@billjr99/pi-openai-compat 1.1.15 → 1.1.16

package/README.md

@@ -56,8 +56,27 @@ If pi is already running when you install, type `/reload` first.
 | **OpenCode Zen** | `https://opencode.ai/zen/v1` | API key from opencode.ai |
 | **Ollama (local)** | `http://localhost:11434/v1` | Keyless |
 | **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 |
 | **Custom** | Any URL you supply | Optional bearer token |
 
+> **llmproxy and the `__` model-id rewrite.**
+> [llmproxy](https://github.com/BillJr99/llmproxy) advertises model ids in the
+> `provider__model` form (e.g. `openrouter__gpt-4`) and exposes virtual models
+> such as `llmproxy__free` and `llmproxy__loadbalanced`. pi rejects model ids that
+> contain `__`, so without special handling **every llmproxy model is silently
+> dropped from `/model`**. The llmproxy template therefore sets
+> `rewriteDoubleUnderscore: true`: the extension rewrites the first `__` of each id
+> to `/` (e.g. `openrouter__gpt-4` → `openrouter/gpt-4`, `llmproxy__free` →
+> `llmproxy/free`) before registering with pi. llmproxy canonicalizes that slash
+> form back to `__` on each request, so routing still works.
+>
+> The flag is a per-provider config field (`rewriteDoubleUnderscore`, default
+> **false**) and is enabled automatically only for the llmproxy template — other
+> providers are unaffected. If you added llmproxy as a **Custom** endpoint instead
+> of via the llmproxy template, set `"rewriteDoubleUnderscore": true` on that
+> provider in `~/.config/pi-openai-compat/config.json` (or re-run `/compat-login`
+> and pick **llmproxy (local)**).
+
 > **Providers whose model catalog lives at a non-standard `/models` path (as of June 2026)**
 > Some providers don't return models at `<base_url>/models`. The extension
 > handles them in one of two ways:
@@ -327,6 +346,15 @@ For Ollama: pull at least one model first (`ollama pull llama3`).
 For OpenRouter: some keys are restricted to free-tier models only.
 For NIM: confirm your account has inference access enabled.
 
+**`/compat-login` reports N models but far fewer appear in `/model`**
+This is the classic llmproxy symptom: pi drops every model id containing `__`, so
+llmproxy's `provider__model` ids (and `llmproxy__free` / `llmproxy__loadbalanced`)
+never show. Make sure you logged in via the **llmproxy (local)** template (it sets
+`rewriteDoubleUnderscore: true` automatically). If you used **Custom**, add
+`"rewriteDoubleUnderscore": true` to that provider in
+`~/.config/pi-openai-compat/config.json` and re-run `/compat-refresh`. After the
+fix the ids appear in slash form (`openrouter/gpt-4`, `llmproxy/free`).
+
 **Models appear in `/model` but requests fail**
 Check `/compat-login` ran successfully (no error message).
 Verify Ollama is still running if using a local endpoint.

package/index.ts

@@ -48,6 +48,13 @@ interface ProviderConfig {
   // matches its template and can't be auto-migrated, so the notice is shown
   // only once rather than on every session_start while the config stays stale.
   staleNotified?: boolean;
+  // When true, rewrite the FIRST "__" of every model id to "/" before
+  // registering with pi. pi rejects model ids containing "__", so providers like
+  // llmproxy (which uses "provider__model" ids) would otherwise have their entire
+  // catalog dropped from /model. The rewrite is the inverse of llmproxy's own
+  // request-side canonicalization, so requests still round-trip. Defaults false;
+  // the wizard turns it on only for the llmproxy template.
+  rewriteDoubleUnderscore?: boolean;
 }
 
 interface ExtensionConfig {
@@ -97,6 +104,12 @@ const TEMPLATES: Record<string, {
   modelsIdField?: string;
   /** Keep only models whose task.name matches this string (case-insensitive). */
   modelsKeepTask?: string;
+  /**
+   * Rewrite the first "__" of each model id to "/" before registering with pi.
+   * pi drops model ids containing "__", so this is required for llmproxy (whose
+   * ids are "provider__model"). Default false; set true only where pi needs it.
+   */
+  rewriteDoubleUnderscore?: boolean;
 }> = {
   openrouter: {
     displayName: "OpenRouter",
@@ -288,6 +301,10 @@ const TEMPLATES: Record<string, {
     baseUrl: "http://localhost:8080/v1",
     keyless: true,
     promptUrl: true,
+    // llmproxy advertises "provider__model" ids (and virtuals like
+    // "llmproxy__free"). pi rejects "__" in model ids, so rewrite the first
+    // "__" to "/" here; llmproxy canonicalizes the slash form back on requests.
+    rewriteDoubleUnderscore: true,
   },
   vercel: {
     displayName: "Vercel AI Gateway",
@@ -535,16 +552,31 @@ async function fetchModels(
 // Provider registration helpers
 // ─────────────────────────────────────────────────────────────────────────────
 
-function buildProviderModels(models: CachedModel[]) {
-  return models.map((m) => ({
-    id: m.id,
-    name: m.id,
-    reasoning: false,
-    input: ["text"] as string[],
-    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-    contextWindow: m.contextWindow ?? 128_000,
-    maxTokens: m.maxTokens ?? 4_096,
-  }));
+/**
+ * Rewrite the FIRST "__" of a model id to "/". Only the first occurrence is
+ * touched: the provider/virtual segment that precedes it never contains "__" or
+ * "/", so this is the exact inverse of the slash→"__" canonicalization llmproxy
+ * applies on the request side, keeping ids round-trippable. Ids without "__" are
+ * returned unchanged.
+ */
+function rewriteFirstDoubleUnderscore(id: string): string {
+  const i = id.indexOf("__");
+  return i === -1 ? id : `${id.slice(0, i)}/${id.slice(i + 2)}`;
+}
+
+function buildProviderModels(models: CachedModel[], rewriteDoubleUnderscore = false) {
+  return models.map((m) => {
+    const id = rewriteDoubleUnderscore ? rewriteFirstDoubleUnderscore(m.id) : m.id;
+    return {
+      id,
+      name: id,
+      reasoning: false,
+      input: ["text"] as string[],
+      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+      contextWindow: m.contextWindow ?? 128_000,
+      maxTokens: m.maxTokens ?? 4_096,
+    };
+  });
 }
 
 function compatKey(key: string): string {
@@ -552,12 +584,16 @@ function compatKey(key: string): string {
 }
 
 function registerProvider(pi: ExtensionAPI, key: string, p: ProviderConfig): void {
+  // Honor the persisted per-provider flag; fall back to the template default for
+  // this key so providers logged in before the flag existed (e.g. an existing
+  // llmproxy provider) still get the rewrite without a re-login.
+  const rewrite = p.rewriteDoubleUnderscore ?? TEMPLATES[key]?.rewriteDoubleUnderscore ?? false;
   pi.registerProvider(compatKey(key), {
     name: `compat/${key.replace(/_/g, "-")}`,
     baseUrl: p.baseUrl,
     apiKey: p.apiKey ?? (isLocalUrl(p.baseUrl) ? "local" : ""),
     api: "openai-completions" as const,
-    models: buildProviderModels(p.cachedModels),
+    models: buildProviderModels(p.cachedModels, rewrite),
   });
 }
 
@@ -800,6 +836,7 @@ export default async function (pi: ExtensionAPI) {
         modelsUrl,
         modelsIdField: tpl.modelsIdField,
         modelsKeepTask: tpl.modelsKeepTask,
+        rewriteDoubleUnderscore: tpl.rewriteDoubleUnderscore,
       };
       saveConfig(config);
       registerProvider(pi, key, config.providers[key]);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@billjr99/pi-openai-compat",
-  "version": "1.1.15",
+  "version": "1.1.16",
   "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",