feat: per-model max output tokens and a conversation-time thinking level backed by agent settings (#28)

Co-authored-by: Alice <alice@prismshadow.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

packages/web/test/catalog-sync.test.ts

@@ -16,6 +16,7 @@ function makeRow(partial: Partial<RowState> & Pick<RowState, "provider" | "model
     original: { provider: partial.provider, modelId: partial.modelId },
     vision: true,
     contextWindow: "",
+    maxTokens: "",
     clientType: "",
     cacheRead: "",
     cacheWrite: "",
@@ -118,6 +119,22 @@ describe("syncRowsWithCatalog", () => {
     expect(rows[0]).toBe(upToDate);
   });
 
+  it("preserves a user-set max output tokens through a preset sync (user-owned, not catalog-owned)", () => {
+    const local = makeRow({
+      provider: "deepseek",
+      modelId: "deepseek-v4-pro",
+      maxTokens: "4096", // user annotation
+      contextWindow: "500000", // stale -> the row does get updated by the sync
+    });
+    const { rows, updated } = syncRowsWithCatalog([local], PRESET);
+    expect(updated).toBe(1);
+    const row = rows[0]!;
+    expect(row.contextWindow).toBe("1000000"); // catalog-owned field reset
+    expect(row.maxTokens).toBe("4096"); // user field survives the {...row, ...fields} merge
+    // Fresh catalog rows default to inherit (no preset output cap exists).
+    expect(rows.find((r) => r.modelId === "glm-5.2")!.maxTokens).toBe("");
+  });
+
   it("keeps locally added models (including user-defined groups) verbatim and in place", () => {
     const mine = makeRow({ provider: "my-gateway", modelId: "my-model", baseUrl: "http://x" });
     const { rows, added } = syncRowsWithCatalog([mine], PRESET);