Reject model IDs unsafe for headers and clarify supported capabilities

README.md

@@ -3,9 +3,9 @@
 Use [uniGPT](https://unigpt.uni-graz.at), the University of Graz's AI
 assistant, inside the [pi](https://pi.dev) coding agent.
 
-Once set up, every model your uniGPT account offers appears in pi's model
-picker, including GPT-4o, GPT-4.1, and the GPT-5 and GPT-6 families. Tools,
-streaming responses, and image input all work.
+Models exposed by your uniGPT API key appear in pi's model picker.
+Streaming and tool calls are supported; recognized vision models also
+support image input.
 
 ## What you need
 

src/provider.ts

@@ -53,8 +53,11 @@ function isRecord(value: unknown): value is Record<string, unknown> {
 }
 
 function isValidId(id: unknown): id is string {
-  // IDs become an HTTP header value, so reject padding and control characters.
-  return typeof id === "string" && id.length > 0 && id.trim() === id && !/[\u0000-\u001f\u007f]/.test(id);
+  // Fetch header values must fit Latin-1; also reject padding and control characters.
+  return typeof id === "string"
+    && id.length > 0
+    && id.trim() === id
+    && !/[\u0000-\u001f\u007f\u0100-\uffff]/.test(id);
 }
 
 export interface ParsedCatalog {

test/provider.test.ts

@@ -58,6 +58,20 @@ test("skips unusable entries but rejects malformed or entirely unusable catalogs
   }
 });
 
+test("catalog IDs must fit Fetch's Latin-1 header values", () => {
+  const valid = ["gpt-4o", "model-\u00ff"];
+  const invalid = ["model-\u0100", "model-\u{1f600}", "model-\ud800"];
+  const parsed = parseCatalog({ data: [...valid, ...invalid].map((id) => ({ id })) });
+  assert.deepEqual(parsed, { ids: valid, skipped: invalid.length });
+  for (const id of parsed.ids) {
+    const headers = new Headers(makeModel(id, BASE_URL).headers);
+    assert.equal(headers.get("deployment-id"), id);
+  }
+  for (const id of invalid) {
+    assert.throws(() => parseCatalog({ data: [{ id }] }), /without any valid model IDs/);
+  }
+});
+
 const known: Record<string, ModelMetadata> = {
   "model-a": { name: "Model A", input: ["text", "image"], contextWindow: 272_000, maxTokens: 128_000 },
 };