVsDirect} vs direct" renders "saved 0.33" when the real saving is two thirds — it halves your own pitch. I deliberately did not convert it to a percentage (the semantics are easy to flip again); it now shows both absolute figures, billed $X vs $Y direct (3.0x), which can't be misread. Added regression tests using the exact live shapes.

2. npx openzoo models doesn't exist

$ npx openzoo models
unknown command: models

Documented 9 times, and it's the first command a new user runs. Replaced with the proxy's own free /v1/models.

3. https://api.openzoo.fun/v1 404s

It resolves to the Vercel website (216.150.1.193), not the gateway, and is presented as "the hosted endpoint" in 5 READMEs. Real upstream is https://x402-tokens.fly.dev/v1 → 200 — which is also what the CLI's own OPENZOO_API_BASE default points at.

Credit where due: the pipe's own comment already hedged "once it exists." The READMEs then stated it flat.

4. Stale / understated numbers

• ~435 models → 480+ (live catalog)
• zoo_ask ~1M tokens → ~9.8M per call / ~128M bound ceiling — the old figure undersold the flagship capability by ~10×

---

Not changed

The localhost:8402 default stays as you had it. Added a note that a Docker install of open-webui needs host.docker.internal:8402 instead, since inside a container localhost is the container and the proxy is on the host. That bit me during setup.

What's good and I left alone

_strip_owui_prefix is genuinely careful — zoo ids contain dots, so naive split-on-first-dot turns nvidia/nemotron-3.5-lightning into 5-lightning, and there's already a regression test recording exactly that. The _OPENAI_CHAT_FIELDS whitelist prevents a real class of 400s. The "kept negative" note on client-side spill is the right call. All four configs parse clean and every active line already pointed at a working local default.

Verification

integrations/{LibreChat,Continue,Hermes}/*.yaml   parse OK
integrations/GrokCLI/models.openzoo.json          parse OK
python3 integrations/OpenWebUI/openzoo_pipe.py    selftest OK

Live, through open-webui:

openzoo pipe works

---
*openzoo: billed $0.021348 · credit · 3364 tokens read*

🤖 Generated with Claude Code