fix(zai): declare glm-5.3-flash image input on the Chat rows

src/providers/registry.ts

@@ -444,6 +444,30 @@ const ZAI_GLM_5X_MODELS = [...ZAI_GLM_53_MODELS, ...ZAI_GLM_52_MODELS];
  * `preserveReasoningContentModels`, where flash DOES belong.
  */
 const ZAI_GLM_5X_SIDECAR_VISION_MODELS = ZAI_GLM_5X_MODELS.filter(id => id !== "glm-5.3-flash");
+/**
+ * Positive input-modality declaration for the Chat-path GLM rows.
+ *
+ * `noVisionModels` already keeps Flash out of the vision sidecar, but that is a NEGATIVE
+ * statement: it stops a detour without telling the catalog what the model can read. With
+ * no `modelInputModalities` entry, `configuredInputModalities` returns undefined and the
+ * catalog falls through to the `["text"]` floor, so every client export (ZCode, Pi, OMP)
+ * listed a native VLM as text-only and its picker refused to attach an image.
+ *
+ * The Responses sibling row below already declares this positively, so the same model was
+ * described two different ways in one registry.
+ *
+ * Authoritative source: `GET https://api.z.ai/api/v1/models` returns `input_modalities:
+ * ["text"]` for glm-5.3 and `["text", "image"]` for glm-5.3-flash (captured in
+ * devlog/_plan/260912_zcode_protocol_and_catalog/evidence/zai-responses-models.json).
+ * docs.z.ai/devpack/latest-model says the same in prose: "GLM-5.3 is a text-only model...
+ * GLM-5.3-FLASH is a multimodal model". Upstream also lists video and file for Flash;
+ * neither the internal vocabulary nor the export vocabulary can express them, so `image`
+ * is where this stops.
+ */
+const ZAI_GLM_5X_INPUT_MODALITIES: Record<string, string[]> = {
+  ...Object.fromEntries(ZAI_GLM_5X_SIDECAR_VISION_MODELS.map(id => [id, ["text"]])),
+  "glm-5.3-flash": ["text", "image"],
+};
 const ZAI_GLM_52_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
 /**
  * GLM-5.3 does NOT share 5.2's five-tier ladder. docs.z.ai/devpack/latest-model folds every
@@ -2604,6 +2628,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     // Z.AI's OpenAI path returns 400 code 1211 for bracketed model ids.
     modelSuffixBracketStrip: true,
     noVisionModels: ZAI_GLM_5X_SIDECAR_VISION_MODELS,
+    modelInputModalities: ZAI_GLM_5X_INPUT_MODALITIES,
     modelReasoningEfforts: ZAI_GLM_5X_REASONING_EFFORTS,
     modelDefaultReasoningEfforts: Object.fromEntries(ZAI_GLM_53_MODELS.map(id => [id, "max"])),
     modelMaxOutputTokens: Object.fromEntries(ZAI_GLM_53_MODELS.map(id => [id, 131_072])),
@@ -2685,6 +2710,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelContextWindows: { "glm-5.3": 1_000_000, "glm-5.3[1m]": 1_000_000, "glm-5.3-flash": 1_000_000, "glm-5.2": 1_000_000, "glm-5.2[1m]": 1_000_000 },
     modelSuffixBracketStrip: true,
     noVisionModels: ZAI_GLM_5X_SIDECAR_VISION_MODELS,
+    modelInputModalities: ZAI_GLM_5X_INPUT_MODALITIES,
     modelReasoningEfforts: ZAI_GLM_5X_REASONING_EFFORTS,
     modelSupportsReasoningSummaries: Object.fromEntries(ZAI_GLM_5X_MODELS.map(id => [id, true])),
     preserveReasoningContentModels: ZAI_GLM_5X_MODELS,

tests/providers/provider-registry-parity.test.ts

@@ -482,6 +482,21 @@ describe("provider registry parity", () => {
     // The sibling it is most often confused with stays text-only, so the assertion above
     // cannot pass by making every GLM row a VLM.
     expect(zai?.noVisionModels ?? []).toContain("glm-5.3");
+    // The global loop above accepts an ABSENT declaration, which is exactly how the Chat
+    // rows shipped: Flash was kept out of the sidecar but never told the catalog it could
+    // read an image, so `configuredInputModalities` returned undefined and every client
+    // export listed a native VLM as text-only. Pin the positive declaration so the two
+    // Chat rows cannot drift back to describing Flash only by what it is not.
+    for (const id of ["zai", "zhipu-bigmodel-coding"] as const) {
+      const row = PROVIDER_REGISTRY.find(entry => entry.id === id);
+      expect(row?.modelInputModalities?.["glm-5.3-flash"]).toEqual(["text", "image"]);
+      expect(row?.modelInputModalities?.["glm-5.3"]).toEqual(["text"]);
+      // The bracketed aliases are looked up by their exact catalog id, not a stripped one,
+      // so they need their own text-only entries.
+      expect(row?.modelInputModalities?.["glm-5.3[1m]"]).toEqual(["text"]);
+      expect(row?.modelInputModalities?.["glm-5.2[1m]"]).toEqual(["text"]);
+      expect(row?.noVisionModels ?? []).not.toContain("glm-5.3-flash");
+    }
     // `glm-5.3-flash` belongs in all three maps. It was seeded into the model list
     // and the context map alone, so it advertised a 1M window with no effort ladder,
     // no default effort and no output cap - and this assertion pinned that gap in