fix(models): mirror model capacity onto the /v1/models top level

src/server/models-capabilities.ts

@@ -10,6 +10,11 @@ import type { CursorEffortTable } from "../integrations/cursor-effort-table";
  * Completions, Responses and Anthropic Messages, streams, and accepts tool calls, so those are
  * constants; context length and vision come from catalog data when known and are omitted
  * otherwise, matching Cursor's optional-field schema.
+ *
+ * Top-level capacity metrics (`context_window`, `context_length`, `max_output_tokens`) are
+ * mirrored directly on each model row for external client discovery (e.g. pi-ai, DSH,
+ * LibreChat) that inspects flat properties rather than Cursor's nested `capabilities.*` shape.
+ * A row that gains a nested capacity value must gain the top-level mirror in the same change.
  */
 
 /**
@@ -132,6 +137,19 @@ export interface ModelCapabilityFields {
     supports_vision?: boolean;
     reasoning_effort?: string[];
   };
+  /**
+   * Mirrored top-level context window for external/legacy client discovery (e.g. pi-ai, DSH)
+   * that reads top-level context_window / context_length instead of nested capabilities.
+   */
+  context_window?: number;
+  /**
+   * Top-level context length alias matching capabilities.context_length for clients expecting context_length.
+   */
+  context_length?: number;
+  /**
+   * Mirrored top-level max output token limit for external/legacy client discovery.
+   */
+  max_output_tokens?: number;
   /**
    * Cursor reads the long-context threshold from `pricing.overrides[].min_prompt_tokens`. That
    * key sits outside its validated capability schema, so it is the one place a threshold can
@@ -153,16 +171,15 @@ export function modelCapabilityFields(input: ModelCapabilityInput): ModelCapabil
   const longContextLength = positiveInt(input.longContextWindow);
   const maxOutputTokens = positiveInt(input.maxOutputTokens);
   const hasLongTier = contextLength !== undefined && longContextLength !== undefined && longContextLength > contextLength;
+  const effectiveContextLength = hasLongTier ? longContextLength : contextLength;
   const modalities = Array.isArray(input.inputModalities)
     ? input.inputModalities.filter(modality => typeof modality === "string" && modality.length > 0)
     : undefined;
   const supportsVision = modalities !== undefined ? modalities.includes("image") : undefined;
   return {
     api_types: [...OPENCODEX_MODEL_API_TYPES],
     capabilities: {
-      ...(hasLongTier
-        ? { context_length: longContextLength }
-        : contextLength !== undefined ? { context_length: contextLength } : {}),
+      ...(effectiveContextLength !== undefined ? { context_length: effectiveContextLength } : {}),
       ...(maxOutputTokens !== undefined ? { max_output_tokens: maxOutputTokens } : {}),
       // Once a gateway advertises api_types, Cursor keeps only rows whose output_modalities
       // include "text"; omitting the key drops the row from the extended catalog.
@@ -174,6 +191,10 @@ export function modelCapabilityFields(input: ModelCapabilityInput): ModelCapabil
       ...(supportsVision !== undefined ? { supports_vision: supportsVision } : {}),
       ...(efforts.length > 0 ? { reasoning_effort: [...efforts] } : {}),
     },
+    ...(effectiveContextLength !== undefined
+      ? { context_window: effectiveContextLength, context_length: effectiveContextLength }
+      : {}),
+    ...(maxOutputTokens !== undefined ? { max_output_tokens: maxOutputTokens } : {}),
     ...(hasLongTier ? { pricing: { overrides: [{ min_prompt_tokens: contextLength }] } } : {}),
   };
 }

structure/catalog.md

@@ -180,6 +180,10 @@ then trusted catalog metadata such as a configured qualified provider/model alia
 This overlay never changes route identity or the upstream wire model, and its catalog fingerprint makes
 a label edit refresh Codex output.
 
+Raw `/v1/models` rows advertise positive safe capacity values in both Cursor's nested
+`capabilities` object and top-level discovery fields used by other clients. A model with a larger
+opt-in context tier uses that effective long window in both shapes; invalid values are omitted.
+
 Supported bare native GPT rows also consume `providers.openai.modelDisplayNames`. Retained sync
 and convergence pass the same map to the observed-state merge. After native normalization and
 ordering, the merge applies the exact nonblank trimmed label and saves

tests/providers/cursor/cursor-local-models-schema.test.ts

@@ -123,6 +123,35 @@ describe("modelCapabilityFields", () => {
     expect(modelCapabilityFields({ maxOutputTokens: 1.9 }).capabilities.supports_reasoning)
       .toBe(false);
   });
+
+  test("mirrors top-level context_window and max_output_tokens for external and legacy client discovery", () => {
+    const fields = modelCapabilityFields({ contextWindow: 200000, maxOutputTokens: 64000 });
+    expect(fields.context_window).toBe(200000);
+    expect(fields.context_length).toBe(200000);
+    expect(fields.max_output_tokens).toBe(64000);
+    expect(fields.capabilities.context_length).toBe(200000);
+    expect(fields.capabilities.max_output_tokens).toBe(64000);
+
+    // Empty or non-positive / unsafe inputs: keys omitted entirely
+    const empty = modelCapabilityFields({});
+    expect("context_window" in empty).toBe(false);
+    expect("context_length" in empty).toBe(false);
+    expect("max_output_tokens" in empty).toBe(false);
+
+    for (const value of [0, -50, Number.NaN, Number.MAX_SAFE_INTEGER + 2]) {
+      const fields = modelCapabilityFields({ contextWindow: value, maxOutputTokens: value });
+      expect("context_window" in fields).toBe(false);
+      expect("context_length" in fields).toBe(false);
+      expect("max_output_tokens" in fields).toBe(false);
+    }
+  });
+
+  test("mirrors the effective long context window at the top level", () => {
+    const fields = modelCapabilityFields({ contextWindow: 272000, longContextWindow: 922000 });
+    expect(fields.capabilities.context_length).toBe(922000);
+    expect(fields.context_window).toBe(922000);
+    expect(fields.context_length).toBe(922000);
+  });
 });
 
 describe("nativeOpenAiContextTier", () => {
@@ -165,6 +194,9 @@ describe("raw /v1/models list advertises Cursor local-agent capabilities", () =>
         supports_vision: true,
         reasoning_effort: ["low", "high", "max"],
       });
+      expect(k3!.context_window).toBe(200000);
+      expect(k3!.context_length).toBe(200000);
+      expect(k3!.max_output_tokens).toBe(64_000);
       // Grok Build's discovery fields stay untouched next to the new keys.
       expect(k3!.supports_reasoning_effort).toBe(true);
       expect(k3!.reasoning_effort).toBe("high");
@@ -186,6 +218,9 @@ describe("raw /v1/models list advertises Cursor local-agent capabilities", () =>
       // Native GPT-5.6: 272k default window, 922k opt-in ceiling → Cursor Context selector.
       expect(solCaps.context_length).toBe(922000);
       expect(solCaps.max_output_tokens).toBe(128_000);
+      expect(sol!.context_window).toBe(922000);
+      expect(sol!.context_length).toBe(922000);
+      expect(sol!.max_output_tokens).toBe(128_000);
       expect(sol!.pricing).toEqual({ overrides: [{ min_prompt_tokens: 272000 }] });
       expect("long_context_threshold_tokens" in sol!).toBe(false);
       expect(solCaps.supports_vision).toBe(true);