@billjr99/pi-openai-compat 1.1.21 → 1.1.23

package/README.md

@@ -129,7 +129,10 @@ dropped) since you logged in, without restarting your session.
 If you have multiple providers registered, you are asked which one to refresh
 (or choose **All providers**); with a single provider it refreshes directly.
 A failed or empty refresh leaves the existing model list untouched, so a flaky
-network call can't blank out a working provider.
+network call can't blank out a working provider.  Capability metadata you have
+hand-edited into `cachedModels` (see [Config file](#config-file)) survives a
+refresh: the fetched list decides which model IDs exist, and any field the
+provider's catalog omits is carried forward from the previous cache.
 
 ### `/compat-logout`
 
@@ -199,6 +202,38 @@ Credentials and cached model lists are stored at:
 API keys are stored in plaintext.  Protect the file with `chmod 600` if
 needed, or delete it to clear all saved credentials.
 
+Each provider's `cachedModels` array holds one entry per model.  Only `id` is
+required; the rest are optional and fall back to conservative defaults when the
+provider's `/models` catalog does not report them:
+
+| Field | Type | Default | Notes |
+|---|---|---|---|
+| `id` | string | required | Model ID as sent in requests. |
+| `contextWindow` | number | `128000` | Context window in tokens. |
+| `maxTokens` | number | `4096` | Maximum output tokens per response. |
+| `reasoning` | boolean | `false` | Enables pi's thinking mode for the model. |
+| `input` | `["text"]` or `["text","image"]` | `["text"]` | Modalities pi may send; `image` lets pi attach image blocks. Any other value is ignored. |
+
+Most catalogs report none of these beyond the ID, so aggregators and proxies
+(CLIProxyAPI, for example) register every model as a 128K, text-only,
+non-reasoning model.  To correct that, edit the entry by hand and `/reload`:
+
+```json
+{
+  "providers": {
+    "cpa": {
+      "cachedModels": [
+        { "id": "claude-sonnet-4-6", "contextWindow": 1000000, "maxTokens": 128000,
+          "reasoning": true, "input": ["text", "image"] }
+      ]
+    }
+  }
+}
+```
+
+`/compat-refresh` keeps these hand-edited fields for any model ID that is still
+present in the refreshed catalog; a field the provider does report always wins.
+
 ### Adding a provider without pi — `add-provider.sh`
 
 `/compat-login` is the normal path, but the repo also ships a standalone script

package/index.ts

@@ -25,10 +25,18 @@ import * as os from "node:os";
 // Types
 // ─────────────────────────────────────────────────────────────────────────────
 
+/** Input modalities pi understands; anything else is dropped at fetch time. */
+type ModelInput = "text" | "image";
+
 interface CachedModel {
   id: string;
   contextWindow?: number;
   maxTokens?: number;
+  // Optional capability metadata. Rarely present in a /models payload, but a
+  // user may hand-edit them into cachedModels for providers whose catalog
+  // omits them (see README "Config file"); /compat-refresh preserves them.
+  reasoning?: boolean;
+  input?: ModelInput[];
 }
 
 interface ProviderConfig {
@@ -67,9 +75,25 @@ type RawModel = {
   name?: string;
   context_window?: number;
   max_tokens?: number;
+  reasoning?: unknown;
+  input?: unknown;
   task?: { name?: string };
 };
 
+/**
+ * Coerce an upstream `input` value into pi's modality list. Unknown or
+ * malformed values yield undefined so buildProviderModels falls back to the
+ * default, rather than caching something pi would choke on (it calls
+ * `model.input.includes("image")` at tool time).
+ */
+function normalizeInput(value: unknown): ModelInput[] | undefined {
+  if (!Array.isArray(value)) return undefined;
+  const kept = value.filter(
+    (v): v is ModelInput => v === "text" || v === "image",
+  );
+  return kept.length > 0 ? kept : undefined;
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Provider templates
 // ─────────────────────────────────────────────────────────────────────────────
@@ -551,7 +575,13 @@ async function fetchModels(
         typeof rawId === "string" ? rawId :
         typeof rawId === "number" ? String(rawId) :
         "";
-      return { id, contextWindow: m.context_window, maxTokens: m.max_tokens };
+      return {
+        id,
+        contextWindow: m.context_window,
+        maxTokens: m.max_tokens,
+        reasoning: typeof m.reasoning === "boolean" ? m.reasoning : undefined,
+        input: normalizeInput(m.input),
+      };
     })
     .filter((m) => Boolean(m.id))
     .sort((a, b) => a.id.localeCompare(b.id));
@@ -567,8 +597,8 @@ function buildProviderModels(models: CachedModel[]) {
     return {
       id,
       name: id,
-      reasoning: false,
-      input: ["text"] as string[],
+      reasoning: m.reasoning ?? false,
+      input: m.input ?? (["text"] as ModelInput[]),
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: m.contextWindow ?? 128_000,
       maxTokens: m.maxTokens ?? 4_096,
@@ -576,6 +606,27 @@ function buildProviderModels(models: CachedModel[]) {
   });
 }
 
+/**
+ * Carry hand-edited capability metadata (reasoning / input / contextWindow /
+ * maxTokens) forward from the previous cache when a fresh /models fetch omits
+ * it. The fetched list decides which ids exist; the old cache only fills gaps,
+ * so a provider that does report a field always wins.
+ */
+function mergeModelMetadata(previous: CachedModel[], fetched: CachedModel[]): CachedModel[] {
+  const prior = new Map(previous.map((m) => [m.id, m]));
+  return fetched.map((m) => {
+    const old = prior.get(m.id);
+    if (!old) return m;
+    return {
+      ...m,
+      contextWindow: m.contextWindow ?? old.contextWindow,
+      maxTokens: m.maxTokens ?? old.maxTokens,
+      reasoning: m.reasoning ?? old.reasoning,
+      input: m.input ?? old.input,
+    };
+  });
+}
+
 function compatKey(key: string): string {
   return `compat-${key}`;
 }
@@ -888,7 +939,7 @@ export default async function (pi: ExtensionAPI) {
           if (models.length > 0) {
             // Only overwrite the cache on a successful, non-empty fetch — a
             // flaky refresh must never blank out a working provider's models.
-            p.cachedModels = models;
+            p.cachedModels = mergeModelMetadata(p.cachedModels, models);
             saveConfig(config);
             registerProvider(pi, key, p);
             refreshed.push(`${p.displayName} (${models.length})`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@billjr99/pi-openai-compat",
-  "version": "1.1.21",
+  "version": "1.1.23",
   "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",