@billjr99/pi-openai-compat 1.1.31 → 1.1.33

package/README.md

@@ -68,6 +68,10 @@ If pi is already running when you install, type `/reload` first.
 | **Token Harbor** | `https://tokenharbor.ai/v1` | `thk_live_...` Universal Key from tokenharbor.ai/dashboard/api-keys |
 | **Atria ASI** | `https://api.atria-asi.ai/v1` | `atr_...` key from api.atria-asi.ai/console/keys (Google sign-in) |
 | **Unbiased AI** | `https://api.unbiased.ai/v1` | `sk_...` key from platform.unbiased.ai (signup reviewed by hand) |
+| **Kilo AI** | `https://api.kilo.ai/api/gateway` | API key from app.kilo.ai → Your Profile (personal account), at the bottom of the page; needs account credits |
+| **ModelScope** | `https://api-inference.modelscope.cn/v1` | `ms-...` SDK token from modelscope.cn/my/myaccesstoken |
+| **Aion Labs** | `https://api.aionlabs.ai/v1` | API key from aionlabs.ai/app/api-keys (aionlabs.ai/docs) |
+| **Agnes AI** | `https://apihub.agnes-ai.com/v1` | `sk-...` key from the Agnes console at platform.agnes-ai.com (wiki.agnes-ai.com) |
 | **Ollama (local)** | `http://localhost:11434/v1` (editable) | Optional bearer token; leave blank for a default local install |
 | **Ollama Cloud** | `https://ollama.com/v1` | Ollama Cloud API key from ollama.com |
 | **llmproxy** | `http://localhost:8080/v1` (editable) | Keyless by default; bearer token if your instance requires one |
@@ -84,7 +88,12 @@ If pi is already running when you install, type `/reload` first.
 > 2. **Built-in fallback list** when no working catalog endpoint exists. Used
 >    when discovery fails or when the upstream simply has no `/models` at all
 >    (e.g. **Hugging Face** returns HTML, **Cloudflare AI Gateway** has no
->    catalog endpoint).
+>    catalog endpoint, **Unbiased AI** has no such route).
+>
+> A fallback list stands in for a missing endpoint, never for a rejected
+> credential: a catalog fetch that fails with 401 or 403 aborts the login
+> instead, so a mistyped key cannot be saved as a provider that looks healthy
+> in `/model` and then fails on every completion.
 >
 > | Provider | Default `/models` symptom | Handling |
 > |---|---|---|
@@ -92,6 +101,7 @@ If pi is already running when you install, type `/reload` first.
 > | **Cloudflare Workers AI** | HTTP 405 (no `GET /v1/models`; real catalog at `/ai/models/search`, id field is `name`, mixed task types) | Live discovery — `modelsUrl: …/ai/models/search`, `modelsIdField: name`, `modelsKeepTask: "Text Generation"` |
 > | **Cloudflare AI Gateway** | HTTP 401 (token missing `AI Gateway: Run`) or HTTP 400 *"Please configure AI Gateway"* (gateway slug doesn't exist, or the upstream isn't configured on it) | Built-in fallback list |
 > | **Hugging Face** | Returns HTML rather than JSON | Built-in fallback list |
+> | **Unbiased AI** | HTTP 404 `unknown_url` with a valid key (no `/v1/models` at all; an *invalid* key returns 401 on every path, including ones that do not exist, so a 401 here proves nothing) | Built-in fallback list |
 >
 > **Auto-heal for older configs:** if you logged in before live discovery
 > existed, your saved provider is missing these `modelsUrl`/`modelsIdField`/

package/add-provider.sh

@@ -72,6 +72,30 @@ const BUILTIN = {
     keyless: false,
     keyHint: "platform.unbiased.ai (signup is reviewed by hand; keys look like sk_...)",
   },
+  kilo: {
+    displayName: "Kilo AI",
+    baseUrl: "https://api.kilo.ai/api/gateway",
+    keyless: false,
+    keyHint: "app.kilo.ai → Your Profile on your personal account (not an organization), at the bottom of the page; needs account credits (docs at kilo.ai/docs/gateway)",
+  },
+  modelscope: {
+    displayName: "ModelScope",
+    baseUrl: "https://api-inference.modelscope.cn/v1",
+    keyless: false,
+    keyHint: "modelscope.cn/my/myaccesstoken (SDK token, format ms-...; docs at modelscope.cn/docs/model-service/API-Inference/intro)",
+  },
+  aion_labs: {
+    displayName: "Aion Labs",
+    baseUrl: "https://api.aionlabs.ai/v1",
+    keyless: false,
+    keyHint: "aionlabs.ai/app/api-keys (docs at aionlabs.ai/docs/api-reference)",
+  },
+  agnes_ai: {
+    displayName: "Agnes AI",
+    baseUrl: "https://apihub.agnes-ai.com/v1",
+    keyless: false,
+    keyHint: "platform.agnes-ai.com → API Key management (keys look like sk-...; docs at wiki.agnes-ai.com)",
+  },
 };
 
 // README "Auth" column text for the built-in providers, used only by patchreadme.
@@ -82,6 +106,10 @@ const BUILTIN_AUTH = {
   tokenharbor: "`thk_live_...` Universal Key from tokenharbor.ai/dashboard/api-keys",
   atria_asi: "`atr_...` key from api.atria-asi.ai/console/keys (Google sign-in)",
   unbiased_ai: "`sk_...` key from platform.unbiased.ai (signup reviewed by hand)",
+  kilo: "API key from app.kilo.ai → Your Profile (personal account), at the bottom of the page; needs account credits",
+  modelscope: "`ms-...` SDK token from modelscope.cn/my/myaccesstoken",
+  aion_labs: "API key from aionlabs.ai/app/api-keys (aionlabs.ai/docs)",
+  agnes_ai: "`sk-...` key from the Agnes console at platform.agnes-ai.com (wiki.agnes-ai.com)",
 };
 
 /**
@@ -158,8 +186,9 @@ function normalizeModels(body, idField, keepTask) {
   else if (json && typeof json === "object") {
     if (Array.isArray(json.data)) raw = json.data;
     else if (Array.isArray(json.result)) raw = json.result;
+    else if (Array.isArray(json.models)) raw = json.models;
   }
-  if (!raw) throw new Error('expected an array or an object with a "data" or "result" array');
+  if (!raw) throw new Error('expected an array or an object with a "data", "result" or "models" array');
   const field = idField || "id";
   return raw
     .filter((m) => {
@@ -436,6 +465,17 @@ fi
 MODEL_COUNT=0
 [ -n "$MODELS_JSON" ] && MODEL_COUNT="$(printf '%s' "$MODELS_JSON" | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>process.stdout.write(String(JSON.parse(s).length)))')"
 
+# Mirrors isAuthFailure() in index.ts: a fallback list stands in for an endpoint
+# that is not there, never for a credential that was refused. Falling back on a
+# 401/403 would write a provider that looks healthy in /model and fails on every
+# completion, which is worse than saying the key was rejected.
+if [ "$MODEL_COUNT" = "0" ] && { [ "$HTTP_CODE" = "401" ] || [ "$HTTP_CODE" = "403" ]; }; then
+  echo "  The provider rejected the API key (HTTP $HTTP_CODE)."
+  echo "  Not using the built-in model list: that would save a provider whose"
+  echo "  every request fails. Re-run with a valid key."
+  exit 1
+fi
+
 if [ "$MODEL_COUNT" = "0" ]; then
   if [ -n "$FALLBACK_MODELS" ]; then
     echo "  Using this template's built-in model list instead."

package/index.ts

@@ -413,11 +413,69 @@ export const TEMPLATES: Record<string, {
     // answers both /v1/chat/completions and Anthropic's /v1/messages and takes
     // either Authorization: Bearer or x-api-key; pi is registered with the
     // OpenAI pair, which is what registerProvider and fetchModels already send.
-    // GET /v1/models is registered and key-gated (a well-formed sk_ key gets
-    // "Invalid API key" rather than a format complaint), so no modelsUrl or
-    // fallbackModels override is needed.
+    //
+    // GET /v1/models does NOT exist: with a valid key it returns 404
+    // "unknown_url". An earlier note here claimed the endpoint was registered
+    // and key-gated, reasoning from a well-formed but invalid key getting
+    // "Invalid API key" rather than a 404. That inference was wrong. The
+    // gateway authenticates before it routes, so a path that certainly does
+    // not exist returns the same 401: GET /v1/definitely-not-a-real-endpoint
+    // answers "Invalid API key" with a bad key and "Missing API key" without
+    // one, exactly as /v1/models does. A 401 here says nothing about routing.
+    //
+    // Discovery therefore finds nothing, which is what fallbackModels is for.
+    // It is safe here because the catch that consumes it now refuses to fall
+    // back on a 401/403 (see isAuthFailure), so a mistyped key still fails the
+    // login instead of being saved as a provider that fails every completion.
+    fallbackModels: ["pareto"],
     keyHint: "platform.unbiased.ai (signup is reviewed by hand; keys look like sk_...)",
   },
+  kilo: {
+    displayName: "Kilo AI",
+    baseUrl: "https://api.kilo.ai/api/gateway",
+    keyless: false,
+    // Multi-vendor gateway; the base path is /api/gateway, not /v1. GET
+    // /api/gateway/models answers 200 without a key and returns an
+    // OpenRouter-shaped {"data": [...]} catalog (~380 models, ids like
+    // anthropic/claude-sonnet-5), so no modelsUrl or fallbackModels override
+    // is needed. Inference itself needs a key and account credits.
+    keyHint: "app.kilo.ai → Your Profile on your personal account (not an organization), at the bottom of the page; needs account credits (docs at kilo.ai/docs/gateway)",
+  },
+  modelscope: {
+    displayName: "ModelScope",
+    baseUrl: "https://api-inference.modelscope.cn/v1",
+    keyless: false,
+    // Alibaba's API-Inference service. GET /v1/models answers 200 without a
+    // key in the plain OpenAI {"data": [...]} shape, so the default discovery
+    // path works as is. Free tier is an account-wide 2,000 requests/day.
+    keyHint: "modelscope.cn/my/myaccesstoken (SDK token, format ms-...; docs at modelscope.cn/docs/model-service/API-Inference/intro)",
+  },
+  aion_labs: {
+    displayName: "Aion Labs",
+    baseUrl: "https://api.aionlabs.ai/v1",
+    keyless: false,
+    // Aion's own API reference documents GET /v1/models as needing no
+    // authentication and returning its array under a top-level "models" key
+    // rather than "data". fetchModels normalizes that shape, so no modelsUrl
+    // or fallbackModels override is needed. Four models as of
+    // September 2026 (aion-2.0, aion-3.0, aion-3.0-mini, aion-rp-llama-3.1-8b).
+    keyHint: "aionlabs.ai/app/api-keys (docs at aionlabs.ai/docs/api-reference)",
+  },
+  agnes_ai: {
+    displayName: "Agnes AI",
+    baseUrl: "https://apihub.agnes-ai.com/v1",
+    keyless: false,
+    // Multimodal gateway. GET /v1/models is key-gated (it answers 401 "Token
+    // not provided", where an unknown /v1 path answers 404), which is exactly
+    // what fetchModels already sends, so no override is needed. The catalog
+    // also carries image and video model ids (agnes-image-*, agnes-video-*);
+    // they are registered too and will show up in /model, but only the text
+    // models serve chat/completions: the free flash tier (agnes-3.0-flash,
+    // agnes-2.5-flash, agnes-2.0-flash) and the paid agnes-2.5-pro pair. No
+    // modelFilter is set because the live id list could not be confirmed
+    // without a key, and a filter that misses would register nothing.
+    keyHint: "platform.agnes-ai.com → API Key management (keys look like sk-...; docs at wiki.agnes-ai.com)",
+  },
   custom: {
     displayName: "Custom Endpoint",
     baseUrl: "",
@@ -610,6 +668,24 @@ export function isLocalUrl(url: string): boolean {
 /** Cap on how much of an upstream error body is surfaced to the user. */
 export const MAX_ERROR_BODY = 500;
 
+/** An Error from fetchModels carrying the upstream HTTP status, when it had one. */
+export type CatalogError = Error & { status?: number };
+
+/**
+ * True when a catalog fetch failed because the credential was rejected.
+ *
+ * This is the distinction that makes fallbackModels safe to use. A fallback
+ * list exists for endpoints that publish no catalog, so it should stand in for
+ * a missing endpoint and never for a bad key: falling back on a 401 would save
+ * a provider that looks healthy in /model and fails on every completion, which
+ * is worse than refusing the login. A network error (no status) is treated as
+ * non-auth, so an offline or unreachable host still gets the fallback.
+ */
+export function isAuthFailure(err: unknown): boolean {
+  const status = (err as CatalogError | null)?.status;
+  return status === 401 || status === 403;
+}
+
 /** Optional per-provider overrides controlling how /models is fetched. */
 interface FetchOverrides {
   /** Full URL to fetch instead of `<baseUrl>/models`. */
@@ -646,28 +722,35 @@ export async function fetchModels(
     const snippet = body.length > MAX_ERROR_BODY
       ? `${body.slice(0, MAX_ERROR_BODY)}… (truncated)`
       : body;
-    throw new Error(`HTTP ${resp.status} from ${url}: ${snippet}`);
+    // Carry the status on the error. Callers need to tell "this endpoint is
+    // not there" from "your key is wrong", and parsing it back out of the
+    // message would break the moment the wording changes.
+    const err = new Error(`HTTP ${resp.status} from ${url}: ${snippet}`) as CatalogError;
+    err.status = resp.status;
+    throw err;
   }
 
   // Normalize the various shapes /models can return:
   //   - OpenAI style:                  {"data": [...]}
   //   - Cloudflare / some gateways:    {"result": [...]}
+  //   - Aion Labs:                     {"models": [...]}
   //   - Together (and a few others):   [...]   (bare JSON array)
   const json = (await resp.json()) as unknown;
   let raw: RawModel[] | undefined;
   if (Array.isArray(json)) {
     raw = json as RawModel[];
   } else if (json && typeof json === "object") {
-    const obj = json as { data?: unknown; result?: unknown };
+    const obj = json as { data?: unknown; result?: unknown; models?: unknown };
     if (Array.isArray(obj.data)) raw = obj.data as RawModel[];
     else if (Array.isArray(obj.result)) raw = obj.result as RawModel[];
+    else if (Array.isArray(obj.models)) raw = obj.models as RawModel[];
   }
   if (!raw) {
     // `url` may be an override (e.g. /catalog/models, /ai/models/search), so
     // keep the wording generic rather than referring specifically to /models.
     throw new Error(
       `Unexpected model catalog payload shape from ${url} ` +
-      `(expected an array or an object with a "data" or "result" array).`
+      `(expected an array or an object with a "data", "result" or "models" array).`
     );
   }
 
@@ -984,7 +1067,7 @@ export default async function (pi: ExtensionAPI) {
           keepTask: tpl.modelsKeepTask,
         });
       } catch (err) {
-        if (tpl.fallbackModels && tpl.fallbackModels.length > 0) {
+        if (tpl.fallbackModels && tpl.fallbackModels.length > 0 && !isAuthFailure(err)) {
           ctx.ui.notify(
             `Could not fetch model list from ${tpl.displayName} (${err}).\nUsing built-in model list instead.`,
             "warning"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@billjr99/pi-openai-compat",
-  "version": "1.1.31",
+  "version": "1.1.33",
   "description": "pi-coding-agent extension: OpenAI-compatible endpoint support (OpenRouter, NVIDIA NIM, Nous Portal, Ollama, custom)",
   "author": "Bill Mongan <https://github.com/BillJr99>",
   "license": "MIT",